Sensor-agnostic mechanical fault identification
Transfer learning with a mapping layer adapts existing fault classifiers to accurately identify faults in mechanical machines using diverse sensors, addressing the limitations of existing systems by enhancing classification accuracy with minimal data.
Patent Information
- Application Number
- JP2022513229
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2019-09-03
- Filing Date
- 2020-09-03
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2040-09-03
AI Technical Summary
Existing fault identification systems for mechanical machines are limited in their ability to accurately classify faults when using different types of sensors, requiring extensive retraining or new classifiers due to differences in sensor characteristics, which can be time-consuming and resource-intensive.
A method and system utilizing transfer learning to modify existing fault classifiers by adding a mapping layer that adapts the classifier to differentiate between different sensor types, allowing accurate fault identification using signals from diverse sensors with minimal additional data.
Enables accurate fault classification in mechanical machines using sensors of varying types with a small dataset, preserving the functionality of existing classifiers and reducing the need for extensive retraining, thus improving efficiency and applicability.
Smart Images

Figure 0007577736000002 
Figure 0007577736000003 
Figure 0007577736000004
Abstract
Description
[Technical field]
[0001] Related Applications No. 62 / 895,247, filed September 3, 2019, entitled SYSTEMS AND METHODS FOR MULTI-SENSOR MAPPING FOR AUTOMATED MACHINE DIAGNOSTICS, is hereby referenced, the disclosure of which is incorporated herein by reference, and priority thereto is claimed pursuant to 37 CFR 1.78(a)(4) and (5)(i).
[0002] The present invention relates generally to monitoring mechanical machines, and more particularly to identifying faults in mechanical machines through such monitoring. [Background technology]
[0003] Various systems and methods are known in the art for fault identification in mechanical machines. Summary of the Invention
[0004] The present invention seeks to provide a new system and method for using transfer learning between different types of sensors monitoring signals emanating from a mechanical machine for the purpose of fault identification and maintenance of the monitored mechanical machine.
[0005] Therefore, according to a preferred embodiment of the present invention, there is provided a method for identifying faults in at least one mechanical machine, the method comprising: causing a first plurality of sensors coupled to a corresponding first plurality of mechanical machines to acquire a first plurality of sets of signals emanating from the first plurality of mechanical machines, the first plurality of mechanical machines sharing at least one characteristic; and feeding the at least first plurality of sets of signals of the first plurality of mechanical machines to an existing fault classifier that has been previously trained to automatically identify faults in the second plurality of mechanical machines based on signals previously acquired by the second plurality of sensors emanating from a second plurality of mechanical machines, the second plurality of sensors being of a different type than the first plurality of sensors and the second plurality of mechanical machines sharing at least one characteristic; and modifying an existing fault classifier by utilizing transfer learning based on a first plurality of signals of the at least one given mechanical machine, thereby providing a modified fault classifier; applying the modified fault classifier to at least one additional set of signals acquired by at least one sensor of the first plurality of sensors and emanating from the at least one given mechanical machine sharing the at least one characteristic, the modified fault classifier being configured to automatically identify at least one fault of the at least one given mechanical machine based on the at least one additional set of signals; and providing, by an output device, a human-detectable output including at least the identification of the fault of the at least one given mechanical machine, wherein at least one of a repair or maintenance action is performed based on the human-detectable output.
[0006] According to one preferred embodiment of the present invention, the method also includes, after causing the first plurality of sensors to acquire the first plurality of signal sets and before providing the first plurality of signal sets to the existing fault classifier, obtaining first plurality of sets of operating condition data for mechanical machines of the first plurality of mechanical machines, each set of operating condition data indicating a state of operation of one mechanical machine of the first plurality of mechanical machines, each state of operation being associated with at least one of the signal sets, wherein providing at least the first plurality of signal sets to the existing fault classifier also includes providing the operating condition data of the first plurality of mechanical machines to the existing fault classifier, and modifying the existing fault classifier by utilizing transfer learning based on at least the first plurality of signal sets also includes modifying the existing fault classifier by utilizing transfer learning further based on the first plurality of sets of operating condition data of the first plurality of mechanical machines.
[0007] Preferably, identifying the fault includes identifying a particular fault of the at least one given mechanical machine and predicting a failure of the at least one given mechanical machine due to the particular fault in the absence of performance of the proposed maintenance action on the at least one given mechanical machine, wherein the at least one given mechanical machine will actually fail in the absence of performance of the proposed maintenance action.
[0008] Preferably, the existing fault classifier includes a neural network, and modifying the existing fault classifier includes adding at least one mapping layer to the neural network, and the neural network is otherwise not modified by the modification other than the addition of the at least one mapping layer.
[0009] Preferably, the neural network including the existing fault classifier includes a data layer and an input layer for receiving data from the data layer, and at least one mapping layer is added between the data layer and the input layer, whereby the at least one mapping layer is configured to receive data from the data layer in the modified fault classifier.
[0010] Preferably, the first plurality of sensors has a first frequency response distribution and the second plurality of sensors has a second frequency response distribution, and the mapping layer is configured to map between the first frequency response distribution and the second frequency response distribution.
[0011] According to one preferred embodiment of the method of the present invention, the first plurality of sensors is operable to sense the same type of signals as are sensed by the second plurality of sensors.
[0012] Preferably, the signals of the same type include one of a vibration signal, a magnetic flux signal, a current, a temperature, and a machine internal pressure signal.
[0013] In accordance with another preferred embodiment of the present invention, the first plurality of sensors and the second plurality of sensors are operable to sense different types of signals from each other.
[0014] Preferably, the mutually distinct types of signals include at least one of: vibration and magnetic flux signals, vibration and current signals, vibration and temperature signals, current and magnetic flux signals, and vibration and machine internal pressure signals.
[0015] Preferably, at least some of the operational conditions of a mechanical machine of the first plurality of mechanical machines, as indicated by the first plurality of sets of operational condition data, are faulty operational conditions.
[0016] In accordance with yet another preferred embodiment of the present invention, the first plurality of sets of signals and the first plurality of sets of operating condition data for the first plurality of mechanical machines include less than 30 fault operating conditions.
[0017] According to another preferred embodiment of the present invention, there is provided a system for identifying faults in at least one mechanical machine, the system comprising: a first plurality of sensors coupled to a corresponding first plurality of mechanical machines and operative to acquire a first plurality of signals emanating from the first plurality of mechanical machines, the first plurality of mechanical machines sharing at least one characteristic; and a data processing unit, receiving the first plurality of signals of the first plurality of mechanical machines, the data processing unit including a pre-existing fault classifier previously trained to automatically classify an operating state of the second plurality of mechanical machines based on signals emanating from a second plurality of mechanical machines and previously acquired by the second plurality of sensors, the second plurality of sensors being of a different type than the first plurality of sensors, the second plurality of mechanical machines sharing at least one characteristic; and a data processing unit, receiving the first plurality of signals of the first plurality of mechanical machines, the data processing unit including a pre-existing fault classifier previously trained to automatically classify an operating state of the second plurality of mechanical machines based on signals emanating from a second plurality of mechanical machines and previously acquired by the second plurality of sensors, the second plurality of sensors being of a different type than the first plurality of sensors, the second plurality of mechanical machines sharing at least one characteristic; and applying the modified fault classifier to at least one additional set of signals acquired by at least one sensor of the first plurality of sensors and emanating from the at least one given mechanical machine sharing the at least one characteristic, the modified fault classifier being configured to automatically identify at least one fault of the at least one given mechanical machine based on the at least one additional set of signals. and an output device in communication with the data processing unit and operative to provide a human-sensible output including at least the identification of the fault of the at least one given mechanical machine, wherein at least one of a repair or maintenance operation is performed based on the human-sensible output.
[0018] According to one preferred embodiment of the present invention, the system also includes a data collection unit operative to obtain first multiple sets of operating condition data for mechanical machines of the first plurality of mechanical machines, each set of operating condition data indicative of a state of operation of one of the first plurality of mechanical machines, each state of operation being associated with at least one of the sets of signals, and the data processing unit operative to receive the operating condition data of the first plurality of mechanical machines and to modify the existing fault classifier further based on the first multiple sets of operating condition data of the first plurality of mechanical machines.
[0019] Preferably, identifying the fault includes identifying a particular fault of the at least one given mechanical machine and predicting a failure of the at least one given mechanical machine due to the particular fault in the absence of performance of the proposed maintenance action on the at least one given mechanical machine, wherein the at least one given mechanical machine will actually fail in the absence of performance of the proposed maintenance action.
[0020] Preferably, the existing fault classifier includes a neural network, and the modification of the existing fault classifier includes adding at least one mapping layer to the neural network, and the neural network is not otherwise modified by the modification other than the addition of the at least one mapping layer.
[0021] Preferably, the neural network including the existing fault classifier includes a data layer and an input layer for receiving data from the data layer, and at least one mapping layer is added between the data layer and the input layer, whereby the at least one mapping layer is configured to receive data from the data layer in the modified fault classifier.
