RU2795745C1 - Sensor-independent machine fault identification - Google Patents

Abstract

FIELD: monitoring machine tools.

SUBSTANCE: invention is related in particular to identification of faults in machine tools by monitoring them. A method for identifying a malfunction of at least one machine tool includes configuring the first plurality of sensors connected to the corresponding first plurality of machine tools to receive the first plurality of signal sets emanating from the first plurality of machine tools, the first plurality of machine tools having at least one common characteristic, supplying at least the first plurality of sets of signals of the first plurality of machine tools to an existing fault classifier, previously trained with the possibility of automatically identifying faults of the second plurality of machine tools based on signals emanating from them and previously received by the second plurality of sensors, the second plurality of sensors is of a different type than the first plurality of sensors, the second plurality of machine tools has at least one said common characteristic, modification of the specified existing fault classifier by using transfer learning based on at least the first plurality of signal sets of the first plurality of machine tools, thereby providing a modified fault classifier, applying the modified fault classifier to at least one additional set of signals received by at least one sensor from the first plurality of sensors and coming from at least one given machine tool having at least one said common characteristic, wherein the modified fault classifier is configured to automatically identify at least one fault of at least one given machine tool based on at least one said additional set of signals, and provide human readable output data using an output device containing at least an identification of malfunction of at least one said given machine tool, and at least one repair or maintenance operation is performed based on human readable output data.

EFFECT: solving the problem of limited use of a fault classifier trained on data received from sensors of a certain type or types, for classifying signals received from sensors of different types or types, due to differences in the characteristics of the sensors, providing the possibility of transferring learning to match between the original type of sensor, on the basis of which the existing classifier was trained, and new different type of sensors, from which new data are obtained that require classification.

24 cl, 11 dwg

Description

RELATED APPLICATIONS

Reference is hereby made to U.S. Provisional Application No. 62/895,247 entitled "MULTI-SENSOR COMPLIANCE SYSTEMS AND METHODS FOR AUTOMATED MACHINE DIAGNOSIS", dated September 3, 2019, the disclosure of which is hereby incorporated by reference and priority hereby claimed under 37 CFR 1.78(a)(4) and (5)(i).

FIELD OF TECHNOLOGY

The present invention relates generally to the monitoring of machine tools, and in particular to the identification of faults in machine tools by tracking them.

BACKGROUND OF THE INVENTION

Various systems and methods for identifying faults in machine tools are known in the art.

DISCLOSURE OF THE INVENTION

The present invention is directed to providing new systems and methods for using transfer learning between various types of sensors that monitor signals from machine tools for the purpose of identifying faults and maintaining the monitored machine tools.

Thus, in accordance with a preferred embodiment of the present invention, there is provided a method for identifying a malfunction of at least one machine tool, comprising causing a first plurality of sensors connected to a respective first plurality of machine tools to receive a first plurality of signal sets emanating from the first plurality of machine tools, wherein the first plurality of machine tools has at least one common characteristic, the supply of at least the first plurality of sets of signals of the first plurality of machine tools to an existing fault classifier, previously trained to automatically identify faults of the second plurality of machine tools based on signals emanating from them and previously received a second set of sensors, the second set of sensors being of a different type than the first set of sensors, the second set of machine tools having said at least one common characteristic, modifying said existing fault classifier by using transfer learning based on at least the first set of signal sets of the first a plurality of machine tools, thereby providing a modified fault classifier, applying the modified fault classifier to at least one additional set of signals received by at least one sensor from the first set of sensors and originating from at least one given machine tool having the specified at least one general characteristic, wherein the modified fault classifier is configured to automatically identify at least one fault of at least one given machine tool based on said at least one additional set of signals, and provide human-perceptible output data using an output device containing at least as identifying a malfunction of said at least one predetermined machine tool, of at least one repair or maintenance operation performed based on human-perceptible output data.

In accordance with one preferred embodiment of the present invention, the method also includes the following steps: causing the first set of sensors to receive the first set of signal sets and before applying the first set of signal sets to the existing fault classifier: obtaining the first set of operating data sets for machine tools from the first set of machine tools, wherein each set of operating state data indicates an operating state of a machine tool from a first set of machine tools, each operating state is associated with at least one of said signal sets, said supply of at least the first set of signal sets to an existing fault classifier also includes supplying said operating state data of the first set of machine tools to the existing fault classifier, said modification of the existing fault classifier by applying transfer learning based on at least the first set of signal sets also includes modifying the existing fault classifier by applying transfer learning further based on the first set of sets operating status data of the first plurality of machine tools.

Preferably, identifying the failure includes identifying a specific failure of said at least one given machine tool and predicting failure of said at least one given machine tool due to said specific failure in the absence of a recommended maintenance operation on it, such that said at least one given machine tool the machine will indeed fail if the recommended maintenance operation is not performed.

Preferably, the existing fault classifier comprises a neural network, and the modification of the existing fault classifier comprises adding at least one matching layer to the neural network, wherein apart from adding said at least one matching layer, the neural network is otherwise unmodified by said modification.

Preferably, a neural network comprising an existing fault classifier comprises a data layer and an input layer for receiving data from the data layer, wherein said at least one matching layer is added between the data layer and the input layer, and said at least one matching layer is configured to receiving data from the data layer in the modified fault classifier.

Preferably, the first plurality of sensors has a first frequency response distribution and the second plurality of sensors has a second frequency response distribution, wherein the matching layer is configured to match between the first and second frequency response distributions.

In accordance with one preferred embodiment of the method according to the present invention, the first set of sensors is configured to receive the same type of signal as the second set of sensors.

Preferably, said same type of signal comprises one of a vibration signal, a magnetic flux signal, a current signal, a temperature signal, and an internal pressure of the machine.

In accordance with another preferred embodiment of the present invention, the first plurality of sensors and the second plurality of sensors are configured to receive mutually different types of signals.

Preferably, the mutually different types of signals comprise at least one of the following signals: vibration and magnetic flux signals, vibration and electric current signals, vibration and temperature signals, electric current and magnetic flux signals, and vibration and machine internal pressure signals.

Preferably, at least some of the operating states of the machine tool of the first plurality of machine tools, as indicated by the first plurality of operating state data sets, are faulty operating states.

According to another preferred embodiment of the present invention, the first plurality of signal sets and the first plurality of operating state data sets of the first plurality of machine tools contain less than 30 faulty operating states.

In accordance with another preferred embodiment of the present invention, there is further provided a system for identifying faults in at least one machine tool, comprising a first plurality of sensors connected to a respective first plurality of machine tools and configured to receive a first plurality of sets of signals emanating from the first plurality of machine tools. , wherein the first plurality of machine tools have at least one common characteristic, a data processing unit configured to: receive a first plurality of signal sets of the first plurality of machine tools, the data processing unit comprising an existing fault classifier pre-trained to automatically classify operating states of the second a plurality of machine tools based on signals emanating from them and previously obtained by a second set of sensors, the second set of sensors being of a different type than the first set of sensors, the second set of machine tools having at least one common characteristic, modifying said existing fault classifier by using transfer learning based on at least a first plurality of signal sets of a first plurality of machine tools, thereby providing a modified fault classifier, and applying the modified fault classifier to at least one additional set of signals received by at least one sensor from the first plurality of sensors and output from at least one given machine tool having said at least one common characteristic, wherein the modified fault classifier is configured to automatically identify at least one fault of said at least one given machine tool based on said at least one additional set of signals , and an output device in communication with the data processing unit and configured to provide human-perceptible output data containing at least an identification of a malfunction of said at least one specified machine tool, at least one repair or maintenance operation performed on based on human-perceptible output data.

In accordance with one preferred embodiment of the present invention, the system also comprises a data acquisition unit configured to obtain a first plurality of machine tool operating state data sets from the first machine tool plurality, each work state data set indicative of a machine tool operating state from of the first set of machine tools, each operating state is associated with at least one of the specified sets of signals, and the data processing unit is configured to receive data on the operating state of the first set of machine tools and with the possibility of further modifying the specified existing fault classifier based on the first set of data sets on the working condition of the first set of machine tools.

Preferably, identifying the failure includes identifying a specific failure of said at least one given machine tool and predicting failure of said at least one given machine tool due to said specific failure in the absence of a recommended maintenance operation on it, such that said at least one given machine tool the machine will indeed fail if the recommended maintenance operation is not performed.

Preferably, the existing fault classifier comprises a neural network, and the modification of the existing fault classifier comprises adding at least one matching layer to the neural network, wherein apart from adding said at least one matching layer, the neural network is otherwise unmodified by said modification.

Preferably, a neural network comprising an existing fault classifier comprises a data layer and an input layer for receiving data from the data layer, wherein said at least one matching layer is added between the data layer and the input layer, and said at least one matching layer is configured to receiving data from the data layer in the modified fault classifier.

Preferably, the first plurality of sensors has a first frequency response distribution and the second plurality of sensors has a second frequency response distribution, wherein the matching layer is configured to match between the first and second frequency response distributions.

