CN118484755A

CN118484755A - Method and device for detecting abnormal operation state of die sinking equipment and storage medium

Info

Publication number: CN118484755A
Application number: CN202410556683.2A
Authority: CN
Inventors: 邹鑫森; 张二利; 陈利英
Original assignee: Suzhou Dingyu Energy Efficient Equipment Co Ltd
Current assignee: Suzhou Dingyu Energy Efficient Equipment Co Ltd
Priority date: 2024-05-07
Filing date: 2024-05-07
Publication date: 2024-08-13

Abstract

The application relates to the technical field of data processing, and provides a method, a device and a storage medium for detecting the running state abnormality of mold opening equipment. The process utilizes a special running state abnormality detection network, generates an abnormal state judgment viewpoint by deeply analyzing the state log field, and further improves the accuracy of abnormality detection. The scheme not only can identify whether the equipment has abnormality, but also can determine specific abnormality types, such as overhigh temperature, overhigh speed and the like. Based on the detailed abnormal labels, the scheme can generate targeted equipment operation and maintenance decision suggestions, so that the operation efficiency and the production quality of equipment are remarkably improved. The method effectively integrates data collection, anomaly detection and decision advice, and provides a comprehensive and efficient solution for operation and maintenance of the die sinking equipment.

Description

Method and device for detecting abnormal operation state of die sinking equipment and storage medium

Technical Field

The present application relates to the field of data processing technologies, and in particular, to a method and apparatus for detecting an abnormal operation state of a mold opening device, and a storage medium.

Background

The mold opening equipment is key equipment for producing various industrial products, and particularly has the function of being free from the mold opening equipment in the manufacturing process of products such as environment-friendly energy-saving ventilation equipment, waste gas treatment equipment, fireproof high-corrosion-resistant pipelines and the like in the field of energy-saving equipment. However, since these devices involve many complex physical and chemical reactions during operation, their operation states are often difficult to accurately grasp directly by observation or measurement.

Traditional equipment maintenance approaches rely primarily on manual periodic inspection or monitoring using simple sensors. However, these methods are time-consuming and labor-consuming, and due to lack of deep analysis, it is often impossible to find and solve hidden problems in the running process of the equipment, such as abnormal states of overheat, overspeed, etc. of the equipment in time. This not only affects the normal operation of the device, but also may cause damage to the device and even cause safety accidents.

With the development of big data and artificial intelligence technology, a more intelligent, efficient and accurate solution is needed for detecting abnormal operation states of mold opening equipment.

Disclosure of Invention

In order to improve the problems, the application provides a method and a device for detecting abnormal operation state of mold opening equipment and a storage medium.

In a first aspect, an embodiment of the present application provides a method for detecting an abnormal operation state of a mold opening device, which is applied to an apparatus for detecting an abnormal operation state, and the method includes:

acquiring an operation state log of the open-model equipment to be subjected to anomaly detection;

Performing abnormality detection on the operation state log of the open-mode equipment to be subjected to abnormality detection by using an operation state abnormality detection network, and generating an abnormality state discrimination view of the operation state log of the open-mode equipment to be subjected to abnormality detection; the abnormal state judging viewpoint characterizes abnormal labels corresponding to all state log fields in the operation state log of the open-mode equipment to be subjected to abnormal detection;

And determining equipment operation and maintenance decision suggestions of the operation state logs of the open-model equipment to be subjected to the anomaly detection based on the anomaly labels corresponding to the various state log fields in the operation state logs of the open-model equipment to be subjected to the anomaly detection.

In a preferred embodiment of the first aspect, the step of debugging the running state anomaly detection network comprises:

Determining a state element vector chain of the target running state log based on a state element vector of the target running state log mined by a state element mining subnet in the running state abnormality detection network; the running state abnormality detection network is obtained by pre-testing an initial running state log and priori training knowledge of each state log field in the initial running state log;

Performing state element optimization according to the state element vector chain to obtain a state element optimization vector chain, wherein a state element optimization vector of a state log field in the state element optimization vector chain is obtained by optimizing the state element of a context state log field of the state log field;

determining the influence weight of each state log field in the target running state log based on the state element vector chain and the state element optimization vector chain;

Selecting a key state log field set to be annotated with the state log field from the target running state log based on the influence weight of each state log field in the target running state log and the state text vector of each state log field in the state element vector chain; the influence weight is used for adjusting the state text vector difference between the two state log fields;

And performing migration learning debugging on the running state anomaly detection network based on the prior training knowledge of the state log fields in the key state log field set and the target running state log.

In a preferred embodiment of the first aspect, the selecting, in the target running state log, a set of key state log fields to be annotated with the state log fields based on the impact weights of the state log fields in the target running state log and the state text vectors of the state log fields in the state element vector chain includes:

Determining a first set of state log fields, the first state log field in the first set of state log fields being a state log field in the target running state log;

for each second state log field in the second state log field set, determining a first target state log field in the first state log field set having the smallest state text vector difference with the second state log field; the state text vector difference is determined according to the state text vectors of the two state log fields; a second state log field in the second state log field set is a state log field in the target running state log, and there is no overlapping field between the second state log field set and the first state log field set;

Adjusting the state text vector difference between the second state log field and the corresponding first target state log field through the influence weight of the first target state log field corresponding to the second state log field to obtain a target difference between the second state log field and the first state log field set, wherein the target difference and the influence weight are in a set quantization relationship;

determining a second target state log field in a second state log field set, which has the largest target difference from the first state log field set;

adding the second target state log field to the first set of state log fields;

the set of critical state log fields is determined based on the set of first state log fields that completed the update.

In a preferred embodiment of the first aspect, the determining the set of critical status log fields based on the set of first status log fields completing the update comprises:

recording the number of the state log fields of the updated first state log field set;

If the number of the state log fields meets the annotation condition, the first state log field set which completes updating is used as the key state log field set;

and if the comment condition is not met based on the number of the state log fields, deleting the second target state log field from the second state log field set to obtain an adjusted second state log field set, jumping to each second state log field in the second state log field set, and determining a first target state log field with the smallest state text vector difference with the second state log field in the first state log field set.

In a preferred embodiment of the first aspect, the operating state anomaly detection network further comprises a first anomaly discrimination operator; the method further comprises, for each second state log field in the second state log field set, before determining a first target state log field in the first state log field set that has a smallest difference in state text vector from the second state log field:

Performing anomaly detection on a result of state element mining on a target running state log by the first anomaly discrimination operator according to the state element mining subnet to obtain a first anomaly state discrimination viewpoint, wherein the first anomaly state discrimination viewpoint characterizes the possibility that each state log field in the target running state log belongs to each anomaly tag in a plurality of anomaly tags;

for each state log field, determining the maximum X target possibilities for the state log field in the possibility that the state log field belongs to a plurality of abnormal labels, wherein X is a positive integer;

according to the X target possibilities determined for each state log field, calculating to obtain annotation indexes of each state log field;

and selecting a state log field with the annotation index larger than the annotation index threshold from the target running state log, and generating the second state log field set according to the state log field with the annotation index larger than the annotation index threshold.

In a preferred embodiment of the first aspect, X is 2, and the X target likelihoods determined for a state log field include a likelihood maximum value and a likelihood next-maximum value of likelihoods that the state log field belongs to a plurality of exception tags; the operation obtains the annotation index of each state log field according to the X target possibilities determined for each state log field, and the method comprises the following steps:

for each state log field, calculating to obtain a summation result of a maximum likelihood value and a secondary likelihood value corresponding to the state log field, and obtaining the reference likelihood of the state log field;

For each state log field, determining the annotation index of the state log field based on the difference result of the subtraction of the reference possibility corresponding to the state log field and the set constant.

In a preferred embodiment of the first aspect, the performing state element optimization according to the state element vector chain to obtain a state element optimized vector chain includes:

Performing regional feature extraction operation on the state element vector chain to obtain a regional state feature relation network, wherein the state text vector value of a reference state log field of a feature extraction core of each round of feature extraction in the regional state feature relation network is 0;

And integrating the regional state characteristic relation network and the state element vector chain to obtain the state element optimization vector chain.

In a preferred embodiment of the first aspect, the performing a region feature extraction operation on the state element vector chain to obtain a region state feature relation network includes:

carrying out moving average processing on the state element vector chain to obtain a state element vector chain to be processed;

Combining the state element vector chain to be processed with the linear feature coverage vector corresponding to the feature extraction kernel to obtain a target state element vector chain;

And carrying out quantization feature mapping on the state element vectors in the target state element vector chain to obtain the regional state feature relation network.

