Method of operating non-stationary running machine and system thereof

Through multi-sensor and wavelet transformation technology, non-stationary signal characteristics are extracted, combined with machine learning analysis, the problem of difficulty in analyzing non-stationary machines in traditional technologies is solved, and accurate monitoring and preventive maintenance of machine health status and efficiency are achieved.

CN118938850APending Publication Date: 2024-11-12AUGURY SYSTEMS LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410971625.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2018-11-09
Filing Date
2019-11-07
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Traditional machine monitoring and diagnostic technologies are difficult to effectively analyze non-stationary signals of non-stationary machines, resulting in the inability to accurately extract the health status and fault information of the machine.

Method used

Multiple sensors are used to obtain non-stationary signals within the operating time frame, and signal characteristics are extracted through wavelet transformation and mask multiplication, combined with signal processing and machine learning technology, the characteristics are analyzed to determine the machine condition, and repair, maintenance and parameter adjustment are performed.

Benefits of technology

It realizes accurate and automatic analysis of non-stationary machine performance, provides real-time monitoring of machine health status and efficiency, and supports preventive maintenance and optimization of operating parameters.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118938850A_ABST
    Figure CN118938850A_ABST
Patent Text Reader

Abstract

A method of operating a non-stationary operating machine and a system thereof, including acquiring a first non-stationary signal generated by a machine operating in a non-stationary manner during an operating time frame; acquiring a second non-stationary signal from the machine during the operating time frame, the second non-stationary signal being different from the first non-stationary signal; applying a wavelet transform to the first non-stationary signal and a different second non-stationary signal; generating a mask of a wavelet transform of one of the first non-stationary signal and the second non-stationary signal; multiplying the mask of the wavelet transform of one of the first non-stationary signal and the second non-stationary signal by the wavelet transform of the other one of the first non-stationary signal and the second non-stationary signal; extracting a feature of at least one of the first non-stationary signal and the second non-stationary signal based on the multiplication; analyzing the features to determine a condition of the machine; in accordance with the condition, at least one of a service operation, a maintenance operation, and an operational parameter modification of the machine is performed.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] This application is a divisional application of application number 201980088345.8 (PCT application number is PCT / IL2019 / 051217), application date November 7, 2019, and invention name “Automatic Analysis of Non-Stationary Machine Performance”.

[0002] Related Applications

[0003] This application hereby claims priority to U.S. Provisional Application No. 62 / 758,054, filed on November 9, 2018, entitled SYSTEM AND METHOD FOR AUTOMATIC DIAGNOSIS OF NON-STATIONARY MACHINES, the disclosure of which is incorporated herein by reference. Technical Field

[0004] The present invention relates generally to analysis of machine performance, and more particularly to automatic analysis of the performance of a non-stationary machine. Background Art

[0005] Various types of systems for automated machine analysis are known in the art. Summary of the invention

[0006] The present invention seeks to provide novel systems and methods for highly accurate automated analysis of the performance of machines having non-stationary characteristics, thereby enabling preventive maintenance and optimization of operating parameters of such machines.

[0007] Therefore, according to a preferred embodiment of the present invention, a method for operating a non-stationary operating machine is provided, the method comprising: acquiring at least one first non-stationary signal generated by at least one machine operating in a non-stationary manner during at least one operating time frame; acquiring at least one second non-stationary signal from the at least one machine during the at least one operating time frame, the second non-stationary signal being different from the first non-stationary signal; applying a wavelet transform to both the at least first non-stationary signal and the different second non-stationary signal; generating a mask of the wavelet transform of one of the first non-stationary signal and the second non-stationary signal; multiplying the mask of the wavelet transform of one of the first non-stationary signal and the second non-stationary signal with the wavelet transform of the other non-stationary signal of the first non-stationary signal and the second non-stationary signal; extracting at least one feature of at least one of the first and second non-stationary signals based on the multiplication; analyzing the at least one feature to determine a condition of the at least one machine; and performing at least one of a repair operation, a maintenance operation, and an operating parameter modification of the at least one machine based on the condition.

[0008] According to a preferred embodiment of the present invention, the first non-stationary signal and the second non-stationary signal are of the same type. Steady operation.

[0009] According to a preferred embodiment of the present invention, the first non-stationary signal and the second non-stationary signal are of different types.

[0010] Preferably, the mask comprises a binary wavelet matrix.

[0011] According to a preferred embodiment of the present invention, the step of multiplying the mask of the wavelet transform of one of the first non-stationary signal and the second non-stationary signal with the wavelet transform of the other of the first non-stationary signal and the second non-stationary signal also includes multiplying the wavelet transform of one of the first non-stationary signal and the second non-stationary signal with the inverse mask of the wavelet transform of the other of the first non-stationary signal and the second non-stationary signal.

[0012] According to a preferred embodiment of the present invention, said extracting comprises summing said multiplied wavelet transforms.

[0013] According to a preferred embodiment of the present invention, the extracting comprises extracting the at least one feature directly from the multiplied wavelet transform.

[0014] According to a preferred embodiment of the present invention, the execution is automatically performed by a controller of the at least one machine.

[0015] According to a preferred embodiment of the present invention, the method further comprises tracking the at least one feature over time by means of a feature tracking algorithm.

[0016] According to a preferred embodiment of the present invention, the step of determining the condition further comprises determining the condition of the at least one machine related to at least one of machine performance and machine efficiency.

[0017] According to another preferred embodiment of the present invention, a system for operating a non-stationary machine is further provided, the system comprising: at least one first sensor for acquiring at least one first non-stationary signal generated by at least one machine operating in a non-stationary manner during at least one operating time frame; at least one second sensor for acquiring at least one second non-stationary signal from the at least one machine during the at least one operating time frame, the second non-stationary signal being different from the first non-stationary signal; a signal processor for: applying a wavelet transform to both the at least first non-stationary signal and the different second non-stationary signal; generating a mask of the wavelet transform of one of the first non-stationary signal and the second non-stationary signal; multiplying the mask of the wavelet transform of one of the first non-stationary signal and the second non-stationary signal with the wavelet transform of the other of the first non-stationary signal and the second non-stationary signal; extracting at least one feature of at least one of the first non-stationary signal and the second non-stationary signal based on the multiplied wavelet transform; and analyzing the at least one feature to determine a condition of the at least one machine; and a machine control module for performing at least one of a repair operation, a maintenance operation, and a modification of an operating parameter of the at least one machine based on the condition.

[0018] According to a preferred embodiment of the present invention, the first sensor and the second sensor are of the same type.

[0019] According to a preferred embodiment of the present invention, the first sensor and the second sensor are of different types.

[0020] According to a preferred embodiment of the present invention, the mask comprises a binary wavelet matrix.

[0021] According to a preferred embodiment of the present invention, the signal processor is used to multiply the wavelet transform of one of the first non-stationary signal and the second non-stationary signal with the mask of the wavelet transform of the other non-stationary signal of the first non-stationary signal and the second non-stationary signal, and is also used to multiply the wavelet transform of one of the first non-stationary signal and the second non-stationary signal with the inverse mask of the wavelet transform of the other non-stationary signal of the first non-stationary signal and the second non-stationary signal.

[0022] According to a preferred embodiment of the present invention, the signal processor is used to extract at least one feature, including summing the multiplied wavelet transforms.

[0023] According to a preferred embodiment of the present invention, the signal processor is configured to extract the at least one feature directly from the multiplied wavelet transform.

[0024] According to a preferred embodiment of the present invention, the machine control module automatically operates based on the condition to perform at least one of the service operation, maintenance operation and modification of operating parameters of the at least one machine.

[0025] According to a preferred embodiment of the present invention, the system further comprises a feature tracking algorithm for automatically tracking the at least one feature extracted by the signal processor over time.

