RU2784925C1 - System and method for detecting and measuring anomalies in signalling originating from components used in industrial processes - Google Patents

Abstract

FIELD: measuring.

SUBSTANCE: invention relates to a method and system for detecting anomalies in sensor data originating from components used in industrial processes. The method includes the following stages: - measuring and/or tracking the measurement data, accordingly, tracking the process parameters (4) of components used in the industrial process (6) by means of measuring apparatus or sensors (2), and identifying same-sized time frames in the measurement and/or process parameters (4) for the time frames wherein the components used in the industrial process (6) function normally, wherein the measurement and/or process parameters (4) contain parameter values for multiple measurement/sensor process parameters (41) and/or variables (42); - converting the values (4) of the parameters of the multiple measurement/sensor process parameters (41) and/or variables (42) into observable binary processing codes for each of the identified same-sized time frames, and assigning binary processing codes to the sequence of stored Markov chain states; - forming a multidimensional data structure containing a set number of values of the variable parameters of the hidden Markov model, wherein the variable parameters of the multidimensional data structure model are determined by means of a machine learning module (8) applied to the sequence of stored Markov chain states with the assigned binary processing codes (91), and wherein the variable parameters (811, 812, …, 81x) of the hidden Markov model of a multidimensional data structure vary and are trained by training the normal state frequency (82) of the occurring emergency events based on the measurement data and/or process parameters (4) of the identified same-sized time frames; - initialising and storing multiple values (83) of the probabilistic state by applying the trained multidimensional data structure with the values of the variable parameters of the hidden Markov model to the pre-discretised binary processing codes with same-sized time frames identical to the time frame for the values (4) of the parameters of the multiple measurement/sensor process parameters (41) and/or variables (42); - determining the logarithmic threshold value of the quantitative indicator of the anomalies by coordinating the logarithmic resulting values of the stored values (83) of the probabilistic state; and - deploying said trained multidimensional data structure with the values of the variable parameters of the hidden Markov model in order to track new measured, accordingly, determined measurement data and/or process parameters (4) from industrial equipment or factories (1) using the threshold value of the quantitative indicator of anomalies so as to detect abnormal values of the sensor data potentially indicating the future system failure, wherein, when the values of sensor data are abnormal, the logarithmic resulting value for the value (83) of the probabilistic state of the new measured, accordingly, determined measurement data and/or process parameters (4) for initiation is formed and compared with the stored values (83) of the probabilistic state based on said logarithmic threshold value of the quantitative indicator of anomalies.

EFFECT: equipment of the industrial enterprise system with a detection and signalling apparatus based on machine learning tools triggering sensors and measuring equipment of the industrial enterprise.

11 cl, 16 dwg

Description

Technical field

The present invention relates to industrial process control and/or monitoring systems. In particular, the invention relates to a method for detecting anomalies or early indicators of equipment failure in industrial equipment or manufacturing plants by monitoring measurement data and/or process parameters emanating from components used in industrial processes, according to claim 1, respectively, The invention relates to a system for detecting anomalies or early indicators of equipment failure in industrial equipment or in a manufacturing plant by monitoring measurement data and/or process parameters emanating from components used in industrial processes, according to claim 15. The invention comprises self-adaptive control equipment closed-loop and/or open-loop for automated control of closed-loop and/or open-loop roller milling systems, more specifically milling plants having a roller frame, but also grinding systems mi and grinding plants in general. Except for its application to control devices for controlling and coordinating roller grinding systems, the present invention also relates generally to systems and methods for detecting and measuring anomalies in signaling from components used in industrial processes. Possible applications of the equipment according to the invention also relate to roller grinding systems with real-time or quasi-real-time measurement and monitoring of operating parameters such as roller temperature, roller gap, roller speed, roller downforce and/or power consumption of one or different roller drives, and/or with real-time or quasi-real-time measurements of ingredients or quality parameters during production conditioning and processing in bread mills for process monitoring (measurement, monitoring) and open-loop and/or closed-loop control plants and processes, such as measured variables such as water content, protein content, damage to starch grains, ash content (minerals) of flour (or milling intermediates), residual starch content, fineness of grinding, etc.

State of the art

In industrial processes and device settings, control systems are used to monitor, control, coordinate and signal equipment and facilities, plants or other equipment and work/technological processes in relation to industrial or chemical processes, etc. Typically, a system that performs control and monitoring uses field devices distributed at key locations in an industrial process, connected to the control circuit via a process control loop. The term "field device" means any device that performs a function in a distributed process control or monitoring system, including all devices used in measurement, such as sensors and measuring devices, in the control, monitoring and signaling of industrial processes and process equipment. Each field device, for example, may include communication means and circuitry used to communicate, particularly for wired or wireless communication, with a process controller, other field devices, or other circuitry, over a process control loop. In some installations, the process control loop is also used to deliver a stabilized current and/or voltage to the field device to supply power to the field device. The process control loop also carries data, either in analog or digital format. Typically, field devices are used to read or manipulate process variables in an industrial process and/or in specific installations as needed in order to keep track of the field device's local environment.