In accordance with one preferred embodiment of the system according to the present invention, the first set of sensors is configured to receive the same type of signal as the second set of sensors.

Preferably, said same type of signal comprises one of a vibration signal, a magnetic flux signal, a current signal, a temperature signal, and an internal pressure of the machine.

In accordance with another preferred embodiment of the system according to the present invention, the first set of sensors and the second set of sensors are configured to receive mutually different types of signals.

Preferably, the mutually different types of signals comprise at least one of the following signals: vibration and flux signals, vibration and electric current signals, vibration and temperature signals, electric current and magnetic flux signals, and vibration and machine internal pressure signals.

Preferably, at least some of the operating states of the machine tool of the first plurality of machine tools, as indicated by the first plurality of operating state data sets, are faulty operating states.

According to another preferred embodiment of the system according to the present invention, the first set of signal sets and the first set of operating state data sets of the first set of machine tools contain less than 30 faulty operating states.

BRIEF DESCRIPTION OF THE DRAWINGS

The present invention will be more fully understood and appreciated from the following detailed description in conjunction with the drawings, in which:

In FIG. 1 is a simplified high-level block diagram illustrating a machine tool troubleshooting system designed and implemented in accordance with a preferred embodiment of the present invention;

In FIG. 2 is a simplified high level block diagram illustrating a system for identifying faults in machine tools constructed and implemented in accordance with another preferred embodiment of the present invention;

In FIG. 3 is a simplified illustration of a modification of a fault classifier made by a system of the type shown in FIG. 1 or fig. 2;

In FIG. 4A and 4B are respective simplified block diagram illustrations of the components of respective systems for training the fault classifier used and modified by the system of FIG. 1 or fig. 2;

In FIG. 5A-5D are graphs depicting the performance of a system of the type shown in FIG. 1 or fig. 2, to identify the failure of machine tools; And

In FIG. 6 is a simplified flowchart illustrating the steps of a machine tool fault identification method in accordance with a preferred embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

Let us now turn to FIG. 1 which shows a simplified high level block diagram of a system for identifying a machine fault, designed and implemented in accordance with a preferred embodiment of the present invention.

As seen in FIG. 1, a system 100 is provided for identifying a machine tool failure. System 100 preferably includes a first plurality of sensors 102 of the first type, referred to here as sensor type S1. All sensors 102 may be of the same type, such as a type S1 sensor as shown here. Alternatively, the first set of sensors 102 may comprise more than one type of sensor, such as sensor types S1, S11 - S1N, as shown in FIG. 2, described in more detail below. Preferably, sensors 102 are connected to a respective first plurality of machine tools 104. Here, by way of example, the first plurality of machine tools 104 is shown to comprise machine tools 1, 2 - N, where N can be any number of machine tools, such as two or more machine tools. machine tools. Typically, sensors 102 are connected to machine tools 104 in a one-to-one arrangement in which one of the sensors 102 is connected to a corresponding one of the machine tools 104. However, other arrangements are possible in which one sensor can sense signals from more than one machine tool 104 The sensors 102 may be in physical contact with the machines 104, such as being directly or indirectly mounted on the machines 104. Alternatively, the sensors 102 may be physically separated from the machines 104, such as located at a predetermined distance from the machines 104, such as when when the sensors 102 are optical sensors.

Sensors 102 may be any type of sensor devices capable of sensing signals from machine tools 104. Machine tools and mechanical systems with moving parts, such as machine tools containing bearings, rotors or shafts, or motors, compressors, pumps, fans, gearboxes, coolers, etc., can generate signals during their operation. Machine tools 104 may refer to any of the above types of mechanical systems, or other types of mechanical systems that generate signals during their operation. Preferably, sensors 102 are configured to sense such signals. The analysis of the sensed signals can be used to determine the state of the machine from which the perceived signal originates, and in some cases to determine the malfunction of the machine from which the signal originates. By way of non-limiting example only, all sensors 102 may be of the same type of vibration sensors, for example, all may be single-axis accelerometers, or all may be multi-axis accelerometers, to sense vibrations emanating from machine tools 104; all sensors 102 may be the same type of magnetic flux sensors, sensing the magnetic flux emanating from the machines 104; all sensors 102 may be the same type of electric current sensors sensing the change in electric current generated by the machines 104; all sensors 102 may be the same type of temperature sensors sensing heat generated by machines 104. It should be understood that alternatively, all sensors 102 may be the same type of any other type of sensor devices configured to sense signals emanating from machines 104 and generated them, including signals related to machine operation such as torque, displacement, input line frequency, etc. Alternatively, sensors 102 may comprise two or more types of sensors, such as, by way of example only, magnetic flux sensors and vibration sensors.

Machine tools 1-N, which are members of the first set of machine tools 104, preferably have one or more common characteristics. 1-N machine tools may or may not be the same machine, provided they share at least one feature in common. For example, general characteristics may refer to a type, model number, manufacturer, physical characteristics or dimensions, performance characteristics or parameters, or other general characteristics that indicate that the observed behavior of one of the machine tools from the specified set of machine tools may be typical of another. a machine tool from the specified set of machine tools.

Preferably, the first plurality of sensors 102 is configured to receive a first plurality of sets of signals from the first plurality of machine tools 104. The signal sets may be "instantaneous signals" sensed by a respective one of the sensors 102 for a short period of time. For example, the signal may be perceived within a few seconds, such as one to four seconds. Alternatively, each set of signals may comprise a plurality of "snapshots" over a period of time, such as four second "snapshots" measured every hour over a period of several hours, days, or even months. Alternatively, each set of signals may contain continuously monitored signals over a longer period or more than one period of time. For example, a signal can be continuously monitored every millisecond.

The signals received by the sensors 102 may be pre-processed, for example, through the analog or digital processing capabilities of the sensor 102 itself or other hardware and/or software signal processing components 108. It should be understood that although the signal processor 108 is shown in FIG. 1 as a separate element from sensors 102, signal processing functionality may be built into one or more sensors 102. Signal pre-processing may include at least one of the following: digitization, compression, feature extraction, and time or frequency domain representation of the signal.

In one embodiment of the present invention, the signal sets received by the first set of sensors 102 may be uploaded to a remote server, such as a cloud server. Signal processing functionality 108 may be performed on a remote server. Preferably, the signal sets received by the first plurality of sensors 102 are accumulated as they are received from the first plurality of machine tools 104. For example, the signal sets may be accumulated on a cloud server.

System 100 may further comprise a data acquisition unit 110. The data acquisition unit 110 may be configured to receive a first plurality of operating state data sets for machine tools from a plurality of machine tools 104, each operating state data set indicating a machine operating state from the first plurality of machine tools 104, and each operating state associated with at least one of said sets of signals received by the first set of sensors 102.

The operating condition data collected in the data acquisition unit 110 is preferably in the form of machine condition diagnoses provided by human experts such as engineers. The signal sets received by the first set of sensors 102 may optionally be provided to the data acquisition unit 110 either by the signal processor 108 and/or directly or indirectly by the sensors 102. Human experts may optionally analyze the signal sets received by the first set of sensors 102, and labeling each set of signals from said sets of signals as representing specific operating states of the respective machine tool 104 from which the signals originate. The human experts may interact with a user interface, such as the data acquisition unit 110 or other device that provides communication between the human expert and the data acquisition unit 110, to enter operating status data. The signals and the labels applied to them can be accumulated and stored in a database located in the data acquisition block 110 . In one embodiment of the present invention, data acquisition unit 110 may be located on a remote server, such as a cloud server.

The identification of faults by human experts may include the identification of one or more specific faults of the machine being monitored 104. Depending on the particular machine being monitored 104, the particular fault identified may include worn rotating machine bearings, mechanical weakness, misalignment, imbalance, electrical faults, or other faults. . Alternatively, fault identification may include identifying the machine 104 as being in a faulty state, that is, in an abnormal state with respect to its normal operating state, but without identifying a specific fault. In this case, the identification of faults by a human expert identifies the machine as not operating normally, but does not identify what is the specific cause of this abnormal operation.

Whether or not the specific operating states represented by the signals are labeled by human experts, the signal sets received by the sensors 102 may comprise both signal sets corresponding to normal good operating states of machine tools 104 and signal sets corresponding to abnormal abnormal operating states of machine tools 104. .

Alternatively, the signal sets received by the sensors 102 optionally comprise signal sets corresponding to abnormal abnormal operating states of the machines 104. In accordance with this embodiment of the present invention, all the machines 104 monitored by the sensors 102 may be in a healthy working condition. Thus, all sets of signals and optionally associated machine operating states can correspond to normal machine operating states.

The sets of signals accumulated from the first plurality of machines 104 are preferably provided to an existing fault classifier 114 contained in the system 100. In the case where the system 100 also includes a data acquisition unit 110 and the signals are tagged, the tagged signals are preferably shared by the signal processing preprocessor 108 and block 110 data collection on the existing classifier 114 faults.