[0026] According to a preferred embodiment of the present invention, the signal processor is operated to determine a condition of the machine, including the signal processor being operated to determine a condition of the at least one machine related to at least one of machine performance and machine efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The present invention will be more fully understood and appreciated through the following detailed description in conjunction with the accompanying drawings, in which:

[0028] Figure 1 is a simplified partially schematic, partially block diagram illustration of a system for automatic analysis of non-stationary machine performance, constructed and operative in accordance with a preferred embodiment of the present invention;

[0029] Figure 2A , Figure 2B and Figure 3 They are shown in Figure 1 A schematic diagram of the characteristics of a signal emitted by a non-stationary machine operating in a system of the type shown;

[0030] Figure 4 is a simplified partially schematic, partially block diagram illustration of a system for automatic analysis of non-stationary machine performance, constructed and operative in accordance with another preferred embodiment of the present invention;

[0031] Figure 5A , Figure 5B and Figure 5C The components are shown respectively Figure 4 A simplified schematic diagram of steps in the training and operation of a machine learning network that is part of a system of the type shown;

[0032] Figure 6 is a simplified flowchart showing the extraction of Figure 1 or Figure 4 a schematic diagram of the steps involved in characterizing a signal generated by a non-stationary machine operating in a system of the type shown;

[0033] Figure 7 is a simplified schematic diagram of a system for automatic analysis of non-stationary machine performance, constructed and operative in accordance with yet another preferred embodiment of the present invention;

[0034] Figure 8 , Fig. 9, Fig.10 and Fig.11 They are shown in Figure 7 Schematic diagram of the characteristics of the signal generated by a non-stationary machine operating in a system of the type shown

[0035] Fig. 12A , Fig. 12B and Fig. 12C The formation Figure 7 a simplified schematic diagram of steps in the training and operation of a machine learning network that is part of a system of the type shown; and

[0036] Fig.13 This is a simplified flow chart that illustrates the process of analyzing Figure 7 The steps involved in characterizing the signals produced by a non-stationary machine operating in a system of the type shown. DETAILED DESCRIPTION

[0037] In reference Figure 1 , Figure 1 is a simplified partially schematic, partially block diagram illustration of a system for automatic analysis of non-stationary machine performance, constructed and operative in accordance with a preferred embodiment of the present invention. Figure 1 As shown, a system 100 for monitoring and automatically analyzing non-stationary machine performance is provided, which is embodied as a motor 102 that operates in a non-stationary manner as an example only. The motor 102 is preferably a non-stationary machine that generates a non-stationary signal during its operation. As used herein, the term non-stationary machine refers to a machine that performs a process having at least one characteristic that varies with time, thereby generating a non-stationary signal during its operation. Such non-stationary machines can include machines driven by time-varying inputs, such as servo motors controlled by a drive that varies the frequency and amplitude of the input power over time; motors that drive mobile equipment, such as elevators and cranes, that provide different driving forces over time; machines that experience time-varying loads at their outputs, such as CNC machining machines that experience load changes and friction changes during their operation, and machines mounted on a mobile base, such as motors mounted on a manufacturing robot.

[0038] It is noteworthy that conventional machine monitoring and diagnostic techniques are not well suited for monitoring and analysis of such non-stationary machines due to the non-stationary signal characteristics they exhibit. For example, conventional machine diagnostics based on spectral analysis of vibration signals generated during machine operation are not well suited for analyzing non-stationary signals because the temporal variations in the spectral content of non-stationary signals are caused by changes in operating parameters, such as speed and load. For example, in the case of motor 102, motor 102 may exhibit non-stationary time-varying speed and its harmonics. The vibration amplitude of the speed frequency and its harmonics, which are typically associated with machine condition and machine failure, will vary in time and therefore cannot be accurately extracted based on conventional spectral analysis of the sampled signal. By further example, motor 102 may exhibit non-stationary time-varying bearing frequencies or line frequencies. The vibration amplitude of these frequencies, which are typically associated with machine condition and machine failure, will vary in time and therefore cannot be accurately extracted based on conventional spectral analysis of the sampled signal.

[0039] According to a preferred embodiment of the present invention, novel systems and methods are provided for monitoring and automatic state analysis of non-stationary machines, based on monitoring of the non-stationary machine by multiple sensors during at least one operating time frame, and fusing the individual non-stationary outputs of the multiple sensors into a fused or combined output. Particularly preferably, the fusion includes modifying the non-stationary signal measured and output by at least one of the sensors by a non-stationary signal measured and output by at least another of the sensors. Such modification can be used to extract signal features that can be accurately analyzed by traditional signal analysis methods suitable for static signal analysis without loss of accuracy. Analysis of the signal features can produce the current health status of the machine, which can include information related to machine performance, machine efficiency, possible faults in progress, and predictions of future performance of the machine. Additionally or alternatively, such methods can be applied to other types of analysis.

[0040] like Figure 1 As shown, the non-stationary machine 102 is preferably monitored by a plurality of sensors, which are exemplified herein by a first vibration sensor 110 that measures vibrations generated by the machine 102 and a second magnetic sensor 112 that measures magnetic flux generated by the machine 102. It should be understood that due to the inherently non-stationary operation of the machine 102, the vibration and magnetic flux signals generated by the machine 102 and measurable by the first sensor 110 and the second sensor 112 are preferably non-stationary signals, regardless of the health of the machine 102.

[0041] It should be understood that vibration sensor 110 and magnetic sensor 112 are shown as examples only, and that additional or alternative sensor types may be used to monitor signals emitted from machine 102, including but not limited to current sensors, electromagnetic sensors, acoustic sensors, temperature sensors, pressure sensors, and fluid flow sensors. It is further understood that the plurality of sensors used to monitor machine 102 may include a plurality of sensors measuring the same type of signal, such as a plurality of vibration sensors or a plurality of magnetic sensors, where the sensors are different from one another and / or mounted at different locations relative to machine 102, and therefore, despite measuring the same type of signal emitted from machine 102, produce slightly different outputs.

[0042] It should also be understood that the multiple sensors monitoring the machine 102 can be contact or non-contact sensors. In certain embodiments of the present invention, the use of non-contact sensors may be preferred due to the difficulty in securely attaching sensors to moving non-stationary equipment. Multiple sensors, such as sensors 110 and 112, preferably monitor the machine 102 during the same at least one operating time frame of the machine 102, so that the non-stationary outputs of each of the multiple sensors represent the same cycle or multiple cycles of machine operation. In a preferred embodiment of the present invention, multiple sensors, such as sensors 110 and 112 monitoring the machine 102, are sampled synchronously with each other. The sampling can be synchronized within a time window dt of the order of dt≈0.01 / f, where f is the frequency of interest, such as the RPM frequency. In another preferred embodiment of the present invention, multiple sensors, such as sensors 110 and 112 monitoring the machine 102, are sampled synchronously with each other within a time window dt of the order of dt≈0.01 / f, where f is the sensor sampling frequency. For example, sensors 110 and 112 may operate at a sampling frequency of 20 kHz and thus the time frame dt over which the sensors are synchronously sampled is given by 0.5 microseconds. It is worth noting that a value of 0.01 is chosen here to limit the error in synchronous sampling to 1%. However, other values ​​may be chosen depending on the desired error limit.

[0043] Preferably, the first non-stationary output provided by the vibration sensor 110 and the second non-stationary output provided by the magnetic sensor 112 are provided to gain filters (GFs) 114 and 116, respectively, which in turn provide corresponding signals to analog-to-digital converters (ADCs) 118 and 120. The digitized signals output by the ADCs 118 and 120 are preferably provided to a memory component 122 such as a CPU and optionally to a cloud database 124. It should be understood that, depending on the operating specifications of the system 100, the synchronously sampled data acquired by the sensors 110 and 112 can be transmitted to a completely cloud-based system, a local hardware system, or a system including local components and cloud-based components. It should also be understood that the specific processing components shown herein, including the GFs 114 and 116, the ADCs 118 and 120, and the CPU 122, are detailed only by way of example, and any suitable signal processing components can be employed in the system of the present invention.

[0044] The first and second non-stationary outputs, preferably synchronously sampled, respectively output by the vibration sensor 110 and the magnetic sensor 112 are preferably provided to a cross-sensor signal processor 130, which may be included in the cloud database 124 or may include local processing components.

[0045] Preferably, the signal processor 130 is operable to receive a first non-stationary output and a second non-stationary output provided by at least a first sensor and a second sensor, embodied herein as a vibration sensor 110 and a vibration sensor 112, and to combine the outputs by modifying the first non-stationary output, such as the output from the vibration sensor 110, based at least on the second non-stationary output, such as the output from the magnetic sensor 112, so as to extract at least one feature of the first non-stationary output, in this case the vibration output.

[0046] The signal processor 130 preferably operates in an automatic manner. In a preferred embodiment of the present invention, Figure 1 As shown, the signal processor 130 can automatically perform wavelet analysis on the non-stationary magnetic and vibration data to extract its features. In another preferred embodiment of the present invention, referring to Figure 4 5 and described below, the signal processor may employ deep learning to extract features of a non-stationary signal input thereto.

[0047] The signal processor 130 may include a magnetic wavelet calculator 132 and a vibration wavelet calculator 134. The magnetic wavelet calculator 132 preferably applies a wavelet transform to the non-stationary magnetic data from the magnetic sensor 112 in order to extract features of the magnetic data. The vibration wavelet calculator 134 preferably applies a wavelet transform to the non-stationary vibration data from the vibration sensor 110 in order to extract features of the vibration data. The wavelet transforms of the magnetic and vibration data produce a temporal correlation of the corresponding magnetic and vibration spectral content, thereby avoiding limitations in spectral analysis due to the non-stationary nature of the signals generated by the machine 102. It should be understood that the magnetic wavelet calculations and vibration wavelet calculations performed by the magnetic wavelet calculator 132 and the vibration wavelet calculator 134 are particularly preferred examples of operations for extracting features from the non-stationary data acquired by the sensors 110 and 112. However, it should be understood that alternative data operations may be implemented to extract features of the non-stationary data acquired by the sensors 110 and 112.

