Improved empirical wavelet transform method and system for permanent magnet synchronous motor fault diagnosis

Abstract

The invention discloses a permanent magnet synchronous motor fault diagnosis method and system with improved empirical wavelet transform, comprising: using an improved Welch method to determine the power spectrum of a fault signal and dividing it into a plurality of continuous spectrum intervals; using Littlewood-Paley method and Meyer method to construct a wavelet filter group in each power spectrum interval, and extracting each modal component in each power spectrum interval; constructing a fault feature vector of the fault signal by using the energy moment of each modal component in each power spectrum interval; constructing a fault feature matrix by using the fault feature vector of each fault signal; using CatBoost method to establish a motor fault diagnosis model, and training the motor fault diagnosis model by using an improved training set and an improved gradient enhancement method; the trained motor fault diagnosis model performs motor fault diagnosis according to the fault feature matrix, and solves the modal aliasing problem caused by unclear signal spectrum boundaries during motor diagnosis.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis of electric power equipment, and in particular, relates to a permanent magnet synchronous motor fault diagnosis method based on improved empirical wavelet transform, which utilizes the improved empirical wavelet transform method to diagnose and analyze faults of permanent magnet synchronous motors under different working conditions. Background Art

[0002] Permanent magnet synchronous motor (PMSM) is widely used in manufacturing, transportation, and renewable energy power generation industries due to its high power factor and fast dynamic performance. Its working condition directly affects the performance of the entire system. Phase loss fault is a common electrical fault of PMSM, which is usually manifested as the phase winding is disconnected from the power supply, and an unbalanced current is generated in the remaining healthy phase, causing large electromagnetic torque pulsation. Failure to diagnose the phase loss category in time and take remedial measures may cause secondary damage to the entire system. Therefore, it is very important to carry out effective condition monitoring and fault diagnosis of permanent magnet synchronous motors.

[0003] In the prior art, motor fault diagnosis methods include: analytical model method, signal analysis method, and data-driven method. The purpose of the analytical model method is to optimize the system hardware by analyzing the redundant components of the system. This method requires the establishment of a detailed equipment model, including model construction, parameter estimation, and state estimation. However, in the actual environment, there are many factors that affect the operating state of the equipment and the correlation is complex. Especially for complex systems, the accuracy of this method is often low; the signal analysis method is to extract the characteristics of the measured signal in the time domain or frequency domain, and then compare the characteristics of the healthy system to make a diagnostic decision. It does not require an accurate physical model. The method is flexible and can be combined with different signal processing technologies. However, this method pays less attention to the dynamic input of the system, and the diagnostic performance will decrease under unknown input disturbances or unbalanced conditions; the data-driven method starts from a large amount of historical data and applies various artificial intelligence technologies to industrial processes. It is suitable for complex systems and has good processing capabilities for various fault types and nonlinear systems. However, this method requires a large amount of high-quality labeled data. Insufficient or unbalanced data will reduce the effect and affect the accuracy and reliability of the model. Moreover, the model is usually a "black box" and it is difficult to explain its diagnostic results. The three methods have certain defects in health monitoring and fault diagnosis for PMSM systems alone. In order to give full play to the advantages of various fault diagnosis methods and make up for their respective shortcomings, hybrid diagnosis methods are usually used in practical engineering applications to reasonably process the input signal to obtain a signal with clear boundaries and modes.

[0004] Motor fault diagnosis must first process the signal. Existing signal processing methods include fast Fourier transform, wavelet transform, empirical mode decomposition, etc. Empirical wavelet transform is based on empirical mode decomposition and wavelet transform. It is a signal analysis method with high time-frequency resolution and the ability to adaptively construct basis functions. However, due to insufficient sampling or analysis methods, different modes of the signal are confused, which can easily lead to misdiagnosis or missed diagnosis of the diagnosis results. Modal aliasing will interfere with the calculation of characteristic parameters such as amplitude and energy, making the assessment of the severity of the fault inaccurate. Therefore, as problems such as complex data structure, large data volume, and uneven data quality become more and more prominent, how to solve modal aliasing and effectively divide the spectrum are crucial. Summary of the invention

[0005] In order to solve the deficiencies in the prior art, the present invention provides a permanent magnet synchronous motor fault diagnosis method and system based on an improved empirical wavelet transform. By improving the empirical wavelet transform control strategy, the improved empirical wavelet transform is used to realize high-speed permanent magnet synchronous motor fault diagnosis, and solve the modal aliasing problem caused by unclear signal spectrum boundaries during motor diagnosis.

[0006] The present invention adopts the following technical solution.