In a preferred embodiment of the first aspect, the determining the impact weight of each state log field in the target running state log based on the state element vector chain and the state element optimization vector chain includes:

for each state log field, acquiring a state text vector of the state log field from the state element vector chain and acquiring a state element optimization vector of the state log field from the state element optimization vector chain;

Calculating to obtain the vector similarity between the state text vector of the state log field and the state element optimization vector of the state log field;

And determining the influence weight of the state log field according to the vector similarity of the state log field.

In a preferred embodiment of the first aspect, the operating state anomaly detection network further comprises a first anomaly discrimination operator; the state element optimization vector chain is obtained by optimizing a regional characteristic proposal model according to the state element vector chain; the performing migration learning debugging on the running state anomaly detection network based on the prior training knowledge of the state log field in the key state log field set and the target running state log includes:

acquiring a first abnormal state judgment view obtained by performing abnormality detection on a state element vector mined by a target running state log according to the state element mining subnet by a first abnormality judgment operator;

Determining a first training cost function based on the first abnormal state discrimination viewpoint and prior training knowledge of the state log fields in the key state log field set;

determining a state element optimization training cost function based on the state element vector chain and the state element optimization vector chain;

Optimizing a training cost function based on the first training cost function and the state elements improves a variable of at least one of the operating state anomaly detection network and the regional feature proposal model.

In a preferred embodiment of the first aspect, the state element vector chain is obtained by performing downsampling processing on a first state element vector chain by using a first residual model, where the first state element vector chain is a state element vector mined by the state element mining subnet for a target running state log;

The method further comprises: performing anomaly detection by a second anomaly discrimination operator based on the state element vector chain to obtain a second anomaly state discrimination viewpoint; determining a second training cost function based on the second abnormal state discrimination perspective and prior training knowledge of the state log fields in the set of key state log fields;

Said optimizing a training cost function based on said first training cost function and said state elements improves a variable of at least one of said run state anomaly detection network and said regional feature proposal model comprising: if the cycle number is smaller than the first preset number, optimizing the training cost function based on the first training cost function, the second training cost function and the state element, and determining a target training cost function; based on the target training cost function, variables of the operating state anomaly detection network, the regional feature proposal model, the first residual model, and the second anomaly discrimination operator are improved.

In a preferred embodiment of the first aspect, the optimizing the training cost function based on the first training cost function and the state element improves a variable of at least one of the operating state anomaly detection network and the regional feature proposal model, further comprising:

If the cycle times are not less than the first preset times, improving the variable of the running state abnormality detection network based on the first training cost function;

And adjusting variables of the region feature proposal model, the first residual model and the second anomaly discrimination operator based on the second training cost function and the weighted cost function of the state element optimization training cost function.

In a preferred embodiment of the first aspect, the method further comprises: and according to the first initial running state log and the learning label of the first initial running state log, performing migration learning debugging on the running state abnormality detection network.

In a preferred embodiment of the first aspect, the performing the migration learning debug on the abnormal running state detection network according to the first initial running state log and the learning label of the first initial running state log includes:

The state element mining subnet performs state element mining on the first initial running state log to obtain a first state element vector chain of the first initial running state log;

performing anomaly detection by a first anomaly discrimination operator according to a first state element vector chain of the first initial running state log to obtain a third anomaly state discrimination viewpoint;

determining a first training cost function for the first initial running state log based on prior training knowledge of the first initial running state log and the third abnormal state discrimination point;

Determining a state element vector chain of the first initial running state log according to a first state element vector chain of the first initial running state log;

performing state element optimization on a state element vector chain of the first initial running state log through the regional characteristic proposal model to obtain a state element optimization vector chain of the first initial running state log;

determining a training cost function for the state element optimization of the first initial running state log based on the state element vector chain of the first initial running state log and the state element optimization vector chain of the first initial running state log;

Determining a third training cost function for the first initial running state log based on the likelihood that each state log field in the third abnormal state discrimination perspective belongs to a priori abnormal label and the likelihood that each state log field belongs to a noise abnormal label;

And improving the variable of at least one of the running state anomaly detection network and the regional feature proposal model according to a first training cost function, a state element optimization training cost function and a third training cost function for the first initial running state log.

In a second aspect, an embodiment of the present application provides an operation state anomaly detection apparatus, including at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes computer-executable instructions stored in the memory such that the at least one processor performs the method of the first aspect.

In a third aspect, embodiments of the present application provide a computer readable storage medium having stored thereon a computer program which, when run, implements the method of the first aspect.

In the embodiment of the application, the abnormality detection of the die sinking equipment and the generation of the equipment operation and maintenance decision suggestion can be accurately performed in real time. In detail, by acquiring the operation state log of the open-model equipment to be subjected to abnormality detection and analyzing by using the operation state abnormality detection network, the scheme can identify various possible equipment abnormalities in real time. The running state anomaly detection network carries out deep analysis on each state log field of the running log to generate an anomaly state discrimination viewpoint, namely a corresponding anomaly label, so that the anomaly detection accuracy is greatly improved.

Furthermore, the scheme not only can identify whether the equipment is abnormal, but also can determine specific abnormal types, such as overhigh temperature, overhigh speed and the like, which are important for subsequent fault removal and repair. Meanwhile, based on the obtained abnormal label, the scheme can generate a targeted equipment operation and maintenance decision suggestion, and provides powerful decision support for equipment maintenance personnel, so that the operation efficiency and the production quality of equipment are greatly improved.

In general, the technical scheme effectively integrates the processes of data collection, anomaly detection and decision suggestion, and provides a comprehensive and efficient solution for the operation and maintenance of the die sinking equipment.

Drawings

Fig. 1 is a flowchart of a method for detecting abnormal operation state of mold opening equipment according to an embodiment of the present application.

Fig. 2 is a schematic structural diagram of an abnormal running state detecting device 200 according to an embodiment of the present application.

Detailed Description

In order to better understand the above technical solutions, the following detailed description of the technical solutions of the present application is made by using the accompanying drawings and specific embodiments, and it should be understood that the specific features of the embodiments and the embodiments of the present application are detailed descriptions of the technical solutions of the present application, and not limiting the technical solutions of the present application, and the technical features of the embodiments and the embodiments of the present application may be combined with each other without conflict.

Fig. 1 shows a method for detecting abnormal operation state of mold opening equipment, which is applied to an abnormal operation state detecting device, and comprises the following steps 110-130.

Step 110, obtaining an operation state log of the open-mode equipment to be subjected to anomaly detection.

And 120, performing anomaly detection on the operation state log of the open-mode equipment to be subjected to anomaly detection by using an operation state anomaly detection network, and generating an anomaly state discrimination view of the operation state log of the open-mode equipment to be subjected to anomaly detection, wherein the anomaly state discrimination view characterizes anomaly labels corresponding to all state log fields in the operation state log of the open-mode equipment to be subjected to anomaly detection.

And 130, determining equipment operation and maintenance decision suggestions of the operation state logs of the open-model equipment to be subjected to anomaly detection based on anomaly labels corresponding to all state log fields in the operation state logs of the open-model equipment to be subjected to anomaly detection.

In the above technical solution, the operation state log of the mold opening device is recorded information generated during the operation of the device, and generally includes information such as various operation parameters (such as temperature, pressure, speed, etc.) of the device, and a timestamp. For example, one running status log might be such that: "2022-03-0110:00:00, temperature: 80 degrees, pressure: 5MPa, speed: 300 rpm).

The abnormal operation state detection network is a model or algorithm for detecting whether the operation state of the device is normal, and is generally based on machine learning or deep learning technology. For example, it is possible to use a deep neural network as the anomaly detection network, input a device running state log, and output an anomaly tag for each state log field.

The process of performing anomaly detection is a process of inputting the running state log into an anomaly detection network to obtain an anomaly tag for each state log field. For example, the above-described operation state log may be input to the anomaly detection network, and the result may be: "temperature: normal, pressure: normal, speed: abnormality).

The abnormal state discrimination viewpoint is a result of determining whether or not each state log field is abnormal by the abnormality detection network, and is generally expressed as a series of abnormality tags. For example, "temperature: normal, pressure: normal, speed: the anomaly is an anomaly state discrimination viewpoint.

The status log field is each specific parameter in the running status log, such as temperature, pressure, speed, etc. Each field corresponds to a value that indicates the specific operating state of the device at a certain moment.

The anomaly flag is a result of determining whether or not each status log field is anomalous by the anomaly detection network, and is generally indicated by "normal" or "anomaly". For example, "temperature: normal, pressure: normal, speed: "normal" and "abnormal" in the abnormality "are abnormality tags.

