Fault diagnosis method, device and equipment for permanent magnet synchronous motor

Abstract

The invention provides a fault diagnosis method, device and equipment for a permanent magnet synchronous motor. The method comprises the following steps: acquiring a magnetic field signal of the motor; reconstructing a phase space of the motor based on the magnetic field signal; constructing a recursive image group corresponding to the magnetic field signal according to the magnetic field signal and the phase space; obtaining a first data set based on the recursive image group; and inputting the first data set into a preset multi-channel convolutional neural network diagnosis model for fault diagnosis to obtain a diagnosis result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric machines, and in particular to a fault diagnosis method, device and equipment for a permanent magnet synchronous motor. BACKGROUND

[0002] Although deep learning methods are widely used in fault diagnosis of permanent magnet synchronous motors, the fault diagnosis method based on one-dimensional time series signals still has defects such as complex signal processing, difficulty in extracting fault information, performance degradation under noise and working condition changes, etc. SUMMARY

[0003] The present application provides a fault diagnosis method, device and equipment for a permanent magnet synchronous motor, which solves the problem of defects such as complex signal processing, difficulty in extracting fault information, performance degradation under noise and working condition changes, etc. in the fault diagnosis method based on one-dimensional time series signals.

[0004] The present application provides a fault diagnosis method for a permanent magnet synchronous motor, which comprises:

[0005] Obtaining a magnetic field signal of the motor;

[0006] Reconstructing a phase space of the motor based on the magnetic field signal;

[0007] Constructing a recurrent image group corresponding to the magnetic field signal according to the magnetic field signal and the phase space;

[0008] Obtaining a first data set based on the recurrent image group;

[0009] Inputting the first data set into a diagnosis model of a preset multi-channel convolutional neural network to perform fault diagnosis and obtaining a diagnosis result.

[0010] In some embodiments, the reconstructing of the phase space of the motor based on the magnetic field signal comprises:

[0011] Performing mean removal processing on the magnetic field signal to obtain a time series;

[0012] Determining a target delay time corresponding to the reconstruction of the magnetic field signal in the phase space based on the time series;

[0013] Obtaining a target embedding dimension corresponding to the reconstruction of the magnetic field signal in the phase space;

[0014] Reconstructing the phase space of the motor based on the target delay time and the target embedding dimension.

[0015] In some embodiments, the determining of the target delay time corresponding to the reconstruction of the magnetic field signal in the phase space based on the time series comprises:

[0016] acquiring a preset given delay time;

[0017] determining an autocorrelation function of the magnetic field signal at different delay times based on the time series and the given delay time;

[0018] normalizing the autocorrelation function to obtain a normalized autocorrelation function;

[0019] determining a variance of the original time series of the motor;

[0020] determining the target delay time based on the normalized autocorrelation function and the variance of the original time series.

[0021] In some embodiments, the acquiring of the target embedding dimension corresponding to the reconstruction of the phase space of the magnetic field signal comprises:

[0022] acquiring a preset given embedding dimension and a first proportion parameter corresponding to a first false nearest neighbor point;

[0023] determining a target nearest neighbor point in the phase space based on the given embedding dimension;

[0024] increasing the given embedding dimension by a preset dimension to construct a new reconstructed phase space of the motor;

[0025] determining a second false nearest neighbor point of the new reconstructed phase space according to the target nearest neighbor point;

[0026] determining a second proportion parameter of the second false nearest neighbor point based on the first proportion parameter;

[0027] in a case where the value of the second proportion parameter is less than a preset threshold, determining the target embedding dimension.

[0028] In some embodiments, the constructing of the recurrent image group corresponding to the magnetic field signal according to the magnetic field signal and the phase space comprises:

[0029] wavelet packet decomposing the magnetic field signal to obtain a length parameter of each sub-band in a frequency band generated by the magnetic field signal;

[0030] determining a distance matrix of all point pairs in the phase space based on the length parameter and the phase space;

[0031] normalizing the distance matrix to obtain a normalized distance matrix;

[0032] taking the normalized distance matrix as a gray value of the recurrent image group to obtain a recurrent image and a corresponding resolution;

[0033] The recursive images corresponding to the magnetic field signals are obtained by transforming the recursive images and corresponding resolutions.

[0034] In some embodiments, the method further comprises:

[0035] obtaining a data set and corresponding labels of each type of motor fault;

[0036] determining a diagnostic model of the multi-channel convolutional neural network based on the data set and corresponding labels of each type of motor fault.

[0037] In some embodiments, the determining a diagnostic model of the multi-channel convolutional neural network based on the data set and corresponding labels of each type of motor fault comprises:

[0038] dividing the data set and corresponding labels of each type of motor fault into a training set and a test set;

[0039] determining a diagnostic model of the multi-channel convolutional neural network based on the training set and the test set.

[0040] In some embodiments, the first data set comprises a plurality of feature graphs; each feature graph represents the characteristics of the magnetic field signals of the motor at different frequency scales; the convolution kernels used in each channel of the diagnostic model of the multi-channel convolutional neural network are independent, and the weights of the results output by each channel are different when integrated in the full connection layer; the input of the first data set into the diagnostic model of the preset multi-channel convolutional neural network for fault diagnosis comprises:

[0041] inputting the plurality of feature graphs into the diagnostic model of the preset multi-channel convolutional neural network for fault diagnosis to obtain a diagnostic result; the diagnostic result comprises winding fault, bearing inner ring fault, bearing outer ring fault, local demagnetization fault, uniform demagnetization fault, and rotor imbalance fault of the motor.

[0042] Embodiments of the present application also provide a fault diagnosis device for a permanent magnet synchronous motor, the device comprising:

[0043] an acquisition unit configured to acquire a magnetic field signal of the motor;

[0044] a reconstruction unit configured to reconstruct a phase space of the motor based on the magnetic field signal;

[0045] a construction unit configured to construct a recursive image group corresponding to the magnetic field signal based on the magnetic field signal and the phase space;

[0046] an acquisition unit configured to acquire a first data set based on the recursive image group;

[0047] A diagnosis unit is configured to input the first data set into a diagnosis model of a preset multi-channel convolutional neural network to perform fault diagnosis and obtain a diagnosis result.

[0048] The embodiment of the present application provides a fault diagnosis device of a permanent magnet synchronous motor, the device comprises a processor and a memory for storing a computer program capable of running on the processor, wherein the processor is configured to execute the steps of the method according to any one of the above embodiments when running the computer program.

[0049] The embodiment of the present application provides a storage medium, and the storage medium stores a computer program; the computer program is executed by a processor to implement the steps of the method according to any one of the above embodiments.

[0050] The embodiment of the present application provides a fault diagnosis method of a permanent magnet synchronous motor, the method comprises: acquiring a magnetic field signal of the motor; reconstructing a phase space of the motor based on the magnetic field signal; constructing a recurrent image group corresponding to the magnetic field signal according to the magnetic field signal and the phase space; acquiring a first data set based on the recurrent image group; inputting the first data set into a diagnosis model of a preset multi-channel convolutional neural network to perform fault diagnosis and obtain a diagnosis result. By using the technical scheme of the present application, the reconstructed phase space of the motor is determined based on the acquired magnetic field signal of the motor, and then the recurrent image group corresponding to the magnetic field signal is constructed; the first data set is acquired based on the recurrent image group; the first data set is input into the diagnosis model of the preset multi-channel convolutional neural network to perform fault diagnosis and obtain a diagnosis result. That is, the fault diagnosis method based on one-dimensional time sequence signal still has defects such as complex signal processing, difficulty in extracting fault information, performance decline under noise and working condition change, and the like. The multi-channel convolutional neural network diagnosis scheme based on the recurrent image group has high diagnosis precision, small calculation amount, strong anti-noise ability, strong adaptability to multiple working conditions, and does not require complex algorithms. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 A flowchart of a fault diagnosis method of a permanent magnet synchronous motor according to an embodiment of the present application is provided;

[0052] Figure 2 A schematic diagram of a PMSM fault diagnosis technical route based on a recurrent image group according to an embodiment of the present application is provided;

[0053] Figure 3 A schematic diagram of a recurrent image group generation process according to an embodiment of the present application is provided;

[0054] Figure 4 A schematic diagram of a fault diagnosis model based on a recurrent image group according to an embodiment of the present application is provided;

[0055] Figure 5 (a) A schematic diagram of an experimental platform according to an embodiment of the present application is provided;

[0056] Figure 5 (b) is a schematic diagram of turn-to-turn short circuit for the embodiment of the present application;

[0057] Figure 5 (c) is a schematic diagram of bearing inner ring fault for the embodiment of the present application;

[0058] Figure 5 (d) is a schematic diagram of bearing outer ring fault for the embodiment of the present application;

[0059] Figure 6 (a) is a scatter plot diagram of fault diagnosis result for the embodiment of the present application;

[0060] Figure 6 (b) is a schematic diagram of confusion matrix of training set for the embodiment of the present application;

[0061] Figure 6 (c) is a schematic diagram of confusion matrix of test set for the embodiment of the present application;

[0062] Figure 7 (a) is a schematic diagram of training process for the embodiment of the present application;

[0063] Figure 8 (b) is a schematic diagram of a fault diagnosis device for a permanent magnet synchronous motor provided by the embodiment of the present application;

[0064] Figure 9 (c) is a schematic diagram of a hardware structure of the fault diagnosis device for the permanent magnet synchronous motor. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be described in detail below with reference to the drawings and specific embodiments.