[0007] The present invention proposes a permanent magnet synchronous motor fault diagnosis method based on improved empirical wavelet transform, comprising:

[0008] Obtain the fault signal and determine the power spectrum of the fault signal using the improved Welch method;

[0009] Divide the power spectrum of the fault signal into multiple continuous spectrum intervals;

[0010] The Littlewood-Paley method and the Meyer method are used to construct wavelet filter banks in each power spectrum interval, and the wavelet filter banks are used to extract the various modal components in each power spectrum interval from the power spectrum in each power spectrum interval; wherein the filtering bandwidth of the wavelet filter bank is dynamically adjusted according to the frequency of each modal component in each power spectrum interval;

[0011] The fault feature vector of the fault signal is constructed by using the energy moment of each modal component in each power spectrum interval; the fault feature matrix is ​​constructed by using the fault feature vector of each fault signal;

[0012] The CatBoost method is used to establish a motor fault diagnosis model, and the motor fault diagnosis model is trained with the improved training set and the improved gradient enhancement method; the trained motor fault diagnosis model performs motor fault diagnosis according to the fault feature matrix.

[0013] Preferably, according to the fault signal, an improved Welch method is used to determine the power spectrum of the fault signal, including:

[0014] The fault signal Divide into non-overlapping Each segment has data, , , the Welch power spectrum obtained by the improved Welch method satisfies the following relationship:

[0015] (1)

[0016] In the formula, is the Welch power spectrum of the fault signal, is the number of segments of the fault signal, is the data volume of each fault signal, is the normalization factor, For the Segment fault signal, is the window function, is the sampling number of the fault signal, is the angular frequency.

[0017] Preferably, the local maximum method is used to determine the Welch power spectrum of the fault signal. The maximum points and the corresponding The main frequency point, is the number of samples of the fault signal, and The main frequency points are sorted in descending order according to the corresponding extreme value points; the sorted main frequency points are used as the original boundaries, and the midpoints of two adjacent main frequency points are used as the segmentation boundaries. The original boundaries and the segmentation boundaries are used to obtain A continuous power spectrum interval.

[0018] Preferably, the wavelet filter banks for each power spectrum interval constructed according to the Littlewood-Paley method and the Meyer method include: empirical scaling function and the empirical wavelet function , respectively satisfying the following relations:

[0019] ,

[0020] ,

[0021] In the formula, , For the The empirical scaling function for each power spectrum interval is For the The empirical wavelet function of the power spectrum interval is is the angular frequency corresponding to the fault signal sampling value, For the The angular frequency of the power spectrum interval, For the The filter bandwidth of the power spectrum interval is is the transition function, is the sampling number of the fault signal;

[0022] The transition function satisfies the following relationship:

[0023] ,

[0024] In the formula, The value range is variables.

[0025] Preferably, the modal components of each power spectrum interval are extracted from the power spectrum of each power spectrum interval using a wavelet filter bank, satisfying the following relationship:

[0026] ,

[0027] In the formula, is the time domain expression of the global trend of the modal component, For the The time domain expression of the modal components in the power spectrum interval is: is the time domain expression of the approximation coefficient of the global trend of the modal component, is angstrom, and its value is , For the The time domain expression of the approximation coefficients of the modal components in the power spectrum interval is: is the empirical scaling function The time domain expression of For the The empirical wavelet function of the power spectrum interval The time domain expression of is the sampling number of the fault signal.

[0028] Preferably, the fault feature vector of the fault signal is constructed using the energy moments of each modal component in each power spectrum interval, including:

[0029] No. modal components The energy moment in each power spectrum interval satisfies the following relationship:

[0030] ,

[0031] In the formula, For the The energy moment of each modal component in each power spectrum interval, is the sampling number of the fault signal, is the sampling point of the fault signal, is the sampling time interval, For the modal components, ;

[0032] For The energy moments of the modal components in each power spectrum interval are normalized to the maximum value to satisfy the following relationship:

[0033] ,

[0034] In the formula, For the The normalized value of the energy moment of each modal component in each power spectrum interval, , Respectively The minimum and maximum values ​​of the energy moments of the modal components in each power spectrum interval;

[0035] The fault signal is constructed by the normalized value of the energy moment of each modal component in each power spectrum interval. The fault feature vector satisfies the following relationship:

[0036] ,

[0037] In the formula, .

[0038] Preferably, the fault feature matrix is ​​constructed using the fault feature vector of each fault signal, including:

[0039] use Fault signal The fault feature vector of the fault is used to construct the fault feature matrix, which satisfies the following relationship:

[0040] ,

[0041] In the formula, is the fault feature matrix, For the The fault feature vector of the fault signal, , is the number of fault signals.