The equipment operation and maintenance decision proposal is a proposal provided for the operation and maintenance of equipment based on the abnormal state discrimination viewpoint. For example, if a speed anomaly is found, the device operation decision recommendation may be: "check and adjust the operating speed of the device".

In some more detailed examples, the force of deep learning may be utilized to address this problem. One exemplary approach is to use a self-encoder (Autoencoder) for anomaly detection, among others.

A self-encoder is an unsupervised learning neural network whose goal is to encode and then decode input data through a hidden layer (also called an encoding layer) so that the output data is as close as possible to the input data. Under normal conditions, the self-encoder can learn the main features of the data and can reconstruct the input data better.

However, when the input data contains anomalies, since the self-encoder does not learn the characteristics of these anomalies, a large error occurs in reconstructing these anomalies. Thus, an abnormality can be detected by comparing the difference of the input data and the reconstructed data. If the difference exceeds a certain threshold, the input data is considered to contain anomalies.

For example, for an operating state log of a mold opening device, various state log fields (e.g., temperature, pressure, speed, etc.) may be extracted as input data to train a self-encoder. During training, the self-encoder learns the characteristics of the normal operating state.

The new running state log may then be input to the trained self-encoder to obtain a reconstructed running state log. The difference between the original log and the reconstructed log is compared, and if the difference exceeds a threshold, then the running state log is considered to contain anomalies.

By the mode, the operation state of the die sinking equipment can be monitored in real time, and early warning is sent out when abnormality is detected, so that the problem is found and solved in advance.

In other more detailed examples, anomaly detection may also be implemented by the neural network model being a Long Short-Term Memory (LSTM).

LSTM is a special Recurrent Neural Network (RNN) adapted to process and predict very long-spaced and delayed events of importance in a time series. In the open device operation state anomaly detection, the operation log of the device can be regarded as a time series data, and then the LSTM is used to learn and predict the normal operation mode of the device.

Specifically, individual status log fields (e.g., temperature, pressure, speed, etc.) may be extracted from a historical running status log and organized into a sequence in chronological order. An LSTM model may then be trained to predict future states from past states.

When the LSTM is used for abnormality detection, a new running state log is input into a trained LSTM model to obtain a predicted next state. The actual next state is then compared to the predicted next state and if the difference between the two exceeds a certain threshold, this running state log is considered to contain anomalies.

For example, if the LSTM model predicts that the temperature at the next time should be 80 degrees, but the actual temperature is 90 degrees, and this difference exceeds a set threshold, then it is considered that an abnormality may occur in the device.

By the method, the operation state of the die sinking equipment can be monitored in real time, and possible abnormality of the equipment in the future can be predicted, so that measures are taken in advance to avoid problems.

In some possible application scenarios, the anomaly tags are not necessarily two-class normal/anomaly. In more complex scenarios, more detailed classification of anomalies may be required in order to more accurately identify and address problems.

For example, for the mold opening apparatus, the following several types of abnormal tags may be defined.

"Normal": the device is operating normally and all parameters are within normal ranges.

"Temperature anomaly": the operating temperature of the device exceeds the normal range, possibly due to a cooling system failure or other reasons.

"Pressure anomaly": the operating pressure of the device exceeds the normal range, possibly due to hydraulic system failure or other reasons.

"Speed anomaly": the operating speed of the device exceeds the normal range, possibly due to a drive system failure or other reasons.

"Multiparameter anomaly": the parameters of the device are simultaneously beyond normal ranges, which may mean that the device has problems or that some important component has failed.

To obtain these anomaly labels, a multi-classification model, such as a multi-class Support Vector Machine (SVM), multi-class logistic regression, CNN or RNN in deep learning, etc., may be trained. In the training process, each state log field in the history log data and the corresponding exception label are required to be input into the model together, so that the model learns various types of exceptions.

In actual operation, the collected new log data is input into the model, and if the model judges that the equipment is likely to be abnormal, a corresponding abnormal label is output. Then, fault diagnosis and maintenance can be performed based on these abnormality tags. By defining different abnormal labels, the problem of the equipment can be more accurately positioned and solved, so that the operation efficiency and the production quality of the equipment are improved.

In addition, for each anomaly type, more specific subcategories may also be defined to enable finer description and handling of the problem. Taking the temperature anomaly as an example, the following sub-categories can be defined.

"Temperature too high": the operating temperature of the device exceeds the upper limit of the normal range, which may be due to cooling system failure or other reasons.

"Temperature too low": the operating temperature of the device is below the lower limit of the normal range, which may be due to insufficient heat sources or other reasons.

"Temperature drastic fluctuation": the operating temperature of the device fluctuates greatly up and down in a short time, which may mean that there is a problem with the temperature control system of the device.

"Temperature jump": the operating temperature of the device suddenly increases or decreases and then returns to the normal range, which may be due to sensor failure or other temporary problems of the device.

To obtain these more specific anomaly labels, a multi-level classification model may be used, or the anomalies may be clustered. For example, a model may be used to determine if there is an anomaly and then a model may be used to determine the particular anomaly type.

The specific implementation method can be selected according to actual conditions. For example, decision trees or random forests may be used for multi-level classification, and clustering algorithms such as K-means or DBSCAN may be used for classifying anomalies.

In actual operation, the collected new log data is input into the model, and if the model judges that the equipment is likely to be abnormal, a corresponding abnormal label is output. Then, fault diagnosis and maintenance can be performed based on these abnormality tags. By defining more specific anomaly tags, the problems of the equipment can be more precisely located and solved, thereby improving the operating efficiency and production quality of the equipment.

Likewise, speed anomalies may have multiple subcategories. For example.

"Speed is too fast": the operating speed of the device exceeds the upper limit of the normal range, which may be caused by a failure of the drive system or a problem with the control system.

"Too slow": the operating speed of the apparatus is below the lower limit of the normal range, which may be caused by insufficient driving force or excessive load, etc.

"Speed severely fluctuates": the operating speed of the device fluctuates up and down considerably in a short time, which may mean that there is a problem with the speed control system of the device.

"Speed jump": the operating speed of the device suddenly increases or decreases and then returns to the normal range, which may be due to sensor failure or other temporary problems of the device.

For these specific anomaly tags, a multi-level classification model may be employed, or anomalies may be clustered. For example, a model may be used to determine if there is an anomaly and then a model may be used to determine the particular anomaly type. The specific implementation method can be selected according to actual conditions. For example, decision trees or random forests may be used for multi-level classification, and clustering algorithms such as K-means or DBSCAN may be used for classifying anomalies.

In the field of energy-saving equipment, the manufacture of products such as environment-friendly energy-saving ventilation equipment, waste gas treatment equipment, fireproof and high-corrosion-resistant pipelines and the like is independent of related mold opening equipment. Thus, in some examples, the mold opening device is exemplified as a mold opening device of a centrifugal ventilator, and the operation state abnormality detection means is responsible for abnormality detection and device operation and maintenance decision suggestion thereof.

When step 110 starts, the operation state abnormality detection device first obtains an operation state log of the centrifugal fan mold opening device to be subjected to abnormality detection. These logs may contain various operating parameters such as the operating speed, temperature, vibration frequency, etc. of the device and corresponding time stamps.

In step 120, the running state abnormality detection device inputs these running state logs into the trained running state abnormality detection network for abnormality detection. The network may identify various types of anomalies and generate a corresponding anomaly tag for each anomaly. The anomaly tags form an anomaly state discrimination viewpoint and characterize anomaly types corresponding to each state log field in the running state log.

In step 130, the running state anomaly detection device determines a device operation and maintenance decision suggestion according to the anomaly tags corresponding to the respective state log fields in the running state log. For example, if the anomaly tag indicates "over-temperature", the equipment operation and maintenance decision recommendation might be "check and adjust cooling system"; if the anomaly flag indicates "speed too fast", the plant operation and maintenance decision recommendation may be "check and adjust motor drive". Therefore, the abnormal running state detection device can provide specific and targeted operation and maintenance decision suggestions for the die opening equipment of the centrifugal fan, so that more accurate and effective fault diagnosis and repair can be conveniently carried out.

It can be seen that, by applying the steps 110-130, the abnormality detection of the mold opening device and the generation of the device operation and maintenance decision suggestion can be performed accurately in real time. In detail, by acquiring the operation state log of the open-model equipment to be subjected to abnormality detection and analyzing by using the operation state abnormality detection network, the scheme can identify various possible equipment abnormalities in real time. The running state anomaly detection network carries out deep analysis on each state log field of the running log to generate an anomaly state discrimination viewpoint, namely a corresponding anomaly label, so that the anomaly detection accuracy is greatly improved.