[0048] Figure 2A A simplified schematic diagram of a noise-filtered magnetic wavelet image is shown in FIG. , such as may be output by the magnetic wavelet calculator 132. It will be appreciated that Figure 2A In the figure, the magnetic wavelet scale is calibrated and the reference pseudo-frequency is plotted. For clarity, Figure 2A The signal amplitudes in are shown relative to a normalized arbitrary unit scale.

[0049] The magnetic wavelet output by the magnetic wavelet calculator 132 is preferably provided to a mask extractor 138. The mask extractor 138 preferably employs an image processing algorithm to extract a magnetic wavelet mask based on the full magnetic wavelet image, but only contains the time correlation of specific frequencies, and all other frequencies are filtered out. For example, the magnetic wavelet mask extracted by the mask extractor 138 may contain only the time correlation of the RPM frequency, filtering out all other frequencies. The magnetic wavelet mask extracted by the mask extractor 138 is preferably a binary wavelet matrix, in which the RPM scale can be converted to pseudo-frequencies. An exemplary preferred algorithm that can be used to create such a wavelet mask is described in WO / 2018 / 198111, which has been assigned to the same assignee as the present invention and is incorporated herein by reference. Figure 2B A simplified example of a binary magnetic wavelet mask is shown, which contains only the Figure 2A The RPM frequency extracted by the full magnetic wavelet, for example, can be output by the mask extractor 138.

[0050] By way of further example, the magnetic wavelet mask extracted by the mask extractor 138 contains primarily only the time correlation of the RPM frequency and the line frequency and its harmonics. These frequencies may correspond to various machine faults, such as mechanical looseness and imbalance in the case of the RPM frequency and its harmonics, and electrical faults in the case of the line frequency and its harmonics.

[0051] The binary magnetic wavelet mask provided by the mask extractor 138 and the full vibration wavelet provided by the vibration wavelet calculator 134 are preferably provided to a wavelet multiplier 150. The wavelet multiplier 150 is preferably used to multiply the binary magnetic wavelet mask by the full vibration wavelet image containing all frequencies and amplitudes. The full vibration wavelet is multiplied by the magnetic wavelet mask to filter out those frequencies that are not related to the mask frequency from the full vibration wavelet. For example, in the case where the magnetic wavelet mask only contains the RPM frequency, the binary wavelet mask is multiplied by the full vibration wavelet to produce a vibration wavelet image containing only the RPM frequency. Figure 3 A simplified illustration of a vibration wavelet image containing only RPM frequencies, such as may be generated by wavelet multiplier 150, is shown.

[0052] By way of further example, in the case where a magnetic wavelet mask is generated based on the RPM and line frequencies and their harmonics, the mask may be multiplied by the full vibration wavelet. In this case, the mask may be applied as an inverse mask to the full vibration wavelet so that only frequencies other than the frequencies obtained by multiplying the RPM frequency and the line frequency and their harmonics by the full vibration wavelet are retained. These frequencies may be referred to as asynchronous frequencies and are closely related to bearing frequencies or frequencies associated with other machine components.

[0053] A numerical summation can then be performed on the vibration wavelet image containing only the RPM frequencies to extract the RPM amplitude of the vibration signal according to the following equation:

[0054]

[0055] Where M is a binary magnetic wavelet matrix with the RPM frequency set to 1 and all other frequencies set to zero, as the output of the mask extractor 138, and V is the total wavelet transform of the vibration signal, as the output of the vibration wavelet calculator 134. The summation represented in equation (1) may be performed by a feature extractor 152. The product A of equation (1) represents the amplitude of the RPM frequency. As described above, the parameter is highly correlated with machine condition and faults. It will be appreciated that the feature extractor 152 may be used to extract any relevant feature from the filtered vibration wavelet image, including the vibration amplitude of the line frequency or its harmonics. The vibration amplitude of the RPM frequency output by the feature extractor 152 may be analyzed to determine the health status of the machine 102. The results of such analysis may be provided to the controller 154, based on which the controller 154 may automatically modify the operating parameters of the machine 102, initiate or schedule a repair operation on the machine 102, or initiate or schedule a maintenance operation on the machine 102. Additionally or alternatively, the results of such analysis may be presented in a human understandable form to at least one human expert, who may assess the machine state based on the results and initiate maintenance, repairs, or operating parameter changes accordingly. Additionally or alternatively, the extracted features may be tracked over time, for example, by a feature tracking algorithm.

[0056] It will be appreciated that the various operations of the signal processor 130 described herein represent particularly preferred embodiments of computational operations in which a first non-stationary signal is modified based on a second non-stationary signal in order to extract at least one feature of the first non-stationary output. Here, for example, a full vibration wavelet derived from the first non-stationary vibration signal is modified by an RPM binary mask derived from the second non-stationary magnetic signal to extract the RPM amplitude of the first vibration signal, which can be analyzed in order to find the machine state.

[0057] It should also be appreciated that modifying the non-stationary vibration signal by extracting features from the non-stationary magnetic signal to derive features of the vibration signal is fundamentally different from simply performing a comparative analysis of the data obtained from the vibration and magnetic sensors 110 and 112. In accordance with a preferred embodiment of the present invention, features of the data from one type of sensor (e.g., magnetic sensor 112) are used as calibration inputs for data from another type of sensor (e.g., vibration sensor 110). This cross-sensor calibration is based on the understanding that while the various types of signals, such as vibration signals and magnetic signals, are not themselves fixed, the relationship between the signals is such that features from one type of signal can be used to calibrate the other type of signal.

[0058] It should be understood that the present invention is not limited to the use of wavelet transforms applied to specific frequencies, such as RPM frequencies, but can be applied to other operations on any relevant part of the spectral content of non-stationary signals. Figure 1 In the described examples, a first signal (as an example a vibration signal) is modified based on features extracted from a second signal (as an example a magnetic signal), and vice versa is also possible, whereby the second signal (for example a magnetic signal) can be modified based on features extracted from the first signal (for example a vibration signal).

[0059] It should also be understood that the various automatic signal processing components of the signal processor 130, including the magnetic wavelet calculator 132, the vibration wavelet calculator 134, the mask extractor 138, the wavelet multiplier 150, and the feature extractor 152, are represented herein as individual components for ease of explanation of their respective functions. However, the various functions performed by the components of the signal processor 130 are not necessarily performed in the order described herein, nor are they necessarily divided into separate modules as shown herein.

[0060] Figure 4 An alternative embodiment of the signal processor 130 is shown in FIG. Figure 4As shown, the signal processor 130 may be replaced by a signal processor 430 including a vibro-magnetic signal fuser 450 operating based on machine learning. The vibro-magnetic signal fuser 450 is preferably operable to automatically fuse the non-stationary outputs of the vibration and magnetic sensors 110 and 112 using machine learning, for example, by modifying at least one first non-stationary output (e.g., a non-stationary vibration signal output by the vibration sensor 110) based on at least one second non-stationary output (e.g., a non-stationary magnetic signal output by the magnetic sensor 112) to extract at least one feature of the first non-stationary output that is insensitive to the stationary level of the non-stationary operation of the machine 102. The at least one feature extracted by the vibro-magnetic signal fuser 450 may then be analyzed by conventional signal analysis techniques that are suitable for analyzing features of stationary signals to generate a condition of the machine 102 and detect anomalies in its operation. It will be appreciated that, therefore, it is advantageously possible to analyze non-stationary signals emitted from machine 102 based on standard signal analysis techniques and accurately find the condition of machine 102 despite the fact that the operating mode of machine 102 is non-stationary, since at least one feature of the non-stationary output that is insensitive to the level of smoothness of the operation of machine 102 is extracted.

[0061] A particularly preferred embodiment of the machine learning technique employed in the vibratory-magnetic signal fusion device 450 is shown in FIG. 5A to FIG. 5C However, it should be understood that the machine learning techniques employed in the vibration-magnetic signal fusion device 450 are not limited to FIG. 5A to FIG. 5C and may include any suitable machine learning techniques that may be implemented to automatically fuse the various non-stationary signals emitted from the machine 102 to extract therefrom a plurality of signal features that are insensitive to the stationary level of operation of the non-stationary machine 102. The extracted signal features may then be automatically analyzed using conventional signal analysis techniques that are designed to classify stationary signal features but may be accurately applied to non-stationary signals emitted from the machine 102.