One of the technical problems of such systems is based on the fact that the control and monitoring of large scale industrial assets (eg, grain mills, food processing plants) typically generates large amounts of process and alarm/failure data. What is often the case is that many alarms/failures are triggered which are nevertheless often ignored or turned off in order to continue the process and control. In addition, alarms/failures can be triggered by light maintenance events on machines and do not cause additional technical concern. In addition, process data, such as from motor currents, can often show atypical values based on thresholds and are not always a cause for concern for individual events. There is a need to provide automated alarm/failure screening into a definitive list of significant events that are abnormal relative to nominal operation. This final list should provide effective preventive maintenance and root cause analysis of downtime events, which are very expensive in industrial processes and should be minimized. It is rather difficult to identify anomalous patterns in advance, so an unsupervised method is required.

In addition, machinery or other industrial installations, such as factories, motorized mills or turbines, etc., are subject to failure for a variety of reasons. As mentioned above, known plant or machinery failures are typically detected by means of sensors, and once the failure is detected, the failure is reported to the operator for correction or signaled to the appropriate assigned alarm devices. However, traditional strategies used to detect damage are typically developed based on known problems that have previously occurred in a machine, plant, or device. These prior occurrences can be determined by automatically inferring sensor profiles that correspond to known abnormal behavior associated with a particular problem. However, for problems that never had a prior occurrence, failures often arrive with no warning or advance indicator. In such situations, repair costs can be significantly higher than if the failure is detected early. In addition, late detection of a failure or an imminent failure may compromise the safety of the machinery. Therefore, it is desirable to provide a system and method that detects unknown abnormal behavior in machinery in an automatic and accurate manner.

Particularly in grain mills and food processing plants, anomaly detection from sensor data is an important application for data mining. Using grinding production as an example, to ensure safe and optimized grinding and prevent major system failures, remote equipment monitoring is a critical part of the production process. The key task of remote monitoring is to detect anomalies, i.e. the need to detect early indicators of failures before they occur. For example, roll pressure and roll temperature are key components in order to ensure consistent production and are monitored for this reason. In the prior art, a lot of effort has been put into automating anomaly detection, but it still represents a very difficult task. There are several technical difficulties, as partially explained above. Sensor data such as temperature, pressure, displacement, flow, vibration, etc. are noisy, sensor values may change intermittently, and the correlation structure may change even on a daily basis. There is a need to include a system and method for intelligent monitoring and detection in order to automatically deal with unwanted noise. Variables and multiple dependencies must be taken into account, so variables should not be parsed separately as this can generate false warnings. In addition, system tracking is often unstable because operating conditions may change over time, such as ambient conditions such as air pressure or air relative humidity/local air humidity. Thus, there is also a need for diagnostic information, such as which variables exhibit anomalies. However, prior art methods are known to typically have severe problems in practice and cannot handle both multiple operating modes and multivariate variable-based anomaly quantification. Most systems cannot efficiently provide information based on variables, which is especially problematic in many industrial applications where the measurement parameters can typically be large in size.

US 2011/288836 discloses a method and system for detecting anomalies in an aircraft engine. The method and system define a behavioral model of an aircraft engine controller using temporal regression modeling the behavior of the controller as a function of a data set associated with the controller and including measurements of previous behaviors and measurements of controller commands and states; continuous recalculation of the behavior model for each set of new data; and monitoring the statistical variation of the behavior pattern to detect a controller behavior anomaly representing an operating anomaly of the engine. US 2016/371600 A discloses systems and methods for monitoring data recorded from systems over time. The technologies described in this document include the ability to detect and classify system events and provide indicators of normal system operation and anomaly detection. The systems and methods of the present disclosure may represent events occurring in the monitored system such that the timing of the events may be captured and used for detection, classification, and/or anomaly detection, which may be particularly useful in solving problems related to complex systems and/or events. US 2017/139398 discloses that a plurality of process facilities and analysis equipment are connected via a fog network. The analysis equipment performs data analysis based on the detection information of the detectors obtained through the fog network, and stores determination information related to the abnormality of each of the plurality of process objects or the abnormality of the production object as a data analysis result. Each of the plurality of process objects determines the abnormality of each of the plurality of process objects or the abnormality of the production object based on the determination information stored in the analysis equipment. EP 3 379 360 discloses that the anomaly detection system 1 includes an arithmetic unit 1H101 that performs predictive model learning processing that predicts the behavior of a monitoring target device based on operating data in the device, anomaly score adjustment processing such that the quantitative the anomaly score for the performance data in normal operation falls within a predetermined range, the anomaly score based on the deviation of the performance data received from the monitoring target device from the prediction result obtained by the predictive model, the anomaly detection processing or the sign of the anomaly based on the adjusted anomaly score; and displaying information about at least one of the anomaly score and the detection result at the output device.

Finally, in prior art systems, alarms/failures are often logged in automated system software and control systems where it is difficult to understand and see trends in data with this form of logging. Therefore, operators rely on their observations of accident/failure events to monitor the health of the plant. There is a need to provide better alarm output and visualization of alarm/failure messages so that operators can easily monitor the operation, safety and health of the plant. Visualization should also allow other people, such as owners and maintenance providers, to gain insight and better communication with operators. It is important to exclude/measure plant downtime in large scale processing plants as it represents a significant loss of revenue.