[0066] In the specific embodiments, various specific technical features in each of the various embodiments can be combined in various ways without contradiction, for example, different specific technical features can form different embodiments through combination. In order to avoid unnecessary repetition, various possible combinations of each specific technical feature in the present application are not described again.

[0067] It should be noted that, in order to avoid the present application being obscured by unnecessary details, only the structures and / or processing steps closely related to the scheme of the present application are shown in the drawings, and other details not closely related to the present application are omitted.

[0068] In addition, it needs to be explained that the term "including", "containing" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, method, article or equipment. In the following description, the terms "first", "second", "..." are only to distinguish different objects, and do not mean that the objects have the same or relationship. It should be understood that the orientation described by the orientation terms "above", "below", "inner", "outer" and the like is the orientation in the normal use state.

[0069] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the specific technical scheme of the application will be further described in detail below with reference to the drawings in the embodiments of the present application. The following embodiments are used to illustrate the present application, but not to limit the scope of the present application.

[0070] Permanent magnet synchronous motor (PMSM) is widely used in various industries as a core device in automation systems due to its high power density, low noise, high efficiency, and strong stability. With the wide application of PMSM in various fields, its health status is facing severe challenges from multiple factors such as working environment, operating state, and frequency converter state. The combined effect of these factors poses a potential threat to the stable operation of PMSM. Therefore, monitoring the health status of PMSM, identifying fault conditions and determining fault types in a timely manner is of great importance in extending the life of the motor, improving the utilization rate of the motor throughout its life cycle, and maintaining the stable operation of the system.

[0071] Common faults of PMSM mainly include three categories: stator faults, permanent magnet faults and mechanical faults. Stator faults mainly manifest as stator winding faults, such as stator inter-turn short circuit. Permanent magnet faults mainly include permanent magnet demagnetization caused by armature reaction, high temperature, material aging, etc., mainly including uniform demagnetization and local demagnetization. Mechanical faults mainly include rotor eccentricity fault and bearing fault. Rotor eccentricity fault is that the degree of asymmetry of the air gap between the stator and the rotor exceeds the allowable operating range of the device. PMSM bearings include bearing inner ring, bearing outer ring, rolling body and retainer, and bearing fault is the failure of the above-mentioned bearing components.

[0072] Whether it is stator fault, permanent magnet fault or mechanical fault of the motor, it will cause changes in one or more signals such as vibration, current, sound, temperature during motor operation, so that sensors can be used to collect these signals to diagnose motor faults.

[0073] At present, one-dimensional time sequence signals are collected through sensors, and a series of signal processing algorithms are used to extract fault feature signals in the data. These methods have defects such as complex signal processing algorithm, high model precision requirement, and sensitivity to disturbance and noise. In recent years, with the continuous development of artificial intelligence technology, the application of artificial intelligence technology in PMSM fault diagnosis is in the ascendant.

[0074] The application will be further described in detail below in combination with the drawings and specific embodiments.

[0075] The application provides a permanent magnet synchronous motor fault diagnosis method, as shown in Figure 1 Figure 1 A flowchart of a permanent magnet synchronous motor fault diagnosis method provided by an embodiment of the application is shown. The method comprises the following steps:

[0076] Step S101: Obtain the magnetic field signal of the motor.

[0077] Step S102: Reconstruct the phase space of the motor based on the magnetic field signal.

[0078] Step S103: Construct a recurrent image group corresponding to the magnetic field signal according to the magnetic field signal and the phase space.

[0079] Step S104: Obtain a first data set based on the recurrent image group.

[0080] Step S105: Input the first data set into a diagnosis model of a preset multi-channel convolutional neural network to perform fault diagnosis and obtain a diagnosis result.

[0081] In this embodiment, the permanent magnet synchronous motor fault diagnosis method can be determined according to actual conditions, which is not limited herein. As an example, the permanent magnet synchronous motor fault diagnosis method can be a graphical permanent magnet synchronous motor intelligent fault diagnosis method.

[0082] In step S101, the magnetic field signal of the motor can be determined according to actual conditions, which is not limited herein. As an example, the magnetic field signal of the motor includes the magnetic field signal on the surface of the motor.

[0083] The specific acquisition process of the magnetic field signal of the motor can be determined according to actual conditions, which is not limited herein. As an example, the magnetic field signal of the motor can be obtained by a sensor to obtain the magnetic field signal on the surface of the motor.

[0084] For ease of understanding, an example is given. The length of the measurement of the magnetic field signal can be denoted as N. The motor surface magnetic field signal with a measurement length of N can be represented by the following formula (1): ​

[0085] x(n) = {x(l), x(2),... x(N)} (1);

[0086] In step S102, reconstructing the phase space of the motor based on the magnetic field signal can be understood as extending the one-dimensional magnetic field signal to a high-dimensional space.

[0087] The specific determination process of reconstructing the phase space of the motor based on the magnetic field signal can be determined according to actual conditions, which is not limited here. As an example, reconstructing the phase space of the motor based on the magnetic field signal can include: performing mean value processing on the magnetic field signal to obtain a time series; determining a target delay time corresponding to the magnetic field signal when reconstructing in the phase space based on the time series; obtaining a target embedding dimension corresponding to the magnetic field signal when reconstructing in the phase space; and reconstructing the phase space of the motor based on the target delay time and the target embedding dimension.

[0088] In step S103, the specific construction process of constructing the recursive image group corresponding to the magnetic field signal based on the magnetic field signal and the phase space can be determined according to actual conditions, which is not limited here. As an example, constructing the recursive image group corresponding to the magnetic field signal based on the magnetic field signal and the phase space can include:

[0089] Wavelet packet decomposition is performed on the magnetic field signal to obtain the length parameter of each sub-band in the frequency band generated by the magnetic field signal; a distance matrix of all point pairs in the phase space is determined based on the length parameter and the phase space; the distance matrix is normalized to obtain a normalized distance matrix; the normalized distance matrix is taken as the gray value of the recursive image group to obtain a recursive image and a corresponding resolution; and the recursive image and the corresponding resolution are transformed to obtain the recursive image group corresponding to the magnetic field signal.

[0090] In step S104, the specific acquisition process of obtaining the first data set based on the recursive image group can be determined according to actual conditions, which is not limited here. As an example, the first data set can include a plurality of feature patterns.

[0091] In step S105, the first data set is input into a preset multi-channel convolutional neural network diagnostic model for fault diagnosis to obtain a diagnosis result. The specific fault diagnosis process in the diagnosis result can be determined according to actual conditions, which is not limited herein. As an example, the first data set includes a plurality of feature patterns; each feature pattern represents the characteristics of the magnetic field signal of the motor at different frequency scales; the convolution kernels used on each channel of the multi-channel convolutional neural network diagnostic model are independent, and the weights of the results output by each channel are different when integrated in the full connection layer; the first data set is input into the preset multi-channel convolutional neural network diagnostic model for fault diagnosis to obtain a diagnosis result, which can include: inputting the plurality of feature patterns into the preset multi-channel convolutional neural network diagnostic model for fault diagnosis to obtain a diagnosis result; the diagnosis result includes winding fault, bearing inner ring fault, bearing outer ring fault, local demagnetization fault, uniform demagnetization fault, and rotor imbalance fault of the motor.