[0062] Now turn to Figure 5A , shows an initial training step 500 in the operation of the vibratory-magnetic signal fusion device 450. A stationary data set 502 is preferably provided as an input to the initial training step 500. The data set 502 preferably includes data from a plurality of sensors S1-S n, where the plurality of sensors are embodied as vibration sensors 110 and magnetic sensors 112 as examples. However, it should be understood that data set 502 may include data from additional or alternative types of sensors, such as current sensors, electromagnetic sensors, acoustic sensors, pressure sensors, fluid flow sensors, and temperature sensors. Data set 502 may include data emitted from machine 102 when operating in a stable manner and generating a stable signal. Additionally or alternatively, data set 502 may include data emitted from a group of machines that share common characteristics with machine 102, although not necessarily the same thereto and operate in a stable manner to generate a stable signal. The group of machines sharing common characteristics from which data set 502 may be obtained may or may not include machine 102.

[0063] A plurality of sensors, such as sensors 110 and 112, preferably monitor the machine 102 during the same at least one operating time frame of the machine 102, such that the respective non-stationary outputs of the sensors represent the same cycle or cycles of the machine operation. In a preferred embodiment of the invention, a plurality of sensors, such as sensors 110 and 112 monitoring the machine 102, are sampled synchronously with respect to each other. The sampling may be synchronized within a time window dt of the order of dt≈0.01 / f, where f is a frequency of interest, such as an RPM frequency. In another preferred embodiment of the invention, a plurality of sensors, such as sensors 110 and 112 monitoring the machine 102, are sampled synchronously with respect to each other within a time window dt of the order of dt≈0.01 / f, where f is the sensor sampling frequency. For example, the sensors 110 and 112 may be operated at a sampling frequency of 20 kHz and thus the time frame dt in which the sensors are synchronously sampled is given by 0.5 microseconds. It is worth noting that a value of 0.01 is chosen here to limit the error in the synchronous sampling to 1%. However, other values ​​may be chosen according to the desired error limit.

[0064] The data set 502 is preferably provided to a first sensor fusion operator 510. The first sensor fusion operator 510 preferably combines the data from the sensors 1-n into a combined single data set, embodied here as the vibration sensor 110 and the magnetic sensor 112. According to one embodiment of the present invention, the first sensor fusion operator 510 may combine the vibration and magnetic signals into a one-dimensional vector, the vector also including a logarithmic function of the data, so that the combined single data set output by the first sensor fusion operator 510 includes the raw sensor data and the logarithmic function thereof. However, it should be understood that the first sensor fusion operator 510 is not limited to including the logarithmic function of the input sensor data, but may additionally or alternatively perform other mathematical operations on the data, including, for example, finding the inverse of the data or finding the square of the data. In addition, the first sensor fusion operator 510 is operable to fuse data from multiple sensors according to other methods known in the art, such as adding the vibration data and the magnetic data and their logarithms and creating a two-dimensional vector, where the second dimension represents the number of sensors, or performing a wavelet transform on the vibration data and the magnetic field data.

[0065] The combined single data set output by the sensor fusion operator 510 is preferably input to a first neural network NN1, represented here by reference numeral 512. NN1 512 is preferably embodied as a multi-layer perceptron (MLP) network. NN1 is preferably operable to automatically extract features of stationary data input thereto, represented here as Xs. It should be understood that the features extracted by NN1 are preferably based on a fusion of data from sensors 1-n, represented here as vibration sensors and magnetic sensors 110 and 112. Particularly preferably, NN1 can be used to extract one or more features from one set of signals, modify a second set of signals based on the extracted features, and then extract one or more features from the second set of modified signals. It should be understood that the input of logarithmic data to NN1 helps NN1 learn the multiplication and division operations of data from various sensor types relative to each other, thereby allowing one signal to be modified by the features of another signal.

[0066] The features extracted by NN1 are preferably provided to a first classifier, classifier 1, indicated by reference numeral 514. Classifier 1 is preferably used to classify the features output by NN1 in order to find the condition of the machine that generated the analysis signal. The machine condition output by classifier 1 is preferably provided to a loss calculator 516, where a loss function is obtained. The loss function can be any suitable loss function. For example, the loss function can represent the difference between the classification of the signal features provided by classifier 1 and the classification of the signal features provided by human analysis, such as a cross entropy loss function.

[0067] The loss function is then preferably fed back to NN1 and classifier 1 so that the weights within NN1 are further updated in an iterative process using back propagation and gradient descent algorithms, thereby training NN1 and classifier 1 to more accurately extract relevant features and classify these features. The iterative process can be continued until the loss function found by the loss calculator 516 is acceptably low, which means that NN1 and classifier 1 are considered to be trained to extract and classify signal features of stationary signals with an acceptable level of accuracy compared to human experts.

[0068] For example, where the combined single data set output by the first sensor fusion operator 510 comprises a one-dimensional vector comprising a signal having a length L and its logarithm, NN1 may comprise a fully connected network comprising a plurality of hidden layers, the exact number of which depends on the signal length L and the desired number of features to be extracted. Assuming L is much greater than 40, the number of neurons in the input layer may correspond to the value of L, L / 4 in the second layer, L / 10 in the third layer, and L / 40 in the last layer. Non-linear activation functions may be included between layers, such as Relu or sigmoid functions, to allow the network to learn non-linear relationships between sensor data points.

[0069] By way of further example, where the combined single data set output by the first sensor fusion operator 510 comprises a two-dimensional vector of vibration and magnetic data and the logarithm of its total length L>40, then NN1 may be replaced by a convolutional neural network CNN, for example, comprising five kernel filters, wherein the first two kernels have a wide dimension of length L / 20 to capture low-frequency sensor fusion content, and the last three kernels have narrow dimensions of size L / 80, L / 200 and L / 400, respectively, to capture high-frequency sensor fusion content. In order to reduce the dimensionality of the data set, a pooling layer may be included in the CNN between the convolutional layers with a maximum pooling of 4 and 2. The activation function is preferably a Relu, Sigmoid or tanH function to allow the CNN to learn nonlinear relationships between sensor data. After the convolutional layer, a classifier 1 may be provided, preferably comprising a fully connected network of a size equal to the size of the output layer of the CNN, to divide the features into classes. These classes may represent the severity of a specific fault in the operation of the machine or the severity of the overall health state of the machine. For example, in the case of multi-class classification, the last layer may contain a softmax activation function, while in the case of binary classification, a sigmoid function may be used. Batch Normalization can be included between the layers of a CNN to improve the performance of CNN training.

[0070] Now turn to Figure 5B , showing an additional training step 520 in the operation of the vibratory-magnetic signal fusion device 450, preferably in Figure 5A After 500 training steps.

[0071] like Figure 5B As seen in , paired stationary data and non-stationary data originating from a given machine, such as machine 102, are preferably provided. The data pairs preferably include randomly paired stationary data for machine 102, including stationary data from a plurality of sensors 1-n, and non-stationary data from machine 102, including non-stationary data from a plurality of sensors 1-n. It will be appreciated that, although the data pairs are randomly paired, it is required that both the stationary and non-stationary paired data correspond to the same machine health state, which may be healthy or unhealthy. The pairs are preferably shuffled so as to be mixed while satisfying the requirement that the components of each pair correspond to the same machine health state.

[0072] It will be appreciated that in order to perform the first training step 500 on stationary data and in order to perform the second training step 520 on stationary data and non-stationary data, it is necessary to distinguish between stationary data and non-stationary data. According to a preferred embodiment of the present invention, stationary data and non-stationary data may be distinguished based on a stationarity test. Such a test may be performed by a machine learning classifier, trained to classify data as stationary or non-stationary based on data labeling. Alternatively, such a test may be performed by a signal processing based classifier for analyzing the operating state of a machine generating the data. For example, such a classifier may be used to perform a wavelet transform on a magnetic signal to determine whether the main frequency peak is constant or non-constant, thereby indicating the stationary characteristics or non-stationary characteristics of the machine operation, respectively. For further example, such a classifier may be used to analyze the time waveform of a magnetic signal using a zero crossing algorithm to determine whether the main frequency peak is constant or non-constant, thereby indicating the stationary characteristics or non-stationary characteristics of the machine operation, respectively.

[0073] Data classified as stationary data is preferably provided to the first sensor fusion operator 510, and data classified as non-stationary data is preferably provided to the second sensor fusion operator 524. The first and second sensor fusion operators 510 and 524 preferably combine the data input thereto from the sensors 1-n, which are embodied here as the vibration sensor 110 and the magnetic sensor 112, into a combined single data set. In addition, the sensor fusion operators 510 and 524 preferably calculate the logarithmic function of the data from the sensors 1-n, respectively, so that the combined single data set output by the sensor fusion operators 510 and 524 includes the raw sensor data and its logarithmic function, respectively. It should be understood that the sensor fusion operators 510 and 524 are not limited to calculating the logarithmic function of the input sensor data, but may additionally or alternatively perform other mathematical operations on the data, including, for example, finding the inverse of the data or finding the square of the data. In addition, the first and second sensor fusion operators 510 and 524 can be used to fuse data from multiple sensors according to other methods known in the art, such as adding the vibration data and magnetic data and their logarithms and creating a two-dimensional vector, where the second dimension represents the number of sensors, or performing a wavelet transform on the vibration data and magnetic data.