In some possible embodiments, the step of debugging the running state anomaly detection network includes steps 210-250.

Step 210, determining a state element vector chain of the target running state log based on a state element vector of the target running state log mined by a state element mining subnet in the running state anomaly detection network; the running state abnormality detection network is obtained by pre-testing an initial running state log and priori training knowledge of each state log field in the initial running state log.

Step 220, performing state element optimization according to the state element vector chain to obtain a state element optimization vector chain, wherein a state element optimization vector of a state log field in the state element optimization vector chain is obtained by optimizing the state element of a context state log field of the state log field.

Step 230, determining the influence weight of each state log field in the target running state log based on the state element vector chain and the state element optimization vector chain.

Step 240, selecting a key state log field set to be annotated with the state log field from the target running state log based on the influence weight of each state log field in the target running state log and the state text vector of each state log field in the state element vector chain; the impact weight is used to adjust the state text vector difference between the two state log fields.

And 250, performing migration learning debugging on the running state anomaly detection network based on the prior training knowledge of the state log fields in the key state log field set and the target running state log.

In this technical solution, some new technical terms are introduced first.

The state element mining subnetwork is a component in the running state anomaly detection network and is responsible for mining key state elements, namely important parameters affecting the running state of the equipment, from the running state log. The state element vector chain is a vector chain formed by state elements mined by the state element mining sub-network, and each state element vector represents the characteristic of one state log field. Initial running state log and target running state log: the initial running state log is the original data used to train the network, while the target running state log is the new data for which anomaly detection is desired. The prior training knowledge is network parameters and structure information obtained by training the initial running state log, and can be used for pre-debugging and migration learning debugging. The state element optimization is an optimization method, and a state element optimization vector chain is generated by optimizing the state element vector chain by considering the context state log field. The influence weight is determined according to the state element vector chain and the state element optimization vector chain and is used for measuring the influence degree of each state log field on the running state of the equipment. The set of critical state log fields is a set of important state log fields to be annotated with state log fields selected based on the impact weight and the state text vector difference. The transfer learning debugging is a debugging method, and the network is adjusted by using priori training knowledge and target running state logs, so that the network can be better adapted to new data.

For example, a die opening device of an exhaust gas treatment device is used as the target device. The running state anomaly detection network has been pre-debugged with an initial running state log and a priori training knowledge of the various state log fields.

At the beginning of step 210, a state element mining subnet in the running state anomaly detection network processes the target running state log. For example, a dimensionality reduction technique such as Principal Component Analysis (PCA) or an automatic encoder (Autoencoder) may be used to convert the multiple state parameters in each log into a state element vector, thereby forming a state element vector chain.

In step 220, state element optimization is performed according to the state element vector chain to obtain a state element optimized vector chain. This process may involve some deep learning method, such as Long Short Term Memory (LSTM) networks, to take into account the context information of the state log field.

And step 230 determines the impact weight for each state log field in the target run state log based on the state element vector chain and the state element optimization vector chain. For example, algorithms such as Support Vector Machines (SVMs) or decision trees may be used for feature selection to determine the impact weight for each state log field.

In step 240, a set of critical state log fields to be annotated with state log fields is selected from the target running state log based on the impact weights and the state element vector chain. This may be accomplished by comparing state text vector differences between different state log fields.

Finally, in step 250, the running state anomaly detection network is subjected to migration learning debugging based on the prior training knowledge of the state log fields in the key state log field set and the target running state log. For example, a pre-trained neural network model (such as convolutional neural network CNN or recurrent neural network RNN) may be utilized to perform migration learning and debugging by fine-tuning its parameters to adapt to new tasks.

In the technical scheme, key factors influencing the running state of the equipment can be more accurately identified by using the technologies of state element mining sub-network, state element optimization, weight influence and the like, and a targeted equipment operation and maintenance decision suggestion can be generated. Meanwhile, through transfer learning and debugging, the network can be adjusted to an optimal state, so that the efficiency and accuracy of anomaly detection are improved. In addition, the technical solution emphasizes the importance of the context information, i.e. the state of one state log field may be affected by other state log fields around it. By considering the context information, the running state of the device can be more comprehensively understood, so that more accurate abnormality judgment and device operation and maintenance decision can be made. In general, the technical scheme realizes deeper and more comprehensive abnormality detection and equipment operation and maintenance decision by introducing new technical terms and methods, and has higher technical effects.

In other possible embodiments, the selecting a set of critical state log fields in the target running state log to be annotated with state log fields based on the impact weights of the state log fields in the target running state log and the state text vectors of the state log fields in the state element vector chain described in step 240 includes 241-step 244.

Step 241, determining a first set of status log fields, wherein the first status log field in the first set of status log fields is a status log field in the target running status log.

Step 242, for each second state log field in the second state log field set, determining a first target state log field in the first state log field set having the smallest difference in state text vector between the first state log field and the second state log field; the state text vector difference is determined according to the state text vectors of the two state log fields; the second state log field in the second set of state log fields is a state log field in the target running state log, and there is no overlapping field of the second set of state log fields with the first set of state log fields.

Step 243, adjusting the state text vector difference between the second state log field and the corresponding first target state log field through the influence weight of the first target state log field corresponding to the second state log field, to obtain a target difference between the second state log field and the first state log field set, where the target difference and the influence weight are in a set quantization relationship.

Step 244, determining a second target state log field in the second state log field set that has the greatest target difference from the first state log field set; adding the second target state log field to the first set of state log fields; the set of critical state log fields is determined based on the set of first state log fields that completed the update.

In this example, the specific implementation of steps 241 to 244 will be described still taking the mold opening device of the exhaust gas treatment device as an example.

Beginning at step 241, a first set of state log fields is determined, which are state log fields in a target running state log. For example, the first set of status log fields may include parameters such as temperature, pressure, and operating speed of the device.

In step 242, for each second state log field in the second set of state log fields, a first target state log field is determined in the first set of state log fields that has the smallest difference in state text vector between the second state log field. The second set of state log fields is also selected from the target running state log and has no overlapping fields with the first set of state log fields. The state text vector difference may be determined by calculating euclidean distance or cosine similarity of the state text vectors of the two state log fields.

In step 243, the state text vector difference between the second state log field and the corresponding first target state log field is adjusted by the influence weight of the first target state log field corresponding to the second state log field, so as to obtain the target difference between the second state log field and the first state log field set. The impact weight may be calculated according to information such as a state element vector chain and a state element optimization vector chain.

Finally, in step 244, a second target state log field having the greatest target difference from the first state log field set is determined from the second state log field set, added to the first state log field set, and a critical state log field set is determined based on the updated first state log field set. This process may be accomplished by sorting or threshold selection, etc.

The technical effect of this solution is mainly manifested in more refined analysis and selection of the status log field. The method can find out key state log fields affecting the running state of the equipment, optimize the key state log fields according to the affecting weights of the key state log fields and the state text vector difference, and further improve the accuracy of anomaly detection and the effectiveness of equipment operation and maintenance decision. In addition, by dynamically updating the state log field set, the scheme can be better adapted to the change of the running state of the equipment, and has stronger practicability and adaptability.

In some preferred embodiments, determining the set of critical state log fields based on the set of first state log fields for which the update was completed in step 244 includes: recording the number of the state log fields of the updated first state log field set; if the number of the state log fields meets the annotation condition, the first state log field set which completes updating is used as the key state log field set; and if the comment condition is not met based on the number of the state log fields, deleting the second target state log field from the second state log field set to obtain an adjusted second state log field set, jumping to each second state log field in the second state log field set, and determining a first target state log field with the smallest state text vector difference with the second state log field in the first state log field set.

In this example, the determination of the set of critical status log fields in step 244 is continued using the mold opening device of the exhaust treatment device as an example.

At the beginning of this process, the number of state log fields of the first set of state log fields for which the update is complete is first recorded. For example, if there are 10 status log fields in the first status log field set, the number of status log fields recorded is 10.

Next, it is determined whether the comment condition is satisfied according to the number of status log fields. The annotation condition may be a preset threshold or ratio, for example, if the number of state log fields reaches 20, or reaches 50% of all state log fields, the annotation condition is considered satisfied.

If the annotation condition is met, the first set of state log fields that complete the update is taken as the set of critical state log fields. This means that these status log fields are considered to have the greatest impact on the device operational status, and should take precedence over anomaly detection and device operational maintenance decisions.

If the annotation condition is not met, deleting the second target state log field from the second set of state log fields, obtaining an adjusted second set of state log fields, then jumping to determine a first target state log field in the first set of state log fields that has the smallest difference in state text vector with the second state log field for each second state log field in the second set of state log fields, and then repeating the preceding steps.