The essence of the invention

The purpose of the present invention is to overcome the disadvantages and technical problems known from the prior art. In particular, the goal is to provide an accurate and efficient control system and method for detecting anomalies in measurement and sensor data from components used in industrial processes. The system must be able to provide automated technology in order to efficiently weed out huge amounts of alarm/failure information into a few triggered critical events that are anomalous relative to nominal operation. The system must be able to perform the control and monitoring process in real time or quasi-real time. More specifically, the object of the invention is to provide intelligent, self-adaptive open-loop/closed-loop control equipment for automated optimization and control of the grinding line of a roller system, which can be used to perform grinding and/or crushing in an optimized and automated manner. , and which increases the functional reliability of the mill and at the same time optimizes the operation by automatically responding to anomalies that occur.

According to the present invention, the objects of the invention can be realized and achieved through the elements and combinations specifically illustrated in the independent and dependent claims. It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and do not limit the invention as described above.

В разновидности варианта осуществления, расстояния могут периодически формироваться, и алгоритм имеет возможность обнаруживать аномалию, если эффективные корреляции посредством примерных способов, описанных выше, являются анормальными.According to the present invention, the above objectives for a system and method for detecting anomalies or early indicators of equipment failure in industrial equipment or manufacturing plants by monitoring sensor or measurement data emanating from components used in industrial processes are achieved, in particular, in that that sensor and/or measurement data of components used in an industrial process are measured by measuring devices or sensors, and time frames or time periods of the same size are identified in the received data stream from sensor and/or measurement data during the time periods when the components used in an industrial process, function normally, wherein the sensor and/or measurement data contain sensor values for a plurality of measurement parameters, in that the sensor values of a plurality of measurement parameters are converted into observable binary processing codes for each of the identified time frames of the same size, and binary processing codes are assigned to a data store or data structure that stores a sequence of stored states of Markov chains, in which a multidimensional data structure is formed containing a specified number of variable parameter values of the hidden Markov model, while the variable parameters of the multivariate model are data structures are determined by a machine learning module applied to a sequence of assigned binary processing codes, and wherein the variables of the hidden Markov model of the multivariate data structure are varied and trained by learning the normal state frequency of occurring fault events based on sensor and/or measurement data of the identified time frames of the same size, in that the set of probabilistic state values is initialized and stored by applying a trained multidimensional data structure with the mentioned values values of the hidden Markov model variables to presampled binary processing codes having a time frame of the same size, identical to the time frame for the measured sensory and/or measurement data values, in that the logarithmic anomaly score threshold is determined by ordering the logarithmic result values of the stored probabilistic values. state, and that said trained multidimensional data structure with variable values of the Latent Markov Model parameters is deployed to track new measured sensor and/or measurement data from industrial equipment or plants using an anomaly score threshold to detect anomalous sensor data values , which may indicate an impending system failure, while for triggering on abnormal values of sensor data, the logarithmic resulting value of the value of ver The probabilistic state of the new measured sensory and/or measurement data is generated and compared with the stored probabilistic state values based on said logarithmic anomaly score threshold. Additionally, various methods are provided for providing binary effective correlation vector distances, such as based on the classical Hamming distance (where n=1, classical Hamming distance) for a window of n rows. For vectors a and b, the distance is equal to the number of 1s in a and b for n row windows, divided by the length of a. Another way is based on the Jaccard distance.

In a variation of the embodiment, the distances may be periodically generated and the algorithm is able to detect an anomaly if the effective correlations by the exemplary methods described above are abnormal.

It is important to note that the inventive system and method, by virtue of principle, work with and without converting an analog signal to a binary signal or code (based on thresholds). However, converting an analog signal to a binary signal or code has, inter alia, advantages such that time series anomaly detection typically relies on thresholds and moving averages and the like to detect anomalies. The result may be the detection of too many anomaly events due to fluctuating/noisy signals (which is typical in an industrial process). For example, for time series and anomaly detection, the threshold value can be adjusted such that it has greater or lesser sensitivity (see FIG. 13, in which anomalies are marked with gray vertical lines). In this approach, each of the events at which the threshold is exceeded is considered as true/1 (and false/0 otherwise) to form a binary sequence. This allows you to check the frequency of exceeding thresholds so that anomalies can be classified based on the abnormal frequency. The advantage is that there is no need to worry about over-sensitivity thresholding and fluctuating/noisy industrial IoT data. Without taking into account these advantages of binary conversion, it is also possible to use an extension of the algorithm to find anomalies in analog process data. The moving average and variance thresholds can be applied to generate a binary sequence. In such a case, the above anomaly detection algorithm can be used. The result is that anomalies are identified when the process atypically exceeds thresholds. Fig. 14 shows the process data anomaly. In FIG. 9, the binary sequence is generated based on the threshold values applied to the process data. In the next step, the described anomaly detection method is applied to the binary sequence, and the anomalous time periods are marked accordingly. It is important to note that the inventive system and method described above for converting analog signal anomalies into binary vectors and then applying statistical HMM anomaly detection, i.e. Hidden Markov Model (HMM) structures of the present invention are technically unique and not provided by prior art systems. The present invention uses the HMM for anomaly detection, while prior art systems use a different technology to flag anomalies using the HMM. In particular, prior art systems do not use the thresholding step that the present invention uses. In addition, prior art systems do not mention the conversion of analog signals to binary sequences, which is also part of the differentiator of claim 1.