The existing fault classifier 114 may be an algorithmic classifier. For example, an existing fault classifier 114 may be stored on a remote server. The system 100 may include a non-volatile computer-readable storage medium that stores computer-executable instructions for the processor to perform the functionality of an existing fault classifier 114. The one or more processors executing the existing fault classifier 114 may be remote processors, such as those located in the cloud, or may be local processors.

Preferably, the existing fault classifier 114 is a fault classifier that has been pre-trained to automatically identify faults of the second plurality of machine tools based on signals emanating from them and previously received by the second plurality of sensors, the second plurality of sensors being of a different type than the first plurality of sensors. 102. All of the second set of sensors may be of the same type, and this type may be different from the type or types of sensors of the first set of sensors 102. Alternatively, the second set of sensors may be of the same type, and this type may differ from at least one of the types of sensors of the first a plurality of sensors 102. Alternatively, the second plurality of sensors may comprise more than one type of sensor, which types may be different types than the sensor or sensors of the first plurality of sensors. Alternatively, the second set of sensors may contain more than one type of sensors, and these types may differ from at least one of the types of sensors of the first set of sensors 102. Preferably, the second set of machine tools has the specified at least one common mechanical characteristic common to the first set machine tools 104.

It should be understood that the fault classifier 114 is referred to herein as "existing" because it may be existing with respect to the signal sets and optional operating status data received by the sensors 102 from the machines 104. The fault classifier 114 may be pre-generated at an earlier point in time, before generating a data set containing signal sets and optional data on the operating status of machines 104.

Preferably, the fault classifier 114 is an accurate classifier configured to accurately identify machine tool faults having the at least one common characteristic based on signals received by the second plurality of sensors. The fault classifier 114 may be such an accurate classifier because it has been pre-trained using machine learning on a large dataset containing signals and possibly related operating status data from a large number of machines. As is well known to those skilled in the art, the more data that is fed into a machine learning fault classifier to train it, the more accurately that classifier can perform up to a given limit. For example, a fault classifier 114 can be trained using data from over 40,000 individual rotary machines containing motors, pumps, fans, coolers, compressors, and gearboxes. A common characteristic of such machines may be the presence of bearings in them. An example of how the failure classifier 114 could be pre-trained is shown in FIG. 3-4B described below.

Thus, the existing fault classifier 114 can be successfully applied to machine tool signals having common characteristics with machine tools based on the data from which the fault classifier 114 has been trained to identify their faults. However, it should be noted that the fault classifier 114 has been pre-trained based on signals received by a sensor of a certain type or types, namely the second set of sensors of a type different from the type of the first set of sensors 102. As a result, the fault classifier 114 is configured to most successfully classify and identify faults when applied to signals received by specific sensors, based on which the fault classifier 114 was trained. However, in the case where the signals supplied to the fault classifier 114 are obtained by different types of sensors than the second plurality of sensors from which the fault classifier 114 was trained, the fault classifier 114 will not be able to accurately classify and identify faults based on these signals. This is due to the difference in sensor characteristics between these sensors, i.e. the second set of sensors, from which the fault classifier 114 was trained, and the sensors, for example, the first set of sensors S1 or the first set of sensors S1-S1N (Fig. 2), which received the present signal requiring classification.

Even though the fault classifier 114 is a high-precision classifier for identifying faults based on signals received by the second sensor set, the fault classifier therefore has limited, if any, use in identifying faults received by different types of sensors, such as the first sensor set S1. or S1-S1N (Fig. 2). If the fault classifier 114 is applied to the signals received by the first set of sensors, the results will be inaccurate.

This can be illustrated by two types of vibration sensors, such as a three-axis accelerometer and a single-axis accelerometer, which sense vibration signals generated by a machine tool. These two types of vibration sensors differ from each other in various parameters such as geometry, mass, internal materials, etc., resulting in differences in the moments of inertia and resonant frequencies of the respective sensors. As a result, the sensors have mutually different frequency responses. The particular signal generated by the monitored machine tool will be perceived and recorded differently by the two sensors due to their inherent differences between the sensors. In addition, if the two sensors are mounted at different locations on the machine being monitored in this way, this difference will be further exacerbated due to the fact that the sensors measure along mutually different measurement axes, and due to the different levels of vibration measured due to the difference in location.

For example, in the case of a 3-axis accelerometer and a 1-axis accelerometer mounted on a cylindrical machine, the two types of accelerometers will measure mutually different levels of vibration, since a 3-axis accelerometer will measure radial vibration along one axis and tangential vibration along two axes, while a single-axis accelerometer will measure radial vibration. or direct vibration.

Therefore, using a fault detection classifier trained with signals received from the same type of sensor, such as vibration signals received from a triaxial accelerometer, will result in incorrect classification results when applied to the same type of signals received from a different type of sensor, such as vibration signals, obtained by a single-axis accelerometer.

This can be further illustrated by referring to the more extreme case of two types of sensors sensing different types of signals, such as a vibration sensor and a magnetic flux sensor sensing vibration and flux signals, respectively, from a particular machine tool. A fault detection classifier trained using data received from one of these types of sensors, for example, vibration signals received by a vibration sensor, will be limited to vibration signal classification and will provide poor results with little relevance when applied to fault identification, for example, for signals magnetic flux generated by the same machine.

In order to provide a fault classifier capable of accurately classifying signals received by a sensor type different from that from which the fault classifier was previously trained, based on signals received by said different type of sensor, for example, the first set of sensors S1 or S1 , S11-S1N (FIG. 2), an entirely new fault classifier can be trained. In this case, the existing fault classifier 114 is not used, and a new fault classifier is developed to identify faults in machine tools such as machine tools 104. However, in order for this new classifier to provide accurate fault identification, it is necessary to feed it a large amount of new data received from sensors of various types, for example, from the first set of sensors S1 or the first set of sensors S1-S1N (Fig. 2), based on what this fault classifier should be trained on. Such a process can be lengthy, and such a large amount of data may not be available. In addition, such a process is also highly limited in terms of productivity and scope due to the many parameters that control the frequency dependence of the machine signals, such as the mounting location of the sensor, its orientation, type of mounting, etc. Also, with this approach, the capabilities of the original existing fault classifier 114 are simply wasted rather than used, since the previous fault classifier 114 is not applied at all.

The present invention advantageously provides a solution to the problem of limited use, if any, of a fault classifier trained on data received from sensors of a certain type or types to classify signals received from sensors of different types or types due to differences in sensor characteristics. Advantageously, the present invention does not require training from scratch of a new classifier based on signals from a different type of sensor. In contrast, the present invention uses a transfer learning approach to match between the original sensor type(s) from which an existing classifier was trained and a new, different sensor type(s) from which new data requiring classification is obtained.

The present invention may use an existing classifier 114 by modifying this existing classifier 114 based on a match between different sensor frequency response distributions of respective sensors of different types, and to perform such matching and modification requires only a small set of data from said different types of sensors, such as a first set of sensors. 102. In a preferred embodiment of the present invention, a modified classifier can be generated based on the original existing classifier 114 and a small new set of data obtained from sensors of a type or types different from the type or types of sensors on which the existing classifier 114 was based. This modified classifier can be configured to accurately identify faults in the signals received by these different types of sensors, such as the first set of sensors 102, despite the small amount of data fed to it. The modified classifier uses the original existing classifier 114 and maps it to different types of sensors, such as the first set of sensors 102, so that it can be accurately applied to signals received by said different types of sensors.

However, it should be noted that the present invention may be useful even if a large high quality data set is available from various types of sensors, such as the first set of sensors 102. Although in this case, since a large high quality data set is available, the new specialized classifier, using transfer learning to modify an existing classifier may still be advantageous in order to take advantage of the capabilities of the original existing classifier. Thus, while the present invention is expected to be most useful in cases where sufficient data in terms of quantity and/or quality is not available to train a new classifier, the present invention may also be useful when a large, high quality dataset is available.

Thus, in the case of an existing accurate fault classifier trained using signals received by sensors of a particular type or types, the present invention proposes a solution for modifying said classifier in such a way that it is capable of accurately classifying signals received by sensors of any other type or types. , other than the specific type or types of sensors from which this fault classifier was trained, if these machine signals have at least one characteristic in common with the machines from which the classifier was previously trained. This can be referred to as a sensor-independent approach, in which the classifier can be calibrated to be applicable to data received by any sensor, regardless of the type and/or structure of the raw data collected by that sensor.

The approach in accordance with the present invention can be applied in the case of the same type of signal, for example, vibration signals received by different types of vibration sensors. The results of this are shown in FIG. 5A-5C. The approach of the present invention can also be applied to different types of signal, such as vibration signals and magnetic signals received by different types of sensors, such as vibration sensors and magnetic flux sensors. The results of this are shown in fig. 5D.