The technical effect of this technical solution is mainly manifested in more flexible selection of the status log field and more refined anomaly detection. The method can dynamically update the state log field set, adapt to the change of the running state of the equipment, flexibly determine the key state log field set according to the preset annotation condition, and further realize more accurate and comprehensive abnormality detection. In addition, by deleting the second target state log field and jumping to the next round of judgment, the scheme can also avoid excessive analysis on irrelevant or secondary state log fields, thereby improving the efficiency of abnormality detection.

In some examples, the operating state anomaly detection network further includes a first anomaly discrimination operator. The method further comprises steps 310-340 before determining, for each second state log field of the set of second state log fields, the first target state log field in the set of first state log fields having the smallest difference in state text vector with the second state log field.

Step 310, performing anomaly detection by the first anomaly discrimination operator according to the result of performing state element mining on the target running state log by the state element mining subnet, so as to obtain a first anomaly state discrimination viewpoint, wherein the first anomaly state discrimination viewpoint characterizes the possibility that each state log field in the target running state log belongs to each anomaly tag in a plurality of anomaly tags.

Step 320, for each state log field, determining the maximum X target probabilities for the state log field among the probabilities that the state log field belongs to a plurality of exception labels, where X is a positive integer.

Step 330, according to the X target probabilities determined for each state log field, the annotation index of each state log field is obtained by operation.

Step 340, selecting a state log field with a comment exponent greater than a comment exponent threshold from the target running state log, and generating the second set of state log fields according to the state log field with a comment exponent greater than the comment exponent threshold.

In this example, the specific implementation of steps 310 through 340 will continue to be described, again taking as an example an open mold device for an exhaust treatment device.

First, the newly emerging technical terms need to be explained: the first anomaly discrimination operator is a component part in the running state anomaly detection network and is responsible for anomaly detection according to the result of the state element mining subnet, a first anomaly state discrimination viewpoint is generated, and the anomaly discrimination operator can be constructed by adopting a classifier algorithm or a decision tree algorithm. The annotation index is an index calculated from the likelihood of targeting of each state log field for evaluating the importance of the state log field. The comment index threshold is a predetermined threshold that is selected for the second set of state log fields only if the comment index of the state log field is greater than the threshold.

In step 310, the first anomaly discrimination operator performs anomaly detection according to the result of the state element mining subnet with respect to the target operation state log, to obtain a first anomaly state discrimination viewpoint. For example, if a deep neural network is used as the first anomaly discrimination operator, it may output a probability distribution indicating the likelihood that each state log field belongs to a plurality of anomaly tags.

In step 320, for each state log field, a maximum X target likelihoods are determined for the state log field among the likelihoods that the state log field belongs to the plurality of exception tags. This process may be accomplished by sorting or threshold selection, etc.

In step 330, the annotation index for each state log field is computed based on the X target likelihoods determined for each state log field. For example, the X target likelihoods may be added or averaged as the annotation index.

Finally, in step 340, a state log field having an annotation index greater than the annotation index threshold is selected from the target running state log and a second set of state log fields is generated from these fields.

The technical effect of the technical scheme is mainly embodied in more accurate and comprehensive abnormality detection. By using the first abnormality discrimination operator, the scheme can perform deeper abnormality detection based on the result of state element mining, and generate a first abnormality state discrimination viewpoint having detailed abnormality information. In addition, by calculating the annotation index and setting the annotation index threshold, the scheme can effectively screen out important state log fields, and further improves the accuracy and efficiency of anomaly detection.

In some examples, X is 2, and the X target likelihoods determined for a state log field include a likelihood maximum and a likelihood next-greatest value of likelihoods that the state log field belongs to a plurality of exception tags. The annotation index for each status log field is computed from the X target likelihoods determined for each status log field as described in step 330, including steps 331-332.

Step 331, for each state log field, calculating to obtain a summation result of the maximum likelihood value and the second maximum likelihood value corresponding to the state log field, and obtaining the reference likelihood of the state log field.

Step 332, for each state log field, determining the annotation index of the state log field based on the difference result of the subtraction of the reference likelihood corresponding to the state log field and the set constant.

In this example, the specific implementation of steps 331 to 332 is still described using the mold opening device of the exhaust gas treatment device as an example.

First, in this example, X is set to 2, meaning that each status log field is concerned with the maximum and next-largest values of the likelihood that it belongs to multiple exception tags. These two values will serve as the basis for the subsequent calculation of the annotation index.

In step 331, for each state log field, the sum of its corresponding likelihood maximum and likelihood next-greatest value is calculated, resulting in a reference likelihood for that state log field. For example, if one status log field has a likelihood of belonging to an "overheated" anomaly tag of 0.7 (the maximum likelihood), and a likelihood of belonging to an "overcurrent" anomaly tag of 0.2 (the next largest likelihood), then its reference likelihood is 0.7+0.2=0.9.

Then, in step 332, the annotation index of each status log field is determined based on the difference between the reference likelihood subtraction and the set constant for that status log field. Assuming a constant of 0.5, the comment index of the status log field in the previous example is 0.9-0.5=0.4.

The technical effect of this solution is mainly manifested in a more accurate evaluation of the importance of the status log field. By focusing only on the maximum value and the next-largest value of the likelihood that each status log field belongs to an anomaly tag, the most likely anomaly can be analyzed more intensively, thereby improving the accuracy of anomaly detection. In addition, by introducing a set constant and calculating a difference result, the calculation mode of the annotation index can be flexibly adjusted, so that the annotation index is more suitable for different abnormality detection requirements. This approach also helps to reduce excessive concerns over unimportant state log fields, further improving the efficiency of anomaly detection.

Under some design considerations, the state element optimization is performed according to the state element vector chain in step 220 to obtain a state element optimization vector chain, which includes steps 221-222.

Step 221, performing region feature extraction operation on the state element vector chain to obtain a region state feature relation network, wherein the state text vector value of the reference state log field of the feature extraction core of each round of feature extraction in the region state feature relation network is 0.

Step 222, integrating the regional status feature relation network and the status element vector chain to obtain the status element optimization vector chain.

In some preferred embodiments, the area feature extraction operation is performed on the state element vector chain in step 221 to obtain an area state feature relation network, which includes steps 2211 to 2213.

Step 2211, performing a moving average process on the state element vector chain to obtain a state element vector chain to be processed.

Step 2212, combining the to-be-processed state element vector chain with the linear feature coverage vector corresponding to the feature extraction kernel to obtain a target state element vector chain.

And 2213, carrying out quantization feature mapping on the state element vectors in the target state element vector chain to obtain the regional state feature relation network.

In this example, the specific implementation of steps 221 to 222 and sub-steps thereof will be continued, taking as an example the mold opening device of the exhaust gas treatment device.

First, the newly emerging technical terms need to be explained: regional feature extraction is a machine learning method that creates a regional state feature relation network that contains relevant or important features by extracting them from a state element vector chain. The moving average process is a data preprocessing technique that helps reduce noise and fluctuations in data by smoothing the data by calculating an average value of each state element vector over a period of time. Linear feature coverage vectors are a tool for feature extraction that can map a chain of state element vectors to be processed into a new feature space, helping to identify or extract more valuable features. The quantized feature map is a feature processing method that obtains a set of discrete, comparable features by performing quantization processing on a state element vector.

In step 221, a region feature extraction operation is performed on the state element vector chain to obtain a region state feature relation network. The state text vector value of the reference state log field of the feature extraction core of each feature extraction in the network is 0, which means that in each feature extraction, a certain state log field is used as a reference for comparison.

Then, in step 222, the regional status feature relation network and the status element vector chain are integrated to obtain a status element optimization vector chain. The method comprises the steps of fusing an original state element vector chain with a regional state characteristic relation network obtained through characteristic extraction to generate a state element optimized vector chain which is more representative and contains more abundant information.

In some preferred embodiments, step 221 may be further refined into steps 2211 through 2213.

In step 2211, a moving average process is performed on the state element vector chain to obtain a state element vector chain to be processed. Next, in step 2212, the state element vector chain to be processed is combined with the linear feature coverage vector corresponding to the feature extraction core, so as to obtain the target state element vector chain. Finally, in step 2213, the state element vectors in the target state element vector chain are subjected to quantization feature mapping to obtain a regional state feature relation network.

The main calculation steps involved in the above technical solution can be represented by the following formulas, for example.