[0092] The embodiment of the present application provides a fault diagnosis method of a permanent magnet synchronous motor, determines a selection strategy of a preset ant colony algorithm through acquired environmental information around a UAV; constructs a path according to the selection strategy to obtain a path planning strategy of the UAV; detects obstacles around the UAV based on the path planning strategy to obtain a detection result for realizing UAV obstacle avoidance. That is, for the dynamic obstacle avoidance problem of the UAV in a complex environment, the improved ant colony algorithm is used for real-time path planning, the dynamic obstacles and the complex terrain are effectively processed, and the navigation efficiency of the UAV is improved.

[0093] In some embodiments, the reconstructing the phase space of the motor based on the magnetic field signal comprises:

[0094] The magnetic field signal is subjected to mean removal processing to obtain a time series;

[0095] The target delay time corresponding to the reconstruction of the phase space of the magnetic field signal is determined based on the time series;

[0096] The target embedding dimension corresponding to the reconstruction of the phase space of the magnetic field signal is acquired;

[0097] The phase space of the motor is reconstructed based on the target delay time and the target embedding dimension.

[0098] In the embodiment, the time series can be denoted as x ′ (n).

[0099] The specific processing process of the magnetic field signal can be determined according to actual conditions, and is not limited herein. As an example, the magnetic field signal can be processed by a preset algorithm to obtain a time sequence. Wherein, the preset algorithm can be determined according to actual conditions, and is not limited herein. As an example, the preset algorithm can refer to the following formula (2):

[0100]

[0101] The specific determination process of the target delay time corresponding to the magnetic field signal in the phase space reconstruction can be determined according to actual conditions, and is not limited herein. As an example, the target delay time corresponding to the magnetic field signal in the phase space reconstruction can include obtaining a given delay time; determining the autocorrelation function of the magnetic field signal at different delay times based on the time sequence and the given delay time; normalizing the autocorrelation function to obtain a normalized autocorrelation function; determining the variance of the original time sequence of the motor; determining the target delay time based on the normalized autocorrelation function and the variance of the original time sequence. Wherein, the target delay time can be understood as a suitable delay time, which can be denoted as τ op , which can also be denoted as τ.

[0102] The specific acquisition process of the target embedding dimension corresponding to the magnetic field signal in the phase space reconstruction can be determined according to actual conditions, and is not limited herein. As an example, the target embedding dimension corresponding to the magnetic field signal in the phase space reconstruction can include obtaining a given embedding dimension and a first proportion parameter corresponding to a first false neighbor point; determining a target neighbor point in the phase space based on the given embedding dimension; increasing the given embedding dimension by a preset dimension to construct a new reconstructed phase space of the motor; determining a second false neighbor point of the new reconstructed phase space according to the target neighbor point; determining a second proportion parameter of the second false neighbor point based on the first proportion parameter; determining the target embedding dimension when the value of the second proportion parameter is less than a preset threshold. Wherein, the target embedding dimension can be understood as a suitable minimum embedding dimension, which can be denoted as m.

[0103] The specific process of reconstructing the phase space of the motor based on the target delay time and the target embedding dimension can be determined according to actual conditions, which is not limited herein. As an example, the reconstructing the phase space of the motor based on the target delay time and the target embedding dimension can be reconstructing the phase space of the motor based on the target delay time and the target embedding dimension by a preset algorithm; wherein the preset algorithm can refer to the following formula (3):

[0104]

[0105] In formula (3), τ is the target delay time; m is the target embedding dimension; X(1), X(2), …, X(m-(m-1)τ) are points in the reconstructed m-dimensional phase space, respectively.

[0106] In some embodiments, the determining the target delay time corresponding to the phase space reconstruction of the magnetic field signal based on the time series comprises:

[0107] obtaining a preset given delay time;

[0108] determining the autocorrelation function of the magnetic field signal at different delay times based on the time series and the given delay time;

[0109] performing normalization processing on the autocorrelation function to obtain a normalized autocorrelation function;

[0110] determining the variance of the original time series of the motor;

[0111] determining the target delay time based on the normalized autocorrelation function and the variance of the original time series.

[0112] In this embodiment, the preset given delay time can be denoted as τ'; the autocorrelation function of the magnetic field signal at different delay times can be denoted as C xx (τ).

[0113] The specific process of determining the autocorrelation function of the magnetic field signal at different delay times based on the time series and the given delay time can be determined according to actual conditions, which is not limited herein. As an example, the determining the autocorrelation function of the magnetic field signal at different delay times based on the time series and the given delay time can be determining the autocorrelation function of the magnetic field signal at different delay times based on the time series and the given delay time by a preset algorithm. Wherein, the preset algorithm can be determined according to actual conditions, which is not limited herein. As an example, the preset algorithm can refer to the following formula (4):

[0114]

[0115] The specific processing process of normalizing the autocorrelation function to obtain the normalized autocorrelation function can be determined according to actual conditions, which is not limited here. As an example, the autocorrelation function can be normalized by a preset algorithm to obtain the normalized autocorrelation function. The preset algorithm can be determined according to actual conditions, which is not limited here. As an example, the preset algorithm can refer to the following formula (5):

[0116]

[0117] In formula (5), is the variance of the original time series, The calculation of can refer to the following formula (6):

[0118]

[0119] The specific determination process of determining the target delay time based on the normalized autocorrelation function can be determined according to actual conditions, which is not limited here. As an example, when the normalized autocorrelation function ρ xx (τ) drops to of the initial value, the corresponding delay time at this time is the appropriate delay time τ op .

[0120] In practical applications, the selection of delay time affects the independence of each coordinate component in the reconstructed phase space. If the delay time is too short, the information correlation between adjacent coordinate components is high, and the dynamic characteristics of the system cannot be effectively revealed; if the delay time is too long, it may lead to the loss of key information in the time series, reducing the quality of the reconstructed phase space. The present application determines the delay time by the autocorrelation method.

[0121] First, the collected motor surface magnetic field signal is de-meaned to obtain the time series x ′ (n), which can refer to the previous formula (2); then, the autocorrelation function result C xx (τ) at different delay times is calculated, which can refer to the previous formula (4); wherein τ is a given delay time. Finally, the normalized autocorrelation function result ρ xx (τ) is calculated, which can refer to the previous formula (5); wherein, is the variance of the original time series, and the calculation formula can refer to the previous formula (6); when the normalized autocorrelation function ρ xx (τ) drops to At this time, the corresponding delay time is the appropriate delay time τ op .

[0122] In some embodiments, the acquiring the target embedding dimension corresponding to the magnetic field signal in the phase space reconstruction comprises:

[0123] acquiring a given embedding dimension and a first proportion parameter of a corresponding first false nearest point;

[0124] determining a target nearest point in the phase space based on the given embedding dimension;

[0125] increasing the given embedding dimension by a preset dimension to construct a new reconstructed phase space of the motor;

[0126] determining a second false nearest point of the new reconstructed phase space according to the target nearest point;

[0127] determining a second proportion parameter of the second false nearest point based on the first proportion parameter;

[0128] if the value of the second proportion parameter is less than a preset threshold, determining the target embedding dimension.

[0129] In the embodiment, the given embedding dimension is denoted as m; the value of the given embedding dimension and the value of the first proportion parameter of the corresponding first false nearest point are determined according to actual conditions, which are not limited herein. As an example, the value of the given embedding dimension can be understood as a smaller value of the embedding dimension, for example, when m = 1, the value of the proportion parameter of the false nearest point is 1.

[0130] The target nearest point can be understood as the nearest point; the specific determination process of determining the target nearest point in the phase space based on the given embedding dimension can be determined according to actual conditions, which are not limited herein. As an example, the distance between each point and other points in the m-dimensional phase space constructed according to formula (7) is calculated to find the nearest point. The formula (7) can be referred to as follows:

[0131] R m (i,j) = ‖X(i) ―X(j)‖, (j≠i) (7);

[0132] In formula (7), R m (i,j) is the distance between the i th point and the j th point in the m-dimensional phase space, and X(i), X(j) is the point in the m-dimensional phase space.

[0133] The preset dimension can be determined according to actual conditions, and is not limited herein. As an example, the preset dimension can be 1. Increasing the given embedding dimension by the preset dimension to construct the new reconstructed phase space of the motor can be increasing the embedding dimension by 1 dimension, to m+1 dimension, to construct a new phase space.