[0022] Preferably, the first plurality of sensors has a first frequency response distribution and the second plurality of sensors has a second frequency response distribution, and the mapping layer is configured to map between the first frequency response distribution and the second frequency response distribution.
[0023] According to one preferred embodiment of the system of the present invention, the first plurality of sensors is operable to sense the same type of signals as are sensed by the second plurality of sensors.
[0024] Preferably, the signals of the same type include one of a vibration signal, a magnetic flux signal, a current, a temperature, and a machine internal pressure signal.
[0025] In accordance with another preferred embodiment of the system of the present invention, the first plurality of sensors and the second plurality of sensors are operable to sense different types of signals from each other.
[0026] Preferably, the mutually distinct types of signals include at least one of: vibration and magnetic flux signals, vibration and current signals, vibration and temperature signals, current and magnetic flux signals, and vibration and machine internal pressure signals.
[0027] Preferably, at least some of the operational conditions of a mechanical machine of the first plurality of mechanical machines, as indicated by the first plurality of sets of operational condition data, are faulty operational conditions.
[0028] In accordance with yet another preferred embodiment of the system of the present invention, the first plurality of sets of signals and the first plurality of sets of operating condition data for the first plurality of mechanical machines include less than 30 fault operating conditions. [Brief description of the drawings]
[0029] The present invention will be more fully understood and appreciated from the following detailed description taken in conjunction with the drawings, in which:
[0030] [Figure 1]1 is a simplified high level block diagram illustrating a system for mechanical machine fault identification, constructed and operative in accordance with a preferred embodiment of the present invention; [Diagram 2] FIG. 1 is a simplified high level block diagram illustrating a system for mechanical machine fault identification, constructed and operative in accordance with another preferred embodiment of the present invention. [Diagram 3] FIG. 3 is a simplified diagram of a fault classifier modification performed by a system of the type shown in FIG. 1 or FIG. 2. [Figure 4A-4B] 3A-3C are simplified respective block diagrams of system components utilized in the systems of FIG. 1 or FIG. 2 for training a modified fault classifier; [Figure 5A-5D] 3 is a graph illustrating the performance of a system of the type shown in FIG. 1 or FIG. 2 for fault identification in a mechanical machine. [Figure 6] 1 is a simplified flow chart illustrating steps involved in a method for mechanical machine fault identification in accordance with a preferred embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0031] Reference is now made to FIG. 1, which is a simplified high level block diagram of a system for mechanical machine fault identification, constructed and operative in accordance with a preferred embodiment of the present invention.
[0032] As seen in FIG. 1, a system 100 for mechanical machine fault identification is provided. The system 100 preferably includes a first plurality of sensors 102 of a first type, shown here as sensor type S1. All of the sensors 102 may be of the same type, e.g., sensor type S1, as shown here. Alternatively, the first plurality of sensors 102 may include sensors of multiple types, e.g., sensor types S1, S11-S1N, as shown in FIG. 2, described in more detail below. The sensors 102 are preferably coupled to corresponding first plurality of mechanical machines 104. Here, by way of example, the first plurality of mechanical machines 104 is shown to include mechanical machines 1, 2-N, where N may be any number of mechanical machines, such as two or more mechanical machines. Typically, the sensors 102 are coupled to the mechanical machines 104 in a one-to-one corresponding arrangement, with one of the sensors 102 coupled to a corresponding one of the machines 104. However, other arrangements are possible, and a single sensor may be arranged to sense signals from multiple mechanical machines 104. The sensor 102 may be in physical contact with the machine 104, such as being directly or indirectly attached to the machine 104. Alternatively, the sensor 102 may be physically separated from the machine 104, such as being located a given distance from the machine 104 if the sensor 102 is an optical sensor.
[0033] The sensor 102 may be embodied as any type of sensing device that operates to sense a signal emanating from the mechanical machine 104. Machines and mechanical systems with moving parts, such as bearings, rotors or shafts, or machines including motors, engines, compressors, pumps, fans, gearboxes, chillers, etc., may generate signals during their operation. The mechanical machine 104 may be of any of the aforementioned types of mechanical systems, or other types of mechanical systems that generate signals during their operation. The sensor 102 preferably operates to sense such signals. Analysis of the sensed signals may be used to ascertain the condition of the machine from which the sensed signal was emitted, and in some cases, to ascertain a fault in the machine from which the signal was emitted. By way of non-limiting example only, the sensors 102 may all be the same type of vibration sensors, such as all single-axis accelerometers or all multi-axis accelerometers, to sense vibrations emanating from the machine 104, the sensors 102 may all be the same type of magnetic flux sensors to sense magnetic flux emanating from the machine 104, the sensors 102 may all be the same type of current sensors to sense variations in current generated by the machine 104, and the sensors 102 may all be the same type of temperature sensors to sense heat generated by the machine 104. Alternatively, it will be understood that the sensors 102 may be of the same type as other kinds of sensing devices capable of sensing signals emanating from and generated by the machine 104, including signals associated with machine operation, such as torque, displacement, input line frequency, etc. Alternatively, the sensors 102 may include two or more types of sensors, such as, by way of example only, magnetic flux sensors and vibration sensors.
[0034] Mechanical machines 1-N that are members of the first plurality of mechanical machines 104 are preferably characterized by one or more shared characteristics. Mechanical machines 1-N may or may not be the same machine, so long as they have at least one shared characteristic in common. For example, the shared characteristic may refer to a type, model number, manufacturer, physical characteristics or dimensions, operating characteristics or parameters, or other shared characteristics that indicate that an observed behavior of one of the mechanical machines of the plurality of mechanical machines may be unique to another mechanical machine of the plurality of mechanical machines.
[0035] The first plurality of sensors 102 preferably operate to obtain a first plurality of sets of signals emanating from the first plurality of mechanical machines 104. A set of signals may be a "signal snapshot" sensed by an appropriate one of the sensors 102 over a short period of time. For example, the signals may be sensed for a period of a few seconds, such as 1-4 seconds. Alternatively, each set of signals may include multiple "signal snapshots" over time, such as, for example, a 4 second "signal snapshot" measured every hour over a period of hours, days, or months. Alternatively, each set of signals may include signals that are continuously monitored over a longer period or multiple periods. For example, the signals may be continuously monitored every millisecond.
[0036] The signals acquired by the sensor 102 may be pre-processed, for example, by analog or digital processing capabilities of the sensor 102 itself, or by other hardware and / or software signal processing components 108. Although the signal processor 108 is shown in FIG. 1 as a separate element from the sensor 102, it will be understood that signal processing functions may be incorporated within one or more of the sensors 102. Signal pre-processing may include at least one of digitization, compression, feature extraction, and representation of the signal in the time or frequency domain.
[0037] In one embodiment of the invention, the set of signals acquired by the first plurality of sensors 102 may be uploaded to a remote server, such as a server in the cloud. The signal processing function 108 may be performed on the remote server. Preferably, the set of signals acquired by the first plurality of sensors 102 are stored as they are acquired from the first plurality of machines 104. For example, the set of signals may be stored on a server in the cloud.
[0038] The system 100 may optionally include a data collection unit 110. The data collection unit 110 may be operable to receive a first plurality of sets of mechanical machine operating condition data of the plurality of mechanical machines 104, each set of operating condition data indicative of a state of operation of one of the first plurality of mechanical machines 104, each state of operation being associated with at least one of the sets of signals acquired by the first plurality of sensors 102.
[0039] The operational condition data collected in the data collection unit 110 is preferably in the form of machine condition diagnostics provided by a human expert, such as an engineer. The set of signals acquired by the first plurality of sensors 102 may be optionally provided to the data collection unit 110 by the signal processor 108 and / or directly or indirectly by the sensors 102. The human expert may optionally analyze the set of signals acquired by the first plurality of sensors 102 and label each signal of the set of signals as representing a particular operational condition of the corresponding mechanical machine 104 from which the signal was emitted. The human expert may, for example, interact with a user interface of the data collection unit 110 or another device enabling communication between the human expert and the data collection unit 110 to input the operational condition data. The signals and the labels applied thereto may be accumulated and stored in a database in the data collection unit 110. In one embodiment of the present invention, the data collection unit 110 may be located in a remote server, such as a server in the cloud.
[0040] Fault identification by the human expert may include identification of one or more specific faults of the monitored machinery 104. Depending on the particular machinery 104 being monitored, the specific faults identified may include rotating machinery bearing wear, mechanical looseness, misalignment, imbalance, electrical faults or other faults. Alternatively, fault identification may include identification of the machinery 104 in a faulty state, i.e., in an abnormal state with respect to its normal, healthy operating state, without identifying a specific fault. In this case, fault identification by the human expert identifies that the machinery is not operating normally, but does not identify the specific cause of the unhealthy operation.
[0041] Regardless of whether the particular operating conditions represented by the signals have been labeled by a human expert, the set of signals acquired by the sensors 102 may include both a set of signals corresponding to healthy, fault-free operating conditions of the machine 104 and a set of signals corresponding to unhealthy, faulty conditions of the machine 104.