In both cases, the original existing classifier can be modified based on the matching of the frequency response distributions of the sensors to create a modified classifier capable of identifying faults in signals received by sensors of a different type than that from which said existing classifier was trained. In both cases, the modified classifier can be applied to signals received by sensors of a different type from those from which the existing classifier was trained, with greater accuracy than could be achieved by applying the existing classifier in its original unmodified form. In addition, in both cases, the modified classifier can be applied to signals received by sensors of a different type from those on which the existing classifier was trained, with greater accuracy than would be achieved with a new classifier trained using signals only, received by these sensors of various types.

The acquisition of a small new set of data from various types of sensors, such as the first set of sensors 102, has been described above with respect to sensors 102 receiving signals and optional associated operating status data from machine tools 104. The acquisition of a small data set by the system of FIG. 1 is preferably carried out using those elements that are enclosed in a dotted box 115.

In the embodiment of the invention shown in FIG. 1, the first set of sensors 102 are preferably all sensors of the same type, i. sensor type S1. However, as mentioned above, the first set of sensors 102 may alternatively comprise more than one type of sensor. In FIG. 2 shows an alternate embodiment of system 100, herein referred to as system 100A, illustrating the introduction of more than one type of sensor into the first set of sensors 102. Now with reference to FIG. 2, the first set of sensors 102 may include sensors S1, S11-S1N on each of the machines 1-N of the plurality of machines 104. It should be understood that the first set of sensors 102 may include any number of sensors S1-S1N, such as two or more sensors. These sensors may be, for example, a combination of vibration sensors, magnetic flux sensors, current sensors, temperature sensors, or sensors for sensing other parameters associated with the operation of machine tools 104, such as torque, displacement, input line frequency, and so on. The sensors S1, S11-S1N may be of mutually different types, but are preferably of the same type with respect to the sets of sensors connected to each of the machines 104. However, it should be understood that the system 100A may tolerate the absence of some data from one of the sensors 102, which can be attributed, for example, using multivariate statistics.

System 100A may generally resemble system 100, except for the many types of sensors contained in the first set of sensors 102 in system 110A, and the description of system 100 as a whole also applies to system 100A.

Continuing to refer to FIG. 1 and 2, a small new dataset from which an existing classifier 114 can be modified by performing a learning match to provide a modified classifier may contain much less data than the amount of data from which the existing classifier 114 was originally trained. For example, the small new data set may contain less than 400, less than 300, less than 200, less than 100, less than 50, or less than 30 sets of each of the first set of signal sets, and optionally the first set of operating status data sets of the first set of machine tools 104. In addition, the small new data set may be an even smaller number of signal sets corresponding to a machine tool faulty operating state of the first set of machine tools 104, such as less than 100, less than 90, less than 80, less than 70, less than 60, less than 50, less than 30, or less 20 sets of signals corresponding to a machine faulty operating state from the first set of machine tools 104. As mentioned above, in some cases, a small new data set may not even contain signals corresponding to a machine faulty operating state from the first set of machine tools 104. This is in contrast to from a much larger dataset from which the existing classifier 114 could be trained, for example several thousand samples.

It should be understood that the first set of sensors 102 is different from the second set of sensors from which the existing classifier 114 was trained, but that the machines 104 preferably share characteristics both with each other and with the machines from which the existing classifier 114 was trained. The machines 104 may or may not be the same as the machines on which the existing classifier 114 was previously trained.

The first set of signal sets received by the first set of sensors 102, and optionally the first set of operating status data collected in the data acquisition unit 110 of the first set of machine tools 104, may be supplied to the existing fault classifier 114. The first set of sets of signals before they are provided to the existing classifier 114 faults can be pre-processed using the processor 108 signal processing.

Then, the existing fault classifier 114 is preferably modified based on said plurality of signal sets and optionally the plurality of operating status data sets of the first plurality of machine tools 104, thereby producing a modified classifier 116.

The modified classifier 116 may be an algorithmic classifier executed by one or more processors, which may be the same or different processors as the processors running the existing classifier 114. For example, the existing classifier 114 and the modified classifier 116 may be implemented in a data processing unit. The one or more processors executing the existing fault classifier 116 may be remote processors, such as those located in the cloud, or may be local processors. For example, the one or more processors executing the modified classifier 116 may be located in one of the sensors 102.

The existing fault classifier 114 is modified by adjusting the classifier to a new data set containing a plurality of signal sets received by sensors 102 and optionally a plurality of operating status data sets of the first plurality of machine tools 104. For example, as seen in FIG. 1, in the case where the first set of sensors 102 contains only the sensor type S1, the modified classifier 116 can be adapted to the sensor type S1. Moreover, as an example, as shown in FIG. 2, in the case where the first sensor set contains a plurality of sensor types S1, S11, etc., the modified classifier 116 can be adapted to this plurality of sensor types. By customizing an existing classifier 114 to create a modified classifier 116, the classifier is calibrated to be applicable to a different dataset from the dataset from which the classifier was originally trained. The tuning includes a correspondence between the original sensor characteristics from which the classifier 114 was previously trained and the sensor characteristics of the first set of sensors 102.

In one embodiment of the present invention, the tuning includes a correspondence between the frequency response distribution of the original sensor set, from which the classifier 114 was previously trained, and the sensor frequency response distribution of the first sensor set 102. This mapping can be performed between the frequency response distributions of the sensors, and not just the frequency responses of the sensors since the frequency responses of both the sensors on which the classifier 114 has been pre-trained and the first set of sensors 102 can be distributed. This distribution may arise due to actual differences between the probes in each set, such as differences in machine characteristics, differences in probe location, probe mounting, exact probe mounting type, and so on.

Compliance may be performed in a controlled manner. In this embodiment, the system 100 may contain a data acquisition unit 110, and the tagged signals are fed to an existing fault classifier 114. Additional details regarding how the existing fault classifier 114 is modified with sensor-to-sensor matching in a controlled manner are presented below with reference to FIG. 3.

Alternatively, the matching may be performed in an unsupervised manner. In this case, the system 100 need not contain a collection unit 110, but signals without corresponding operating states are fed to the existing fault classifier 114. The existing fault classifier 114 can be modified using unsupervised learning to match the differences between the frequency response distributions of the sensors from which classifier 114 was previously trained and new sensors, such as the first set of sensors 102.

Alternatively, the matching may be performed in a semi-supervised manner, in which tagged signals are applied to the existing fault classifier 114 and matching between sensors to create a modified classifier 116 is performed in an unsupervised manner. Additional details relating to unsupervised matching are provided below with reference to FIG. 3-4B.

The modified classifier 116, adapted by matching to be applicable to data from the first set of sensors 102, is now ready to be used to classify signals received by sensors of the same type or types as the first set of sensors 102.

An example of using a modified classifier 116 to identify faults in signals received by a sensor of the same type as the first set of sensors 102 is further shown in FIG. 1 and 2. As seen in FIG. 1 and 2, at least one sensor 122, such as sensor S1 in FIG. 1 and sensors S1-S1N in FIG. 2 may receive at least one set of signals emanating from at least one given machine tool, here designated as machine tool X, indicated by 124. The machine tool 124 may have at least one characteristic in common with the first set of machine tools 104 , as well as with many machine tools, on the basis of the signals from which the existing fault classifier 114 was trained. The machine tool 124 may or may not be a member of the first set of machine tools 104. It should be understood that although machine tool 124 is shown here implemented as a single machine, this is for simplicity only, and system 100 may include any number of machine tools 124, such as one, two, or more machine tools 124, which may or may not be identical with each other, provided that the machines 124 have the specified at least one common characteristic, as described above.

At least one sensor 122 is one of the first plurality of sensors 102, which means that the sensor 122 is of the same type as the first plurality of sensors 102. As shown in FIG. 1, in the case where the first plurality of sensors 102 contains one type of sensors, i. e. sensor S1, sensor 122 also belongs to the type of sensor S1. As shown in FIG. 2, in the case where the first plurality of sensors 102 contains a plurality of sensor types, i. e. sensors S1-S1N, at least one sensor 122 also contains the same plurality of sensor types, ie. sensors S1-S1N. The at least one sensor 122 is preferably configured to receive at least one set of signals from a given machine tool 124. The signal sets may be "instantaneous signals" sensed by at least one sensor 122 for a short period of time. For example, the signal may be perceived within a few seconds, such as one to four seconds. Alternatively, the waveform sets may comprise a plurality of "snapshots" over a period of time, such as four second "snapshots" measured every hour over a period of several hours, days, or even months. Alternatively, signal sets may contain continuously monitored signals over a longer period or more than one time period. For example, a signal can be continuously monitored every millisecond.

The signals received by at least one sensor 122 may be pre-processed, for example, using the analog or digital processing capabilities of the sensor 122 itself or other hardware and/or software signal processing components 128. It should be understood that although the signal processor 128 is shown in FIG. 1 and 2 as a separate element from at least one sensor 122, signal processing functionality may be built into one or more sensors 122. Signal pre-processing may include at least one of the following: digitization, compression, feature extraction, and signal time representation. or frequency domain.