[0042] Preferably, the training set is improved as follows:

[0043] Define the label value corresponding to each row element in the fault feature matrix; cluster each element in the fault feature matrix to obtain a cluster sequence , take the average value of the label values ​​corresponding to each element in the cluster as the new label value of all elements in the cluster;

[0044] Set the prior probability corresponding to the new label value of each element in the cluster and the weight corresponding to the prior probability, and determine the numerical characteristics of each element in the cluster according to the following relationship:

[0045] ,

[0046] In the formula, For the Cluster Neidi The numerical characteristics of the elements, , is the number of fault signals, , Respectively Cluster Neidi elements and Cluster Neidi The fault characteristics of each element, For the Cluster The label value is Cluster The new tag values ​​for all elements in is the prior probability corresponding to the new label value of each element in the cluster, is the weight corresponding to the prior probability, Brackets for Iverson.

[0047] Preferably, the improvement to the gradient enhancement method is as follows:

[0048] In the In the iteration, the numerical features of the improved training set are randomly sorted and a numerical feature is randomly removed from the sorting result. After that, a new training set is obtained, and the motor fault diagnosis model is trained using the new training set to establish an underfitting tree model. , calculate the arrangement in The previous numerical features were in the underfitting tree model The first-order gradient of the loss function on is used to train the first The first iteration Model trees to build a fitted tree model , with an underfit tree model and fit tree model The cumulative result is taken as the model tree and enter the +1 iteration.

[0049] Preferably, the AOA optimization algorithm is used to determine the hyperparameters of the motor fault diagnosis model, including:

[0050] The initial random solution is obtained by initializing the AOA parameters; when When the fitness function is calculated to determine the current best solution, the value of the mathematical accelerator optimization function is updated based on the current best solution. and the value of the probability function ;when Execute the exploration phase operations, including: Then perform the division operation, if Then perform the multiplication operation; when Execute the development phase operations at the same time, including: Then perform the subtraction operation, if Then run the addition operation; after executing the exploration phase operation or the development phase operation, execute ;when When , the optimal solution of the hyperparameters of the motor fault diagnosis model is returned; where , , for A random number in the range, , are the current number of iterations and the maximum number of iterations respectively.

[0051] The present invention also proposes a permanent magnet synchronous motor fault diagnosis system based on improved empirical wavelet transform, comprising:

[0052] An improved empirical wavelet transform decomposition module is used to obtain fault signals, and the power spectrum of the fault signal is determined by using an improved Welch method; the power spectrum of the fault signal is divided into multiple continuous spectrum intervals; the Littlewood-Paley method and the Meyer method are used to construct a wavelet filter group for each power spectrum interval, and the wavelet filter group is used to extract each modal component of each power spectrum interval from the power spectrum of each power spectrum interval; wherein the filtering bandwidth of the wavelet filter group is dynamically adjusted according to the frequency of each modal component in each power spectrum interval;

[0053] A feature extraction module is used to construct a fault feature vector of a fault signal using the energy moment of each modal component in each power spectrum interval; and to construct a fault feature matrix using the fault feature vectors of each fault signal;

[0054] The Catboost classifier module is used to establish a motor fault diagnosis model using the CatBoost method, and train the motor fault diagnosis model using an improved training set and an improved gradient enhancement method; the trained motor fault diagnosis model performs motor fault diagnosis according to a fault feature matrix.

[0055] The present invention is also a terminal, comprising a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the method.

[0056] The present invention is also a computer-readable storage medium having a computer program stored thereon, which implements the steps of the method when executed by a processor.

[0057] The beneficial effects of the present invention are that, compared with the prior art, the present invention at least includes that, in view of the relevant problems existing in the fault detection of the six-phase permanent magnet synchronous motor, based on the fault diagnosis strategy of signal analysis, from the perspective of optimizing the signal spectrum, the Welch power spectrum curve is used to replace the FFT curve, and the spectrum is segmented on a relatively smooth curve to obtain the time-frequency characteristics of the fault signal. At the same time, in order to achieve the purpose of optimizing the model, the feature set and the label combination are combined to train the CatBoost classification model by combining data drive, and the model with better performance is selected and saved through the arithmetic optimization algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Schematic diagram of spectrum segmentation of empirical wavelet transform in an embodiment of the present invention;

[0059] Figure 2 Schematic diagram of spectrum segmentation of improved empirical wavelet transform in an embodiment of the present invention