[0042] Alternatively, the set of signals acquired by the sensors 102 does not necessarily include a set of signals that correspond to an unhealthy fault condition of operation of the machine 104. According to this embodiment of the invention, all of the machines 104 being monitored by the sensors 102 may be in healthy operating conditions. Thus, the sets of signals, and optionally associated conditions of the machine operation, may all correspond to healthy conditions of the machine operation.
[0043] The accumulated set of signals from the first plurality of machines 104 is preferably provided to an existing fault classifier 114 included in the system 100. If the system 100 also includes a data collection unit 110 and the signals are labeled, the labeled signals are preferably combined and provided to the existing fault classifier 114 by the signal pre-processor 108 and the data collection unit 110.
[0044] The existing fault classifier 114 may be an algorithmic classifier. For example, the existing fault classifier 114 may be stored on a remote server. The system 100 may include a non-transitory computer-readable storage medium having stored thereon computer-executable instructions for performing, by a processor, the functions of the existing fault classifier 114. The one or more processors that execute the existing fault classifier 114 may be a remote processor, e.g., located in the cloud, or may be a local processor.
[0045] The pre-existing fault classifier 114 is preferably a fault classifier previously trained to automatically identify faults of the second plurality of mechanical machines based on signals emanating from the second plurality of mechanical machines and previously acquired by a second plurality of sensors, the second plurality of sensors being of a different type than the first plurality of sensors 102. The second plurality of sensors may all be of the same type as each other, which may be different from the type of sensors of the first plurality of sensors 102. Alternatively, the second plurality of sensors may all be of the same type as each other, which may be different from at least one of the types of sensors of the first plurality of sensors 102. Alternatively, the second plurality of sensors may include multiple types of sensors, which may all be of a different type than one or more sensors of the first plurality of sensors. The second plurality of sensors may alternatively include multiple types of sensors, which may be different from at least one of the types of sensors of the first plurality of sensors 102. The second plurality of mechanical machines preferably share at least one mechanical characteristic shared by the first plurality of mechanical machines 104 .
[0046] It is understood that the fault classifier 114 is referred to herein as “pre-existing” because the fault classifier 114 may be pre-existing with respect to the set of signals and optional operational condition data acquired by the sensors 102 from the machine 104. The fault classifier 114 may have been previously generated at an earlier point in time prior to generating the data set including the set of signals and optional operational condition data of the machine 104.
[0047] Preferably, the fault classifier 114 is an accurate classifier configured to accurately identify faults in mechanical machines sharing at least one common characteristic based on signals acquired by the second plurality of sensors. The fault classifier 114 may be such an accurate classifier because it has been previously trained using machine learning on a large data set including signals and associated operating condition data acquired from potentially a large number of machines. As is well known to those skilled in the art, the greater the amount of data that is fed to a machine learning fault classifier for the purposes of its training, the more accurately the classifier can perform, up to a given limit. For example, the fault classifier 114 may have been trained using data sensed from over 40,000 individual rotating machines including motors, pumps, fans, chillers, compressors, and gearboxes. The common characteristic shared by such machines may be that they contain bearings. An example of how the fault classifier 114 has been previously trained is shown in Figures 3-4B, described below.
[0048] Therefore, the existing fault classifier 114 may be successfully applied to signals emanating from a mechanical machine having shared characteristics with the mechanical machine based on the data on which the fault classifier 114 was trained to identify the faults. However, it should be noted that the fault classifier 114 was previously trained based on signals acquired by a particular type of sensor, i.e., a second plurality of sensors of a different type than the first plurality of sensors 102. As a result, the fault classifier 114 may best classify and identify faults when applied to signals acquired by the particular sensors on which the fault classifier 114 was trained. However, if the signals provided to the fault classifier 114 are acquired by a different type of sensor than the second plurality of sensors on which the fault classifier 114 was trained, the fault classifier 114 may not be able to accurately classify and identify faults based on these signals. This is because the sensor characteristics differ between the sensor, i.e., the second plurality of sensors on which the fault classifier 114 was trained, and the sensor, e.g., the first plurality of sensors S1 or the first plurality of sensors S1-S1N (FIG. 2), that acquired the current signal requiring classification.
[0049] Even though the fault classifier 114 is a highly accurate classifier for identifying faults based on signals acquired by the second plurality of sensors, the fault classifier has limited, if any, use to identify faults acquired by a different type of sensor, e.g., the first plurality of sensors S1 or S1-S1N (FIG. 2). If the fault classifier 114 is applied to signals acquired by the first plurality of sensors, the results will not be accurate.
[0050] This may be illustrated by reference to the case of two types of vibration sensors, such as a triaxial accelerometer and a single-axis accelerometer, that sense vibration signals generated by a mechanical machine. The two types of vibration sensors differ from each other in various parameters such as shape, mass, internal materials, etc., resulting in differences in the moment of inertia and resonant frequency of each sensor. As a result, the frequency responses of the sensors differ from each other. A particular signal generated by the mechanical machine being monitored will be sensed and recorded differently by the two sensors due to the inherent differences between the sensors. Furthermore, this difference is further exacerbated if the two sensors are mounted at different locations on the machine being monitored, as the sensors will measure along different measurement axes relative to each other, resulting in different vibration levels being measured due to the difference in location.
[0051] For example, if a triaxial accelerometer and a single-axis accelerometer are mounted on a cylindrical machine, the two types of accelerometers will measure different vibration levels from each other since the triaxial accelerometer measures radial vibrations along one axis and tangential vibrations along two axes while the single-axis accelerometer measures radial or direct vibrations.
[0052] Therefore, using a fault detection classifier trained on signals acquired from one type of sensor, e.g., a vibration signal acquired by a three-axis accelerometer, will lead to inappropriate classification results when applied to the same type of signal acquired by a different type of sensor, e.g., a vibration signal acquired by a single-axis accelerometer.
[0053] This may be further illustrated by reference to the more extreme case of two types of sensors that sense different types of signals, such as a vibration sensor and a magnetic flux sensor that respectively sense vibration and magnetic flux signals emanating from a particular mechanical machine. A fault detection classifier trained using data acquired by one of the types of sensors, such as a vibration signal acquired by a vibration sensor, will be limited to classifying vibration signals and will produce results that are of little relevance if applied to identify faults in, for example, magnetic flux signals generated by the same machine.
[0054] In order to provide a fault classifier that can accurately classify signals acquired by sensors of different types than the fault classifier was previously trained on, a completely new fault classifier can be trained based on signals acquired by sensors of different types, e.g., the first plurality of sensors S1 or S1, S11 to S1N (FIG. 2). In this case, the existing fault classifier 114 is not utilized and a new fault classifier is developed to identify faults in a mechanical machine, such as the machine 104. However, in order for this new classifier to provide accurate fault identification, a large amount of new data acquired by sensors of different types, e.g., the first plurality of sensors S1 or the first plurality of sensors S1 to S1N (FIG. 2), needs to be provided on which the fault classifier needs to be trained. Such a process may be time consuming and such a large amount of data may not be available. Moreover, such a process is significantly limited in performance and scope of application due to the many parameters that control the frequency dependence of the machine signals, such as the mounting position, orientation, mounting type of the sensors, etc. Moreover, in this approach, the functionality of the original existing fault classifier 114 is simply wasted rather than utilized since the previous fault classifier 114 is never applied.
[0055] The present invention advantageously provides a solution to the problem that a fault classifier trained on data obtained from a particular type or types of sensors is of limited, if any, use in classifying signals obtained from sensors of different types due to differences in sensor characteristics. Advantageously, the present invention does not require training a new classifier "from scratch" based on signals from sensors of different types. Rather, the present invention utilizes a transfer learning approach for mapping between the original sensor type on which the existing classifier was trained and the new, different sensor type from which new data requiring classification is obtained.
[0056] The present invention may utilize an existing classifier 114 by modifying the existing classifier 114 based on a mapping between different sensor frequency response distributions of each different sensor type, and requires only a small data set from a different type of sensor, e.g., the first plurality of sensors 102, to perform such mapping and modification. Thus, in a preferred embodiment of the present invention, a modified classifier may be generated based on the original existing classifier 114 and a small new data set obtained from a different type of sensor than the type of sensor on which the existing classifier 114 was based. This modified classifier may be able to accurately identify faults in signals obtained by a different type of sensor, e.g., the first plurality of sensors 102, despite the small data set fed to it. The modified classifier utilizes the original existing classifier 114 and maps it to a different type of sensor, e.g., the first plurality of sensors 102, so that it can be accurately applied to signals obtained by the different type of sensor.
[0057] However, it should be noted that the present invention may be useful even when a large, high-quality dataset is available from a different type of sensor, e.g., the first plurality of sensors 102. In this case, a large, high-quality dataset is available so that a new dedicated classifier can be trained to provide suitable results, but the use of transfer learning to modify an existing classifier may still be advantageous in order to take advantage of the functionality of the original, existing classifier. Thus, while the present invention is expected to be most useful when sufficient data is not available in terms of quantity and / or quality to train a new classifier, the present invention may also be useful when a large, high-quality dataset is available.