[0074] A single data set comprising a combination of logarithmic functions output by the first sensor fusion operator 510 based on stationary data is preferably provided to NN1 512. It should be understood that as in Figure 5A The result of training NN1 512 in the first training step 500 is shown. NN1 512 has been trained to extract relevant features from the stationary data set output by the first sensor fusion operator 510. The features of the stationary data extracted by NN1 are represented as X s .

[0075] The single data set including the combination of logarithmic functions output by the second sensor fusion operator 524 based on the non-stationary data is preferably provided to a second NN, here represented as NN2 530. NN2 530 preferably has substantially the same architecture as NN1, but has different weights within the layers of the NN. NN2 is preferably used to extract relevant features from the non-stationary data set output by the second sensor fusion operator 524. It should be understood that the features extracted by NN2 are preferably based on the fusion of non-stationary data from sensors 1-n, which are embodied here as vibration sensors and magnetic sensors 110 and 112. It can be understood that the input of logarithmic data to NN2 helps NN2 learn the multiplication and division operations of data from various sensor types. The features of the non-stationary data set extracted by NN2 are represented here as X ns .

[0076] To train NN2 against NN1, the stationary features X found by NN1s and the non-stationary features X found by NN2 ns Preferably, it is provided to a loss calculator 532, where a loss function is found that represents the difference between stationary features and non-stationary features. Any suitable loss function may be calculated by the loss calculator 532. Preferably, the calculated loss is based on (X s –X ns ) 2 The MSE loss is calculated. The loss function is then preferably fed back to NN2 to further update the weights within NN2 so as to train NN2 to automatically and accurately extract features from the non-stationary data set that differ little from those extracted from the stationary data set. The iterative process can continue until the loss function found by the loss calculator 532 is acceptably low, meaning that NN2 is considered to be trained to extract signal features of non-stationary signals with an acceptable level of accuracy compared to signal features of stationary signals acquired from the same machine or a similar group of machines.

[0077] Now go to Figure 5C , Figure 5C is an operation step 540 in the operation of the vibration-magnetic signal fusion device 450, preferably in Figure 5A and Figure 5B After training steps 500 and 520.

[0078] like Figure 5C As shown, during operation of the machine 102 in a non-stationary manner, a non-stationary signal 542 emanating from the machine 102 is measured by a plurality of sensors S1-Sn, here implemented by way of example as a vibration sensor 110 and a magnetic sensor 112. The non-stationary signal 542 is preferably provided to a second sensor fusion operator 524, which provides a combined sensor data set to a NN2 530. It will be appreciated that the NN2 530 has been trained according to the second training step 520 to extract features from the non-stationary data that are distinguishable from features of stationary data from the same or similar machines by a minimum loss function. As a result, due to the commonality between these non-stationary signal features and the stationary signal features, the features extracted by NN2 from the fused non-stationary sensor signal can be provided to a classifier 1514 in order to classify the condition of the machine 102 based on these features, based on the classifier 1 trained in the first training step 500. It will be appreciated that NN2 is therefore preferably operable to extract at least one feature of the first non-stationary output that is insensitive to the level of stationarity of the non-stationary operation of the machine 102 and that can therefore be accurately classified by the classifier 1, despite the fact that the classifier 1 has been trained to classify features of stationary signals.

[0079] The classification results performed by the classifier 1 in step 540 may be provided to the controller 154, which may automatically modify the operating parameters of the machine 102 based on the results, initiate or schedule a maintenance operation on the machine 102, or initiate or schedule a maintenance operation on the machine 102. Additionally or alternatively, the results of such classification may be presented to at least one human expert in a human-understandable form, who may assess the machine state based on the results and initiate maintenance, repairs, or operating parameter changes accordingly. It should be understood that the classifier 1 may be trained in the first training step 500 to allow classification of specific faults in the machine 102, such as bearing wear. Additionally or alternatively, the classifier 1 may be trained in a non-fault-specific manner to allow classification of the state of the machine 102 as healthy or unhealthy, where the nature of the abnormal operation of the machine 102 is not specified by the classifier 1.

[0080] It should be understood that FIG. 5A to FIG. 5C In the example, multiple sensors S1–S n The initial input signal may include raw data acquired by a plurality of sensors or may include processed data based on the raw data, wherein the raw data is converted before being further processed by the network.

[0081] Reference now Figure 6 , which is a simplified flowchart showing the Figure 1 or Figure 4 The steps involved in extracting features of signals generated by a non-stationary machine operating within a system of the type shown.

[0082] like Figure 6 As shown, the method 600 for extracting features of signals from a non-stationary machine may begin with a first step 602 , where at least first and second sets of signals are acquired from a non-stationary machine during operation of the non-stationary machine.

[0083] The first set of signals and the second set of signals are preferably obtained over the same at least one operating time frame range of the machine 102, so that the corresponding non-stationary outputs of the sensors represent the same cycle or multiple cycles of the machine operation. In a preferred embodiment of the present invention, multiple sensors are sampled synchronously with each other. The sampling can be synchronized within a time window dt of the order of dt≈0.01 / f, where f is the frequency of interest, such as the RPM frequency. In another preferred embodiment of the present invention, multiple sensors are sampled synchronously with each other within a time window dt of the order of dt≈0.01 / f, where f is the sensor sampling frequency. It is worth noting that the value of 0.01 is selected here to limit the error in the synchronous sampling to 1%. However, other values ​​can be selected depending on the required error limit.

[0084] Due to the non-stationary operation of the machine monitored thereby, the first and second sets of signals preferably have a non-stationary nature.

[0085] It is understood that the first group of signals and the second group of signals may be signals of different types, such as vibration signals and magnetic signals, respectively. Alternatively, the first group of signals and the second group of signals may be signals of the same type, such as both magnetic signals or both vibration signals, but are acquired by different types of sensors and are therefore different in some aspects thereof.

[0086] As shown in the second step 604, one or more features are preferably extracted from the first set of signals. For example, one or more features may be extracted from a non-stationary magnetic signal output by a magnetic sensor or a non-stationary vibration signal output by a vibration sensor. For example, a magnetic wavelet transform may be applied to the non-stationary magnetic signal to derive a binary magnetic wavelet mask containing only frequencies related to the RPM of the non-stationary machine, as described above with reference to Figure 1 For further example, the NN may automatically extract one or more features from a non-stationary magnetic signal output by a magnetic sensor or from a non-stationary vibration signal output by a vibration sensor, as described above with reference to Figures 4 to 5C described.

[0087] As seen in the third step 606, the second set of signals is preferably modified using the features extracted from the first set of signals in the second step 604. Thus, the first set of signals is used to transform or calibrate the second set of signals before further features are extracted from the second set of signals. For example, the binary magnetic wavelet mask derived in the second step 604 can be used to calibrate the vibration signal by multiplying the binary magnetic wavelet mask by a vibration wavelet based on the vibration signal so as to filter out all frequencies from the vibration wavelet except for frequencies associated with the non-stationary machine speed frequency. By way of further example, the NN can automatically modify one type of non-stationary signal based on features extracted from another type of non-stationary signal.

[0088] As shown in the fourth step 608, one or more features are then preferably extracted from the second set of signals as modified in the third step 606. For example, the modified vibration wavelets derived in the third step 606 can be summed to extract the vibration amplitude of the RPM frequency. By further example, the NN can automatically extract features of the non-stationary signal modified thereby. As seen in the fifth step 610, the extracted one or more features can be analyzed to find the health status of the monitored non-stationary machine. It should be understood that the one or more features extracted by the method 600 are preferably representative of machine health and are insensitive to the level of smoothness of the machine operation. For example, in the case of signal features extracted by the NN, the extracted features can be classified and the machine condition can be derived based on a classifier trained using data obtained from the same or similar machine operating in a stable rather than non-stationary manner.

[0089] As shown in the sixth step 612, a repair or maintenance operation may be performed or planned to be performed on the machine, or an operating parameter of the machine may be changed based on the health status of the machine determined in the fifth step 610. It should be understood that the sixth step 612 may be performed in an automated manner by a controller coupled to the monitored machine. Additionally or alternatively, the health status of the monitored machine may be communicated to a human expert in a human-perceivable form, and the human expert may be involved in directing repairs, maintenance, or operational changes to the monitored machine.

[0090] Reference now Figure 7 , which is a simplified partial schematic, partial block diagram of a system for automatic analysis of non-stationary machine performance constructed and operative in accordance with yet another preferred embodiment of the present invention.

[0091] like Figure 7 As shown, a system 700 is provided for monitoring and automatically analyzing the performance of a non-stationary machine, implemented here by way of example only, as a robotic machine 702 or servo motor operating in accordance with a programmable operating mode or recipe embedded in a controller 704 of the machine. The machine 702 operates in a non-stationary manner due to time-varying modifications of machine parameters (e.g., input power and rotational speed) by the controller 704.