In one embodiment of the present invention, signals received by at least one sensor 122 may be uploaded to a remote server, such as a cloud server. Signal processing functionality 128 may be performed on a remote server.

A set of signals received by at least one sensor, here exemplarily implemented as sensor 122, originating from at least one given machine tool, here exemplarily implemented as machine tool X, may be provided to a modified classifier 116. The modified the classifier 116 may be configured to automatically identify at least one fault of the machine tool X based on a set of signals output from it. It should be understood that the modified classifier 116 can be accurately applied to signals received by at least one sensor 122, since the modified classifier 116 is adapted to classify signals received by sensors such as the first set of sensors 102.

The identification of faults by the modified classifier 116 may include the identification of one or more specific faults of the monitored machine 104. Depending on the particular machine 104 being monitored, the particular fault identified may include worn rotating machine bearings, mechanical looseness, misalignment, imbalance, electrical faults, or other faults. .

Alternatively, the fault identification may include identifying machine X as being in a faulty state, that is, in an abnormal state with respect to its normal operating state, but without identifying a specific fault. In this case, the fault identification identifies the machine as not operating normally, but does not identify what is the specific cause of this abnormal operation.

Systems 100 and 100A may further comprise an output device 130. The output device 130 may be configured to receive a fault identification output from the modified classifier 116 and provide human-readable output data containing at least a fault identification of at least one given machine tool 124. The human-readable output data may comprise at least one from visual, tactile or audio output. Preferably, at least one repair or maintenance operation is performed based on said human-perceptible output.

For example, in the case where the modified classifier 116 is executed by a remote processor, the fault identification output of the modified classifier 116 may be transmitted to the output device 130 .

The output device 130 may also be configured to predict the failure of at least one target machine tool 124 due to a fault identified by the modified classifier 116 in the absence of a recommended maintenance operation such that at least one target machine tool 124 will actually fail. if the recommended maintenance operation is not performed.

In some cases, maintenance 132 of a given machine tool 124 may be performed in response to human-perceptible output provided by output device 130. The output device 130 may optionally be operatively connected to the machine controller 124 and the operation of the machine 124 may be adjusted in response to the identification of the fault. For example, machine tool 124 may be turned off, run at reduced power, or otherwise adjusted. Such adjustment may be automatic or may be controlled by a human expert in response to human-perceptible output provided by output device 130.

Additional details about the existing classifier 114 itself and how the existing classifier 114 can be adapted or calibrated to create a modified classifier 116 are now presented with reference to FIG. 3-4B, which show a simplified illustration of classifier modification performed by any of the type systems shown in FIG. 1 and 2, and simplified respective block diagrams of the components of two possible systems for training the classifier used in the system of FIG. 1 or 2, respectively. In one embodiment of the present invention, systems 100 and 100A may comprise a non-volatile readable storage medium that stores computer executable instructions for one or more processors to perform a classifier modification method, as detailed below with respect to FIG. 4A and 4B. Such processors may be remote processors or local processors.

As seen in FIG. 3, an existing fault classifier 114 may be an artificial neural network (ANN) classifier 300 for supervised fault and anomaly detection. For example, the ANN 300 may be configured to identify a specific malfunction or anomaly in machine tools having a rotating component. Without loss of generality, a fault classifier 114 having a multilayer perceptron architecture is shown in FIG. 3. However, it should be understood that the systems and methods in accordance with the present invention can be applied to any type of fault classifier regardless of its structure. In particular, the number layers and neurons shown in FIG. 3 are only given as a very simplified example for the purpose of illustrating the principles of the present invention. In addition, the existing fault classifier 114 may be an unsupervised fault classifier as described above, such as an autoencoder, a deep web of trust, a classifier based on K-means clustering or hidden Markov models, or any other type of model capable of conducting an unsupervised learning.

On top panel 302 in FIG. 3 shows an example of an existing classifier 114 in its original, unmodified form, after it has been trained, and before it has been modified in any way to match between sensor types in accordance with the present invention.

The existing classifier 114 may have a layered architecture. The data layer 308 may be a layer containing input data in the form of sets of sensor signals. The input layer 310 may be the source layer on which sets of sensor signals are injected into the network 300. At each of the multiple subsequent hidden layers 312, the signal sets are combined with appropriate weights and an activation function applied to the combined output of the layer before being forwarded to the next layer. layer of hidden layers 312. This process is repeated for each layer of hidden layers 312 until the output layer 314 is reached. The output layer 314 provides a fault indicator. This indicator represents the fault identification. Fault identification may include identifying a present particular fault or anomaly, or predicting a future particular fault. The fault can contain any type of machine anomaly.

The general mathematical expression for the ANN architecture shown in panel 302 of FIG. 3, within network layers

where s is the input of the neuron, which for the first layer 310 is the sensor signal, σ is the activation function, ω is the weight, b is the bias term, and i and j are indices that operate on the layer's input data points and neurons, respectively.

Training the network 300 to create an existing classifier 114 can be better understood with additional reference to FIG. 4A and 4B.

Referring first to FIG. 4A, a system 400 for training an existing classifier 114 may comprise a second plurality of sensors 402. The sensors 402 are preferably coupled to a respective second plurality of machine tools 404. All of the second plurality of sensors 402 may all be of the same second type, referred to here as sensor type S2. Alternatively, the sensors 402 may comprise a plurality of sensor types coupled to each respective one of the second plurality of machine tools 404. The case of a second plurality of sensors 402 comprising a plurality of sensor types is shown in system 400A in FIG. 4b. System 400A generally resembles system 400 except for the inclusion of multiple types of sensors 402. Referring now to FIG. 4B, it should be understood that the second set of sensors 402 shown in FIG. 4B may include any number of sensors S2-S2N, such as two or more sensors. These sensors may be, for example, a combination of vibration sensors, magnetic flux sensors, current sensors, temperature sensors, or sensors for sensing other parameters associated with the operation of machine tools 404, such as torque, displacement, input line frequency, and so on. Sensors S2-S2N may be of mutually different types, but are preferably of the same type with respect to each of the sets of sensors connected to each of the machines 404.

Sensors 402, such as sensor S2 or sensors S2-S2N, may be different from the first set of sensors 102 in FIG. 1 or be different from the sensor types S1-S1N of the first plurality of sensors 102 in FIG. 2. Any combination of the embodiments shown in FIG. 1 and 2, as well as 4A and 4B, that is, the first set of sensors 102 may contain only one type of sensor S1, and the second set of sensors 402 may contain only one type of sensor S2, different from S1 (the embodiment shown in Fig. 1 and Fig. 4A); the first set of sensors 102 may contain a sensor of only one type S1, and the second set of sensors 402 may contain a plurality of sensor types S2-S2N, at least some of which are different from S1 (the embodiment shown in Fig. 1 and Fig. 4B); the first set of sensors 102 may contain a plurality of sensor types S1-S1N, and the second set of sensors 402 may contain only one type of sensors S2, which differs from at least some of the S1-S1N (the embodiment shown in Fig. 2 and Fig. 4A) ; the first plurality of sensors 102 may comprise a plurality of sensor types S1-S1N, and the second plurality of sensors 402 may comprise a plurality of sensor types S2-S2N, at least some of which are different from S1-S1N (the embodiment shown in FIG. 2 and FIG. 4B).

In some embodiments, the first and second sets of sensors 102 and 402 may contain some common sensors, for example, the first set of sensors 102 may contain vibration sensors and magnetic sensors, and the second set of sensors 402 may contain only magnetic sensors.

Here, by way of example, the second plurality of machine tools 404 is shown as containing machine tools 1, 2 - N, where N can be any number of machine tools, such as two or more machine tools. Typically, sensors 402 are connected to machine tools 404 in a one-to-one correspondence arrangement, with one of the sensors 402 connected to a corresponding one of the machine tools 404. However, other arrangements are possible in which one sensor can be configured to sense signals from more than one machine tool 404. Sensors 402 may be in physical contact with machine tools 404, such as being directly or indirectly mounted on machine tools 404. Alternatively, sensors 402 may be physically separated from machine tools 404, such as being located at a predetermined distance from machine tools 404, such as in the case where the sensors 402 are optical sensors.

In one embodiment of the present invention, the sensors from the second plurality of sensors 402 may have generally the same frequency response. The frequency response of the second set of sensors 402 may be referred to as the frequency response distribution of the sensors 402. The first set of sensors 102 may also have generally the same frequency response. The frequency response of the first set of sensors 102 may be referred to as the frequency response distribution of the sensors 102. Due to the fact that the sensors 402 are of a different type than the sensors 102, the frequency response distribution of the first set of sensors 102 is different from the frequency response distribution of the second set of sensors 402. For example, the first the plurality of sensors 102 may be single-axis vibration sensors and the second plurality of sensors 402 may be multi-axis vibration sensors, or vice versa, or the first plurality of sensors may be single-axis vibration sensors and the second plurality of sensors may be magnetic flux sensors. By way of example only, sensors 102 and 402 or sensors 402 and 102 can measure vibration and flux signals; vibration and electric current signals; vibration and temperature signals; electric current and magnetic flux signals; and machine vibration and internal pressure signals, respectively.