[0060] Figure 3 Flowchart of CatBoost tree construction and AOA algorithm in an embodiment of the present invention;

[0061] Figure 4 A block diagram of a fault diagnosis system for a six-phase permanent magnet synchronous motor according to an embodiment of the present invention;

[0062] Figure 5 This is a fault current spectrum segmentation diagram in an embodiment of the present invention;

[0063] Figure 6 This is the ACC curve diagram of the model training in the embodiment of the present invention;

[0064] Figure 7 This is a comparison chart of the accuracy of the verification set in an embodiment of the present invention;

[0065] Figure 8 1 is a comparison chart of the accuracy of different models in the embodiments of the present invention. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only embodiments of a part of the present invention, rather than all embodiments. Based on the spirit of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the protection scope of the present invention.

[0067] The present invention provides a permanent magnet synchronous motor fault diagnosis method based on improved empirical wavelet transform, which is composed of improved empirical wavelet analysis (IEWT) and a gradient boosting algorithm library (CatBoost). The IEWT includes: using improved Welch to determine the power spectrum of the fault signal, power spectrum segmentation, wavelet filter group construction and mode extraction.

[0068] include:

[0069] Step 1, obtaining a fault signal, and using an improved Welch method to determine the power spectrum of the fault signal;

[0070] In the embodiment, the fault signal Divide into non-overlapping Each segment has data, , , the Welch power spectrum obtained by the improved Welch method satisfies the following relationship:

[0071] (1)

[0072] In the formula, is the Welch power spectrum of the fault signal, is the number of segments of the fault signal, is the data volume of each fault signal, is the normalization factor, For the Segment fault signal, is the window function, is the sampling number of the fault signal, is the angular frequency.

[0073] In the existing Welch method, in order to make full use of all the data of the signal, the signal is divided into overlapping segments, and a window function is applied to each segment to calculate the periodogram of each segment, and a smooth power spectrum is obtained by averaging the periodogram; the fault signal is processed by the FFT method, and the obtained power spectrum has multi-modal discrimination capability, but the power spectrum is not smooth enough, resulting in unclear boundaries when the spectrum is segmented, thereby affecting the fault diagnosis result of the permanent magnet synchronous motor. However, the existing Welch method cannot solve the problem of modal aliasing due to the overlap of signal segmentation. Therefore, the present invention proposes to divide the fault signal into non-overlapping segments and then determine the Welch power spectrum, so as to obtain an improved Welch power spectrum curve, which takes into account the advantages of the Welch power spectrum curve being smooth and easy to segment, and also has the multi-modal discrimination characteristics of the FFT power spectrum curve, so as to perform power spectrum segmentation on a relatively smooth curve without reducing the power spectrum processing speed.

[0074] The principle of empirical wavelet transform is to divide the FFT power spectrum of the fault signal into multiple continuous intervals, then construct a wavelet filter group on each interval for filtering, and finally obtain the frequency modulation component through signal reconstruction. However, empirical wavelet transform cannot avoid the problem of modal aliasing, which will lead to incorrect diagnosis results when applied to permanent magnet synchronous motor fault diagnosis; the present invention first starts from the processing of power spectrum and uses the improved Welch method to determine the power spectrum of the fault signal, which is the basis for solving the problem of modal aliasing.

[0075] Step 2: divide the power spectrum of the fault signal into multiple continuous power spectrum intervals.

[0076] Specifically, the local maximum method is used to determine the Welch power spectrum of the fault signal. The maximum points and the corresponding The main frequency point, is the number of samples of the fault signal, and The main frequency points are sorted in descending order according to the corresponding extreme value points; the sorted main frequency points are used as the original boundaries of the spectrum, and the midpoints of two adjacent main frequency points are used as the segmentation boundaries of the spectrum. A continuous power spectrum interval;

[0077] Since the power spectrum curve of the fault signal determined by the improved Welch method proposed in the present application is not only relatively smooth but also has clear boundaries, the maximum point that can be determined only by the local maximum method is the main frequency point, and each power spectrum interval corresponds to an eigenmode component; The maximum points are arranged in descending order, then The main frequency points are arranged in descending order according to the corresponding extreme value points;

[0078] The number of continuous power spectrum intervals corresponds to the number of samples of the fault signal, ensuring that the modal component is extracted for each sample, thereby ensuring full utilization of the sampled signal data, making up for the possible decrease in data utilization due to non-overlapping segmentation in the improved Welch method, effectively avoiding missed diagnosis, and thus improving the accuracy of fault diagnosis.