In step 2211, a moving average process is performed on the state element vector chain. If the state element vector chain is expressed as v= { V1, V2,..once., vn }, then the moving average process can be expressed as: v' = { SMA (V1), SMA (V2),. SMAs (vn) };

wherein SMA (vi) represents a sliding average of vi over a window of time.

In step 2212, the state element vector chain to be processed is combined with the linear feature coverage vector corresponding to the feature extraction kernel. Assuming that the linear feature coverage vector is expressed as l= { L1, L2,..once., ln }, this step can be expressed as: v "= { combination (V', L) };

Wherein combination (V', L) represents an operation of combining a chain of to-be-processed state element vectors and a linear feature coverage vector.

In step 2213, the state element vectors in the target state element vector chain are subjected to quantization feature mapping to obtain a regional state feature relation network. This step can be expressed as: q= { quantize (V ") |v" ∈v "};

Wherein quantize (v ") represents the operation of the quantized feature map for v".

Finally, in step 222, the regional status feature relation network and the status element vector chain are integrated to obtain a status element optimization vector chain. This step can be expressed as: v_optimized=integer (Q, V);

Where integration (Q, V) represents the operation of integrating the regional state feature relation network with the state element vector chain.

The technical effect of the technical scheme is mainly reflected in deeper feature extraction and more effective state element optimization. By means of regional feature extraction, moving average processing, linear feature coverage vector combination, quantization feature mapping and other methods, the scheme can extract more valuable features from an original state element vector chain, and integrate the features into the state element optimization vector chain, so that more accurate and comprehensive anomaly detection is achieved.

In some alternative embodiments, the determining of the impact weight for each state log field in the target running state log based on the state element vector chain and the state element optimization vector chain described in step 230 includes steps 231-233.

Step 231, for each state log field, acquiring a state text vector of the state log field from the state element vector chain and acquiring a state element optimization vector of the state log field from the state element optimization vector chain.

And step 232, calculating to obtain the vector similarity between the state text vector of the state log field and the state element optimization vector of the state log field.

Step 233, determining the influence weight of the state log field according to the vector similarity of the state log field.

In this example, the specific implementation of steps 231 through 233 will be continued, taking the mold opening device of the exhaust gas treatment device as an example.

First, the newly emerging technical terms need to be explained: vector similarity is an indicator of how similar two vectors are to each other. There are various ways to calculate vector similarity, such as cosine similarity, euclidean distance, etc.

In step 231, for each state log field, a state text vector for the state log field is obtained from the state element vector chain, while a state element optimization vector for the state log field is obtained from the state element optimization vector chain. For example, if the state log field of interest is "modulo current," then the state text vector and the state element optimization vector corresponding to this field can be obtained from two vector chains.

Then, in step 232, a vector similarity between the state text vector of the state log field and the state element optimization vector of the state log field is calculated. For example, cosine similarity may be used to calculate the similarity of the two vectors: similarity=cos (theta) = (a) B)/(|A| B is I B ||) is provided;

Wherein A and B represent a state text vector and a state element optimization vector, respectively, and theta represents an included angle between the two vectors.

Finally, in step 233, the impact weight of the status log field is determined according to the vector similarity of the status log field. For example, the similarity may be directly used as the weight, or the similarity may be converted into the weight by some mapping function.

The technical effect of this solution is mainly represented by a more accurate evaluation of the impact weight of the status log field. By comparing the similarity between the state text vector and the state element optimization vector, the scheme can consider the information difference of the state log field in the original state and the optimized state, so that the influence weight of the state log field can be accurately determined. This helps to improve the accuracy and robustness of anomaly detection.

In some examples, the run state anomaly detection network further comprises a first anomaly discrimination operator; the state element optimization vector chain is obtained by optimizing the regional characteristic proposal model according to the state element vector chain. Then the migration learning debug is performed on the running state anomaly detection network based on the prior training knowledge of the state log fields in the set of critical state log fields and the target running state log as described in step 250, including steps 251-254.

Step 251, obtaining a first abnormal state discrimination view obtained by performing abnormality detection on the state element vector mined by the target running state log according to the state element mining subnet by the first abnormal discrimination operator.

Step 252, determining a first training cost function based on the first abnormal state discrimination viewpoint and a priori training knowledge of the state log fields in the key state log field set.

And 253, determining a state element optimization training cost function based on the state element vector chain and the state element optimization vector chain.

Step 254, optimizing a training cost function based on the first training cost function and the state elements, improving a variable of at least one of the operating state anomaly detection network and the regional feature proposal model.

In this example, the specific implementation of steps 251 to 254 will be continued, taking still the example of an open mold device of the exhaust gas treatment device.

First, the newly emerging technical terms need to be explained: training a cost function: also known as a loss function or objective function, is the amount that needs to be minimized (or maximized) in the machine learning model optimization process. It reflects the gap between the model predictions and the true values.

In step 251, a first abnormal state discrimination viewpoint obtained by performing abnormality detection on the state element vector mined by the target operation state log by the first abnormality discrimination operator according to the state element mining subnet is obtained. For example, the first anomaly discrimination operator may find anomalies in the "modulo current" field.

Then, in step 252, a first training cost function is determined based on the first abnormal state discrimination perspective and a priori training knowledge of the state log fields in the set of critical state log fields. For example, a two-class cross entropy loss function may be constructed from the first abnormal state discrimination viewpoint and the abnormal label of the critical state log field.

Next, in step 253, a state element optimization training cost function is determined based on the state element vector chain and the state element optimization vector chain. This cost function may be used to measure the gap between the state element optimization vector chain and the target state, for example, a cross entropy loss function may also be used.

Finally, in step 254, the training cost function is optimized based on the first training cost function and the state elements to improve the variables of at least one of the operating state anomaly detection network and the regional feature proposal model. For example, model parameters may be adjusted by an optimization algorithm such as gradient descent to minimize the training cost function.

The technical effect of the technical scheme is mainly realized in more accurate transfer learning debugging. By combining multiple perspectives and training the cost function, the solution is able to more fully account for various factors, thereby more accurately adjusting the operating condition anomaly detection network and the regional feature proposal model. This helps to improve the accuracy and generalization ability of anomaly detection while also making the model more adaptive and flexible.

In some examples, the state element vector chain is obtained by performing downsampling processing on a first state element vector chain through a first residual model, where the first state element vector chain is a state element vector mined by the state element mining subnet for a target running state log. The method further comprises steps 410-420.

Step 410, performing anomaly detection by a second anomaly discrimination operator based on the state element vector chain to obtain a second anomaly state discrimination viewpoint.

Step 420, determining a second training cost function based on the second abnormal state discrimination viewpoint and a priori training knowledge of the state log fields in the set of critical state log fields.

Optimizing the training cost function based on the first training cost function and the state elements as described in step 254 improves the variables of at least one of the operating state anomaly detection network and the regional feature proposal model, including steps 2541-2542.

Step 2541, if the number of loops is less than the first preset number of loops, optimizing the training cost function based on the first training cost function, the second training cost function and the state element, and determining a target training cost function.

Step 2542, based on the target training cost function, improving variables of the running state anomaly detection network, the regional feature proposal model, the first residual model and the second anomaly discrimination operator.

In this example, the specific implementation of steps 410 to 420 and 2541 to 2542 will be described further using the mold opening device of the exhaust gas treatment device as an example.

First, the newly emerging technical terms need to be explained: the second anomaly discrimination operator is similar to the first anomaly discrimination operator, and is also a tool or model for anomaly detection based on a chain of state element vectors. The first residual model is a model for performing downsampling processing on an input state element vector chain, so as to obtain a new state element vector chain. Downsampling processing is typically used to reduce the complexity and computational requirements of the data.

In step 410, the second anomaly discrimination operator performs anomaly detection based on the state element vector chain to obtain a second anomaly state discrimination viewpoint. For example, the second anomaly discrimination operator may find anomalies in the "Cooling Water flow" field.

Then, in step 420, a second training cost function is determined based on the second abnormal state discrimination perspective and a priori training knowledge of the state log fields in the set of critical state log fields. For example, a new class-two cross entropy loss function may be constructed from the second abnormal state discrimination viewpoint and the abnormal label of the critical state log field.

Next, in step 2541, if the number of loops is less than the first preset number of loops, the training cost function is optimized based on the first training cost function, the second training cost function, and the state element, and a target training cost function is determined. For example, the average of these three training cost functions may be taken as the target training cost function.

Finally, in step 2542, the variables of the operating state anomaly detection network, the regional feature proposal model, the first residual model, and the second anomaly discrimination operator are modified based on the target training cost function. For example, parameters of these models may be adjusted by optimization algorithms such as gradient descent to minimize the target training cost function.