[0134] The specific determination process of the second false neighboring point in the new reconstructed phase space according to the target neighboring point can be determined according to actual conditions, and is not limited herein. As an example, the distance R m+1 (i,j) of the corresponding points in the m+1 dimensional space can be calculated according to the following formula (8):

[0135]

[0136] The neighboring point is considered to be a second false neighboring point. In this application, r tol is 0.4.

[0137] The specific determination process of the second false neighboring point in the new reconstructed phase space according to the target neighboring point can be determined according to actual conditions, and is not limited herein. As an example, the distance R

[0138] The preset threshold value can be determined according to actual conditions, and is not limited herein. As an example, the preset threshold value can be 0.15.

[0139] The specific determination process of the second false neighboring point in the new reconstructed phase space according to the target neighboring point can be determined according to actual conditions, and is not limited herein. As an example, when the ratio is less than a certain threshold value (for example, 0.15), the corresponding m is the appropriate minimum embedding dimension.

[0140] In practical applications, a small embedding dimension m is first given. When m=1, the ratio of the false neighboring point is 1. The distance between each point and other points in the m-dimensional phase space constructed according to the above formula (7) is calculated to find its nearest neighboring point, wherein R m (i,j) is the distance between the i-th point and the j-th point in the m-dimensional phase space, and X(i), X(j) is the point in the m-dimensional phase space. The embedding dimension is increased by 1 dimension, to m+1 dimension, to construct a new phase space. The distance R m+1 (i,j) of the corresponding points in the m+1 dimensional space is calculated again. If the above formula (8) is satisfied, the neighboring point is considered to be a false neighboring point. Wherein, r tolTake 0.4. Calculate the proportion of pseudo-closeness points in the m-dimensional phase space, gradually increase m, and repeat the above steps. When the proportion is less than a certain threshold (0.15 in this application), the corresponding m is the appropriate minimum embedding dimension.

[0141] In some embodiments, the constructing the recurrent image group corresponding to the magnetic field signal according to the magnetic field signal and the phase space comprises:

[0142] Wavelet packet decomposition is performed on the magnetic field signal to obtain a length parameter of each sub-band in a frequency band generated by the magnetic field signal;

[0143] A distance matrix of all point pairs in the phase space is determined based on the length parameter and the phase space;

[0144] The distance matrix is normalized to obtain a normalized distance matrix;

[0145] The normalized distance matrix is taken as a gray value of the recurrent image group to obtain a recurrent image and a corresponding resolution;

[0146] The recurrent image and the corresponding resolution are transformed to obtain the recurrent image group corresponding to the magnetic field signal.

[0147] It should be noted that wavelet packet decomposition (WPD) is an extended form of wavelet transform, which breaks through the limitation of traditional wavelet decomposition that only low-frequency components are recursively decomposed, and allows multi-level decomposition of high-frequency and low-frequency components at the same time, thereby realizing more fine signal analysis in the time-frequency domain. The core idea is to find the best balance between time-frequency resolution by adaptively selecting the optimal sub-band decomposition method.

[0148] The calculation formula of wavelet packet decomposition can refer to the following formula (9):

[0149]

[0150] In formula (9), and are the approximation coefficients and the detail coefficients of the jth layer, respectively.

[0151] The WPD algorithm realizes the global decomposition of the signal, and its decomposition process presents a binary tree structure. In each layer of decomposition, the algorithm synchronously decomposes the low-frequency coefficients and the high-frequency coefficients of the previous layer, thereby realizing the extraction of the information characteristics of the original signal in multiple frequency scales.

[0152] The magnetic field signal and the length parameter can be determined according to actual conditions, which are not limited herein. As an example, the magnetic field signal can include a magnetic field signal time sequence with a length of N; and the length parameter can be denoted as L J .

[0153] The specific decomposition process of the length parameter of each sub-band in the frequency band generated by the magnetic field signal through the wavelet packet decomposition of the magnetic field signal can be determined according to actual conditions, which is not limited herein. As an example, for a magnetic field signal time sequence with a length of N, after J-layer wavelet packet decomposition, the number of sub-bands is 2J, that is, 2J frequency bands are generated, and when downsampling is used and the length of the wavelet basis itself is ignored, the length of each sub-band is L J , which can be referred to as the following formula (10):

[0154]

[0155] The distance matrix of all point pairs in the phase space can be denoted as R m (i,j); and the specific determination process of the distance matrix of all point pairs in the phase space based on the length parameter and the phase space can be determined according to actual conditions, which is not limited herein. As an example, for a sub-band with a length of L J , after reconstruction operation with embedding dimension m and delay time τ op , a new phase space can be reconstructed, and the distance matrix R m (i,j) of all point pairs in the phase space can be obtained.

[0156] The specific processing process of the distance matrix through the normalization processing to obtain the normalized distance matrix can be determined according to actual conditions, which is not limited herein. As an example, the normalization processing of the distance matrix to obtain the normalized distance matrix can be normalization processing of the distance matrix through a preset algorithm to obtain the normalized distance matrix; and the preset algorithm can be determined according to actual conditions, which is not limited herein. As an example, the preset algorithm can be referred to as the following formula (11):

[0157]

[0158] The distance matrix is normalized to [0, 1] according to formula (11), wherein max(R) and min(R) are the maximum and minimum values of the distance matrix, respectively, is the distance between the i-th point and the j-th point in the normalized m-dimensional phase space.

[0159] The normalized distance matrix is taken as the gray value of the recursive image group, and the recursive image and the corresponding resolution can be obtained by directly mapping the distance value to the color intensity and drawing the recursive graph.

[0160]

[0161] The specific process of the transformation processing of the recursive image and the corresponding resolution to obtain the recursive image group corresponding to the magnetic field signal can be determined according to actual conditions, which is not limited here. As an example, the recursive feature map always has a main diagonal line, that is, The recursive graph is symmetrical about the main diagonal line. Therefore, in order to reduce the amount of calculation, the half of the graph of the 2 J Recursive images are directly removed along the diagonal line, at this time the recursive image becomes a set of isosceles right triangles, and the recursive image at this time is cut along the midpoint of the right angle side, each image can obtain a small isosceles right triangle and a quadrilateral after processing, and a set of rectangular graphs can be obtained by rotating the small triangle along the point A of each image by 180 degrees clockwise, and the resolution of each graph can be referred to as shown in the following formula (13):

[0162]

[0163] After a series of steps such as wavelet packet decomposition, phase space reconstruction, recursive image production and image transformation, a series of recursive image groups with the same resolution are generated, each of which represents the information of a frequency band, and the recursive image groups are referred to as recursive image groups.

[0164] When downsampling is used, with the increase of the wavelet packet decomposition layer number, the length of the generated approximation coefficient and detail coefficient sequence is shortened, and satisfies the following formula (14):

[0165]

[0166] Therefore, the recursive image group realizes the graphical representation of the motor health status on multiple scales by decomposing the original signal into a series of recursive images of different frequency bands, and effectively solves the problem of insufficient long-time sequence representation of traditional recursive graphs.

[0167] As can be seen from formula (13), with the continuous increase of the decomposition layer number, the pixels of each image in the recursive image group are continuously reduced, but the total number of pixel points of all recursive images of each layer remains basically unchanged, so the change of the wavelet packet decomposition layer number has little effect on the calculation amount of the fault diagnosis algorithm based on the recursive image group image.

[0168] In some embodiments, the method further comprises:

[0169] obtaining a data set and corresponding labels of each type of motor fault;

[0170] determining a diagnostic model of the multi-channel convolutional neural network based on the data set and corresponding labels of each type of motor fault.

[0171] In this embodiment, the data set and corresponding labels of each type of motor fault can be determined according to actual conditions, which are not limited herein. As an example, the data set of each type of motor fault can include 500 groups; the corresponding labels can be 1, 2, 3, 4, 5, and 6. Among them, 1 can represent winding fault; 1 can represent winding fault; 2 can represent bearing inner ring fault; 3 can represent bearing outer ring fault; 4 can represent local demagnetization fault; 5 can represent uniform demagnetization fault; and 6 can represent rotor imbalance.

[0172] The specific determination process in the determination of the diagnostic model of the multi-channel convolutional neural network based on the data set and corresponding labels of each type of motor fault can be determined according to actual conditions, which are not limited herein. As an example, the determination of the diagnostic model of the multi-channel convolutional neural network based on the data set and corresponding labels of each type of motor fault can include: dividing the data set and corresponding labels of each type of motor fault into a training set and a test set; determining the diagnostic model of the multi-channel convolutional neural network based on the training set and the test set.