[0079] Step 3: Use the Littlewood-Paley method and the Meyer method to construct a wavelet filter bank for each power spectrum interval, and use the wavelet filter bank to extract the modal components of each power spectrum interval from the power spectrum of each power spectrum interval.

[0080] Specifically, the wavelet filter banks for each power spectrum interval constructed according to the Littlewood-Paley method and the Meyer method include: empirical scaling function and the empirical wavelet function , respectively satisfying the following relations:

[0081] ,

[0082] ,

[0083] In the formula, , For the The empirical scaling function for each power spectrum interval is For the The empirical wavelet function of the power spectrum interval is is the angular frequency corresponding to the fault signal, For the The angular frequency of the power spectrum interval, For the The filter bandwidth of the power spectrum interval is is the transition function, is the sampling number of the fault signal;

[0084] The transition function satisfies the following relationship:

[0085] ,

[0086] In the formula, The value range is Variables;

[0087] In the embodiment, in order to simplify the calculation, it is defined .

[0088] In the wavelet filter group constructed by the present invention, by introducing a transition function, the filtering bandwidth of the wavelet filter group is dynamically adjusted according to the frequency of each modal component in each power spectrum interval, thereby effectively suppressing the influence of modal aliasing on fault diagnosis.

[0089] Using the wavelet filter bank, the modal components of each power spectrum interval are extracted according to the following relationship:

[0090] ,

[0091] In the formula, is the time domain expression of the global trend of the modal component, For the The time domain expression of the modal components in the power spectrum interval is: is the time domain expression of the approximation coefficient of the modal component in the first power spectrum interval, is angstrom, and its value is , For the The time domain expression of the approximation coefficients of the modal components in the power spectrum interval is: is the empirical scaling function of the first power spectrum interval The time domain expression of For the The empirical wavelet function of the power spectrum interval The time domain expression of is the sampling number of the fault signal;

[0092] For the same fault current signal, Figure 1 Schematic diagram of spectrum segmentation of empirical wavelet transform. Figure 2 Schematic diagram of spectrum segmentation for improving empirical wavelet transform; Figure 1 and Figure 2 In the above, the spectrum is divided into 5 power spectrum intervals, corresponding to 5 modal components. Figure 1 and Figure 2 The dotted lines in the figure are the boundary values ​​of each power spectrum interval, and the circles are the maximum values ​​of each power spectrum interval. Figure 1 The "1" in the table indicates an invalid frequency band, and the "2" indicates a subsequent frequency band. Figure 1 and Figure 2 The comparison results of spectrum segmentation effects of different wavelet changes are obtained. Figure 1It can be seen that the empirical wavelet transform method will mistakenly divide an invalid frequency band due to the large sidelobe value near the main frequency, resulting in one of the two characteristic frequency points in the subsequent frequency band not being divided. The improved empirical wavelet transform method performs spectrum segmentation on the Welch power spectrum curve, and the resulting spectrum line is smoother and has less spectrum leakage. When using the local maximum algorithm to obtain extreme points, it will not be disturbed by the sidelobe value, and can more accurately locate the characteristic frequency points, making the spectrum segmentation more accurate, ensuring the subsequent extraction of the correct modal components, and effectively suppressing modal aliasing.

[0093] Step 4: construct the fault feature vector of the fault signal using the energy moment of each modal component in each power spectrum interval; and construct the fault feature matrix using the fault feature vector of each fault signal.

[0094] Since the motor fault signal is not complex, the feature extraction stage can only use the energy moments of each modal component in each power spectrum interval to characterize the fault signal. modal components The energy moment in each power spectrum interval satisfies the following relationship:

[0095] ,

[0096] In the formula, For the The energy moment of each modal component in each power spectrum interval, is the sampling number of the fault signal, is the sampling point of the fault signal, is the sampling time interval, For the modal components, ;

[0097] For The energy moments of the modal components in each power spectrum interval are normalized to the maximum value to satisfy the following relationship:

[0098] ,

[0099] In the formula, For the The normalized value of the energy moment of each modal component in each power spectrum interval, , Respectively The minimum and maximum values ​​of the energy moments of the modal components in each power spectrum interval;

[0100] The fault signal is constructed by the normalized value of the energy moment of each modal component in each power spectrum interval. The fault feature vector satisfies the following relationship:

[0101] ,

[0102] Fault signal The fault signal set is formed, and the fault feature vector of each fault signal is used to construct the fault feature matrix for motor fault diagnosis and classification. , For the The fault feature vector of the fault signal, , is the number of fault signals.