Preferably, the sensors 402 are configured to sense signals from the machine tools 404. The machine tools 1-N, which are members of the second plurality of machine tools 404, preferably have one or more characteristics that are common to each other and to the machine tools 104 1-N machine tools may or may not be the same machine, provided they share at least one feature in common. For example, general characteristics may refer to a type, model number, manufacturer, physical characteristics or dimensions, performance characteristics or parameters, or other general characteristics that indicate that the observed behavior of one of the machine tools among a plurality of machine tools may be typical of another machine tool. machine from this set.

Preferably, the second plurality of sensors 402 is configured to receive a second plurality of signal sets from the second plurality of machine tools 404. The signal sets may be "instantaneous signals" sensed by a respective one of the sensors 402 for a short period of time. For example, the signal may be perceived within a few seconds, such as one to four seconds. Alternatively, each set of signals may comprise a plurality of "snapshots" over a period of time, such as four second "snapshots" measured every hour over a period of several hours, days, or even months. Alternatively, each set of signals may contain continuously monitored signals over a longer period or more than one period of time. For example, a signal can be continuously monitored every millisecond.

The signals received by the sensors 402 may be pre-processed, for example, using the analog or digital processing capabilities of the sensor 402 itself or other hardware and/or software processing components 408. For example, the signal may be at least one of the following signals: digitized, compressed, features can be extracted from the signal, and the signals can be represented in the time or frequency domain.

Systems 400 and 400A may optionally include a data acquisition unit 410. Preferably, the data acquisition unit 410 is configured to receive a second plurality of machine tool operating state data sets from the second machine tool plurality, each operating state data set indicating a machine tool operating state from said second machine tool set 404, and each operating state the state is associated with at least one of the sets of signals received by the second set of sensors 402.

The operating condition data collected in the data acquisition unit 410 is preferably in the form of machine condition diagnoses provided by human experts such as engineers. The sets of signals received by the first set of sensors 402 may optionally be provided to the data acquisition unit 410 either via the signal processor 408 and/or directly or indirectly via the sensors 402.

These human experts can analyze the signal sets received by the second set of sensors 402 and label each signal set from said signal sets as representing particular operating states of the respective machine tool 404 from which the signals originate. The human experts may interact with a user interface, such as the data acquisition unit 410 or other device that provides communication between the human expert and the data acquisition unit, to enter operating status data. The signals and the labels applied to them can be accumulated and stored in a database in block 410 of the data collection. In one embodiment of the present invention, data acquisition unit 410 may be located on a remote server, such as a cloud server.

The identification of faults by human experts may include the identification of one or more specific faults of the monitored machines 404. Depending on the particular machine 404 being monitored, the particular fault identified may include worn rotating machine bearings, mechanical looseness, misalignment, imbalance, electrical faults, or other faults. . Alternatively, fault identification may include identifying the machine 404 as being in a faulty state, that is, in an abnormal state with respect to its normal operating state, but without identifying a particular fault. In this case, the identification of faults by a human expert identifies the machine as not operating normally, but does not identify what is the specific cause of this abnormal operation.

It should be understood that in this example, systems 400/400A preferably comprise a data collector 410 and signals are preferably labeled by human experts in order to train the fault classifier 300 in a supervised manner. However, in other embodiments of the present invention, the data collector 410 may be omitted and the signals need not be labeled by human experts. In this case, the existing fault classifier 114 may be initially trained using unsupervised learning.

Returning now to panel 302 in FIG. 3, the network 300 can be trained by feeding it the signal sets received by the second set of sensors 402 as input to layer 308, and feeding it the associated operating state data, as labeled by human experts, as the required corresponding output. at the output layer 314. Training the network 300 may be performed using backpropagation and gradient descent algorithms on predefined data labels. Training parameters such as loss function, learning rate, optimizer type, etc. are preferably chosen according to the output metric. For example, a cross entropy loss function, an Adam optimizer, and an L2 regularization term can be chosen to detect binary faults. The training process is first applied to a dataset for which all sensors 402 are compatible across the entire dataset, which means that it is preferable not to mix the sensors 402 for each vector at a data point.

After the training is completed, the parameters of the network 300 set based on the training are preferably fixed. These parameters contain activation and weight functions. The network 300 now constitutes an existing classifier, such as the existing classifier 114, based on the sensor signal sets received by the second set of sensors 402. The existing classifier 114, in the form of the trained network 300 shown in panel 302, is now capable of accurately classifying new input data, having the same or similar structure and derived from the same or similar sources as the data from which the existing classifier 114 was previously trained. In this case, the existing classifier 114 in the form of a trained network from machines having a common characteristic with machines 404 (FIGS. 4A and 4B) and sensed by sensors of the same type as the second set of sensors 402 (FIGS. 4A and 4B). However, as detailed above, the existing classifier 114 in the form of a trained network 300 is not designed to accurately classify new inputs sensed by sensors of a different type than a second set of sensors 402, such as the first set of sensors 102 (FIGS. 1 and 2). even though these new inputs come from machines that share a common characteristic with machines 404, such as the first set of machines 104. In this case, the accuracy of the classifier will be significantly reduced due to the different frequency responses of these different types of sensors.

In order to make the existing classifier 114 able to accurately classify sets of signals from a sensor of a different type, for example, the first set of sensors 102, than those on which this classifier was previously trained, for example, the second set of sensors 402, the existing classifier 114 can be modified . A modification of the existing classifier 114 as a network 300 in accordance with the preferred embodiment of the present invention is shown in the bottom panel 440 of FIG. 3.

As seen in the bottom panel 440 in FIG. 3, network 300 may be modified to create a modified network 442. Preferably, network 300 is modified by adding at least one additional layer 450 to network 300. Said at least one additional layer 450 may be referred to as a "match layer" and may be implemented with the ability to study the difference in frequency characteristics between the original type of sensor, i.e. a second set of sensors 402 from which the classifier 114 was trained as a network 300, and a new type of sensor, i. e. the first set of sensors 102 from which new data has been collected.

The mapping layer 450 may be configured by freezing the structure and parameters of all layers of the network 300 and retraining the classifier with the new data set received from the first set of sensors 102, as provided by the elements enclosed in dashed box 115 and described above with respect to FIG. 1 and 2. Thus, the network 442 is forced to optimize the weights of the matching layer 450 for the difference in frequency response of the two sensors, since the parameters of all other layers of the network 442 are already configured, frozen, and cannot be changed. As a result, the mapping layer 450 is forced to learn mappings between the original sensor signal sets provided by the sensors 402 (FIGS. 4A and 4B) and the new sensor signal sets provided by the sensors 102 (FIGS. 1 and 2).

Alternatively, the structure and parameters of all layers of the network 300 do not have to be frozen, but can be tuned during classifier retraining with the new data set received from the first set of sensors 102. In this case, the parameters of the network 300 serve as the starting point for the tuned parameters of the modified classifier. 442.

The mapping layer 450 is preferably configured to take into account the nature of the difference in frequency response between different types of sensors, such as the first and second plurality of sensors 102 and 402.

In the case of a match between different sensors of the same type, for example, if both the first and second plurality of sensors 102 and 402 are vibration sensors but of different types having different frequency responses, the frequency response difference is typically a linear function of the signal frequency.

In the case of correspondence between different sensors of different types, for example, if the first and second plurality of sensors 102 and 402 are vibration and magnetic sensors, respectively, having different frequency responses, the frequency response difference is generally a non-linear function, therefore, a non-linear correspondence. Accordingly, the activation function may, but need not, take a non-linear form.

In some cases, the matching layer 450 may be facilitated by providing the matching layer 450 with more complex forms of new sensor data, such as more complex waveforms received by the first set of sensors 102. For example, inverse data, logarithmic data, or other forms of data may be provided. The input signal modification process can be optimized with respect to the classifier accuracy, whereby the new sensor data is modified in accordance with the feedback about the classifier accuracy characteristics, and then the conformance layer 450 is retrained with the modified data.

Preferably, the mapping layer 450 is included in the network 442 as the first layer after the data layer 308 and is therefore configured to receive input as a new data set. The mapping layer 450 is upstream of the original input layer 310 and precedes the original input layer 310 with respect to input. The location of the match layer 450 in the network 442 is important because it is the location of the match layer 450 as the initial layer in the network that allows the match layer 450 to learn the match between sensor types and adapt incoming new data so that it is in the proper form for transition to other hidden layers 312 located downstream in network 442.