A machine learning module, for example, can process a sequence of assigned binary processing codes by applying maximum likelihood parameter estimation to train a multidimensional data structure with variable parameters of a hidden Markov model, with the elements of the sequence of stored Markov chain parameter states assumed to be measurements independent of each other. , and wherein the parameters of the multivariate data structure model are varied by maximizing the multiplied probabilities product to obtain the parameters of the trained multivariate data structure model. The model parameters of a multidimensional data structure, for example, may iteratively vary as long as a predetermined convergence threshold is not exceeded. In order to determine said anomaly score threshold, an averaging process, for example, can be applied based on different frequencies of occurrence of sensory and/or measurement data alarms of the identified time frames. The invention has, among other things, such an advantage that it provides a new method and system for the automated detection of the corresponding initiation upon anomalies in data emanating from components used in industrial processes. It provides an efficient automated system for managing and monitoring large scale industrial assets (eg grain mills, food processing plants) that typically generate large volumes of process and alarm/failure data that are difficult to process.

It should be noted that the machine is stopped for every relevant alarm (eg, roller temperature limit violations, pressure thresholds exceeded between rollers, etc.), regardless of the anomaly detection systems of the invention. However, the present invention provides a new system and method for unsupervised detection of anomalies, such as those associated with industrial multivariate time series data. Unsupervised discovery can, among other things, be essential in unknown-to-unknown scenarios where operators are unaware of potential failures and do not observe prior occurrences of such unknown failures. The inventive system can also provide data quality assessment, missing value calculation and shaping, validation, and evaluation of additional or new features. The present invention allows unknown faults to be determined based on a comparison of a normal operating profile (eg, all sensors indicate values in the normal range) with reported differences in the current operating state. Sensors can be associated with various measurable elements of a piece of machinery, such as vibration, temperature, pressure and changes in environmental conditions, etc. In some cases, the definition of unknown failures is associated with the detection of a failure that should occur (for example, with early detection). In some cases, the definition of unknown faults is related to early detection, as well as other cases in which a fault may have occurred in the past, but has an impact on current operation. Additionally, the present invention provides efficient filtering and differentiation of alarms/failures that are triggered by light maintenance events on machines and are not a cause for concern. It also applies to process data, such as from motor currents, which can often show atypical values based on thresholds and are not always a cause for concern for individual events. The present invention provides efficient and automated sifting of alarm/failure data streamed from sensors and measuring devices into a final list of important events that are anomalous relative to nominal operation. This definitive list provides the basis for a new way of effective preventive maintenance and root cause analysis of downtime events that are very expensive in industrial processes and must be minimized. The invention makes it possible to identify anomalous patterns in advance, so that an uncontrolled fully automated method becomes technically possible by controlling and monitoring, through the invention, the correct operation of the machine. Thus, the invention provides unsupervised anomaly detection, in particular associated with industrial multivariate time series data. Unsupervised discovery is essential in unknown-to-unknown scenarios where operators are unaware of potential failures and do not observe prior occurrences of such unknown failures. The invention has the ability to determine unknown failures by comparing a normal operating profile or machine/engine (eg all sensors indicate values in the normal range) with reported differences in the current state of the machine/engine. Sensors may be associated with various measurable elements of a piece of machinery, such as, for example, vibration, temperature, pressure, and changes in environmental conditions, etc. In some cases, the definition of unknown failures (eg, evaluation) is associated with the detection of a failure that should occur (eg, early detection). In some other cases, the detection of unknown failures is associated with early detection also in the case in which the failure occurs in the past. In addition, the present invention allows a new way to gain understanding and see trends in sensor data and/or alarms/faults with this form of logging, which also renders continuous monitoring of alarms/faults by operators redundant. The present invention also provides new alarm/fault monitoring so that operators can easily monitor plant operation and health. The new monitoring also allows others, such as owners and maintenance providers, to gain automated insights and better communication with operators. The invention makes it possible to exclude/measure plant downtime in large-scale processing plants, since it represents a significant loss of income.

In a variation of the embodiment, the sensitivity of selected time frames, for example, may be automatically adjusted based on dynamic threshold adjustments. This variation of the embodiment has, among other things, the advantage that the rate of convergence can be optimized by learning the variables of the Hidden Markov Model of the multidimensional data structure.

In another variation of the embodiment, anomalous time frames are, for example, evaluated for a plurality of assets of identical industrial process lines, with the anomalous time frames being applied to root cause analysis of plant downtime to trigger at an anomaly score. Additionally, as a variation, maintenance alarms, for example, can be generated based on said root cause analysis of plant downtime. This variation of the embodiment has, among other things, the advantage that it provides a reliable application of the present invention for various assets and industrial process lines. Another advantage is that this variation of the embodiment provides implementation of cloud and/or network applications and signaling for automated maintenance and/or provision of services.