[0103] The improved empirical wavelet transform method proposed in this invention can accurately separate each mode while retaining the signal frequency domain information, effectively suppressing the modal aliasing phenomenon caused by spectrum leakage in the empirical wavelet transform method. The energy moment of each mode is used to characterize the fault signal, and a fault classification data set is constructed to provide data support for the construction of subsequent machine learning classification models.

[0104] Step 5, using the CatBoost method to establish a motor fault diagnosis model, and training the motor fault diagnosis model with the improved training set and the improved gradient enhancement method; the trained motor fault diagnosis model performs motor fault diagnosis according to the fault feature matrix.

[0105] The core of the CatBoost method is to linearly combine weak classifiers, which is an improvement on the gradient boosting decision tree algorithm. Its uniqueness lies in that it can automatically process category features without tedious coding. Compared with other mainstream improved algorithms, the present invention proposes an effective category feature processing method and ranking promotion mechanism.

[0106] Specifically, the improvements to the training set are as follows:

[0107] Define the label value corresponding to each row element in the fault feature matrix; cluster each element in the fault feature matrix to obtain a cluster sequence , take the average value of the label values ​​corresponding to each element in the cluster as the new label value of all elements in the cluster;

[0108] In the embodiment, the fault feature matrix Row elements correspond to the same label value ; Cluster the elements in the fault feature matrix. A cluster is a range of elements with the same fault feature. Within this range, the average value of the label value corresponding to each element is used as the label value of each cluster. The label value of each cluster is the new label value of all elements in the cluster. In fact, the labels of each element in the fault feature matrix are redefined.

[0109] Set the prior probability corresponding to the new label value of each element in the cluster and the weight corresponding to the prior probability, and determine the numerical characteristics of each element in the cluster according to the following relationship:

[0110] ,

[0111] In the formula, For the Cluster Neidi The numerical characteristics of the elements, , is the number of fault signals, , Respectively Cluster Neidi elements and Cluster Neidi The fault characteristics of each element, For the Cluster The label value is Cluster The new tag values ​​for all elements in is the prior probability corresponding to the new label value of each element in the cluster, is the weight corresponding to the prior probability, Brackets for Iverson.

[0112] In the prior art, when numerical features are obtained using the label values ​​of each row, when the feature distribution of an individual case or a test set is inconsistent with that of a training set, label leakage problems will occur. The converted numerical features of the elements are the label values, which will affect the reliability of subsequent fault diagnosis. Therefore, on the basis of redefining the labels of each element in the fault feature matrix, the prior probabilities corresponding to the new label values ​​of each element in the cluster and the weights corresponding to the prior probabilities are set. The new label values, prior probabilities and weights corresponding to the prior probabilities of each element in the cluster are used to convert the fault features of the elements in the fault feature matrix into numerical features to form an improved training set.

[0113] In the embodiment, the improvement of the training set is encoded, so that the category features of the fault can be processed and transformed without occupying the preprocessing time, and the label leakage problem in the process of gradient enhancement is solved.

[0114] In the embodiment, the prior probability value is 0.5, and the weight value corresponding to the prior probability is 0.5.

[0115] Specifically, the improvements to the gradient enhancement method are as follows:

[0116] In the In the iteration, the numerical features of the improved training set are randomly sorted and a numerical feature is randomly removed from the sorting result. After that, a new training set is obtained, and the motor fault diagnosis model is trained using the new training set to establish an underfitting tree model. , calculate the arrangement in The previous numerical features were in the underfitting tree model The first-order gradient of the loss function on is used to train the first The first iteration Model trees to build a fitted tree model , with an underfit tree model and fit tree model The cumulative result is taken as the model tree and enter the +1 iteration.

[0117] In the present invention, another important improvement is the model tree training generation method based on the sorting promotion mechanism. In the previous gradient enhancement algorithm, the structure and leaf node value of the next tree are usually determined according to the gradient of the loss function on the current tree model. However, in the process of model accumulation, each tree is trained on the same sample, which will cause deviations in point-by-point gradient estimation and overfitting of the model. The sorting promotion mechanism proposed by the improved CatBoost algorithm alleviates the above problems and effectively reduces gradient deviation. Catboost uses a symmetric binary tree as a base classifier, has a regularization effect, and has an extremely fast prediction speed. It also cross-combines categorical features based on a greedy strategy to give full play to the advantages of the features.

[0118] In the embodiment, the trained motor fault diagnosis model performs motor fault diagnosis according to the fault feature matrix, such as Figure 3 As shown, the test set is converted into a feature combination and then input into the CatBoost tree structure for training and testing to obtain multiple prediction values, which are summed up as the motor fault diagnosis result; wherein, the prior probability and the weight of the prior probability are added to the CatBoost tree structure.