[0173] In some embodiments, the determination of the diagnostic model of the multi-channel convolutional neural network based on the data set and corresponding labels of each type of motor fault comprises:

[0174] dividing the data set and corresponding labels of each type of motor fault into a training set and a test set;

[0175] determining the diagnostic model of the multi-channel convolutional neural network based on the training set and the test set.

[0176] In this embodiment, the training set and the test set can be determined according to actual conditions, which are not limited herein. As an example, the data set is 500 groups; the training set is 400 groups; and the test set is 100 groups.

[0177] In practical applications, 3500 groups of data sets of various motor health conditions were established, each group of data being 0.2s of data (i.e. 1000 sample point data), 500 groups of data of each type of motor fault, 400 groups of data were selected as a training set and 100 groups of data were selected as a test set from each type of fault data set by random selection, and the number of motor labels and data sets was established.

[0178] In some embodiments, the first data set includes a plurality of feature graphs; each of the feature graphs represents a feature of a magnetic field signal of the motor at different frequency scales; the convolution kernels used on each channel of the multi-channel convolutional neural network are independent, and the weights of the results output by each channel are different when integrated in the full connection layer; the first data set is input into a preset diagnostic model of a multi-channel convolutional neural network for fault diagnosis to obtain a diagnostic result, which includes:

[0179] The plurality of feature graphs are input into a preset diagnostic model of a multi-channel convolutional neural network for fault diagnosis to obtain a diagnostic result; the diagnostic result includes winding fault, inner ring bearing fault, outer ring bearing fault, local demagnetization fault, uniform demagnetization fault, and rotor imbalance fault of the motor.

[0180] In this embodiment, the plurality of feature graphs can be determined according to actual conditions, which are not limited herein. As an example, the plurality of feature graphs can be 2J feature graphs.

[0181] Each of the feature graphs representing a feature of a magnetic field signal of the motor at different frequency scales can be understood as each of the feature graphs as input being a representation of the original motor surface magnetic field signal at different frequency scales.

[0182] The convolution kernels used on each channel of the diagnostic model of the multi-channel convolutional neural network are independent, and the weights of the results output by each channel are different when integrated in the full connection layer, which can be understood as in the fault diagnosis process, since the convolution kernels used on each channel are random and independent of each other, the weights of the results output by each channel are different when integrated in the full connection layer.

[0183] The specific diagnosis process in the diagnosis result obtained by inputting the plurality of feature patterns into the preset multi-channel convolutional neural network diagnosis model for fault diagnosis can be determined according to actual conditions, which is not limited herein. As an example, the diagnosis result obtained by inputting the plurality of feature patterns into the preset multi-channel convolutional neural network diagnosis model for fault diagnosis can include winding fault, bearing inner ring fault, bearing outer ring fault, local demagnetization fault, uniform demagnetization fault, and rotor imbalance fault of the motor.

[0184] In practical applications, CNN is particularly suitable for image-based fault classification methods due to its advantages of feature extraction, weight sharing, and translation invariance. The present application proposes a PMSM fault diagnosis model based on recursive image groups and CNN. A typical CNN structure consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. After the operation steps, the original motor surface magnetic field signal generates 2 J feature patterns, so the CNN fault diagnosis model contains 2 J inputs. Each original signal corresponds to a fault, so the fault diagnosis model has only one output result, and the feature information is synthesized through the fully connected layer. Each input of the fault diagnosis method proposed in the present application is a representation of the original motor surface magnetic field signal at different frequency scales, and in the fault diagnosis process, the convolution kernel used in each channel is random and independent of each other, so the weight of the result output by each channel is different when synthesized in the fully connected layer. This is one of the advantages of the CNN fault diagnosis model based on recursive image groups, that is, it can realize the independent setting of the weight of the original signal at each scale in the fault diagnosis process.

[0185] At the same time, since sampling noise, environmental noise, etc. are mainly concentrated in the high-frequency part, noise is mainly concentrated in one or a few images in the recursive image group. Due to the independence of the weight, the method proposed in the present application has excellent anti-noise performance in theory.

[0186] The present application first decomposes the original one-dimensional time series motor magnetic field signal into multiple scales through wavelet packet decomposition, then generates a threshold-free recursive image group using the decomposed wavelet coefficients, and finally sends the recursive image group after a certain rule of cutting, flipping transformation into the corresponding neural network diagnosis model for training and diagnosis. The experimental results show that this method has high diagnosis accuracy, small computational complexity, strong anti-noise ability, strong adaptability to multiple working conditions, and can realize independent setting of multiple scales without complex algorithms.

[0187] In practical applications, as an example, the fault diagnosis method of the permanent magnet synchronous motor can be a graphical permanent magnet synchronous motor intelligent fault diagnosis method. The technical solutions of the present application are implemented as follows:

[0188] Suppose the measured length of the motor surface magnetic field signal is N, refer to the above formula (1).

[0189] Set the delay time as τ and the embedding dimension as m, and the reconstructed phase space shown in the above formula (3) can be obtained. Wherein: X(1), X(2), …, X(N-(m-1)τ) are respectively points in the reconstructed m-dimensional phase space.

[0190] The selection of the delay time τ and the embedding dimension m has a very important influence on the reconstructed phase space.

[0191] The selection of the delay time affects the independence of each coordinate component in the reconstructed phase space. If the delay time is too short, the information correlation between adjacent coordinate components is high, and the dynamic characteristics of the system cannot be effectively revealed; if the delay time is too long, the key information in the time series may be lost, reducing the quality of the reconstructed phase space. The present application determines the delay time by the autocorrelation method.

[0192] Firstly, the collected motor surface magnetic field signal is subjected to mean removal processing to obtain a time series x ′ (n), which can refer to the above formula (2).

[0193] Then, the autocorrelation function result C xx (τ) at different delay times is calculated, as shown in the above formula (4). Wherein, τ is the given delay time.

[0194] Finally, the normalized autocorrelation function result is calculated, as shown in the above formula (5).

[0195] Wherein, is the variance of the original time series, and the calculation formula is shown in the above formula (6).

[0196] When the normalized autocorrelation function drops to the initial value, the corresponding delay time at this time is the appropriate delay time.

[0197] The embedding dimension m determines the dimension of the reconstructed phase space. If the embedding dimension is too low, the dynamic characteristics of the system may not be completely restored, resulting in information loss. On the contrary, if the embedding dimension is too high, although the complexity of the system can be better revealed, the calculation cost and data demand will be increased. The present application determines the embedding dimension by the false nearest neighbor method.

[0198] First, a small embedding dimension is given, (the proportion of false nearest neighbors is 1 when m = 1), the distance between each point and other points in the constructed phase space is calculated according to the above formula (7), and the nearest neighbor is found, wherein, R m (i,j) is the distance between the i th point and the j th point in the m-dimensional phase space, X(i), X(j) is the point in the m-dimensional phase space.

[0199] The embedding dimension is increased by 1 dimension to m+1 dimension, a new phase space is constructed, and the distance of the corresponding points in the m+1-dimensional space is calculated again, if the above formula (8) is satisfied, it is considered as a nearest neighbor. In the application, r tol 0.4 is taken.

[0200] The proportion of false nearest neighbors in all points in the m-dimensional phase space is calculated, and the above steps are repeated by gradually increasing m. When the proportion is less than a certain threshold (0.15 is taken in the application), the corresponding m is the appropriate minimum embedding dimension.

[0201] Wavelet packet decomposition (WPD) is an extended form of wavelet transform, which breaks through the limitation of traditional wavelet decomposition that only low-frequency components are recursively decomposed, and allows high-frequency and low-frequency components to be simultaneously decomposed at multiple levels, thereby realizing more fine signal analysis in time-frequency domain. The core idea is to find the best balance between time-frequency resolution by adaptively selecting the optimal sub-band decomposition method.

[0202] The calculation formula of wavelet packet decomposition can refer to the above formula (9).

[0203] Among them, and are the approximation coefficients and detail coefficients of the j th layer respectively.

[0204] The WPD algorithm realizes the global decomposition of the signal, and the decomposition process presents a binary tree structure. In each layer of decomposition, the algorithm synchronously decomposes the low-frequency coefficients and high-frequency coefficients of the previous layer, thereby realizing the extraction of information features of the original signal at multiple frequency scales.

[0205] For the convenience of understanding, the content of the application can be understood in combination with Figure 2 , Figure 2 is a schematic diagram of the PMSM fault diagnosis technology route based on recursive image group of the embodiment of the application.