[0092] It is noteworthy that conventional machine monitoring and diagnostic systems are not well suited for monitoring and analyzing such non-stationary machines due to the characteristics of the non-stationary signals they display. According to preferred embodiments of the present invention, novel systems and methods are provided for monitoring and automatic state analysis of a machine operating in a non-stationary process according to a time-varying recipe. Such state analysis is preferably based on monitoring the non-stationary machine with multiple sensors, fusing the non-stationary outputs of the sensors over a recipe interval to create a fused output, and extracting signal characteristics reflecting the state of the machine operating according to the recipe based on the fused output.

[0093] In certain embodiments of the present invention, after finding a recipe interval, it may be advantageous to adjust the sampling characteristics of a plurality of sensors in accordance with the found recipe interval. Signal features may then be extracted and automatically learned that are associated with the machine during the recipe interval. Such learning may be used to establish a baseline model of the machine's performance during the recipe interval. Subsequent signal features extracted from the signals generated by the machine may then be compared to the baseline model of the machine's performance in order to detect anomalies and deviations that indicate a change in the health of the machine. Appropriate corrective action may be taken based on the anomalies found, such as machine repair, maintenance, or changes in operating parameters.

[0094] like Figure 7As shown, the robot 702 is preferably monitored by a plurality of sensors, which are implemented here by way of example as a first vibration sensor 710 that measures, for example, vibration emitted from a motor 711 driving the arm of the robot 702 and a second magnetic sensor 712 that measures magnetic flux emitted from the motor 711. It should be understood that due to the non-stationary operation mode of the robot 702, regardless of the health state of the robot 702, the vibration signal and magnetic flux signal emitted from the motor 711 of the robot 702 and measured by the first and second sensors 710 and 712 are preferably non-stationary signals.

[0095] Exemplary data measured by sensors 710 and 712 representing vibration and magnetic signals generated by a processing machine such as robot 702 is shown in FIG. Figure 8 and Fig. 9 middle. Figure 8 Data is shown for two vibration waveforms during a four second recording of the motor 711 of the robot 702 operating during two different time periods. Fig. 9 Shows the Figure 8 The two vibration waveforms shown are synchronized to obtain the two magnetic waveform data. Figure 8 and Fig. 9 As can be seen in FIG. 7 , the robot 702 operates according to a recipe having three operating areas. Figure 8 and Fig. 9 Each operating region is associated with increased vibration and magnetic signals, denoted A, B, and C in the recipe. The three operating regions A, B, and C within the recipe produce different signals due to different input parameters, as modified by the controller 704 of the machine 702.

[0096] It is noted that vibration sensors and magnetic sensors 710 and 712 are shown by way of example only, and additional or alternative sensor types may be used to monitor the performance of machine 702, including but not limited to current sensors, electromagnetic sensors, acoustic sensors, and temperature sensors. It should also be understood that the multiple sensors monitoring machine 702 may include multiple sensors measuring the same type of signal, such as multiple vibration sensors or multiple magnetic sensors, where the sensors are different from each other and, therefore, produce slightly different outputs despite measuring the same parameter of machine 702.

[0097] It should also be appreciated that the plurality of sensors monitoring machine 702 may be contact or non-contact sensors. In certain embodiments of the invention, the use of non-contact sensors may be preferred due to the difficulty in securely attaching sensors to moving, non-stationary equipment.

[0098] A plurality of sensors, such as sensors 710 and 712, preferably monitor the machine 702 during the same at least one operating time frame of the machine 702, so that the respective non-stationary outputs of the sensors represent the same cycle or cycles of the machine operation. In a preferred embodiment of the present invention, a plurality of sensors, such as sensors 710 and 712 monitoring the machine 702, are sampled synchronously with each other. The sampling may be synchronized within a time window dt of the order of dt≈0.01 / f, where f is the frequency of interest, such as the RPM frequency. In another preferred embodiment of the present invention, a plurality of sensors, such as sensors 710 and 712 monitoring the machine 702, are sampled synchronously with each other within a time window dt of the order of dt≈0.01 / f, where f is the sensor sampling frequency. For example, the sensors 110 and 112 may operate at a sampling frequency of 20 kHz and thus the time frame dt in which the sensors are synchronously sampled is given by 0.5 microseconds. It is worth noting that a value of 0.01 is selected here to limit the error in the synchronous sampling to 1%. However, other values ​​may be selected according to the desired error limit. As described in detail below, the sensor sampling frequency can be adjusted according to the recipe interval after the recipe interval is found.

[0099] Preferably, the first non-stationary output provided by the vibration sensor 710 and the second non-stationary output provided by the magnetic sensor 712 are provided to gain filters (GFs) 714 and 716, respectively, which in turn provide corresponding signals to analog-to-digital converters (ADCs) 718 and 720. The digitized signals output by the ADCs 718 and 720 are preferably provided to a memory component 722 such as a CPU and optionally to a cloud database 724. It should be understood that, depending on the operating specifications of the system 700, the synchronously sampled data acquired by the sensors 710 and 712 can be transmitted to a completely cloud-based system, a local hardware system, or a system including local components and cloud-based components. It should also be understood that the specific processing components shown herein, including GFs 714 and 116, ADCs 718 and 120, and CPU 722, are detailed only by way of example, and any suitable signal processing components can be employed in the system of the present invention.

[0100] The synchronously sampled first non-stationary output and second non-stationary output respectively output by the vibration sensor 710 and the magnetic sensor 712 are preferably provided to the machine recipe analyzer 730, which may be included in the cloud database 724, or may include a local processing component.

[0101] Preferably, the machine recipe analyzer 730 includes a recipe interval finder 760 for finding recipe intervals that control the operation of the robotic machine 702. The recipe interval finder 760 may find recipe intervals based on repeating intervals in the outputs of the detection sensors 710 and 712. Preferably, although not necessarily, the recipe interval finder 760 may employ cross sensor fusion, as described above with reference to Figures 1 to 6 As described above, in order to find repeating signal features in the fused magnetic signal and vibration signal generated by the machine 702.

[0102] Figure 8 The level of cross-correlation between the two sets of vibration data indicates the recipe interval, such as Fig.10 The cross-correlation algorithm can be applied at multiple time intervals and a ranking can be given based on the correlation magnitude weighted by the rotation speed according to the present invention.

[0103]

[0104] Where S i represents sensor data from any type of sensor (e.g., a vibration sensor or a magnetic sensor), S tot is the sum of the products of the energy magnitudes of the two sensors, and W is the RPM weighting function.

[0105] The recipe interval finder 760 may selectively provide feedback control to the sensors 710 and 712 such that, after determining a recipe interval, the recipe interval finder 760 may automatically operate to reconfigure sampling parameters of the sensors 710 and 712 based on the recipe interval thus found. For example, after determining a recipe interval, the sampling duration and / or frequency of one or both of the sensors 710 and 712 may be adjusted to allow the sensors 710 and 712 to monitor the robot 702 throughout the recipe interval, within a defined period within a recipe interval, or within multiple periods of a recipe interval.

[0106] After the recipe interval finder 760 finds the recipe interval and adjusts the sampling parameters of the sensors 710 and 712 accordingly, the data acquired by the sensors 710 and 712 over the recipe interval may be processed to detect regions or zones of interest within the signals acquired over the recipe interval, as found by the recipe interval finder 760. These regions may be selected based on being particularly representative of machine performance, and may, as examples, correspond to regions that include increased vibration or increased magnetic flux signals.

[0107] Region detection can be performed based on applying a Hilbert transform to the data within the recipe interval and analyzing its envelope function. The Hilbert transform can be applied to data from individual sensors or extracted from cross-sensor processing of data from individual sensors. Alternatively, region detection can be performed based on convolution masking with a Gaussian function in order to capture the exponential signal rise and fall associated with the vibration region. Fig.11 An example of a vibration signal divided into regions is shown in FIG. Fig.11 As shown in , the dashed line 780 represents the fitting algorithm applied to region detection and the individual regions are delineated by vertical lines 782.

[0108] The data acquired by sensors 710 and 712 within the recipe interval is preferably analyzed in order to extract its characteristic features and thus learn the characteristics of the machine that executes the recipe within the recipe interval. In the case of performing zone detection, the signal features can be calculated region by region, each specific region, rather than region by region over the entire recipe interval. Multiple regions can be classified based on their characteristic features, such as regional energy, regional duration, and torque changes within the region. It can be understood that it is very advantageous to divide the recipe interval into representative regions and to find signal features in each region rather than the entire recipe interval, because feature extraction within the entire recipe interval usually results in low-resolution, low-representative features, because the operating mode within the recipe interval varies greatly. However, it should be understood that in certain embodiments of the present invention, if the changes within the recipe interval are of a small enough scale to make such a zone division unnecessary, the recipe interval may not need to be divided into regions.