[0119] Specifically, the AOA optimization algorithm is used to determine the hyperparameters of the motor fault diagnosis model. The specific process is as follows: Figure 3 As shown, including:

[0120] The initial random solution is obtained by initializing the AOA parameters; when When the fitness function is calculated to determine the current best solution, the value of the mathematical accelerator optimization function is updated based on the current best solution. and the value of the probability function ;when Execute the exploration phase operations, including: Then perform the division operation, if Then perform the multiplication operation; when Execute the development phase operations at the same time, including: Then perform the subtraction operation, if Then run the addition operation; after executing the exploration phase operation or the development phase operation, execute ;when When , the optimal solution of the hyperparameters of the motor fault diagnosis model is returned; where , , for A random number in the range, , are the current number of iterations and the maximum number of iterations respectively.

[0121] The AOA optimization algorithm selects appropriate hyperparameters to make the model performance better. Inspired by the hierarchy of arithmetic operators and their dominance from outside to inside, it determines the best element that meets specific criteria from a set of candidate alternatives and can solve optimization problems without calculating derivatives.

[0122] The present invention also proposes a permanent magnet synchronous motor fault diagnosis system based on improved empirical wavelet transform, comprising:

[0123] An improved empirical wavelet transform decomposition module is used to obtain fault signals, and the power spectrum of the fault signal is determined by using an improved Welch method; the power spectrum of the fault signal is divided into multiple continuous spectrum intervals; the Littlewood-Paley method and the Meyer method are used to construct a wavelet filter group for each power spectrum interval, and the wavelet filter group is used to extract each modal component of each power spectrum interval from the power spectrum of each power spectrum interval; wherein the filtering bandwidth of the wavelet filter group is dynamically adjusted according to the frequency of each modal component in each power spectrum interval;

[0124] A feature extraction module is used to construct a fault feature vector of a fault signal using the energy moment of each modal component in each power spectrum interval; and to construct a fault feature matrix using the fault feature vectors of each fault signal;

[0125] The Catboost classifier module is used to establish a motor fault diagnosis model using the CatBoost method, and train the motor fault diagnosis model using an improved training set and an improved gradient enhancement method; the trained motor fault diagnosis model performs motor fault diagnosis according to a fault feature matrix.

[0126] In the embodiment, the six-phase PMSM fault diagnosis system of the permanent magnet synchronous motor fault diagnosis system using the improved empirical wavelet transform proposed by the present invention is as follows: Figure 4 As shown, Figure 4The motor drive module includes a permanent magnet synchronous motor (rotation angular velocity is ), a six-phase full-bridge inverter and an SPWM modulator; in the control module, the phase current and encoder sampling module transmits the collected machine-end current and each phase current of the permanent magnet synchronous motor to the fault diagnosis module, and the phase current and encoder sampling module transmits signals to the speed current dual closed-loop control module. The output signal of the speed current dual closed-loop control module is transmitted to the first coordinate conversion module ( ) are respectively input to the fault-tolerant current loop control module and the XY subspace harmonic suppression module. The output signals of the fault-tolerant current loop control module and the XY subspace harmonic suppression module are converted by the second coordinate conversion module ( ) and then input into the SPWM modulator; in the fault diagnosis module, the signal output by the IEWT decomposition module is transmitted to the feature extraction module, the signal output by the feature extraction module is transmitted to the Catboost classifier module, the signal output by the Catboost classifier module is transmitted to the mode switching module, and the signal output by the mode switching module is used to control the working modes of the first coordinate conversion module and the second coordinate conversion module.

[0127] The fault diagnosis system proposed in the present invention adds a fault diagnosis module combining signal analysis and data drive on the basis of the motor drive module and control module to realize the prediction and identification of faults. Combined with the fault-tolerant control algorithm, the operation reliability, fault-tolerant operation capability and energy conversion efficiency of the motor are enhanced. Among them, the fault diagnosis module includes two modes: offline training and online identification.

[0128] The offline learning process builds a fault switching model and records the sampling signals of the six-phase PMSM under different working conditions. The IEWT algorithm is used to decompose the sampling signal, extract the submodal energy moment to construct the fault feature set, and train the CatBoost classifier to save the optimal model. The online recognition process reads the six-phase PMSM signal online through the window function. After signal decomposition and feature extraction, the CatBoost model trained in the offline state is called to diagnose the fault category. The control unit then selects the corresponding fault-tolerant control scheme to avoid secondary damage to the system caused by long-term phase loss of the motor.