[0206] For a magnetic field signal time sequence with a length of N, after J-layer wavelet packet decomposition, the number of sub-bands is 2 J , that is, 2 J frequency bands are generated, and when downsampling is used and the length of the wavelet basis itself is ignored, the length of each sub-band is L J , then refer to the above formula (10).

[0207] From the previous analysis, we can see that the length is L J The subband is embedded with dimension m and delay time τ op After the reconstruction operation, it can be reconstructed into a new phase space, and the distance matrix R of all point pairs in the phase space is obtained. m (i,j).

[0208] According to the previous formula (11), the distance matrix is ​​normalized to [0,1], where max(R) and min(R) are the maximum and minimum values ​​of the distance matrix, respectively. is the distance between the i-th point and the j-th point in the normalized m-dimensional phase space.

[0209] The normalized distance matrix is ​​used as the grayscale value of the recursive graph, and the distance value is directly mapped to the color intensity. The recursive graph is drawn to obtain 2 J recursive images, the resolution of each image can refer to the previous formula (12).

[0210] From the above analysis, we can see that the recursive feature map always has a main diagonal, that is The recursive graph is symmetrical about the main diagonal. Therefore, to reduce the amount of calculation, 2 J Each recursive image removes half of the image along the diagonal line. At this time, the recursive image becomes a set of isosceles right triangles. The recursive image is cut along the midpoint of the right angle. After processing, each image can obtain a small isosceles right triangle and a quadrilateral. Then, the cut small triangles are rotated 180 degrees clockwise along the point A of each image to obtain a set of rectangular figures. The resolution of each figure is shown in the previous formula (13). The transformation process is as follows: Figure 3 shown. Figure 3 Schematic diagram of the recursive image group generation process in an embodiment of the present invention.

[0211] like Figure 2 and Figure 3 As shown, the original one-dimensional signal undergoes a series of steps such as wavelet packet decomposition, phase space reconstruction, recursive image production, and image transformation to produce a series of recursive image groups with the same resolution. Each recursive image represents the information of a frequency band. This application refers to these recursive image groups as recursive image groups.

[0212] When downsampling is used, as the number of wavelet packet decomposition layers increases, the length response of the generated approximate coefficient and detail coefficient sequences shortens and satisfies the previous formula (14).

[0213] Therefore, the recursive image group realizes the graphical representation of the motor health status on multiple scales by decomposing the original signal into a series of recursive images of different frequency bands, and effectively solves the problem of insufficient long-time sequence representation of the traditional recursive image.

[0214] As can be seen from the above formula (13), with the continuous increase of the decomposition layer number, the pixels of each image in the recursive image group are continuously reduced, but the total number of pixel points of all recursive images of each layer remains basically unchanged, so the change of the wavelet packet decomposition layer number has little effect on the calculation amount of the fault diagnosis algorithm based on the recursive image group image.

[0215] The CNN is particularly suitable for image-based fault classification methods due to its advantages of feature extraction, weight sharing, translation invariance, etc. The application research proposes a PMSM fault diagnosis model based on recursive image group and CNN.

[0216] A typical CNN structure is composed of an input layer, a convolution layer, a pooling layer, a full connection layer, and an output layer.

[0217] In the present application, the motor original surface magnetic field signal generates 2 J feature patterns after the above operation steps, so the CNN fault diagnosis model contains 2 J inputs. Each original signal corresponds to a fault, so the fault diagnosis model has only one output result, and the application comprehensively processes the feature information through the full connection layer. The complete fault diagnosis model is shown in Figure 4 . Figure 4 FIG. 1 is a schematic diagram of the fault diagnosis model based on the recursive image group of the embodiment of the application.

[0218] As can be seen from the above analysis, each input of the fault diagnosis method proposed in the application is a representation of the original motor surface magnetic field signal on different frequency scales, and since the convolution kernel used on each channel is random and independent of each other in the fault diagnosis process, the weight of each channel output is different when the full connection layer is synthesized. This is one of the advantages of the CNN fault diagnosis model based on the recursive image group, that is, it can realize the independent setting of the weight of the original signal on each scale in the fault diagnosis process.

[0219] At the same time, since the sampling noise, environmental noise, etc. are mainly concentrated in the high frequency part, the noise is mainly concentrated in one or a few images in the recursive image group. Due to the independence of the weight, the method proposed in the application has excellent anti-noise performance in theory.

[0220] As an example, the motor used in the present application is a 5.5kW four-pole three-phase embedded permanent magnet synchronous motor, and the specific motor and experimental parameters are shown in Table 1. The experimental platform mainly consists of five parts: 1) PMSM, 2) torque sensor, 3) frequency converter, 4) magnetic field sensor, 5) brake and brake controller, as shown in Figure 5 (a), 5(b), 5(c), 5(d). Figure 5 (a) is a schematic diagram of the experimental platform of the embodiment of the present application; Figure 5 (b) is a schematic diagram of the inter-turn short circuit of the embodiment of the present application; Figure 5 (c) is a schematic diagram of the bearing inner ring fault of the embodiment of the present application;

[0221] Figure 5 (d) is a schematic diagram of the bearing outer ring fault of the embodiment of the present application. Table 1 shows the motor and test parameters.

[0222] Table 1

[0223]

[0224]

[0225] The present application collects 7 types of motor surface magnetic field data in different states, which are normal motor, stator winding fault motor, bearing inner ring fault motor, bearing outer ring fault motor, local demagnetization motor, uniform demagnetization motor and rotor imbalance motor. The fault setting method is as follows:

[0226] Stator winding fault motor: the A-phase tap of the motor winding is short-circuited to make the ABC three-phase winding unbalanced, and the number of turns of the A-phase winding is 3 turns less than that of the other two phases, and the normal phase winding is 46 turns.

[0227] Bearing inner / outer ring fault motor: the motor bearing mainly consists of bearing inner ring, bearing outer ring, rolling body and retainer. The inner / outer ring fault is a 1mm wide and 1.5mm deep straight groove cut on the inner / outer ring of the motor bearing.

[0228] Local demagnetization motor: the present application uses a motor with one permanent magnet removed and demagnetized at high temperature to simulate the local demagnetization fault of the motor rotor.

[0229] Uniform demagnetization motor: the present application uses a method of replacing all the original motor permanent magnets with slightly weaker permanent magnets to simulate the uniform demagnetization fault.

[0230] Rotor imbalance motor: the present application adds weight pads to the rotor after the motor rotor is manufactured, so that the balance value on both sides is above ±15g.

[0231] The AC magnetic field measurement sensor used in the motor surface magnetic field measurement of the present application is AMS-2K. First, the motor surface magnetic field signals of motors in different health states under no-load and rated speed conditions are collected, the collection position is about 1 cm away from the side of the motor, parallel to the axis direction, and the sampling rate is 2 kHz. Then, the collected surface magnetic field signals are subjected to wavelet packet decomposition and recursive image group generation to establish a data set.

[0232] The present application establishes 3500 groups of data sets of various motor health conditions, 500 groups of data sets for each type of motor fault. 400 groups of data are selected from each type of fault data set as a training set and 100 groups of data are selected as a test set by random sampling. The number of motor labels and data sets is shown in Table 2. Table 2 is the number of data sets and labels of motors in different health states.

[0233] Table 2

[0234]

[0235]

[0236] The present application selects the wavelet basis function as dmey and the wavelet packet decomposition layer is 3 layers. A total of 8 sub-bands are decomposed. In the convolutional neural network part, 7 hidden layers are set, including 2 convolutional layers and 2 max pooling layers. The size of the convolution kernel is 5x5. The first convolutional layer contains 3 outputs, and the second convolutional layer contains 6 outputs. The network structure is shown in Figure 3 .

[0237] To reduce the randomness of the CNN fault diagnosis result, the experimental results of the present application are the average of 10 consecutive experiments.

[0238] Figure 6 (a) is a scatter diagram of the fault diagnosis result of the embodiment of the present application. From Figure 6 (a), it can be seen that there is no obvious overlap between different fault type samples of the method of the present application, and all fault samples have greater inter-class distance and clear separation boundary.

[0239] Figure 6 (b) is a schematic diagram of the confusion matrix of the training set of the embodiment of the present application. Figure 6 (c) is a schematic diagram of the confusion matrix of the test set of the embodiment of the present application. From Figure 6 (b) and Figure 6As can be seen from (c), the method proposed in the application can achieve 100% identification accuracy for all faults on the training set, and the overall identification accuracy on the test set is more than 98%. The results show that the method proposed in the application has good identification ability and diagnostic accuracy for the fault type of the PMSM.