[0109] Additionally or alternatively, the analysis of the data acquired by sensors 710 and 712 during the recipe interval may be based on other machine learning algorithms. Features extracted from the data may include statistical features, such as probability density function moments, cumulants, or entropy of the data; operating characteristics, such as speed and input power; features such as main waveform components, waveform power spectrum, signal energy, signal integrity, etc.; machine performance features, such as machine efficiency, for example based on a signal analysis algorithm of the type described in WO / 2018 / 198111; and features associated with environmental features that may be acquired from external sensors, for example, in the case of a machine involving fluid flow, the ambient temperature acquired from an external temperature sensor may be compared with the machine temperature and the machine pressure.

[0110] The signal characteristics associated with the machine 702 executing the recipe, preferably including the regional characteristics, can be learned by the recipe learning algorithm 790. The automatic learning of the signal characteristics associated with the recipe can be enhanced by the input of external data, such as, for example, data from the controller 704. The controller data can be correlated with the data from the sensors 710 and 712 to provide additional information related to the recipe according to which the machine 702 operates, thereby improving the accuracy of the recipe interval detection.

[0111] After the recipe learning algorithm 790 discovers the recipe characteristics and features, a performance model of the machine 702 during the recipe execution can be established. Such a model can be constructed by the recipe model builder 792 and can constitute a baseline model representing the machine performance within a given recipe range. The model can be stored in a database, such as a cloud database 724. The model can be used as a basis, and subsequent features of the signals generated by the machine 702 during the recipe execution can be compared with the basis to detect anomalies and deviations. For example, anomalies in machine performance can be detected by anomaly detector 794. Anomalies in recipe characteristics may be due to, for example, machine degradation. Alternatively, anomalies may be caused by malicious intervention in machine operation, such as in the case of a malicious security vulnerability installed on the controller 704 in order to modify the machine recipe. Anomalies detected by the anomaly detector 794 can be output, and based on these anomalies, the health status or malicious intervention of the machine 702 can be discovered. Appropriate corrective measures can then be performed automatically or based on human intervention, such as changing machine operating parameters, stopping machine operation, or repairing the machine, such as through a machine control module 796.

[0112] The model constructed by the recipe model builder 792 may include data from low-power sensors coupled to the machine 702 that operate at a lower power than but are related to sensors 710 and 712. Such low-power sensing may be performed continuously for the machine 702 rather than at specific recipe intervals, and may be used to detect real-time anomalies in sensor readings from the machine performance model during recipe execution. In the event that such anomalies are discovered, the lower-power sensors may activate higher-power sensors, such as sensors 710 and 712, to collect additional data, based on which the anomaly detector 794 may apply more advanced anomaly detection algorithms.

[0113] The recipe learning algorithm 790, the recipe model builder 792 and the anomaly detector 794 correspond to a preferred embodiment of a machine learning algorithm for automatically constructing an operating model of the machine 702 that executes a given recipe and comparing the signals emitted from the machine 702 with the model to evaluate anomalies in the operation of the machine 702.

[0114] According to a particularly preferred embodiment of the present invention, the recipe learning algorithm 790, the recipe model builder 792 and the anomaly detector 794 can adopt deep learning, and particularly preferably can adopt a recursive NN (RNN), such as a long-short term memory (LSTM) network, in order to detect anomalies in the non-stationary signals emitted from the machine 702 during the execution of the resulting recipe.

[0115] Fig. 12A , Fig. 12B and Fig. 12C A particularly preferred embodiment of a possible implementation of an LSTM network that may be used within the machine recipe analyzer 730 is shown in FIG. Fig. 12A , a non-stationary recipe data set 8100 such as generated by the machine 702 is considered to include multiple signals, here embodied as a first signal set S1 including a time-varying vibration signal and a second signal set S2 including a time-varying magnetic signal. It will be appreciated that the time series of the vibration signal and the magnetic signal preferably span the entire recipe interval and may span multiple recipe intervals, as preferably discovered by the recipe interval finder 760. It will also be appreciated that the first and second signal sets S1 and S2 may include raw sensor data, but more preferably include features extracted from the raw data, such as line frequency or RPM or regional features. For example, the first signal set S1 may include the vibration level of the motor 711, and the second signal set S2 may include the current line frequency of the motor, although not necessarily in the region. It will be appreciated that while the vibration level of the motor 711 represents the mechanical state of the motor 711, the line frequency of the motor 711 based on the magnetic signal is directly related to the motor speed, and thus directly related to the operating state of the motor.

[0116] The signal sets S1 and S2 are preferably provided to a sensor fusion time series operator 8102. The sensor fusion time series operator 8102 preferably fuses the different types of signals into a single combined signal, the combined signal comprising alternating signals in the signal type arranged in time. Thus, magnetic signal Mi-1 is interspersed with a corresponding synchronous vibration signal Vi-1, followed by magnetic signal Mi interspersed with a corresponding synchronous vibration signal Vi, and so on. Where signals S1 and S2 correspond to signals representing the mechanical and operational states of the machine, respectively, the sensor fusion time series operator 8102 is operable to fuse these signals.

[0117] The interleaved data set output by the sensor fusion time series operator 8102, which may include interleaved data representing the machine mechanical state and the machine operating state, is preferably provided to the LSTM network 8104. Referring to FIG. Fig. 12B , one possible mode of operation of the LSTM network 8104 can be best understood. Fig. 12BAs shown, the LSTM network 8104 receives the interleaved magnetic signal and the vibration signal. At each neuron 8105, the LSTM network predicts the corresponding shift signal, so that each LSTM neuron is operable to predict the next adjacent data point in the fused time series.

[0118] The predicted signal value is then compared to the actual measured value of the signal in the loss function module 8106. The loss function module 8106 is preferably used to compare the predicted signal value with the actual corresponding measured signal value, and iteratively update the weights of the LSTM network 8104 to minimize the difference between the predicted value and the actual value. The LSTM network 8104 uses back propagation and gradient descent algorithms to iteratively train the loss function, which represents the difference between the time series predicted by the LSTM network 804 and the true time series. For example, the loss function module 8106 can find the MSE loss.

[0119] According to another preferred embodiment of the present invention, the LSTM network 8104 can be an LSTM autoencoder, which is operable to encode the interspersed data set and output the signal features to the loss function module 8106. The signal features output by the LSTM network 8104 are preferably based on the fusion of the input signals, which in this case are embodied as magnetic signals and vibration signals, and feature extraction based thereon. The loss function module 8106 is preferably used to compare the decoded signal features output by the LSTM network 8104 with the input interspersed data set. The difference between the decoded signal features and the interspersed input data set is then fed back to the LSTM network 8104, and the weights of the encoder are iteratively adjusted using back-propagation and gradient descent algorithms until the loss function is small enough to be acceptable and the LSTM network 8104 is considered to be trained to encode the interspersed data set to an acceptable level of accuracy.

[0120] Now go to Fig. 12C ,exist Fig. 12A and Fig. 12B After the training of the LSTM network 8104 shown in , the magnetic signal Mi generated by the machine 702 during its operation in the recipe interval can be provided to the trained LSTM network 8104. The LSTM network 8104 is preferably used to predict the corresponding vibration signal Corresponding to the magnetic signal input, based on Fig. 12A and Fig. 12B The process shown learns the correspondence between the vibration signal and the magnetic signal.

[0121] Predicting vibration signatures Can be compared with the magnetic signal M i The actual vibration signal V is measured synchronously iFor example, the actual vibration signal and the predicted vibration signal may be compared by anomaly detector 8108, which may be a particularly preferred embodiment of anomaly detector 794. In the event that the predicted signal differs from the actual measured signal by more than a given threshold difference, the measured signal is considered to be abnormal and indicates an abnormal operating state of machine 702.

[0122] It is worth noting that the systems and methods described herein for conditionally monitoring a process machine operating in accordance with a given recipe are not limited to the case of a single machine. Rather, the present invention can be implemented with respect to multiple machines contributing to a common process. Such a process can be a parallel recipe process, in which multiple machines share the same operating mode at the same point in time, or a series recipe process, in which machines operate in series, with the output of one machine forming the input of another machine in a series manner.

[0123] In the case of multiple machines participating in a common process, the signals are preferably acquired synchronously by multiple sensors coupled to each of the multiple machines. Preferably, in the case of a single machine 702, the recipe features, including recipe intervals, recipe areas, and recipe area features, are extracted as described above, and the recipe baseline model is constructed on all or part of the machines executing the process, rather than for a single machine.

[0124] In the case of multiple machines, the cross-correlation to establish recipe intervals can generally be performed as described above in the case of a single machine, but for multiple sensors across multiple machines, more specifically, for sensors of a given type coupled to multiple machines, according to the following formula:

[0125]

[0126] Where n represents the sensor type for any type of sensor, such as vibration or magnetic, i represents the machine index, and j ranges from i+1 to i n The other parameters are as described above with reference to formula (2).