In an alternative embodiment of the present invention, an existing fault classifier can be initially trained in an unsupervised manner by providing it with a large amount of sensor data in order for the classifier to learn how to identify anomalies in the sensor data related to the operating state of the machine. For example, the NN autoencoder can be used to learn a low-dimensional representation of sensor data and clustering-based classification, and then used to identify outliers (anomalies). Such an existing unsupervised fault classifier 114 can then be modified in an unsupervised manner in accordance with the preferred embodiment of the present invention to create a modified fault classifier 116 adapted to identify anomalies in sensor data obtained by a sensor of a different type than that from which the existing one was originally trained. classifier 114 faults. A modified fault classifier 116 can be created, for example, by adding an initial matching layer to the autoencoder NN, freezing the parameters of all network layers except the matching layer, and training the matching layer to learn the matching between the original and new sensor types. During the training of the match layer for training purposes, another match layer, such as an inverse match layer, can be additionally added after the decoder at the output of the NN autoencoder. The output can then be classified using clustering classification to identify outliers (anomalies).

The improvement in classifier performance as a result of conformance learning in accordance with the present invention is illustrated in the following illustrative graphs:

In FIG. 5A shows an example of the performance of a modified classifier in accordance with the present invention as a precision-recall curve. In this example, a fault classifier was initially trained and tested on a dataset containing over 100,000 labeled signal sets, of which 7,000 signal sets indicated machine faults, received from rotary motion machines containing bearings via single axis vibration sensors. After training, the fault classifier was tested on a data set containing 20,000 signal sets, 1400 of which corresponded to machine faults. A fault classifier trained in this way can be called an existing fault classifier.

The existing classifier was then modified in accordance with the preferred embodiment of the present invention to be able to classify faults based on signals received by triaxial vibration sensors from rotary machines containing bearings. The new waveform dataset acquired by 3-axis vibration sensors and used to modify the existing fault classifier contained a training and verification dataset consisting of 256 waveforms, of which only 20 waveforms were associated with machine failure conditions, namely bearing wear. As described above, the existing fault classifier has been modified based on this very small set of signals.

Then, a modified fault classifier to test its performance was applied to a test set consisting of 7149 signal sets, of which 486 signal sets were associated with machine fault conditions, to detect faults, and the signals were received by three-axis vibration sensors from rotary motion machines containing bearings (line 502). In comparison, a completely new classifier was trained from scratch with the same 256 signal sets and applied to the same test set of 7149 signal sets. (line 504). For completeness of comparison, the original existing classifier in its unmodified form was also applied to an exemplary dataset consisting of 7149 signal sets (line 506). As can be seen from a consideration of FIG. 5A, the performance of the modified classifier in fault identification is the best despite the extremely small set of data provided. The indicators listed in Fig. 5A are average accuracy figures, although other approximate figures can be used.

In FIG. 5B shows an example of the performance of a modified classifier in accordance with the present invention as a precision-recall curve. In this example, the fault classifier was initially trained and tested on a dataset containing over 100,000 labeled signal sets, of which 7,000 signal sets indicated machine faults, received from rotary motion machines containing bearings via single axis vibration sensors. After training, the fault classifier was tested on a data set containing 20,000 signal sets, 1400 of which corresponded to machine faults. A fault classifier trained in this way can be called an existing fault classifier.

The existing classifier was then modified in accordance with the preferred embodiment of the present invention to be able to classify faults based on signals received by triaxial vibration sensors from rotary machines containing bearings. The new waveform dataset acquired by 3-axis vibration sensors and used to modify the existing fault classifier contained 924 waveform sets, of which only 70 waveform sets were related to machine failure conditions, namely bearing wear. As described above, the existing fault classifier has been modified based on this very small set of signals. The modified fault classifier was then applied to a set of 6841 waveforms containing 436 waveform sets corresponding to machine faults to test its performance to detect faults, the signals being received by three-axis vibration sensors from rotary motion machines containing bearings (line 510 ). In comparison, a completely new classifier was trained from scratch with the same 924 signal sets and also applied to 6841 sample data sets (line 512). For completeness of comparison, the original existing classifier in its unmodified form was also applied to 6841 exemplary data sets (line 514). As can be seen from a consideration of FIG. 5B, the performance of the modified classifier in fault identification is the best despite the small set of data provided. The indicators listed in Fig. 5B are average accuracy figures, although other approximate figures can be used.

In FIG. 5C shows an example of the performance of a modified classifier in accordance with the present invention as a precision-recall curve. In this example, the fault classifier was initially trained and tested on a dataset containing over 100,000 labeled signal sets, of which 7,000 signal sets indicated machine faults, received from rotary motion machines containing bearings via single axis vibration sensors. After training, the fault classifier was tested on a data set containing 20,000 signal sets, 1400 of which corresponded to machine faults. A fault classifier trained in this way can be called an existing fault classifier.

The existing classifier was then modified to be able to classify faults based on signals received by three-axis vibration sensors from rotary machines containing bearings. The new waveform dataset obtained by the 3-axis vibration sensors and used to modify the existing fault classifier contained 5924 waveform sets, of which only 350 waveform sets were related to machine failure conditions, namely bearing wear. As described above, the existing fault classifier has been modified based on this small set of signals. Then, a modified fault classifier to test its performance was applied to a set of 1481 exemplary signal sets containing 156 machine fault signal sets to detect faults, the signals being received by three-axis vibration sensors from rotary motion machines containing bearings (line 520). In comparison, a completely new classifier was trained with the same 5924 signal sets and also applied to 1481 sample data sets (line 522). For completeness of comparison, the original existing classifier in its unmodified form was also applied to these 1481 exemplary data sets (line 524). As can be seen from the examination of Fig. 5C, the performance of the modified classifier in fault identification is the best. The indicators listed in Fig. 5C are average accuracy figures, although other approximate figures can be used.

In FIG. 5D shows an example of the performance of a modified classifier in accordance with the present invention as a precision-recall curve. In this example, a fault classifier was trained and tested on a dataset containing more than 100,000 labeled signal sets, of which 7,000 signal sets indicated machine faults received from electric motors containing bearings through single-axis vibration sensors. After training, the fault classifier was tested on a data set containing 20,000 signal sets, 1400 of which corresponded to machine faults. A fault classifier trained in this way can be called an existing fault classifier.

The existing classifier has been modified to classify faults based on magnetic flux signals received by magnetic pickups from electric motors containing bearings. The new signal data set obtained by magnetic pickups and used to modify the existing fault classifier contained 3411 signal sets, of which only 222 signal sets were associated with machine failure conditions, namely bearing wear. As described above, the existing fault classifier has been modified based on this small set of signals. Next, an existing fault classifier to test its performance was applied to a test set of 754 signals containing only 50 exemplary sets of signals corresponding to machine faults to detect faults, the signals being received by magnetic pickups from electric motors containing bearings (line 530) . In comparison, a completely new classifier was trained with the same 3411 magnetic signal sets and also applied to the same 754 sample magnetic data sets (line 532). For completeness of comparison, the original existing classifier in its unmodified form was also applied to the same 754 exemplary magnetic data sets (line 534). As can be seen from the examination of Fig. 5D, the performance of the modified classifier in fault identification is the best. The indicators listed in Fig. 5D are average accuracy figures, although other approximate figures can be used.

In the above examples in Fig. 5A-5D, the data set from which classifier 114 was originally trained consisted of data received from over 40,000 individual rotary machines, containing mainly motors, pumps, fans, gearboxes, coolers, and compressors. Bearing records, some of which were measured more than once, were kept for three years. Each bearing was measured in three axes using a single axis piezoelectric vibration sensor. In case the new data set contained data from a 3-axis vibration sensor, it was a 3-axis MEMS vibration sensor.

As can be seen from the above data, the proportion of faults in the data set was approximately 7%. The classifier was designed to detect failure of a single bearing, so that even in cases where additional good bearings in the machine gave signals indicative of bearing wear, due to the proximity of good bearings to bad bearings and due to the propagation of an acoustic wave between these bearings, the classifier was made with the ability to correctly classify serviceable bearings as such. Data labeling was done by over 10 human experts.

Let us now turn to FIG. 6, which is a simplified flowchart illustrating the steps of a learning transfer-based machine fault identification method in accordance with a preferred embodiment of the present invention.

In FIG. 6 shows a method 600 for identifying a machine fault. As can be seen, in the first step 602, the first set of signals from the plurality of machine tools can be received by means of the first plurality of sensors. The sensors from the first plurality of sensors are preferably of the first mutually identical type. Alternatively, the sensors from the first plurality of sensors may be of a plurality of types. The sensors from the first plurality of sensors may have generally the same sensor frequency response, which may be referred to as the first sensor frequency response distribution.

As can be seen, in the second step 604, a first set of machine run data corresponding to the first set of signals obtained in step 602 may optionally be obtained. The machine run state data may comprise an identification of a run state associated with each set of signals from the first set of signals. The operating state may be faulty or good.