[0240] Figure 7 Fig. 1 is a schematic diagram of the training process in the embodiment of the application, and Fig. 2 is a schematic diagram of the diagnosis process in the embodiment of the application. Figure 7 As can be seen from (c), the method proposed in the application has a faster convergence speed during training.

[0241] The application first proposes a permanent magnet synchronous motor fault diagnosis method based on recursive image group and multi-channel convolutional neural network. First, the one-dimensional time sequence magnetic field signal collected on the outer surface of the motor is subjected to multi-scale decomposition through wavelet packet decomposition; second, the wavelet coefficients after decomposition are combined with the phase space reconstruction method to generate a threshold-free recursive image group, and the recursive image group is transformed into a recursive image group image after being cut and flipped according to certain rules, and then the recursive image group image is sent into the corresponding convolutional neural network diagnosis model for training and diagnosis; finally, a PMSM experimental platform containing seven different health states of PMSM rotor demagnetization, bearing fault, rotor misalignment, etc. is built to verify the method proposed in the application. Including the following aspects:

[0242] 1) The concept of recursive image group is proposed by combining wavelet packet decomposition with recursive image, which converts one-dimensional time sequence signals collected from sensors into recursive image groups, realizing the graphical representation of motor health status at multiple scales;

[0243] 2) The recursive image group is based on wavelet packet decomposition, so as the number of wavelet packet decomposition layers increases, the lengths of the approximation coefficients and the detail coefficients gradually decrease, thereby solving the problem that traditional recursive images are difficult to represent long time sequence features;

[0244] 3) A multi-channel convolutional neural network fault diagnosis method based on recursive image group for permanent magnet synchronous motor is designed, which can realize independent setting of weights at multiple scales without combining various attention mechanisms, and the algorithm is simple. When the wavelet packet decomposition adopts down-sampling decomposition, the size of the algorithm is not limited by the number of wavelet packet decomposition layers;

[0245] 4) The combination of recursive image group image and fault diagnosis realizes the extraction of multi-scale fault features, and the feature extraction ability is stronger. At the same time, the method itself has strong anti-noise ability, reducing the problem of feature information loss caused by denoising algorithms;

[0246] 5) The recursive image group is to convert the acquisition sensor signal into a series of recursive images of different scales. In fault diagnosis, the weights of feature images of multiple scales are independent of each other. Therefore, in theory, as long as the sampling rate and sampling time are consistent, the method can realize the simultaneous input of multiple different sensor data without excessive adjustment, which is of great benefit to better distinguish different types of motor faults.

[0247] The application first performs multi-scale decomposition on the original one-dimensional time series motor magnetic field signal through wavelet packet decomposition, then generates a threshold-free recursive image group using the decomposed wavelet coefficients, and finally sends the recursive image group after a certain rule of shearing and flipping transformation into the corresponding neural network diagnosis model for training and diagnosis. The experimental results show that the method has high diagnostic accuracy, small computational complexity, strong noise resistance, strong adaptability to multiple working conditions, and can realize independent setting of multiple scales without complex algorithms.

[0248] Based on the same inventive concept as the foregoing, Figure 8 A structure diagram of a fault diagnosis device for a permanent magnet synchronous motor according to an embodiment of the application is shown in Figure 8 As shown, the device 800 includes:

[0249] A first acquisition unit 801 configured to acquire a magnetic field signal of the motor.

[0250] A reconstruction unit 802 configured to reconstruct a phase space of the motor based on the magnetic field signal.

[0251] A construction unit 803 configured to construct a recursive image group corresponding to the magnetic field signal according to the magnetic field signal and the phase space.

[0252] A second acquisition unit 804 configured to acquire a first data set based on the recursive image group.

[0253] A diagnosis unit 805 configured to input the first data set into a diagnosis model of a preset multi-channel convolutional neural network to perform fault diagnosis and obtain a diagnosis result.

[0254] In some embodiments, the reconstruction unit 802 is further configured to perform mean removal processing on the magnetic field signal to obtain a time series, determine a target delay time corresponding to the magnetic field signal when the phase space is reconstructed based on the time series, acquire a target embedding dimension corresponding to the magnetic field signal when the phase space is reconstructed, and reconstruct the phase space of the motor based on the target delay time and the target embedding dimension.

[0255] In some embodiments, the reconstruction unit 802 is further configured to obtain a preset given delay time; determine an autocorrelation function of the magnetic field signal at different delay times based on the time sequence and the given delay time; perform normalization processing on the autocorrelation function to obtain a normalized autocorrelation function; determine a variance of the original time sequence of the motor; and determine the target delay time based on the normalized autocorrelation function and the variance of the original time sequence.

[0256] In some embodiments, the reconstruction unit 802 is further configured to obtain a preset given embedding dimension and a first scale parameter corresponding to a first false neighbor point; determine a target neighbor point in the phase space based on the given embedding dimension; construct a new reconstructed phase space of the motor by increasing the given embedding dimension by a preset dimension; determine a second false neighbor point of the new reconstructed phase space according to the target neighbor point; determine a second scale parameter of the second false neighbor point based on the first scale parameter; and determine the target embedding dimension in a case where a value of the second scale parameter is less than a preset threshold.

[0257] In some embodiments, the construction unit 803 is further configured to perform wavelet packet decomposition on the magnetic field signal to obtain a length parameter of each subband in a frequency band generated by the magnetic field signal; determine a distance matrix of all pairs of points in the phase space based on the length parameter and the phase space; perform normalization processing on the distance matrix to obtain a normalized distance matrix; take the normalized distance matrix as a grayscale value of the recursive image group to obtain a recursive image and a corresponding resolution; and perform transformation processing on the recursive image and the corresponding resolution to obtain the recursive image group corresponding to the magnetic field signal.

[0258] In some embodiments, the second obtaining unit 804 is further configured to obtain a data set and a corresponding label of each type of motor fault.

[0259] The reconstruction unit 802 is further configured to determine a diagnostic model of the multi-channel convolutional neural network based on the data set and the corresponding label of each type of motor fault.

[0260] In some embodiments, the reconstruction unit 802 is further configured to divide the data set and the corresponding label of each type of motor fault into a training set and a test set; and determine a diagnostic model of the multi-channel convolutional neural network based on the training set and the test set.

[0261] In some embodiments, the first data set comprises a plurality of feature maps; each of the feature maps represents a feature of the magnetic field signal of the motor at a different frequency scale; the convolution kernels used in each channel of the diagnostic model of the multi-channel convolutional neural network are independent, and the weights of the results output by each channel are different when integrated in the full connection layer; the diagnostic unit 805 is further configured to input the plurality of feature maps into a preset diagnostic model of a multi-channel convolutional neural network to perform fault diagnosis and obtain a diagnostic result; the diagnostic result comprises winding fault, inner ring bearing fault, outer ring bearing fault, local demagnetization fault, uniform demagnetization fault, and rotor imbalance fault of the motor.

[0262] It should be noted that the fault diagnosis device of the permanent magnet synchronous motor provided by the embodiments of the present application and the fault diagnosis method of the permanent magnet synchronous motor provided by the foregoing embodiments of the present application belong to the same inventive concept, and the meanings of the words appearing here have been described in detail in the foregoing, which will not be described again here.

[0263] The embodiments of the present application also provide a storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the foregoing method embodiments, and the foregoing storage medium includes a mobile storage device, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media capable of storing program codes.

[0264] The embodiments of the present application also provide a fault diagnosis device of a permanent magnet synchronous motor, which comprises a processor and a memory for storing a computer program capable of running on the processor, wherein the processor is configured to execute the computer program to perform the steps of the foregoing method embodiments stored in the memory.