[0127] Notably, such a multi-machine process model can be in addition to a single baseline model built for each machine in the process, such as robot 702. This improves the accuracy of identifying machine performance anomalies, as multi-machine process anomaly detection can be used to enhance single machine anomaly detection.

[0128] Sensors of the same or different types respectively coupled to different machines of the plurality of machines may be correlated with each other. As described above, such correlation may be used to find the type of process performed by the machine and its recipe interval in order to appropriately configure the sensor sampling parameters.

[0129] If an LSTM network is used within the machine recipe analyzer 730 to analyze the operation of multiple machines that jointly perform a process, for example, the signals emitted from the machines can be combined into: a total magnetic vector including magnetic signals from all multiple machines and a total vibration vector including vibration signals from all multiple machines, the vibration signals being acquired synchronously with respect to the magnetic signals. The vibration vector can then be interleaved with the magnetic vector by the sensor fusion time series operator 8102 to allow the LSTM network 8104 to learn the characteristics of the signals of a group of machines performing a given recipe, thereby detecting deviations therefrom, as described above. It should be understood that, additionally or alternatively, the signals from multiple machines that jointly perform a process can be averaged or otherwise combined and then fused by the sensor fusion time series operator 8102 and provided to the LSTM network 8104.

[0130] Reference now Fig.13 , which is a simplified flowchart illustrating the extraction of Figure 7 The steps involved in characterizing the signals produced by a non-stationary machine operating in a system of the type shown.

[0131] like Fig.13 As shown, a method 1300 for extracting features of signals generated by a non-stationary machine executing a process according to a recipe may begin with a first step 1302, wherein signals are synchronously acquired from a plurality of sensors coupled to at least one non-stationary machine operating according to a recipe. According to a preferred embodiment of the present invention, the plurality of sensors may be coupled to a single non-stationary machine executing a recipe embedded in its controller. According to another embodiment of the present invention, the at least one non-stationary machine may include a group of non-stationary machines contributing to a common process, in a serial or parallel manner, and the plurality of sensors may include a plurality of sensors coupled to each machine in the group of non-stationary machines.

[0132] As shown in the second step 1304, a recipe interval for a recipe executed by the machine or group of machines is preferably found. As shown in the third step 1306, parameters of a plurality of sensors may be adjusted based on the recipe interval found in the second step 1304. For example, the sensor sampling frequency or sampling duration may be adjusted to allow the sensor to sample the machine or group of machines throughout the recipe, within multiple cycles of the recipe, or at a given point in time within multiple cycles.

[0133] As shown in the fourth step 1310, a baseline model of machine performance during recipe execution can be constructed based on the discovered recipe characteristics. It should be understood that such a baseline model can be constructed for a single machine executing a process according to a recipe, a group of machines contributing to a common process according to a recipe, or both.

[0134] According to a preferred embodiment of the present invention, establishing a machine performance model may include fusing data from multiple sensors in an interspersed arrangement of data time points from different sensors, and providing the interspersed data set to an LSTM network for learning the relationship between data points.

[0135] As shown in the fifth step 1314, after the machine performance model is created during the execution of the recipe, subsequent machine signal features measured during the execution of the recipe by the machine or a group of machines can be compared with the baseline model to detect anomalies associated therewith. Such anomalies can indicate the health status of the machine or the safety status of the machine.

[0136] In the case where the fourth step 1310 uses an LSTM network to learn the recipe features, the fifth step 1314 may involve the LSTM network predicting a signal from one type of sensor at one point in time that is expected to correspond to a measurement signal acquired from a different type of sensor at substantially the same point in time. The predicted signal is then compared to the actual measurement signal and anomalies are identified if there is a deviation between the signals.

[0137] As shown in the sixth step 1316, in the event that a machine performance anomaly is detected, actions may be performed on the machine or a group of machines, including repair or maintenance activities and changes in machine operating parameters, including stopping the machine from operating.

[0138] Notably, method 1300 is preferably automated such that machine performance models may be automatically created and anomalies in machine characteristics associated therewith may be automatically detected notwithstanding that operation of the monitored machine or group of machines is non-stationary.

[0139] Those skilled in the art will appreciate that the present invention is not limited to what is specifically shown and described above. The scope of the present invention includes combinations and sub-combinations of the various features described above and modifications thereof, all of which are not in the prior art.

Claims

1. A method for operating a non-steady-running machine, characterized in that: The method comprises: acquiring at least one first non-stationary signal generated by at least one machine operating in a non-stationary manner during at least one operating time frame; acquiring at least one second non-stationary signal from the at least one machine during the at least one operating time frame, the second non-stationary signal being different from the first non-stationary signal; applying a wavelet transform to both the at least first non-stationary signal and the different second non-stationary signal; generating a mask of the wavelet transform of one of the first non-stationary signal and the second non-stationary signal; multiplying the mask of the wavelet transform of one of the first non-stationary signal and the second non-stationary signal by the wavelet transform of the other of the first non-stationary signal and the second non-stationary signal; extracting at least one feature of at least one of the first and second non-stationary signals based on the multiplication; analyzing the at least one characteristic to determine a condition of the at least one machine; and Based on the condition, at least one of a service operation, a maintenance operation, and a modification of an operating parameter of the at least one machine is performed.

2. The method according to claim 1, characterized in that: The first non-stationary signal and the second non-stationary signal are of the same type.

3. The method according to claim 1, characterized in that: The first non-stationary signal and the second non-stationary signal are of different types.

4. The method according to claim 1, characterized in that: The mask comprises a binary wavelet matrix.

5. The method according to claim 1, characterized in that: The step of multiplying the mask of the wavelet transform of one of the first and second non-stationary signals with the wavelet transform of the other of the first and second non-stationary signals also includes multiplying the wavelet transform of one of the first and second non-stationary signals with the inverse mask of the wavelet transform of the other of the first and second non-stationary signals.

6. The method according to claim 1, characterized in that: The extracting comprises summing the multiplied wavelet transforms.

7. The method according to claim 6, characterized in that: The extracting includes extracting the at least one feature directly from the multiplied wavelet transforms.

8. The method according to claim 1, characterized in that: The execution is automatically performed by a controller of the at least one machine.

9. The method according to claim 1, characterized in that: The method also includes tracking the at least one feature over time via a feature tracking algorithm.

10. The method according to claim 1, characterized in that: The step of determining the condition further includes determining a condition of the at least one machine related to at least one of machine performance and machine efficiency.

11. A system for operating a non-steady-running machine, characterized in that: The system comprises: at least one first sensor for acquiring at least one first non-stationary signal generated by at least one machine operating in a non-stationary manner during at least one operating time frame; at least one second sensor for acquiring at least one second non-stationary signal from the at least one machine during the at least one operating time frame, the second non-stationary signal being different from the first non-stationary signal; A signal processor for: applying a wavelet transform to both the at least first non-stationary signal and the different second non-stationary signal; generating a mask of the wavelet transform of one of the first non-stationary signal and the second non-stationary signal; multiplying the mask of the wavelet transform of one of the first non-stationary signal and the second non-stationary signal by the wavelet transform of the other of the first non-stationary signal and the second non-stationary signal; extracting at least one feature of at least one of the first non-stationary signal and the second non-stationary signal based on the multiplied wavelet transform; and analyzing the at least one characteristic to determine a condition of the at least one machine; and A machine control module is configured to perform at least one of a service operation, a maintenance operation, and a modification of an operating parameter of the at least one machine based on the condition.

12. The system according to claim 11, characterized in that: The first sensor and the second sensor are of the same type.

13. The system of claim 11, wherein: The first sensor and the second sensor are of different types.

14. The system of claim 11, wherein: The mask comprises a binary wavelet matrix.

15. The system of claim 11, wherein: The signal processor is used to multiply the wavelet transform of one of the first non-stationary signal and the second non-stationary signal by the mask of the wavelet transform of the other non-stationary signal of the first non-stationary signal and the second non-stationary signal, and is also used to multiply the wavelet transform of one of the first non-stationary signal and the second non-stationary signal by the inverse mask of the wavelet transform of the other non-stationary signal of the first non-stationary signal and the second non-stationary signal.

16. The system of claim 11, wherein: The signal processor is used to extract at least one feature, including summing the multiplied wavelet transforms.

17. The system of claim 16, wherein: The signal processor is used to extract the at least one feature directly from the multiplied wavelet transform.

18. The system of claim 11, wherein: The machine control module automatically operates to perform at least one of the service operation, maintenance operation, and modification of an operating parameter of the at least one machine based on the condition.

19. The system of claim 11, wherein: The system also includes a feature tracking algorithm for automatically tracking the at least one feature extracted by the signal processor over time.

20. The system of claim 11, wherein: The signal processor is operative to determine a condition of the machine, including the signal processor being operative to determine a condition of the at least one machine related to at least one of machine performance and machine efficiency.