[0058] Thus, for an existing accurate fault classifier trained using signals acquired by a particular type or types of sensors, the present invention provides a solution to modify the classifier so that it can accurately classify signals acquired by any other type or types of sensors different from the particular type or types of sensors on which the fault classifier was trained, where these signals have at least one shared characteristic with the machine on which the classifier was previously trained. This is sometimes referred to as a sensor-agnostic approach, where the classifier can be calibrated to be applicable to data acquired by any sensor, regardless of the type and / or structure of the source data collected by the sensor.
[0059] The inventive approach may be applicable in the case of signals of the same type, e.g. vibration signals, acquired by different types of vibration sensors, the results of which are shown in Figures 5A-5C. The inventive approach may also be applicable in the case of signals of different types, e.g. vibration and magnetic signals, acquired by different types of sensors, e.g. vibration sensors and magnetic flux sensors, the results of which are shown in Figure 5D.
[0060] In either case, the original existing classifier may be modified based on a mapping between the frequency response distributions of the sensors to create a modified classifier that is capable of identifying faults in signals acquired by a different type of sensor than the existing classifier was trained on. In either case, the modified classifier may be applied to signals acquired by a different type of sensor than the existing classifier was trained on with greater accuracy than would be achieved by applying the existing classifier in its original, unmodified form. Furthermore, in either case, the modified classifier may be applied to signals acquired by a different type of sensor than the existing classifier was trained on with greater accuracy than would be achieved by applying a new classifier trained using only signals acquired by the different type of sensor.
[0061] The acquisition of small new data sets from different types of sensors, such as the first plurality of sensors 102, has been described above with respect to the sensors 102 acquiring signals and optional associated operating condition data from the machine 104. The acquisition of small data sets by the system of Figure 1 is preferably performed by elements enclosed in dashed box 115.
[0062] In the embodiment of the invention shown in FIG. 1, the first plurality of sensors 102 are preferably all the same type of sensor, i.e., sensor type S1. However, as previously mentioned, the first plurality of sensors 102 may alternatively include multiple types of sensors. FIG. 2 illustrates an alternative embodiment of the system 100, here shown as system 100A, showing multiple types of sensors included in the first plurality of sensors 102. Turning now to FIG. 2, the first plurality of sensors 102 may include sensors S1, S11-S1N on each of machines 1-N of the plurality of machines 104. It will be appreciated that the first plurality of sensors 102 may include any number of sensors S1-S1N, such as two or more sensors. These sensors may be, for example, vibration sensors, magnetic flux sensors, current sensors, temperature sensors, or combinations of sensors for sensing other parameters associated with the operation of the machine 104, such as torque, displacement, input line frequency, etc. The sensors S1, S11-S1N may be of different types from one another, but preferably are of the same type with respect to the set of sensors coupled to each of the machines 104. However, it is understood that the system 100A may tolerate some cases of missing data from one of the sensors 102, where the missing data may be imputed, for example, by using multivariate statistics.
[0063] System 100A may be generally similar to system 100, except for the types of sensors included in first plurality of sensors 102 of system 110A, and the description of system 100 generally applies to system 100A as well.
[0064] 1 and 2, the small new data set on which the existing classifier 114 may be modified by mapping learning to provide a modified classifier may include much less data than the amount of data on which the existing classifier 114 was initially trained. For example, the small new data set may include 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 plurality of signal sets and optionally a first plurality of sets of operating condition data of the first plurality of mechanical machines 104. Additionally, within the small new data set, there may be even fewer signal sets corresponding to a faulty operating condition of one of the first plurality of mechanical machines 104, such as, for example, less than 100, less than 90, less than 80, less than 70, less than 60, less than 50, less than 30, or less than 20 sets of signals corresponding to a faulty operating condition of one of the first plurality of mechanical machines 104. As noted above, in some cases, the small new data set may not even include a signal corresponding to a faulty operating condition of one mechanical machine of the first plurality of mechanical machines 104. This is in contrast to a much larger data set, such as thousands of samples, on which the existing classifier 114 may have been trained.
[0065] It is understood that while the first plurality of sensors 102 is different from the second plurality of sensors on which the existing classifier 114 was trained, the machines 104 preferably share common characteristics both with each other and with the machines on which the existing classifier 114 was trained. The machines 104 may or may not be the same as those machines on which the existing classifier 114 was previously trained.
[0066] The first plurality of sets of signals acquired by the first plurality of sensors 102, and optionally the first plurality of sets of operating condition data collected by the data collection unit 110 of the first plurality of mechanical machines 104, may be provided to the existing fault classifier 114. The first plurality of sets of signals may be pre-processed by the signal processor 108 before being provided to the existing fault classifier 114.
[0067] The existing fault classifier 114 is then preferably modified based on the multiple sets of signals and, optionally, the multiple sets of operating condition data for the first multiple mechanical machines 104, thereby generating a modified classifier 116.
[0068] The modified classifier 116 may be an algorithmic classifier executable by one or more processors, which may be the same or different processors that execute the existing classifier 114. For example, the existing classifier 114 and the modified classifier 116 may be embodied in a data processing unit. The one or more processors that execute the existing fault classifier 116 may be a remote processor, e.g., located in the cloud, or may be a local processor. For example, the one or more processors that execute the modified classifier 116 may be located within one of the sensors 102.
[0069] The existing fault classifier 114 is modified by tuning the classifier to a new data set including a set of signals acquired by the sensors 102 and, optionally, a set of operating condition data of the first plurality of mechanical machines 104. For example, as seen in FIG. 1, if the first plurality of sensors 102 includes only sensor type S1, the modified classifier 116 may be adapted to sensor type S1. As a further example, as seen in FIG. 2, if the first plurality of sensors includes multiple sensor types S1, S11, etc., the modified classifier 116 may be adapted to those multiple sensor types. By tuning the existing classifier 114 to generate the modified classifier 116, the classifier is calibrated to be applicable to a different data set than the data set on which the classifier was originally trained. The tuning involves a mapping between the original sensor characteristics on which the classifier 114 was previously trained and the sensor characteristics of the first plurality of sensors 102.
[0070] In one embodiment of the invention, the tuning involves a mapping between the sensor frequency response distributions of the original plurality of sensors on which the classifier 114 was previously trained and the sensor frequency response distributions of the first plurality of sensors 102. Because the frequency responses of both the sensors on which the classifier 114 was previously trained and the first plurality of sensors 102 may be distributed, the mapping may be between sensor frequency response distributions rather than simply sensor frequency responses. This distribution may occur due to actual variations between sensors within each plurality, such as changes in machine characteristics, changes in sensor location, sensor mounting, exact type of sensor mounting, etc.
[0071] The mapping may be performed in a supervised manner. In this embodiment, the data collection unit 110 may be included in the system 100, and the labeled signals are fed to an existing fault classifier 114. Further details regarding how the existing fault classifier 114 is modified by the mapping between sensors in a supervised manner are provided hereafter with reference to FIG.
[0072] Alternatively, the mapping may be performed in an unsupervised manner. In this case, the data collection unit 110 does not need to be included in the system 100 and signals without associated operating conditions are fed to the existing fault classifier 114. The existing fault classifier 114 may be modified by unsupervised learning to map differences between the sensor frequency response distributions of sensors on which the classifier 114 was previously trained and a new sensor, e.g., the first plurality of sensors 102.
[0073] Alternatively, the mapping may be performed in a semi-supervised manner, where the labeled signals are fed to an existing fault classifier 114, and the mapping between the sensors to generate a modified classifier 116 is performed in an unsupervised manner. Further details regarding unsupervised mapping are provided hereafter with reference to Figures 3-4B.
[0074] The modified classifier 116, adapted by the mapping so that it can be applied to data from the first plurality of sensors 102, is ready to be used to classify signals acquired by sensors of the same type as the first plurality of sensors 102.
[0075] An example of the use of the modified classifier 116 to identify faults in signals acquired by sensors of the same type as the first plurality of sensors 102 is further illustrated in Figures 1 and 2. As seen in Figures 1 and 2, at least one sensor 122, e.g., sensor S1 in Figure 1 and sensors S1-S1N in Figure 2, can acquire at least one set of signals emanating from at least one given mechanical machine, denoted here as mechanical machine X, denoted by reference numeral 124. The mechanical machine 124 may share at least one characteristic with the first plurality of mechanical machines 104, as well as the plurality of mechanical machines based on the signals on which the existing fault classifier 114 was trained. The mechanical machine 124 may or may not be a member of the first plurality of mechanical machines 104. While the mechanical machine 124 is shown here embodied as a single machine, this is for simplicity only, and it will be understood that the system 100 may include any number of mechanical machines 124, such as one, two, or more mechanical machines 124 that may or may not be identical to one another, provided that the machines 124 share at least one characteristic as described above.