In yet another variation of the embodiment, to determine said anomaly score threshold, a frequency pattern is generated for each of the identified same-sized time frames using pattern recognition to initialize a plurality of Markov chain sequences of stored parameter states, with each stored parameter state representing is a function of a plurality of measurement parameters, whereby by means of the applied pattern recognition, the weighting factors and/or the average value and/or the variance of each of the plurality of sequences of stored parameter states are determined, and irrelevant time frames are removed from the used set of identified time frames of the same size. This variation of the embodiment has, among other things, the advantage that the speed of convergence can be optimized by learning the variable parameters of the hidden Markov model of the multidimensional data structure. Thus, pattern recognition and weighting coefficients provide a correlation measure of the anomalies of each variable in a noisy data sample by comparing the measured data sample with the reference data, even when some variables are highly correlated. Therefore, spurious relationships introduced by noise can be removed by focusing on significant relationships for each variable. Environment selection, for example, can be performed in an adaptive manner by fitting a sparse graphical Gaussian model as a maximum likelihood estimate. The anomaly correlation score for each measurement parameter can then be generated by the distances between the fitted conditional distributions.

In a variation of the embodiment, a strobe signal is generated, as a digital signal or pulse, providing an appropriate time window in which an emerging abnormal time frame of new measured sensory data is selected from among a plurality of measured measurement data time frames, and normal time frames should be excluded or discarded, and wherein the selection of the occurring abnormal time frame initiates the appropriate generation of an alarm and a transition to the assigned alarm and/or monitoring and/or control/coordination devices. This variation of the embodiment has, inter alia, the advantage that it provides effective machine-to-machine signaling by generating appropriate coordinating signaling controlling the operation of associated devices triggered by detected anomalies or early indicators of equipment failure in industrial equipment or manufacturing plants.

In a further variation of the embodiment, the above method and system for automated detection of anomalies or early indicators of equipment failure in industrial equipment or manufacturing plants is applied to an intelligent, self-adaptive closed-loop and open-loop control method for closed-loop control equipment and/or open-loop control for self-optimized control of the grinding plant and/or the grinding line of the grinding plant's roller system, the grinding line comprising a plurality of processing units which can be separately actuated by means of closed-loop and open-loop control equipment and separately adjusted during operation on based on process operating parameters, wherein the closed-loop and open-loop control equipment comprises a pattern recognition module based on the above-described method for detecting anomalies ii, wherein the operation of the control equipment is initiated by signaling the pattern recognition module, and the operation of the grinding plant is coordinated and adapted by the control equipment based on the transmitted trigger signal. As a variation, closed-loop and/or open-loop control equipment, for example, may comprise a batch controller having a predetermined processing sequence in the processing units that can be controlled by the process recipe and control equipment, wherein the target amount of end product is produced from one or more feedstocks by a working process recipe, wherein the processing units are controlled based on the batch process operating parameters specifically associated with the working process recipe, and where the functional batch controller is adapted or optimized based on the one or more occurring and detected anomalous time frames of new measured sensory data by the control equipment. The control equipment, for example, may further comprise a second pattern recognition module for recognizing process work recipes having multidimensional batch process parameter templates, where the process work recipe contains at least one or more raw materials, a predetermined grinding process sequence in grinding line processing units and batch process operating parameters stored in association with respective grinding line processing units, wherein closed-loop and open-loop control equipment comprises memory equipment for storing statistical process operating recipes having statistical batch process parameters, at the same time, the statistical parameters of the periodic technological process of the recipe of the technological process set a multidimensional pattern typical for the technological process a set of batch parameters for an optimized batch process in the normal range, where the entry of a new working process recipe results in the initiation and/or selection of one or more stored statistical working process recipes as the nearest batch parameter template(s) by pattern recognition by a pattern recognition module based on the associated multi-dimensional batch process parameter patterns, wherein new batch parameter patterns having new batch process parameters are generated by closed loop and open loop control equipment upon detection of one or more occurring anomalous time frames of new measured sensory data by control equipment based on triggered near steps blocks of normal batch process parameters, wherein the processing units are actuated and controlled by closed loop and open loop control equipment appropriately based on generated process operating recipes having associated batch process parameters. This variation of embodiment has, among other things, the advantage that it provides intelligent, self-adaptive open-loop/closed-loop control equipment for automated optimization and control of the grinding line of the roller system, which can be used to perform grinding and/or or crushing in an optimized and automated way, and which increases the functional reliability of the mill and at the same time optimizes the operation or reacts automatically to anomalies that occur.

It should be noted that sensors and measuring devices as field devices are used to read or control process variables of industrial equipment or plant in an industrial process. However, in some installations, it may be desirable to keep track of the local environment of the field device. Additionally, it should be noted that the anomaly detection system or method, for example, may include an additional threshold value estimation module to store an upper threshold value and a lower threshold value according to the received technical condition data for one or more respective sensor signals. The threshold value evaluation module compares the received health data with threshold values and generates, regardless of the evaluation by the anomaly detection system or method, an anomaly indicator for the specific health data if the corresponding data value is outside the interval specified by the respective upper and lower thresholds. . In other words, evaluating sensor data based on thresholds can provide a quick way to detect an anomaly indicator. If a particular health data value is immediately outside the tolerance range defined by the upper lower thresholds, the corresponding anomaly indicator is detected regardless of what the anomaly detection result provides. For example, the thresholds may be pre-set (eg, by an operator) based on prior experience, or the thresholds, as fast method values, may be recognized by a machine learning module from sensor statistics.