The technical effect of the technical scheme is mainly realized in more comprehensive and accurate transfer learning debugging. By introducing the second anomaly discrimination operator and the first residual model, the scheme can more comprehensively consider various factors, so that the running state anomaly detection network, the regional feature proposal model and the like can be adjusted more accurately. In addition, by setting the circulation times, the training process can be flexibly adjusted according to actual needs, so that the training efficiency and stability are improved.

In some preferred embodiments, optimizing the training cost function based on the first training cost function and the state elements in step 254 improves a variable of at least one of the operating state anomaly detection network and the regional feature proposal model, further comprising: if the cycle times are not less than the first preset times, improving the variable of the running state abnormality detection network based on the first training cost function; and adjusting variables of the region feature proposal model, the first residual model and the second anomaly discrimination operator based on the second training cost function and the weighted cost function of the state element optimization training cost function.

In this example, the specific implementation of step 254 is continued, again taking as an example an open mold device for an exhaust treatment device.

In a preferred embodiment, step 254 includes the following two parts: and if the cycle times are not less than the first preset times, improving the variable of the running state abnormality detection network based on the first training cost function. For example, if the number of loops reaches a first preset number of times, then only the first training cost function is used to adjust parameters of the operating condition anomaly detection network. And optimizing the weighted cost function of the training cost function based on the second training cost function and the state element, and adjusting variables of the regional feature proposal model, the first residual model and the second anomaly discrimination operator. For example, the weighted average of the second training cost function and the state element optimization training cost function may be used as a new objective cost function, and then the parameters of these models may be adjusted by an optimization algorithm such as gradient descent to minimize the new objective cost function.

The technical effect of the technical scheme is mainly realized in the aspect of more flexibly and effectively performing migration learning debugging. By selecting different training cost functions according to the cycle times, the scheme can flexibly adjust the training process, thereby improving the training efficiency and stability. Meanwhile, by using the weighted cost function, the scheme can integrate various training information more effectively, so that model parameters can be adjusted more accurately. This helps to improve the accuracy and generalization ability of anomaly detection while also making the model more adaptive and flexible.

In other possible embodiments, the method further comprises: and according to the first initial running state log and the learning label of the first initial running state log, performing migration learning debugging on the running state abnormality detection network.

In other preferred embodiments, the performing the migration learning debug on the running state anomaly detection network according to the first initial running state log and the learning label of the first initial running state log includes: the state element mining subnet performs state element mining on the first initial running state log to obtain a first state element vector chain of the first initial running state log; performing anomaly detection by a first anomaly discrimination operator according to a first state element vector chain of the first initial running state log to obtain a third anomaly state discrimination viewpoint; determining a first training cost function for the first initial running state log based on prior training knowledge of the first initial running state log and the third abnormal state discrimination point; determining a state element vector chain of the first initial running state log according to a first state element vector chain of the first initial running state log; performing state element optimization on a state element vector chain of the first initial running state log through the regional characteristic proposal model to obtain a state element optimization vector chain of the first initial running state log; determining a training cost function for the state element optimization of the first initial running state log based on the state element vector chain of the first initial running state log and the state element optimization vector chain of the first initial running state log; determining a third training cost function for the first initial running state log based on the likelihood that each state log field in the third abnormal state discrimination perspective belongs to a priori abnormal label and the likelihood that each state log field belongs to a noise abnormal label; and improving the variable of at least one of the running state anomaly detection network and the regional feature proposal model according to a first training cost function, a state element optimization training cost function and a third training cost function for the first initial running state log.

In this example, the specific implementation of the above steps is continued, still taking as an example the mold opening device of the exhaust gas treatment device.

First, the newly emerging technical terms need to be explained: the first initial operating state log is an original data set for transfer learning and contains a series of log information about the operating state of the open device. Learning labels is the process of labeling an original dataset, typically including assigning one or more labels to each data sample. Here, learning the callout may include specifying whether each status log field has an exception. The third abnormal state discrimination point is a result of abnormality detection based on the first state element vector chain of the first initial running state log.

Next, the following is a specific example of the above-described technical means.

In this embodiment, first, a state element mining subnet is used to perform state element mining on a first initial running state log to obtain a first state element vector chain. Then, the first abnormality discrimination operator is used for carrying out abnormality detection according to the vector chain, and a third abnormal state discrimination viewpoint is obtained.

And determining a first training cost function for the first initial running state log based on the prior training knowledge of the first initial running state log and the third abnormal state discrimination viewpoint. At the same time, it is also necessary to determine a state element vector chain from the first state element vector chain of the first initial running state log.

Next, the state element optimization vector chain is performed on the state element vector chain by using the regional characteristic proposal model, and the state element optimization vector chain is obtained. Based on this optimization vector chain and the state element vector chain, a state element optimization training cost function for the first initial running state log is determined.

In addition, a third training cost function for the first initial running state log needs to be determined based on the likelihood that each state log field belongs to the a priori anomaly label and the likelihood that each state log field belongs to the noise anomaly label in the third anomaly state discrimination viewpoint.

Finally, optimizing the training cost function and the third training cost function according to the first training cost function, the state element, and improving the variable of at least one of the running state anomaly detection network and the regional characteristic proposal model.

The technical effect of the technical scheme is mainly that the running state abnormality detection network and the regional characteristic proposal model can accurately adjust model parameters by introducing more training cost functions and more comprehensive training information, so that the accuracy and generalization capability of abnormality detection are improved. Meanwhile, through detailed analysis and processing of the initial running state log, the scheme can further understand the data characteristics, so that the model performance is further optimized.

Furthermore, the process of the above technical solution can be explained by the following series of steps and related formulas.

State element mining: first, there is a state element mining subnet that converts a first initial running state log into a first state element vector chain. If the function representing the state element mining subnet is f and the first initial running state Log is Log, this procedure can be represented as: v=f (Log);

where V is the resulting first state element vector chain.

Abnormality detection: then, the first abnormality discrimination operator performs abnormality detection with V as an input, and obtains a third abnormality state discrimination viewpoint. If the function representing the first anomaly discrimination operator is g, this process can be represented as: p=g (V);

wherein P is a third abnormal state discrimination viewpoint.

Determining a first training cost function: based on the a priori training knowledge K and P of the first initial run state log, a first training cost function J1 for the first initial run state log may be determined. If the function representing this process is h, this process can be represented as: j1 =h (K, P);

State element optimization: then, the regional feature proposal model performs a state element optimization on V to obtain a state element optimization vector chain V'. If the function representing the regional feature proposal model is m, this process can be represented as: v' =m (V);

Determining a state element optimization training cost function: based on V and V', it may be determined to optimize the training cost function J2 for the state elements of the first initial running state log. If the function representing this process is n, this process can be represented as: j2 =n (V, V');

Determining a third training cost function: based on the likelihood of each state log field in P belonging to the a priori anomaly tags, p_e, and the likelihood of belonging to the noise anomaly tags, p_n, a third training cost function J3 for the first initial run state log may be determined. If the function representing this process is q, this process can be represented as: j3 =q (p_e, p_n);

model variable improvement: finally, according to J1, J2, and J3, the variables of at least one of the operating condition anomaly detection network and the regional feature proposal model can be improved. If the function representing this process is r, this process can be represented as: r (J1, J2, J3);

In this way, the state elements of the original log data are mined, the abnormality detection is carried out through the abnormality discrimination operator, the training cost function is determined by combining the priori training knowledge, and the variable of at least one item in the running state abnormality detection network and the regional characteristic proposal model is improved through optimizing the training cost function, so that the efficient abnormality detection of the running state log is realized.

Fig. 2 is a schematic structural diagram of an operation state anomaly detection device 200 according to an embodiment of the present application. The operation state abnormality detection apparatus 200 shown in fig. 2 includes a processor 210, and the processor 210 may call and execute a computer program from a memory to implement the method in the embodiment of the present application.

Optionally, as shown in fig. 2, the operation state abnormality detection device 200 may further include a memory 230. Wherein the processor 210 may call and run a computer program from the memory 230 to implement the method in an embodiment of the application.

Wherein the memory 230 may be a separate device from the processor 210 or may be integrated into the processor 210.

Optionally, as shown in fig. 2, the operation state anomaly detection apparatus 200 may further include a transceiver 220, and the processor 210 may control the transceiver 220 to interact with other devices, and specifically may send information or data to the other devices, or receive information or data sent by the other devices.

Optionally, the running state anomaly detection apparatus 200 may implement the storage engine or a component (such as a processing module) in the storage engine or a corresponding flow corresponding to a device in which the storage engine is deployed in each method of the embodiments of the present application, which is not described herein for brevity.