As can be seen, in the third step 606, the existing fault classifier, previously trained to detect faults in machines based on the signals from the machines and received by the second set of sensors, is supplied with the indicated sets of signals and optional corresponding operating state data. The second set of sensors, on the basis of which the existing fault classifier was trained, may be of the same type, but different from the type of the first set of sensors. Alternatively, the second set of sensors from which the existing fault classifier has been trained may be of a variety of types, at least some of which are different from the first set of sensors. The sensors from the second plurality of sensors may have generally the same sensor frequency response, which may be referred to as a second sensor frequency response distribution. The second sensor frequency response distribution may be different from the first sensor frequency response distribution of the first plurality of sensors.

As can be seen, in the fourth step 608, the existing fault classifier is preferably modified using the transfer learning approach and based on the new data fed to it. The modification may include adding at least one matching layer to the existing fault classifier, wherein the matching layer can be trained on the frequency response difference between the first sensor frequency response distribution and the second sensor frequency response distribution.

As can be seen, in the fifth step 610, an additional set of signals can subsequently be received from at least one machine by a sensor of the same type as the first set of sensors.

As can be seen, in the sixth step 612, to identify faults in the machines through which the additional set of signals was generated, a modified classifier created in step 608 can be applied to this additional set of signals.

As can be seen, in the seventh step 614, maintenance or repair of the machine can be performed based on the identified faults.

It should be understood that all of the various machines described as generating signals in steps 602, 606 and 610 may be the same or different machines having a common characteristic as described above.

Those skilled in the art should understand that the present invention is not limited to what has been specifically shown and described above. The scope of the present invention includes both combinations and subcombinations of the various features described above, as well as modifications thereof, all of which are not prior art.

Claims (49)

1. A method for identifying a malfunction of at least one machine tool, including: providing the first set of sensors connected to the corresponding first set of machine tools, the first set of sets of signals emanating from the first set of machine tools, and the first set of machine tools has at least one common characteristic; supplying at least the first set of signal sets of the first set of machine tools to an existing fault classifier, previously trained to automatically identify faults of the second set of machine tools based on signals emanating from them and previously received by the second set of sensors, the second set of sensors being of a different type than the first set of sensors, the second set of machine tools has the specified at least one common characteristic; modifying said existing fault classifier by using transfer learning based on at least the first plurality of signal sets of the first plurality of machine tools, thereby providing a modified fault classifier; applying the modified fault classifier to at least one additional set of signals received by at least one sensor from the first set of sensors and coming from at least one given machine tool having the specified at least one common characteristic, and the modified fault classifier is configured to automatically identifying at least one malfunction of said at least one given machine tool based on said at least one additional set of signals; And providing a human-readable output, by means of an output device, comprising at least identifying said malfunction of said at least one given machine tool, at least one repair or maintenance operation performed based on said human-readable output. 2. The method according to claim 1, further comprising after said ensuring that the first set of sensors receives the first set of signal sets and before said supply of the first set of signal sets to the existing fault classifier: obtaining a first set of operating state data sets for machine tools from the first set of machine tools, each set of operating state data indicating an operating state of a machine tool from the first set of machine tools, each operating state associated with at least one of said sets of signals; said supply of at least a first plurality of signal sets to an existing fault classifier also includes supplying said operating status data of the first plurality of machine tools to the existing fault classifier; said modification of the existing fault classifier by applying transfer learning based on at least the first plurality of signal sets also includes modifying the existing fault classifier by applying transfer learning additionally based on the first plurality of operating condition data sets of the first plurality of machine tools. 3. The method according to claim. 1 or 2, characterized in that the specified fault identification includes the identification of a specific fault of the specified at least one specified machine tool and the prediction of failure of the specified at least one specified machine tool due to the specified specific fault in the absence of the recommended operation for its maintenance such that said at least one given machine tool will actually fail if said recommended maintenance operation is not performed. 4. The method according to claim 2 or 3, wherein said existing fault classifier comprises a neural network, and modifying the existing fault classifier includes adding at least one matching layer to the neural network, wherein in addition to adding said at least one matching layer, the neural network otherwise is unmodified by said modification. 5. The method of claim 4, wherein said neural network comprising an existing fault classifier comprises a data layer and an input layer for receiving data from the data layer, wherein said at least one matching layer is added between the data layer and the input layer, and said at least one matching layer is configured to receive data from the data layer in the modified fault classifier. 6. The method of claim 4 or 5, wherein the first plurality of sensors has a first frequency response distribution and the second plurality of sensors has a second frequency response distribution, wherein the matching layer is configured to match between the first and second frequency response distributions. 7. A method according to any one of the preceding claims, wherein the first plurality of sensors is configured to sense the same type of signal as the second plurality of sensors. 8. The method of claim 7, wherein said same signal type comprises one of a vibration signal, a magnetic flux signal, a current signal, a temperature signal, and an internal pressure of the machine. 9. The method according to any one of paragraphs. 1-6, wherein the first plurality of sensors and the second plurality of sensors are configured to sense mutually different types of signals. 10. The method according to claim 9, characterized in that said mutually different types of signals comprise at least one of: vibration and magnetic flux signals; vibration and electric current signals; vibration and temperature signals; electric current and magnetic flux signals; And vibration and internal pressure signals of the machine. 11. The method according to any one of paragraphs. 2-10, wherein at least some of said machine tool operating states of the first set of machine tools, as indicated by the first set of operating state data sets, are faulty operating states. 12. The method of claim 11, wherein said first plurality of signal sets and the first plurality of operating state data sets of the first plurality of machine tools contain less than 30 faulty operating states. 13. A system for identifying a malfunction of at least one machine tool, comprising: a first plurality of sensors connected to a respective first plurality of machine tools and configured to receive a first plurality of sets of signals from the first plurality of machine tools, the first plurality of machine tools having at least one common characteristic; data processing unit configured to: receiving the first set of signal sets of the first set of machine tools, wherein the data processing unit comprises an existing fault classifier, previously trained to automatically classify the operating states of the second set of machine tools based on signals emanating from them and previously received by the second set of sensors, the second set of sensors being to a different type than the first set of sensors, the second set of machine tools has the specified at least one common characteristic; modifying said existing fault classifier by using transfer learning based on at least the first plurality of signal sets of the first plurality of machine tools, thereby providing a modified fault classifier, and applying the modified fault classifier to at least one additional set of signals received by at least one sensor from the first set of sensors and coming from at least one given machine tool having the specified at least one common characteristic, and the modified fault classifier is configured to automatically identifying at least one malfunction of said at least one given machine tool based on said at least one additional set of signals, and an output device in communication with said data processing unit and configured to provide human-perceptible output data containing at least an identification of said malfunction of said at least one given machine tool, of at least one repair or maintenance operation performed on based on said human-perceptible output. 14. The system of claim. 13, further comprising a data acquisition unit configured to obtain a first set of data sets on the operating state for machine tools from the first set of machine tools, and each set of data on the operating state indicates the operating state of the machine tool from the first set machine tools, each operating state is associated with at least one of the specified sets of signals; said data processing unit is configured to receive operating status data of the first plurality of machine tools and to further modify said existing fault classifier based on the first plurality of operating status data sets of the first plurality of machine tools. 15. The system according to claim 13 or 14, wherein said identification of said fault includes identifying a particular fault of said at least one given machine tool and predicting failure of said at least one given machine tool due to said particular fault in the absence of performing the recommended operation on it. maintenance such that said at least one given machine tool will actually fail if said recommended maintenance operation is not performed. 16. The system according to claim 14 or 15, in which said existing fault classifier comprises a neural network, and modifying the existing fault classifier includes adding at least one matching layer to the neural network, and in addition to adding said at least one matching layer, the neural network otherwise is unmodified by said modification. 17. The system according to claim 16, in which the neural network containing the existing fault classifier contains a data layer and an input layer for receiving data from the data layer, wherein said at least one correspondence layer is added between the data layer and the input layer, and at least one matching layer is configured to receive data from the data layer in the modified fault classifier. 18. The system of claim 16 or 17, wherein the first plurality of sensors has a first frequency response distribution and the second plurality of sensors has a second frequency response distribution, wherein the matching layer is configured to match between the first and second frequency response distributions. 19. The system according to any one of paragraphs. 13-18, wherein the first plurality of sensors is configured to sense the same type of signal as the second plurality of sensors. 20. The system of claim 19, wherein said same type of signal comprises one of a vibration signal, a magnetic flux signal, a current signal, a temperature signal, and an internal pressure of the machine. 21. The system according to any one of paragraphs. 13-18, characterized in that said first set of sensors and said second set of sensors are configured to receive mutually different types of signals. 22. The system according to claim. 21, characterized in that these mutually different types of signals contain at least one of: vibration and magnetic flux signals; vibration and electric current signals; vibration and temperature signals; electric current and magnetic flux signals; And vibration and internal pressure signals of the machine. 23. The system according to any one of paragraphs. 14-22, wherein at least some of said machine tool operating states of the first set of machine tools, as indicated by the first set of operating state data sets, are faulty operating states. 24. The system of claim 23, wherein said first plurality of signal sets and the first plurality of operating state data sets of the first plurality of machine tools contain less than 30 faulty operating states.

Images