[0265] Figure 9 A hardware structure diagram of the fault diagnosis device of the permanent magnet synchronous motor according to the embodiments of the present application is shown in FIG. 9. The fault diagnosis device 900 of the permanent magnet synchronous motor comprises at least one processor 901 and a memory 902. Optionally, the fault diagnosis device 900 of the permanent magnet synchronous motor can further comprise at least one communication interface 903. The various components in the fault diagnosis device 900 of the permanent magnet synchronous motor are coupled together through a bus system 904, and it can be understood that the bus system 904 is used to realize the connection and communication between the components. The bus system 904 includes a data bus, a power supply bus, a control bus and a state signal bus. However, for the purpose of clear illustration, all the buses are marked as the bus system 904 in FIG. 9. Figure 9

[0266] ​It can be appreciated that the memory 902 can be a volatile memory or nonvolatile memory, and can also include both volatile and nonvolatile memory. Among them, the nonvolatile memory can be a Read Only Memory (ROM), a Programmable Read-Only Memory (PROM), an Erasable Programmable Read-Only Memory (EPROM), an Electrically Erasable Programmable Read-Only Memory (EEPROM), a Ferromagnetic Random Access Memory (FRAM), a Flash Memory, a magnetic surface memory, an optical disc, or a Compact Disc Read-Only Memory (CD-ROM); the magnetic surface memory can be a disk memory or a tape memory. The volatile memory can be a Random Access Memory (RAM) used as an external cache. By way of example and not limitation, many forms of RAM can be used, such as Static Random Access Memory (SRAM), Synchronous Static Random Access Memory (SSRAM), Dynamic Random Access Memory (DRAM), Synchronous Dynamic Random Access Memory (SDRAM), Double Data Rate Synchronous Dynamic Random Access Memory (DDR SDRAM), Enhanced Synchronous Dynamic Random Access Memory (ESDRAM), Sync Link Dynamic Random Access Memory (SLDRAM), and Direct Rambus Random Access Memory (DRRAM).The memory 902 described in the embodiments of the present application is intended to include, but is not limited to, these and any other suitable type of memory.

[0267] The memory 902 in the embodiments of the present application is used to store various types of data to support the operation of the fault diagnosis device 900 of the permanent magnet synchronous motor. Examples of these data include: any computer program for operating on the fault diagnosis device 900 of the permanent magnet synchronous motor, the program implementing the method of the embodiments of the present application can be contained in the memory 902.

[0268] The method disclosed in the above embodiments of the present application can be applied in the processor 901 or implemented by the processor 901. The processor can be an integrated circuit chip with processing capability. In the implementation process, each step of the above method can be completed by the integrated logic circuit of hardware in the processor or the instruction in the form of software. The above processor can be a general processor, a digital signal processor (DSP), or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, etc. The processor can implement or execute the disclosed methods, steps and logic block diagrams in the embodiments of the present application. The general processor can be a microprocessor or any conventional processor, etc. In combination with the steps of the method disclosed in the embodiments of the present application, the above-mentioned method can be directly embodied as a hardware decoding processor to execute, or be executed by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium, which is located in the memory, and the processor reads the information in the memory to complete the above-mentioned method steps in combination with the hardware thereof.

[0269] In the exemplary embodiments, the fault diagnosis device 900 of the permanent magnet synchronous motor can be implemented by one or more application specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers (MCUs), microprocessors (Microprocessors), or other electronic elements, for executing the above-mentioned method.

[0270] In several embodiments provided by the present application, it should be understood that the disclosed devices and methods can be implemented in other manners. The embodiments described above are merely exemplary, for example, the division of the units is only a logical function division, and there can be other division manners in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed coupling, or direct coupling or communication connection between the components can be indirect coupling or communication connection through some interfaces, devices or units, and can be electrical, mechanical or in other forms. The units described as separated components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e. can be located in one place, or distributed on a plurality of network units; some or all of the units can be selected according to actual needs to achieve the purposes of the embodiments. In addition, the units in the embodiments of the present application can be integrated in a processing module, or each unit can be a separate unit, or two or more units can be integrated in a unit; the integrated unit can be implemented in the form of hardware, or in the form of hardware plus software function unit.

[0271] The above description is merely illustrative of the preferred embodiments of the present application and is not intended to limit the scope of the application.

Claims

1. A method for diagnosing a fault of a permanent magnet synchronous motor, characterized in that: The method comprises: Acquiring a magnetic field signal of the motor; reconstructing a phase space of the motor based on the magnetic field signal; constructing a recursive image group corresponding to the magnetic field signal according to the magnetic field signal and the phase space; acquiring a first data set based on the recursive image group; The first data set is input into a preset multi-channel convolutional neural network diagnostic model for fault diagnosis to obtain a diagnostic result.

2. The method according to claim 1, characterized in that The reconstructing the phase space of the motor based on the magnetic field signal includes: performing de-meaning processing on the magnetic field signal to obtain a time series; Determining a target delay time corresponding to the magnetic field signal during phase space reconstruction based on the time series; Obtaining a target embedding dimension corresponding to the magnetic field signal during phase space reconstruction; A phase space of the motor is reconstructed based on the target delay time and the target embedding dimension.

3. The method according to claim 2, characterized in that The determining, based on the time series, a target delay time corresponding to the magnetic field signal during phase space reconstruction includes: Get the preset given delay time; determining the autocorrelation function of the magnetic field signal at different delay times based on the time series and the given delay time; performing normalization processing on the autocorrelation function to obtain a normalized autocorrelation function; The target delay time is determined based on the normalized autocorrelation function.

4. The method according to claim 2, characterized in that The obtaining of the target embedding dimension corresponding to the magnetic field signal during phase space reconstruction includes: Obtaining a preset given embedding dimension and a first scale parameter of a corresponding first pseudo neighboring point; determining a target neighboring point in the phase space based on the given embedding dimension; Adding a preset dimension to the given embedding dimension to construct a new reconstructed phase space of the motor; Determine a second pseudo adjacent point in the new reconstructed phase space according to the target adjacent point; determining a second scale parameter of the second pseudo adjacent point based on the first scale parameter; When the value of the second scale parameter is less than a preset threshold, the target embedding dimension is determined.

5. The method according to claim 1, characterized in that The constructing a recursive image group corresponding to the magnetic field signal according to the magnetic field signal and the phase space includes: performing wavelet packet decomposition on the magnetic field signal to obtain a length parameter of each sub-band in a frequency band generated by the magnetic field signal; determining a distance matrix for all pairs of points in the phase space based on the length parameter and the phase space; Normalizing the distance matrix to obtain a normalized distance matrix; Using the normalized distance matrix as the grayscale value of the recursive image group to obtain a recursive image and a corresponding resolution; The recursive images and the corresponding resolutions are transformed to obtain a recursive image group corresponding to the magnetic field signal.

6. The method according to claim 5, characterized in that The method further comprises: Obtain a dataset and corresponding labels for each type of motor fault; The diagnostic model of the multi-channel convolutional neural network is determined based on the data set and corresponding labels of each type of motor fault.

7. The method according to claim 6, characterized in that The method of determining the diagnostic model of the multi-channel convolutional neural network based on the data set and corresponding labels of each type of motor fault includes: Divide the data set of each type of motor fault and the corresponding label into a training set and a test set; A diagnostic model of the multi-channel convolutional neural network is determined based on the training set and the test set.

8. The method according to any one of claims 1 to 7, characterized in that The first data set includes a plurality of feature graphs; each feature graph represents the characteristics of the magnetic field signal of the motor at different frequency scales; the convolution kernel used on each channel in the diagnostic model of the multi-channel convolutional neural network is independent, and the output results of each channel have different weights when integrated in the fully connected layer; The step of inputting the first data set into a preset multi-channel convolutional neural network diagnostic model to perform fault diagnosis and obtain a diagnostic result includes: The multiple characteristic patterns are input into a preset multi-channel convolutional neural network diagnostic model for fault diagnosis to obtain a diagnostic result; the diagnostic result includes a winding fault, a bearing inner ring fault, a bearing outer ring fault, a local demagnetization fault, a uniform demagnetization fault, and a rotor imbalance fault of the motor.

9. A fault diagnosis device for a permanent magnet synchronous motor, characterized in that: The device comprises: a first acquiring unit, configured to acquire a magnetic field signal of the motor; a reconstruction unit, configured to reconstruct a phase space of the motor based on the magnetic field signal; a construction unit, configured to construct a recursive image group corresponding to the magnetic field signal according to the magnetic field signal and the phase space; a second acquiring unit, configured to acquire a first data set based on the recursive image group; The diagnostic unit is used to input the first data set into a preset multi-channel convolutional neural network diagnostic model to perform fault diagnosis and obtain a diagnostic result.

10. A fault diagnosis device for a permanent magnet synchronous motor, characterized in that: The device comprises: a processor and a memory for storing a computer program that can be run on the processor, wherein the processor executes the steps of the method according to any one of claims 1 to 8 when running the computer program.

Images