It should be appreciated that the processor of an embodiment of the present application may be an integrated circuit chip having signal processing capabilities. In implementation, the steps of the above method embodiments may be implemented by integrated logic circuits of hardware in a processor or instructions in software form. The Processor may be a general purpose Processor, a digital signal Processor (DIGITAL SIGNAL Processor, DSP), an Application SPECIFIC INTEGRATED Circuit (ASIC), an off-the-shelf programmable gate array (Field Programmable GATEARRAY, FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware components. The disclosed methods, steps, and logic blocks in the embodiments of the present application may be implemented or performed. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The steps of the method disclosed in connection with the embodiments of the present application may be embodied directly in the execution of a hardware decoding processor, or in the execution of a combination of hardware and software modules in a decoding processor. The software modules may be located in a random access memory, flash memory, read only memory, programmable read only memory, or electrically erasable programmable memory, registers, etc. as well known in the art. The storage medium is located in a memory, and the processor reads the information in the memory and, in combination with its hardware, performs the steps of the above method.

It will be appreciated that the memory in embodiments of the application may be volatile memory or nonvolatile memory, or may include both volatile and nonvolatile memory. The nonvolatile Memory may be a Read-Only Memory (ROM), a Programmable ROM (PROM), an Erasable PROM (EPROM), an Electrically Erasable EPROM (EEPROM), or a flash Memory. The volatile memory may be random access memory (RandomAccess Memory, RAM) which acts as external cache memory. By way of example, and not limitation, many forms of RAM are available, such as static random access memory (STATIC RAM, SRAM), dynamic random access memory (DYNAMIC RAM, DRAM), synchronous Dynamic Random Access Memory (SDRAM), double data rate Synchronous dynamic random access memory (Double DATA RATE SDRAM, DDR SDRAM), enhanced Synchronous dynamic random access memory (ENHANCED SDRAM, ESDRAM), synchronous link dynamic random access memory (SYNCHLINK DRAM, SLDRAM), and Direct memory bus RAM (DR RAM). It should be noted that the memory of the systems and methods described herein is intended to comprise, without being limited to, these and any other suitable types of memory.

It should be appreciated that the above memory is exemplary but not limiting, and for example, the memory in the embodiments of the present application may also be static random access memory (STATIC RAM, SRAM), dynamic random access memory (DYNAMIC RAM, DRAM), synchronous Dynamic Random Access Memory (SDRAM), double data rate synchronous dynamic random access memory (doubledata RATE SDRAM, DDR SDRAM), enhanced synchronous dynamic random access memory (ENHANCED SDRAM, ESDRAM), synchronous link dynamic random access memory (SYNCH LINK DRAM, SLDRAM), direct Rambus RAM (DR RAM), and the like. That is, the memory in embodiments of the present application is intended to comprise, without being limited to, these and any other suitable types of memory.

On the basis of the above, a computer readable storage medium is provided, on which a computer program is stored, which computer program, when run, implements the method described above.

In conclusion, by acquiring and analyzing the operation state log of the mold opening equipment in real time, the abnormality detection is effectively executed, so that the accurate grasp of the operation condition of the equipment is realized. The process utilizes a special running state abnormality detection network, generates an abnormal state judgment viewpoint by deeply analyzing the state log field, and further improves the accuracy of abnormality detection. More importantly, the scheme not only can identify whether the equipment is abnormal, but also can determine specific abnormal types, such as over-high temperature, over-high speed and the like. Based on the detailed abnormal labels, the scheme can generate targeted equipment operation and maintenance decision suggestions for equipment maintenance personnel, so that the operation efficiency and the production quality of the equipment are remarkably improved. The technical scheme effectively integrates data collection, anomaly detection and decision advice, and provides a comprehensive and efficient solution for operation and maintenance of the die sinking equipment.

The foregoing is merely exemplary of the present application and is not intended to limit the present application. Various modifications and variations of the present application will be apparent to those skilled in the art.

Claims

1. A method for detecting an abnormality in an operation state of a mold opening device, the method being applied to an operation state abnormality detecting device, the method comprising:

2. The method of claim 1, wherein the step of debugging the operating state anomaly detection network comprises:

3. The method of claim 2, wherein the selecting a set of critical state log fields in the target running state log to be annotated with state log fields based on the impact weights of the state log fields in the target running state log and the state text vectors of the state log fields in the state element vector chain comprises:

adding the second target state log field to the first set of state log fields;

4. The method of claim 3, wherein the determining the set of critical state log fields based on the first set of state log fields completing the update comprises:

5. A method according to claim 3, wherein the operating condition anomaly detection network further comprises a first anomaly discrimination operator; the method further comprises, for each second state log field in the second state log field set, before determining a first target state log field in the first state log field set that has a smallest difference in state text vector from the second state log field:

selecting a state log field with the annotation index larger than the annotation index threshold from the target running state log, and generating the second state log field set according to the state log field with the annotation index larger than the annotation index threshold;

Wherein X is 2, and the X target likelihoods determined for a state log field include a likelihood maximum value and a likelihood next-maximum value of likelihoods that the state log field belongs to a plurality of exception tags; the operation obtains the annotation index of each state log field according to the X target possibilities determined for each state log field, and the method comprises the following steps: for each state log field, calculating to obtain a summation result of a maximum likelihood value and a secondary likelihood value corresponding to the state log field, and obtaining the reference likelihood of the state log field; for each state log field, determining the annotation index of the state log field based on the difference result of the subtraction of the reference possibility corresponding to the state log field and the set constant.

6. The method according to claim 2, wherein the performing the state element optimization according to the state element vector chain to obtain a state element optimized vector chain includes:

Integrating the regional state feature relation network and the state element vector chain to obtain the state element optimization vector chain;

The step of extracting the regional characteristics of the state element vector chain to obtain a regional state characteristic relation network comprises the following steps: carrying out moving average processing on the state element vector chain to obtain a state element vector chain to be processed; combining the state element vector chain to be processed with the linear feature coverage vector corresponding to the feature extraction kernel to obtain a target state element vector chain; and carrying out quantization feature mapping on the state element vectors in the target state element vector chain to obtain the regional state feature relation network.

7. The method of claim 2, wherein the determining the impact weight for each state log field in the target running state log based on the state element vector chain and the state element optimization vector chain comprises:

8. The method of claim 2, wherein the operating condition anomaly detection network further comprises a first anomaly discrimination operator; the state element optimization vector chain is obtained by optimizing a regional characteristic proposal model according to the state element vector chain; the performing migration learning debugging on the running state anomaly detection network based on the prior training knowledge of the state log field in the key state log field set and the target running state log includes:

optimizing a training cost function based on the first training cost function and the state elements, improving a variable of at least one of the operating state anomaly detection network and the regional feature proposal model;

The state element vector chain is obtained by performing downsampling processing on a first state element vector chain through a first residual error model, and the first state element vector chain is a state element vector mined by the state element mining subnet aiming at a target running state log; the method further comprises: performing anomaly detection by a second anomaly discrimination operator based on the state element vector chain to obtain a second anomaly state discrimination viewpoint; determining a second training cost function based on the second abnormal state discrimination perspective and prior training knowledge of the state log fields in the set of key state log fields; said optimizing a training cost function based on said first training cost function and said state elements improves a variable of at least one of said run state anomaly detection network and said regional feature proposal model comprising: if the cycle number is smaller than the first preset number, optimizing the training cost function based on the first training cost function, the second training cost function and the state element, and determining a target training cost function; based on the target training cost function, improving variables of the running state anomaly detection network, the regional feature proposal model, the first residual model and the second anomaly discrimination operator;

Wherein said optimizing a training cost function based on said first training cost function and said state elements improves a variable of at least one of said run state anomaly detection network and said regional feature proposal model, further comprising: if the cycle times are not less than the first preset times, improving the variable of the running state abnormality detection network based on the first training cost function; and adjusting variables of the region feature proposal model, the first residual model and the second anomaly discrimination operator based on the second training cost function and a weighted cost function of the state element optimization training cost function;

wherein the method further comprises: according to the first initial running state log and the learning label of the first initial running state log, performing migration learning debugging on the running state abnormality detection network;

The performing migration learning debugging on the running state anomaly detection network according to the first initial running state log and the learning label of the first initial running state log includes:

9. An operation state abnormality detection device, characterized by comprising at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executing computer-executable instructions stored in the memory causes the at least one processor to perform the method of any one of claims 1-8.

10. A computer readable storage medium, characterized in that it has stored thereon a computer program which, when run, performs the method of any of claims 1-8.