Additional aspects of the invention are to be realized and carried out by means of the elements and combinations specifically illustrated in the appended claims. It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and do not limit the invention as described above.

Brief description of the drawings

In the following, the present invention will be explained in more detail, by way of example, with reference to the drawings, in which:

Fig. 1 shows a diagram schematically illustrating a monitoring and adaptation workflow in an industrial plant, with intelligent, self-adaptive control equipment for self-optimized control of an industrial plant.

Fig. 2 shows a diagram schematically illustrating a monitoring and adaptation workflow in an industrial plant. Data from the sensors is transmitted periodically, for example, every 3 minutes.

Fig. 3-12 illustrate plant failure/suspension (downtime) events in the premium grinding section of the exemplary plant of FIG. 2. Serious failure/suspension events due to mechanical errors in the grinding/cleaning/first cleaning sections are illustrated in the overall operating status view. Frequent failures at the individual sensor level are shown for 2017 and 2018 sample data, respectively. Anomalies are illustrated in the grind level section and for individual sensors. Alarms for optimization and preventive maintenance are also illustrated.

Fig. 3 shows a diagram schematically illustrating the monitoring of the approximate grind yield over the 2017-2018 time period.

Fig. 4 shows a diagram schematically illustrating the monitoring of an exemplary F1 output over the 2017-2018 time period.

Fig. 5 shows a diagram schematically illustrating the monitoring of an exemplary summary of grinding suspensions in the grinding section over the 2017-2018 time period. The total number of suspension events in 2018 is, in this example, 80 with a total duration of 2 days 27 hours 11 minutes. The longest suspension in the grinding section is 14 hours 51 minutes on 01/01/2018, the total number of suspension events in 2017 is 275 with a total duration of 9 days 8 hours 58 minutes. The sum excludes missed events.

Fig. 6a-6o show a diagram schematically illustrating error monitoring of an exemplary grinding plant over the time period November 01 to November 30, 2017.

Fig. 7 shows a diagram schematically illustrating a summary of the error monitoring frequency of an exemplary grinding plant over the time period 2017-2018. The drawing shows the plant shutdown events for failures due to mechanical errors in the second cleaning section (MUEPS001), the grinding section (MUEPS002) and the first cleaning section (RE1PS001) as a number per week. The chart above shows missing data, larger than 10 minutes. The vertical bars indicate when failures occur and are incremented (by 15 hours) to allow visibility of short failure events on the timeline. Thicker vertical lines indicate longer failure events or multiple short failure events close to each other. Failures shorter than 3 minutes are excluded (data sampled every 3 minutes).

Fig. 8a and 8b show a diagram schematically illustrating exemplary first 10 alarms for failures by duration in 2018.

Fig. 9 shows a diagram schematically illustrating error monitoring in a cleaning section of an exemplary plant over a time period of 2017-2018. The inventive system and method provides for the generation of an alarm if preventive maintenance or additional monitoring is required for the color sorter. The inventive anomaly detection method has the ability to identify equipment that requires preventive maintenance or monitoring. The graph shows some preliminary results, with abnormal weeks marked in orange for the cleanup section. Fig. 9 shows preliminary results showing abnormal failure rates. The vertical bars indicate when failures occur and are enlarged slightly to allow visibility of short failure events on the timeline. The colored bar indicates the classification of each fault signal. The time period marked in gray is assumed to be typical plant operation. Time periods marked in blue are classified as normal operation, while weeks marked in orange are classified as abnormal. Missing data periods of time are classified as normal operation and are not indicated. It should be noted that the color sorter, the WT high level sensor, the flow balancing unit 203, and the cleaning section are just examples for various machines. Thus, in FIG. 9, the color sorter may also be collectively referred to as "machine 1", the high-level sensor WT as "machine 2", the flow balancing unit 203 as "machine 3", and the cleaning section as "machine 4" of the overall system.

Fig. 10 is a diagram schematically illustrating the monitoring of errors in the cleaning section of an exemplary grinding plant by the location of a sensor in the plant. Number references in the form A-xxxx designate sensors and measuring devices that capture measurement data during operation of the mill cleaning section and are used at various locations in processing.

Fig. 11 is a diagram schematically illustrating the monitoring of errors in the cleaning section of an exemplary grinding plant by means of scales and flow balancing modules.

Fig. 12 shows a diagram schematically illustrating error/failure correlation in sensor and measurement data. The inventive system and new method for control equipment detecting anomalies in plant operation can handle a high number of correlation and sensory values. Chord plot o FIG. 12 shows failures occurring simultaneously, implying a possible correlation between mechanical failures.

Fig. 13 shows a diagram schematically illustrating a variant of an embodiment of the present invention with binary conversion of workflow data. Typically, time series based anomaly detection algorithms rely on thresholds and moving averages and the like to detect anomalies. The result may be the detection of too many anomaly events due to fluctuating/noisy signals (which is typical in an industrial process). In the present invention, it is possible to adjust the threshold value so that it has greater or lesser sensitivity (see Fig. 13, in which anomalies are marked with vertical lines). Each of the events at which the threshold is exceeded is considered as true/1 (and false/0 otherwise) to form a binary sequence. This allows the frequency of threshold exceedances to be checked so that anomalies can be classified based on the abnormal frequency. The technical advantage is that there is no need to worry about the hypersensitivity of threshold and fluctuating/noisy industrial IoT data.