[0076] The at least one sensor 122 is one of the first plurality of sensors 102, meaning that the sensor 122 is of the same type as the first plurality of sensors 102. As shown in FIG. 1, if the first plurality of sensors 102 includes a single type of sensor, i.e., sensor S1, the sensor 122 is also of sensor type S1. As shown in FIG. 2, if the first plurality of sensors 102 includes multiple types of sensors, i.e., sensors S1-S1N, the at least one sensor 122 also includes the same multiple types of sensors, i.e., sensors S1-S1N. The at least one sensor 122 preferably operates to acquire at least one set of signals emanating from a given mechanical machine 124. The set of signals may be a "signal snapshot" sensed by the at least one sensor 122 over a short period of time. For example, the signals may be sensed for a period of several seconds, such as 1-4 seconds. Alternatively, a set of signals may include multiple "signal snapshots" over time, such as, for example, a 4 second "signal snapshot" measured every hour over a period of hours, days, or months. Alternatively, a set of signals may include signals that are continuously monitored over a longer period or multiple periods. For example, signals may be continuously monitored every millisecond.
[0077] The signals acquired by the at least one sensor 122 may be pre-processed, for example, by analog or digital processing capabilities of the sensor 122 itself or by other hardware and / or software signal processing components 128. Although the signal processor 128 is shown in Figures 1 and 2 as a separate element from the at least one sensor 122, it will be understood that signal processing functions may be incorporated within one or more of the sensors 122. Signal pre-processing may include at least one of digitization, compression, feature extraction, and signal representation in the time or frequency domain.
[0078] In one embodiment of the invention, signals acquired by at least one sensor 122 may be uploaded to a remote server, such as a server in the cloud. Signal processing functions 128 may be performed on the remote server.
[0079] A set of signals acquired by at least one sensor emanating from at least one given mechanical machine, embodied here as sensor 122 by way of example, and embodied here as mechanical machine X by way of example, may be provided to the modified classifier 116. The modified classifier 116 may be configured to automatically identify at least one fault of the mechanical machine X based on the set of signals emanating therefrom. It will be appreciated that the modified classifier 116 may be accurately applied to signals acquired by at least one sensor 122, since the modified classifier 116 is adapted to classify signals acquired by sensors of the type of the first plurality of sensors 102.
[0080] The identification of the fault by the modified classifier 116 may include identification of one or more specific faults of the monitored machinery 104. Depending on the particular machinery 104 being monitored, the specific faults identified may include rotating machinery bearing wear, mechanical looseness, misalignment, imbalance, electrical faults or other faults.
[0081] Alternatively, fault identification may involve identification of machine X in a faulty state, i.e., in an abnormal state with respect to its normal, healthy operating state, without identifying a specific fault. In this case, fault identification identifies that the machine is not operating normally, but does not identify the specific cause of the unhealthy operation.
[0082] The systems 100 and 100A may further include an output device 130. The output device 130 may be operable to receive the fault identification output by the modified classifier 116 and provide a human perceptible output including at least an identification of a fault of at least one given mechanical machine 124. The human perceptible output may include at least one of a visual, tactile, or auditory output. Preferably, at least one of a repair or maintenance action is performed based on the human perceptible output.
[0083] For example, if the modified classifier 116 is executable by a remote processor, the fault identification output by the modified classifier 116 may be communicated to the output device 130 .
[0084] The output device 130 is also operable to provide a prediction of failure of the at least one given mechanical machine 124 due to a failure identified by the modified classifier 116 in the absence of performance of the proposed maintenance action on the at least one given mechanical machine, and the at least one given mechanical machine 124 will actually fail in the absence of performance of the proposed maintenance action.
[0085] In some cases, maintenance 132 may be performed on a given machine 124 in response to a human perceptible output provided by output device 130. Output device 130 may optionally be operatively coupled to a controller of machine 124, such that operation of machine 124 may be adjusted in response to fault identification. For example, machine 124 may be switched off, operated at reduced power, or otherwise adjusted. Such adjustments may be automatic or may be directed by a human expert in response to a human perceptible output provided by output device 130.
[0086] Further details of the existing classifier 114 itself and how it may be adapted or calibrated to generate a modified classifier 116 are provided with reference to Figures 3-4B, which are respectively a simplified diagram of classifier modification performed by either of the types of systems shown in Figures 1 and 2, and a simplified respective block diagram of components of two possible systems for training a classifier utilized in the systems of Figures 1 or 2. In one embodiment of the invention, systems 100 and 100A may include a non-transitory computer-readable storage medium having stored thereon computer-executable instructions for performing, by one or more processors, a method of classifier modification, as described in more detail below with respect to Figures 4A and 4B. Such a processor may be a remote processor or a local processor.
[0087] As seen in FIG. 3, the existing fault classifier 114 may be an artificial neural network (ANN) classifier 300 for monitored fault and anomaly detection. For example, the ANN 300 may be configured to identify specific faults or performance anomalies of a machine having rotating components. Without loss of generality, the fault classifier 114 having a multi-layer perceptron architecture is shown in FIG. 3. However, it is understood that the systems and methods of the present invention may be applied to any type of fault classifier, regardless of its structure. More specifically, the few layers and neurons shown in FIG. 3 are only highly simplified examples for the purpose of illustrating the principles of the present invention. Furthermore, the existing fault classifier 114 may be a classifier based on an autoencoder, a deep belief network, clustering, K-means or a hidden Markov model, or any other type of model capable of performing unsupervised learning, as described above.
[0088] The top panel 302 of FIG. 3 shows an example of an existing classifier 114 in its original, unmodified form, after its training, and before its modification for the purposes of mapping between sensor types in accordance with the present invention.
[0089] The existing classifier 114 may have a multi-layer architecture. The data layer 308 may be a layer that contains input data in the form of a set of sensor signals. The input layer 310 may be an initial layer where the sensor signal set is input to the network 300. In each of the subsequent multiple hidden layers 312, the signal set is fused with an activation function that is applied to the respective weighting and combining layer output before forwarding to the next layer of hidden layers 312. This process repeats itself for each layer of hidden layers 312 until it reaches the output layer 314. The output layer 314 generates a fault score. This score represents the identification of the fault. The identification of the fault may include the identification of a current specific fault or anomaly, or the prediction of a future impending specific fault. The fault may include any type of machine anomaly.
[0090] The general mathematical formula for the ANN architecture shown in panel 302 of FIG. 3 in the network layer is:
number
[0091] The training of the network 300 to generate the existing classifier 114 may be better understood with further reference to FIGS. 4A and 4B.
[0092] Turning first to FIG. 4A, a system 400 for training an existing classifier 114 may include a second plurality of sensors 402. The sensors 402 are preferably coupled to corresponding second plurality of mechanical machines 404. The second plurality of sensors 402 may all be of the same second type, shown here as sensor type S2. Alternatively, the sensors 402 may include multiple types of sensors coupled to respective corresponding ones of the second plurality of mechanical machines 404. The case of a second plurality of sensors 402 including multiple types of sensors is shown in system 400A of FIG. 4B. System 400A is generally similar to system 400, except that multiple types of sensors 402 are included. Turning now to FIG. 4B, it is understood that the second plurality 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, vibration sensors, magnetic flux sensors, current sensors, temperature sensors, or a combination of sensors for sensing other parameters associated with the operation of the machine 404, such as torque, displacement, input line frequency, etc. Sensors S2-S2N may be of different types from one another, but preferably are of the same type for each of the sets of sensors coupled to each of the machines 404.
[0093] The sensor 402, e.g., sensor S2 or sensors S2-S2N, may be different from the first plurality of sensors 102 of FIG. 1, or may be different from the sensor types S1-S1N of the first plurality of sensors 102 of FIG. 2. Any combination of the embodiments shown in FIGS. 1 and 2 and 4A and 4B is possible, i.e., the first plurality of sensors 102 may include only one type of sensor S1 and the second plurality of sensors 402 may include only one type of sensor S2 different from S1 (embodiment of FIG. 1 and FIG. 4A), the first plurality of sensors 102 may include only one type of sensor S1 and the second plurality of sensors 402 may include multiple types of sensors S2-S2N, at least some of which are different from S1 (embodiment of FIG. 1). and 4B), the first plurality of sensors 102 may include multiple types of sensors S1-S1N and the second plurality of sensors 402 may include only one type of sensor S2 that is different from at least some of S1-S1N (embodiment of FIGS. 2 and 4A), the first plurality of sensors 102 may include multiple types of sensors S1-S1N and the second plurality of sensors 402 may include multiple types of sensors S2-S2N that are at least some of S1-S1N different from (embodiment of FIGS. 2 and 4B).
[0094] In some embodiments, the first and second pluralities of sensors 102 and 402 may include some sensors in common, for example, the first pluralities of sensors 102 may include vibration and magnetic sensors and the second pluralities of sensors 402 may include only magnetic sensors.
[0095] Here, by way of example, the second plurality of mechanical machines 404 is shown to include mechanical machines 1, 2 through N, where N may be any number of mechanical machines, such as two or more mechanical machines. Typically, the sensors 402 are coupled to the mechanical machines 404 in a one-to-one corresponding arrangement, with one of the sensors 402 coupled to a corresponding one of the machines 404. However, other arrangements are possible, and a single sensor may be positioned to sense signals from the multiple mechanical machines 404. The sensor 402 may be in physical contact with the machine 404, such as being directly or indirectly attached to the machine 404. Alternatively, the sensor 402 may be physically separated from the machine 404, such as being located a given distance from the machine 404, for example if the sensor 402 is an optical sensor.
[0096] In an embodiment of the invention, the sensors of the second plurality of sensors 402 may generally have the same frequency response as one another. The frequency response of the second plurality of sensors 402 may be referred to as the frequency response distribution of the sensors 402. The first plurality of sensors 102 may also generally have the same frequency response as one another. The frequency response of the first plurality of sensors 102 may be referred to as the frequency response distribution of the sensors 102. The frequency response distribution of the first plurality of sensors 102 is different from the frequency response distribution of the second plurality of sensors 402 because the sensors 402 are of a different type than the sensors 102. For example, the first 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. Sensors 102 and 402 or sensors 402 and 102, respectively, may measure, by way of example only, vibration and magnetic flux signals, vibration and current signals, vibration and temperature signals, current and magnetic flux signals, and vibration and machine internal pressure signals.
[0097] The sensor 402 is preferably operative to sense a signal emanating from the mechanical machine 404. The mechanical machines 1-N that are members of the second plurality of mechanical machines 404 are preferably characterized by one or more shared characteristics with each other and with the mechanical machine 104. The mechanical machines 1-N may or may not be the same machine, so long as they have at least one shared characteristic in common. For example, the shared characteristic may refer to a type, model number, manufacturer, physical characteristics or dimensions, operating characteristics or parameters, or other shared characteristics that indicate that an observed behavior of one of the mechanical machines of the plurality of mechanical machines may be unique to another mechanical machine of the plurality.
[0098] The second plurality of sensors 402 preferably operates to obtain a set of a second plurality of signals emanating from the second plurality of mechanical machines 404. A set of signals may be a "signal snapshot" sensed by an appropriate one of the sensors 402 over a short period of time. For example, the signals may be sensed for a period of a few seconds, such as 1-4 seconds. Alternatively, each set of signals may include multiple "signal snapshots" over time, such as, for example, a 4 second "signal snapshot" measured every hour over a period of hours, days, or months. Alternatively, each set of signals may include signals that are continuously monitored over a longer period or multiple periods. For example, the signals may be continuously monitored every millisecond.
[0099] The signals acquired by the sensor 402 may be pre-processed, for example, by analog or digital processing capabilities of the sensor 402 itself or by other hardware and / or software processing components 408. For example, the signals may be digitized and / or compressed, features may be extracted from the signals, the signals may be represented in the time or frequency domain.
[0100] Systems 400 and 400A may optionally include a data collection unit 410. Data collection unit 410 is preferably operable to receive a second plurality of sets of mechanical machine operating condition data of the second plurality of mechanical machines, each set of operating condition data indicative of a state of operation of one of said second plurality of mechanical machines 404, each state of operation being associated with at least one of the sets of signals acquired by the second plurality of sensors 402.
[0101] The operational condition data collected by the data collection unit 410 is preferably in the form of machine condition diagnostics supplied by a human expert, such as an engineer. The set of signals acquired by the first plurality of sensors 402 may optionally be provided to the data collection unit 410 by the signal processor 408 and / or directly or indirectly by the sensors 402.
[0102] These human experts can analyze the set of signals acquired by the second plurality of sensors 402 and label each signal of the set of signals as representing a particular operating condition of the corresponding mechanical machine 404 from which the signal originated. The human experts can, for example, interact with a user interface of the data collection unit 410 or another device enabling communication between the human experts and the data collection unit to input the operating condition data. The signals and the labels applied thereto can be accumulated and stored in a database within the data collection unit 410. In one embodiment of the present invention, the data collection unit 410 can be located on a remote server, such as a server in the cloud.
[0103] Fault identification by the human expert may include identification of one or more specific faults of the monitored machinery 404. Depending on the particular machinery 404 being monitored, the specific faults identified may include rotating machinery bearing wear, mechanical looseness, misalignment, imbalance, electrical faults or other faults. Alternatively, fault identification may include identification of the machinery 404 in a faulty state, i.e., in an abnormal state with respect to its normal, healthy operating state, without identifying a specific fault. In this case, fault identification by the human expert identifies that the machinery is not operating normally, but does not identify the specific cause of the unhealthy operation.
[0104] It should be understood that in this example, system 400 / 400A preferably includes a data collector 410, and signals are preferably labeled by a human expert to train fault classifier 300 in a supervised manner. However, in other embodiments of the invention, data collector 410 may be omitted, and signals do not need to be labeled by a human expert. In this case, existing fault classifier 114 may be trained first by using unsupervised learning.
[0105] Returning now to panel 302 of FIG. 3, the network 300 may be trained by feeding, at layer 308, a set of signals acquired by the second plurality of sensors 402 as input data, and at output layer 314, associated operational condition data labeled by a human expert as the required corresponding output. Training of the network 300 may be performed using backpropagation and gradient descent algorithms with respect to predefined data labeling. Training parameters such as loss function, learning rate, optimizer type, etc. are preferably selected with respect to the output scores. For example, in the case of binary fault detection, a cross-entropy loss function, Adam optimizer and L2 regularization term may be selected. The training process is first applied to a data set where all sensors 402 are consistent within the entire data set, meaning that there is preferably no mixing between sensors 402 for each vector in the data points.
[0106] Once training is complete, the parameters of the network 300 established based on the training are preferably fixed. These parameters include activation functions and weights. The network 300 now configures an existing classifier, such as the existing classifier 114, based on a set of sensor signals acquired by the second plurality of sensors 402. The existing classifier 114, in the form of the trained network 300 shown in panel 302, is now configured to accurately classify new input data having the same or similar structure and acquired from the same or similar source as the data on which the existing classifier 114 was trained. In this case, the existing classifier 114, in the form of the trained network 300, is configured to accurately classify new input signals originating from a machine having characteristics shared with the machine 404 (FIGS. 4A and 4B) and sensed by the same type of sensor as the second plurality of sensors 402 (FIGS. 4A and 4B). However, as detailed above, an existing classifier 114 in the form of a trained network 300 may be unable to accurately classify new input signals sensed by a different type of sensor than the second plurality of sensors 402, such as the first plurality of sensors 102 (FIGS. 1 and 2), even though these new input signals originate from machines having characteristics shared with the machines 404, such as the first plurality of machines 104. In this case, the accuracy of the classifier is significantly degraded due to differences in frequency responses of the various types of sensors.
[0107] The existing classifier 114 may be modified to enable it to accurately classify a set of signals from a different type of sensor than the one the classifier was previously trained on, e.g., the first plurality of sensors 102, e.g., the second plurality of sensors 402. The modification of the existing classifier 114 in the form of a network 300, in accordance with a preferred embodiment of the present invention, is shown in the lower panel 440 of FIG.
[0108] As seen in the lower panel 440 of Figure 3, the network 300 may be modified to generate a modified network 442. The network 300 is preferably modified by adding at least one additional layer 450 to the network 300. The at least one additional layer 450 may be referred to as a "mapping layer" and is configured to learn frequency response differences between the original sensor type on which the classifier 114 in the form of the network 300 was trained, i.e., the second plurality of sensors 402, and the new sensor type on which new data was collected, i.e., the first plurality of sensors 102.
[0109] The configuration of the mapping layer 450 may be accomplished by freezing the structure and parameters of all layers of the network 300 and retraining the classifier with a new data set obtained from the first plurality of sensors 102, as provided by the elements enclosed in the dashed box 115 described above with respect to Figures 1 and 2. In this way, the network 442 is forced to optimize the weight values of the mapping layer 450 with respect to the frequency response difference of the two sensors, since the parameters of all other layers of the network 442 are already configured and frozen and cannot be changed. As a result, the mapping layer 450 is forced to learn a mapping between the original set of sensor signals provided by the sensor 402 (Figures 4A and 4B) and the new set of sensor signals provided by the sensor 102 (Figures 1 and 2).
[0110] Alternatively, the structure and parameters of all layers of the network 300 do not necessarily need to be frozen, but rather may be adjusted during retraining of the classifier with a new data set acquired from the first plurality of sensors 102. In this case, the parameters of the network 300 serve as a starting point for the adjusted parameters of the modified classifier 442.
[0111] The configuration of the mapping layer 450 is preferably selected with respect to the nature of the frequency response differences between different types of sensors, for example the first and second pluralities of sensors 102 and 402 .
[0112] In the case of mapping between different sensors of the same type, for example where the first and second pluralities of sensors 102 and 402 are both vibration sensors but of different types having different frequency responses, the difference in frequency responses is generally a linear function of the signal frequency.
[0113] In the case of mapping between different sensors of different types, for example the first and second pluralities of sensors 102 and 402 are vibration and magnetic sensors, respectively, with different frequency responses, the difference in frequency responses is generally a nonlinear function and thus a nonlinear mapping is necessary. Correspondingly, the activation function may, but does not necessarily, take on a nonlinear form.