Fig. 14 shows a diagram schematically illustrating process data anomaly detection, wherein in a first step a binary sequence is generated based on thresholds applied to the process data (see FIG. 14). In the second step, the anomaly detection described herein is applied to the binary sequence and the anomalous time periods are flagged accordingly. Converting analog signal anomalies to binary vectors and then applying the HMM (Hidden Markov Model) based statistical anomaly detection structure of the present invention is technically unique and cannot be extracted from any of the prior art systems.

Fig. 15 shows other diagrams fundamentally illustrating process data anomaly detection based on downtime and error sensor data, with the left column showing downtime measurements in different sections over time, and the right column showing in the top diagram measured machine failures across all machines, in the middle diagram, the time spent in the error state by the machine and in the lower diagram, the error rate measured per day.

Fig. 16 shows another diagram principally illustrating error/failure correlation in sensor and measurement data for the example shown in FIG. 15. FIG. 16 illustrates how the inventive system and method for detecting anomalies or early indicators of equipment failure in industrial equipment or manufacturing plants triggered by sensory or measurement data originating from components used in industrial processes provides appropriate coordinating signals based on the detected and measured alert frequencies and correlations and anomalies. Thus, the inventive system provides a technically completely new way of initiating correlations between alarm events and/or visualization of alarm events in time and/or detection of anomalies in the form of abnormal downtime/alarms, and/or playback of alarms and corresponding generation of electronic alarms.

Claims (23)