[0114] In some cases, the mapping of the mapping layer 450 may be aided by providing the mapping layer 450 with a more complex form of new sensor data, such as a more complex form of the signals acquired by the first plurality of sensors 102. For example, inverse data, logarithmic data, or other forms of data may be provided. The process of modifying the input signals may be optimized with respect to the accuracy of the classifier, whereby the new sensor data is modified per classifier accuracy performance feedback, and the mapping layer 450 is then retrained with the modified data.
[0115] The mapping layer 450 is preferably incorporated into the network 442 as the first layer after the data layer 308 and is configured to receive input data in the form of 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 data. The position of the mapping layer 450 in the network 442 is important because it is its position as an initial layer of the network where the mapping layer 450 can learn mappings between sensor types and adapt incoming new data into an appropriate format for continuing on to other hidden layers 312 downstream in the network 442.
[0116] In an alternative embodiment of the present invention, an existing fault classifier may be initially trained in an unsupervised manner by providing a large amount of sensor data for the classifier to learn how to identify anomalies in sensor data related to the operating condition of the machine. For example, an autoencoder NN may be used to learn a low-dimensional representation of the sensor data, and then clustering-based classification may be used to identify outliers (anomalies). Such an existing unsupervised fault classifier 114 may then be modified in an unsupervised manner in accordance with a preferred embodiment of the present invention to generate a modified fault classifier 116 adapted to identify anomalies in sensor data acquired by a different type of sensor than the one on which the existing fault classifier 114 was initially trained. The modified fault classifier 116 may be generated, for example, by adding an initial mapping layer to the autoencoder NN, freezing the parameters of all layers of the network other than the mapping layer, and training the mapping layer to learn the mapping between the original sensor type and the new sensor type. During training of the mapping layer, another mapping layer, such as an inverse mapping layer, may additionally be added to the NN output after the decoder for the training process. The output may then be classified using clustering-based classification to identify outliers (anomalies).
[0117] The improvement in classifier performance as a result of the mapping learning of the present invention is shown in the example graph below. An example of the performance in the form of a precision-recall curve of the modified classifier of the present invention is shown in FIG. 5A. In this example, the fault classifier was first trained and validated on a dataset containing over 100,000 labeled signal sets, of which 7000 signal sets were indicative of machine faults acquired by a single-axis vibration sensor from a rotating machine including a bearing. Following training, the fault classifier was tested on a dataset containing 20,000 signal sets, of which 1,400 corresponded to machine faults. The fault classifier trained in this way may be referred to as the existing fault classifier.
[0118] The existing classifier was then modified in accordance with a preferred embodiment of the present invention to enable it to classify faults based on signals acquired by a three-axis vibration sensor from a rotating machine including a bearing. The new data set of signals acquired by the three-axis vibration sensor and used to modify the existing fault classifier consisted of a training and validation data set of 256 signal sets, of which only 20 sets of signals were associated with a faulty machine condition, i.e., bearing wear. Based on this very small signal set, the existing fault classifier was modified as described above.
[0119] The modified fault classifier was then applied to a test set of 7,149 signal sets, of which 486 were associated with faulty machine states, of signals acquired by a three-axis vibration sensor from a rotating machine containing bearings for fault detection, to test its performance (line 502). For comparison, a completely new classifier was trained "from scratch" on the same 256 signal sets and applied to the same test set of 7,149 signal sets (line 504). For completeness of comparison, the original existing classifier in unmodified form was also applied to the 7149 exemplary data set (line 506). As is evident from an examination of FIG. 5A, the performance of the modified classifier in fault identification is the best, despite the very small data set provided. The scores listed in FIG. 5A are average accuracy scores, although other exemplary scores may be used.
[0120] An example of the performance in the form of a precision recall curve of the modified classifier of the present invention is shown in FIG. 5B. In this example, the fault classifier was originally trained and validated on a dataset containing over 100,000 labeled signal sets, of which 7000 signal sets were indicative of machine faults acquired by a single-axis vibration sensor from a rotating machine including a bearing. Following training, the fault classifier was tested on a dataset containing 20,000 signal sets, of which 1,400 corresponded to machine faults. Fault classifiers trained in this way may be referred to as existing fault classifiers.
[0121] The existing classifier was then modified in accordance with a preferred embodiment of the present invention to enable it to classify faults based on signals acquired by a three-axis vibration sensor from a rotating machine including bearings. The new data set of signals acquired by the three-axis vibration sensor and used to modify the existing fault classifier consisted of 924 signal sets, of which only 70 sets of signals were associated with a faulty machine condition, i.e., bearing wear. Based on this very small signal set, the existing fault classifier was modified, as described above. To test its performance, the modified fault classifier was then applied to a set of 6841 signal sets, including 436 signal sets corresponding to machine faults acquired by a three-axis vibration sensor from a rotating machine including bearings for fault detection (line 510). For comparison, a completely new classifier was trained "from scratch" on the same 924 signal sets and was also applied to the 6841 exemplary data set (line 512). For completeness of comparison, the original existing classifier in its unmodified form was also applied to the 6841 exemplary data set (line 514). As is evident from an examination of Fig. 5B, the performance of the modified classifier in fault identification is the best, despite the small dataset fed in. The scores listed in Fig. 5B are average accuracy scores, although other exemplary scores may be used.
[0122] An example of the performance in the form of a precision-recall curve of the modified classifier of the present invention is shown in FIG. 5C. In this example, the fault classifier was originally trained and validated on a dataset containing over 100,000 labeled signal sets, of which 7000 signal sets were indicative of machine faults acquired by a single-axis vibration sensor from a rotating machine including a bearing. Following training, the fault classifier was tested on a dataset containing 20,000 signal sets, of which 1,400 corresponded to machine faults. Fault classifiers trained in this way may be referred to as existing fault classifiers.
[0123] The existing classifier was then modified so that it could classify faults based on signals acquired by a three-axis vibration sensor from a rotating machine including bearings. The new data set of signals acquired by the three-axis vibration sensor and used to modify the existing fault classifier consisted of 5924 signal sets, of which only 350 sets of signals were associated with a faulty machine condition, i.e., bearing wear. Based on this small signal set, the existing fault classifier was modified as described above. To test its performance, the modified fault classifier was then applied to a set of 1481 exemplary signal sets, including 156 signal sets corresponding to machine faults acquired by a three-axis vibration sensor from a rotating machine including bearings for fault detection (line 520). For comparison, a completely new classifier was trained on the same 5924 signal sets and also applied to the 1481 exemplary data set (line 522). For completeness of comparison, the original existing classifier in its unmodified form was also applied to the 1481 exemplary data set (line 524). As is evident from an examination of Fig. 5C, the performance of the modified classifier in fault identification is the best. The scores listed in Fig. 5C are average accuracy scores, although other exemplary scores may be used.
[0124] An example of the performance in the form of a precision-recall curve of the modified classifier of the present invention is shown in FIG. 5D. In this example, the fault classifier was first trained and validated on a dataset containing over 100,000 labeled signal sets, of which 7,000 signal sets were indicative of machine faults acquired by a single-axis vibration sensor from an electric motor including a bearing. Following training, the fault classifier was tested on a dataset containing 20,000 signal sets, of which 1,400 corresponded to machine faults. The fault classifier trained in this way may be referred to as the existing fault classifier.
[0125] The existing classifier was modified so that it could classify faults based on magnetic flux signals acquired by a magnetic sensor from an electric motor including bearings. The new data set of signals acquired by the magnetic sensor and used to modify the existing fault classifier consisted of 3411 signal sets, of which only 222 sets of signals were associated with a faulty machine condition, i.e., bearing wear. Based on this small signal set, the existing fault classifier was modified as described above. Next, to test its performance, the existing fault classifier was applied to a test signal set of 754 signals, including only the 50 exemplary signal sets corresponding to machine faults acquired by a magnetic sensor from an electric motor including bearings for fault detection (line 530). For comparison, a completely new classifier was trained on the same 3411 sets of magnetic signals and was also applied to the same 754 exemplary magnetic data set (line 532). For completeness of comparison, the original existing classifier in unmodified form was also applied to the same 754 exemplary magnetic data set (line 534). As is evident from an examination of FIG. 5D, the performance of the modified classifier in fault identification is the best. The scores listed in FIG. 5D are average accuracy scores, although other example scores may be used.
[0126] In the example of FIGS. 5A-5D above, the data set on which the classifier 114 was originally trained consisted of sensed data from over 40,000 individual rotating machines, including primarily motors, pumps, fans, gearboxes, chillers, and compressors. Bearing records, some of which were measured multiple times, were recorded over a three-year period. Each bearing was measured along three axes by a single-axis piezoelectric vibration sensor. Where the new data set consisted of data from a three-axis vibration sensor, this was a MEMS three-axis vibration sensor.
[0127] As is evident from the data above, the failure rate of the dataset was approximately 7%. The classifier was designed to detect failures of a single bearing, and even when additional non-faulty machine bearings emitted signals indicative of bearing wear, the classifier was able to correctly classify the non-faulty bearing due to its proximity to the faulty bearing and due to acoustic wave propagation between the bearings. Data labeling was performed by more than 10 human experts.
[0128] Reference is now made to FIG. 6, which is a simplified flowchart illustrating the steps involved in a method for mechanical machine fault identification based on transfer learning, in accordance with a preferred embodiment of the present invention.