1. A method for detecting anomalies or early indicators of equipment failure in industrial equipment or manufacturing plants (1) by monitoring measurement data and/or process parameters (4) coming from components used in industrial processes (6), characterized in that that it contains steps in which: - measuring and/or monitoring measurement data, respectively, tracking parameters (4) of the technological process of the components used in the industrial process (6), by means of measuring devices or sensors (2), and identifying time frames of the same size in the measurement parameters (4) and /or technological process for time frames in which the components used in the industrial process (6) are functioning normally, and the parameters (4) of measurements and/or the technological process contain parameter values for a plurality of measurement/sensor parameters (41) and/or variables (42) technological process, - converting the values (4) of the parameters of the set of measuring/sensor parameters (41) and/or variables (42) of the technological process into observed binary processing codes for each of the identified time frames of the same size and assigning binary processing codes to the sequence of stored states of the Markov chains, - forming a multidimensional data structure containing a predetermined number of values of the variable parameters of the hidden Markov model, while the variable parameters of the model of the multidimensional data structure are determined by the machine learning module (8) applied to the sequence of stored states of the Markov chains with assigned binary processing codes (91), and at the same time, the variable parameters (811, 812, ..., 81x) of the hidden Markov model of the multidimensional data structure vary and are trained by learning the frequency (82) of the normal state of emerging emergency events based on the measurement data and/or parameters (4) of the technological process of the identified time frames of the same size, - initialize and store a set of values (83) of a probabilistic state by applying a trained multidimensional data structure with values of variable parameters of the hidden Markov model to presampled binary processing codes having a time frame of the same size, identical to the time frame for the values (4) of the parameters of the set of measurement/sensor parameters (41) and/or process variables (42), determining a logarithmic anomaly score threshold by ordering the logarithmic result values of the stored probabilistic state values (83), and deploying said trained multidimensional data structure with variable parameter values of the hidden Markov model in order to track new measured respectively determined measurement data and/or process parameters (4) from industrial equipment or plants (1) using an anomaly score threshold, to detect anomalous sensor data values that may indicate an impending system failure, while in order to trigger at anomalous sensor data values, the logarithmic result value for the probabilistic state value (83) of the new measured, respectively, determined measurement data and/or process parameters (4) process is formed and compared with the stored values (83) probabilistic state on the basis of the mentioned logarithmic threshold score of anomalies. 2. A method for detecting anomalies or early indicators of equipment failure in industrial equipment or manufacturing plants (1) according to claim 1, wherein processing binary codes (91) are generated based on threshold values (92) applied to measurement data and/ or parameters (4) of the technological process. 3. A method for detecting anomalies or early indicators of equipment failure in industrial equipment or manufacturing plants (1) according to claim 2, characterized in that for detection of anomalies having too many anomaly events due to fluctuating/noisy signals, the sensitivity for event detection is automatically adjusted based on dynamic adjustments of one of the threshold values (92), whereby if the threshold value (92) is exceeded, the event takes on a value of 1 (" true"), and 0 (false) otherwise, to form a binary sequence. 4. Method for detecting anomalies or early indicators of equipment failure in industrial systems according to one of paragraphs. 1-3, characterized in that anomalous time frames are measured for a plurality of assets of identical industrial process lines, while in order to trigger at the anomaly quantity, anomalous time frames are applied to the root cause analysis of plant downtime, and maintenance and repair alarms are generated based on of the aforementioned root cause analysis of plant downtime. 5. A method for detecting anomalies or early indicators of equipment failure in industrial equipment or manufacturing plants (1) according to one of paragraphs. 1-4, characterized in that the machine learning module (8) processes the sequence of assigned binary processing codes (91) by applying parameter estimation according to the maximum likelihood principle for training a multidimensional data structure with variable parameters of the hidden Markov model, while the elements of the sequence of stored parameter states Markov chains are assumed to be measurements independent of each other, and the parameters of the multivariate data structure model are varied by maximizing the multiplied product of probabilities to obtain the parameters of the trained multivariate data structure model. 6. A method for detecting anomalies or early indicators of equipment failure in industrial equipment or manufacturing plants (1) according to claim 5, characterized in that the parameters of the model of the multidimensional data structure (82) iteratively vary until a predetermined threshold value convergence is not exceeded. 7. A method for detecting anomalies or early indicators of equipment failure in industrial equipment or manufacturing plants (1) according to one of paragraphs. 1-6, characterized in that, in order to determine the mentioned threshold value of the quantitative indicator of anomalies, the averaging process is applied on the basis of different frequencies of occurring emergency events (61) of the measurement parameters (4) and/or the technological process of the identified time frames. 8. A method for detecting anomalies or early indicators of equipment failure in industrial equipment or manufacturing plants (1) according to one of paragraphs. 1-7, characterized in that in order to determine the mentioned threshold value of the quantitative indicator of anomalies, the frequency pattern (612) of the frequency (611) of occurring emergency events (61) is formed for each of the identified time frames of the same size using pattern recognition to initialize a plurality of circuit sequences Markov of stored parameter states, where each stored parameter state is a function of a set of measurement data and/or process parameters (4), wherein, through the applied pattern recognition, weights and/or average and/or variance of each from the set of sequences of stored parameter states, and irrelevant time frames are removed from the used set of identified time frames of the same size. 9. A method for detecting anomalies or early indicators of equipment failure in industrial equipment or manufacturing plants (1) according to one of paragraphs. 1-8, characterized in that a strobe signal is generated as a digital signal or pulse, providing an appropriate time window in which the emerging anomalous time frame of new measured measurement data and/or process parameters (4) is selected from among the set of measured time frames measurement data and/or process parameters (4), and normal time frames shall be excluded or discarded, and the selection of the resulting abnormal time frame initiates the appropriate generation of an alarm and transition to the assigned alarm and/or monitoring and/or control/ coordination (14). 10. A method for detecting anomalies or early indicators of equipment failure in industrial equipment or manufacturing plants (1) according to claim 9, characterized in that an electronic control and coordinating alarm is generated, while based on the detected emerging anomalous time frame, new measured ones are selected measurement data from among the plurality of measured time frames of measurement data, and normal time frames should be excluded or discarded, and at the same time, the selection of at least one emerging anomalous time frame initiates the corresponding generation of an alarm and a transition to the regulation of the operation of industrial equipment and / or plants - manufacturers (1) or component through control/coordination devices (14). 11. System (7) for detecting anomalies or early indicators of equipment failure in industrial equipment or manufacturing plants (6) by monitoring measurement data and/or process parameters (4) coming from components used in industrial processes (6) , which is different: - by the fact that the system (7) contains sensors or measuring devices (2/13) for measuring measurement data and/or process parameters (4) of the components used in the industrial process (6) and detection devices for identifying time frames of the same size in the measurement data and/or process parameters (4) for time frames in which the components used in the industrial process (6) function normally, and the measurement data and/or process parameters (4) contain parameter values for a plurality of measurement/ sensor parameters (41) and/or process variables (42), - the fact that the system (7) contains a differentiator (9) for converting the values (4) of the parameters of the set of measuring/sensor parameters (41) and/or variables (42) of the technological process into observable binary processing codes for each of the identified time frames of the same size and assignment of binary codes for processing the sequence of stored states of Markov chains, - by the fact that the system (7) contains a machine learning module (8) for generating a multidimensional data structure containing a specified number of values of the variable parameters of the hidden Markov model, while the variable parameters of the model of the multidimensional data structure are determined by the module (8) of machine learning applied to sequences of stored states of Markov chains with assigned binary processing codes (91), while the variable parameters of the hidden Markov model of the multidimensional data structure are varied and trained by learning the frequency (82) of the normal state of occurring emergency events based on the measurement data and/or parameters (4) technological process of identified time frames of the same size, - by the fact that the machine learning module (8) contains a means for initializing and saving a set of values (83) of a probabilistic state by applying a trained multidimensional data structure with values of variable parameters of the hidden Markov model to presampled binary processing codes having a time frame of the same size, identical to the time frame frame for the values (4) of the parameters of the set of measuring/sensor parameters (41) and/or variables (42) of the technological process, - in that the machine learning module (8) comprises means for determining a logarithmic anomaly score threshold by ordering the logarithmic result values of the stored probabilistic state values (83), and - by the fact that the machine learning module (8) contains means deploying said trained multidimensional data structure with the values of variable parameters of the hidden Markov model in order to track new measured, respectively, certain measurement data and/or process parameters (4) from industrial equipment or plants (1) using an anomaly score threshold to detect abnormal sensor data values that may indicate an impending system failure, while to trigger sensor data abnormal values, the logarithmic result value for the probabilistic state value (83) of the new measured parameters (4 ) measurement and/or process is generated and compared with the stored values (83) of the probabilistic state based on said logarithmic anomaly score threshold.

Images