[0129] 6 shows a method 600 for identifying machine faults. As seen in a first step 602, a first set of signals emanating from a plurality of mechanical machines may be acquired by a first plurality of sensors. The sensors of the first plurality of sensors are preferably of a first, mutually same type. Alternatively, the sensors of the first plurality of sensors may be of multiple types. The sensors of the first plurality of sensors may generally have the same sensor frequency response, which may be referred to as a first sensor frequency response distribution.
[0130] As seen in a second step 604, a first set of machine operating condition data may optionally be obtained in response to the first set of signals obtained in step 602. The machine operating condition data may include an identification of an operating condition associated with each set of signals in the first set of signals. The operating conditions may be fault conditions or non-fault conditions.
[0131] As seen in a third step 606, the signal set and optional corresponding operating condition data are provided to an existing fault classifier previously trained to identify faults in the machine based on signals emanating from the machine and acquired by a second plurality of sensors. The second plurality of sensors on which the existing fault classifier was trained may be of the same type as each other, but different in type from the first plurality of sensors. The second plurality of sensors on which the existing fault classifier was trained may alternatively be of multiple types, at least some of which are different from the first plurality of sensors. The sensors of the second plurality of sensors may generally have 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.
[0132] As seen in a fourth step 608, the existing fault classifier is modified based on the new data provided, preferably using a transfer learning approach. The modification may include the addition of at least one mapping layer to the existing fault classifier, which may learn frequency response differences between the first sensor frequency response distribution and the second sensor frequency response distribution.
[0133] As seen in a fifth step 610, an additional set of signals may then be obtained from the at least one machine with the same sensor type as the first plurality of sensors.
[0134] As seen in a sixth step 612, the modified classifier generated in step 608 may be applied to additional signal sets to identify faults in the machine from which the additional signal sets were generated.
[0135] As seen in a seventh step 614, based on the identified faults, maintenance or repairs to the machine may be performed.
[0136] It is understood that all of the various machines described as generating signals in steps 602, 606, and 610 may be the same machine or different machines with shared characteristics, as described above.
[0137] It will be appreciated by those skilled in the art that the present invention is not limited to what has been particularly shown and described above, and the scope of the present invention includes both combinations and sub-combinations of the various features described above, as well as modifications thereof, all of which are not present in the prior art.
Claims
1. 1. A method for identifying a fault in at least one mechanical machine, comprising: causing a first plurality of sensors coupled to a corresponding first plurality of mechanical machines to acquire a first plurality of sets of signals emanating from the first plurality of mechanical machines, the first plurality of mechanical machines sharing at least one characteristic, the shared characteristic indicating that an observed behavior of one mechanical machine having the shared characteristic may be characteristic of another mechanical machine having the shared characteristic; feeding at least the first set of signals of the first plurality of mechanical machines to an existing fault classifier previously trained to automatically identify faults of the second plurality of mechanical machines based on signals previously acquired by a second plurality of sensors emanating from a second plurality of mechanical machines, the second plurality of sensors being of a different type than the first plurality of sensors, the second plurality of mechanical machines sharing the at least one characteristic; modifying the existing fault classifier by utilizing transfer learning based on at least the first plurality of sets of signals of the first plurality of mechanical machines, thereby providing a modified fault classifier; applying the modified fault classifier to at least one additional set of signals acquired by at least one sensor of the first plurality of sensors and emanating from at least one given mechanical machine sharing the at least one characteristic, the modified fault classifier being configured to automatically identify at least one fault of the at least one given mechanical machine based on the at least one additional set of signals; providing, by an output device, a human-perceptible output including at least an identification of said fault of said at least one given mechanical machine, wherein at least one of a repair or maintenance action is performed based on said human-perceptible output; A method comprising:
2. The method comprises: after causing the first plurality of sensors to acquire the first plurality of signal sets and prior to providing the first plurality of signal sets to the existing fault classifier; obtaining a first plurality of sets of operating condition data for mechanical machines of the first plurality of mechanical machines, each set of operating condition data indicative of a state of operation of a mechanical machine of the first plurality of mechanical machines, each state of operation being associated with at least one of the sets of signals; Also includes providing at least the first plurality of sets of signals to the existing fault classifier also includes providing the operational condition data of the first plurality of mechanical machines to the existing fault classifier; 2. The method of claim 1 , wherein the modifying the existing fault classifier by utilizing transfer learning based on at least the first plurality of sets of signals also comprises modifying the existing fault classifier by utilizing transfer learning further based on the first plurality of sets of operating condition data of the first plurality of mechanical machines.
3. 3. The method of claim 1 or claim 2, wherein the identification of the fault comprises identifying a specific fault of the at least one given mechanical machine and predicting a failure of the at least one given mechanical machine due to the specific fault in the absence of performance of a proposed maintenance action on the at least one given mechanical machine, the at least one given mechanical machine actually failing in the absence of performance of the proposed maintenance action.
4. 4. The method of claim 2 or claim 3, wherein the existing fault classifier comprises a neural network, and wherein the modifying the existing fault classifier comprises adding at least one mapping layer to the neural network, and wherein the neural network is not otherwise modified by the modification other than the addition of the at least one mapping layer.
5. 5. The method of claim 4, wherein the neural network including the existing fault classifier includes a data layer and an input layer for receiving data from the data layer, and the at least one mapping layer is added between the data layer and the input layer, whereby the at least one mapping layer is configured to receive the data from the data layer in the modified fault classifier.
6. 6. The method of claim 4 or claim 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, and the mapping layer is configured to map between the first and second frequency response distributions.
7. A method according to any preceding claim, wherein the first plurality of sensors are operable to sense the same type of signal as is sensed by the second plurality of sensors.
8. The method of claim 7 , wherein the same type of signals include one of a vibration signal, a magnetic flux signal, a current, a temperature, and a machine internal pressure signal.
9. The method of any preceding claim, wherein the first and second plurality of sensors are operable to sense different types of signals from each other.
10. The different types of signals are: Vibration and magnetic flux signals, Vibration and current signals, vibration and temperature signals, current and magnetic flux signals, and Vibration and machine internal pressure signals, The method of claim 9 , comprising at least one of:
11. The method of claim 2, wherein at least some of the operating conditions of one mechanical machine of the first plurality of mechanical machines are faulty operating conditions, as indicated by the first plurality of sets of operating condition data.
12. The method of claim 11 , wherein the first plurality of sets of signals and the first plurality of sets of operational condition data for the first plurality of mechanical machines include fewer than 30 of the faulty operational conditions.
13. 1. A system for identifying faults in at least one mechanical machine, comprising: a first plurality of sensors coupled to a corresponding first plurality of mechanical machines and operative to acquire a first plurality of signal sets emanating from the first plurality of mechanical machines, the first plurality of mechanical machines sharing at least one characteristic, the shared characteristic indicating that an observed behavior of one mechanical machine having the shared characteristic may be characteristic of another mechanical machine having the shared characteristic; A data processing unit, receiving the first plurality of signal sets of the first plurality of mechanical machines, the data processing unit comprising a pre-existing fault classifier previously trained to automatically classify the operating states of the second plurality of mechanical machines based on signals emanating from a second plurality of mechanical machines and previously acquired by a second plurality of sensors, the second plurality of sensors being of a different type than the first plurality of sensors, and the second plurality of mechanical machines sharing the at least one characteristic; modifying the existing fault classifier by utilizing transfer learning based on at least the first plurality of sets of signals of the first plurality of mechanical machines, thereby providing a modified fault classifier; applying the modified fault classifier to at least one additional set of signals acquired by at least one sensor of the first plurality of sensors and emanating from at least one given mechanical machine sharing the at least one characteristic, the modified fault classifier being configured to automatically identify at least one fault of the at least one given mechanical machine based on the at least one additional set of signals; a data processing unit operative to: an output device in communication with said data processing unit and operable to provide a human-perceivable output including at least an identification of said fault of said at least one given mechanical machine, wherein at least one of repair or maintenance actions is performed based on said human-perceivable output; A system comprising:
14. The system also includes a data collection unit operative to obtain a first plurality of sets of operating condition data for mechanical machines of the first plurality of mechanical machines, each set of operating condition data indicative of a state of operation of a mechanical machine of the first plurality of mechanical machines, each state of operation being associated with at least one of the sets of signals; 14. The system of claim 13, wherein the data processing unit is operative to receive the operational condition data of the first plurality of mechanical machines and to modify the existing fault classifier further based on the first plurality of sets of operational condition data of the first plurality of mechanical machines.
15. 15. The system of claim 13 or claim 14, wherein the existing fault classifier includes a neural network, and the modification of the existing fault classifier includes adding at least one mapping layer to the neural network, and the neural network is not otherwise modified by the modification other than the addition of the at least one mapping layer.
Citation Information
Patent Citations
Remote maintenance system
JP2002342185A
Device abnormality monitoring method and system
JP2010287011A
System and method for determining state of vehicle components
JP2016118546A
Operation data classification device
JP2017021502A
Landmark-based classification model updating
US11620518B2