[0129] Figure 5 The fault current spectrum segmentation diagram is shown in Figure 1. After the frequency band division is completed, the corresponding wavelet filter is constructed to extract 8 groups of modal components IMF1~8 from low frequency to high frequency of the original signal. Among them, IMF1 is the global trend of the signal, IMF2~4 are mainly low-frequency components, IMF5~7 reflect the time-frequency information of the fault process, and IMF8 is mainly high-frequency noise. The modal components obtained by IEWT can well reflect the time of occurrence of the fault signal and the distribution of the fault frequency. The energy moment of each mode is extracted to characterize the original signal, which is used to determine the working condition of the motor.

[0130] Based on the AOA optimization algorithm, 80% of the input data set is used as the training set to perform hyperparameter tuning on CatBoost. A 5-fold cross-validation method is used to analyze the cross-validation indicator accuracy: ACC, and the optimal value of the model is selected according to the indicator size. Among them: the optimal value of the model in the interval of [50,300] is 115 when the number of iterations is in the interval [50,300]; the optimal value of the model in the interval of [4,10] is 6; the optimal value of the model in the interval of [0.02,0.2] is 0.074 when the learning rate is in the interval of [0.6,1.0]; the optimal value of the model in the interval of [0.6,1.0] is 0.637; the optimal value of the model in the interval of [0.6,1.0] is 0.894 when the response surface is in the interval of [0.6,1.0]. The change curves of ACC on the test set and the validation set during the corresponding training process are shown in the following figure. Figure 6 As shown in the figure, the accuracy ACC of the model on the validation set reached the optimal value of 0.9975 at 115 iterations.

[0131] Figure 7 The changes in the model cross-validation accuracy before and after tuning are given. After using the optimized parameters, the model's validation set classification accuracy is improved in each round of K=5-fold cross-validation. The AOA tuning algorithm improves the performance of the CatBoost classifier.

[0132] Three integrated algorithms (RF, XGBoost, CatBoost) and two single machine learning algorithms (SVM, BP) are used to train and predict the same fault data set. The accuracy comparison results of different models are as follows: Figure 8 As shown in the figure. The prediction accuracy of the three integrated algorithms: RF, XGboost, and CatBoost are 0.9896, 0.9895, and 0.9943 respectively, which are better than the single machine learning algorithms SVM and BP. Accuracy: 0.9719, 0.9832. Among them, the accuracy of the CatBoost algorithm is better than the other four classification algorithms, and the prediction ability in each category is balanced, and the model has strong generalization ability.

[0133] In summary, based on the six-phase PMSM phase loss fault diagnosis problem, an improved empirical wavelet transform algorithm is proposed to extract fault signal features and construct a fault data set for CatBoost training and optimization of the classification model. A fault diagnosis strategy combining signal analysis and data-driven is adopted to effectively identify the phase loss fault type of the six-phase PMSM.

[0134] This embodiment aims to solve the problem of phase failure during motor operation due to driver damage, loose stator winding connection, etc., and improves the empirical wavelet algorithm based on the Welch power spectrum to perform spectrum segmentation on a smoother curve. A permanent magnet synchronous motor fault model is constructed, fault signals under different working conditions are collected, and a fault data set is constructed using the IEWT algorithm. Finally, a CatBoost fault classification model is constructed by selecting appropriate hyperparameters, and its processing results are compared with traditional classification algorithms. The proposed method effectively suppresses the erroneous segmentation phenomenon caused by excessive side lobes near the main frequency in the traditional empirical wavelet transform, improves the accuracy of fault classification, performs well under different fault conditions, and realizes the diagnosis of phase failure of permanent magnet synchronous motors.

[0135] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0136] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the above. More specific examples (a non-exhaustive list) of computer-readable storage media include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the above. The computer-readable storage medium used herein is not to be interpreted as a transient signal itself, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0137] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0138] The computer program instructions for performing the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer, partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect through the Internet). In some embodiments, by using the state information of the computer-readable program instructions to personalize an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), the electronic circuit may execute the computer-readable program instructions, thereby implementing various aspects of the present disclosure.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A permanent magnet synchronous motor fault diagnosis method based on improved empirical wavelet transform, characterized in that: include: Obtain the fault signal and determine the power spectrum of the fault signal using the improved Welch method; Divide the power spectrum of the fault signal into multiple continuous spectrum intervals; The Littlewood-Paley method and the Meyer method are used to construct the wavelet filter bank of each power spectrum interval, and the wavelet filter bank is used to extract each modal component of each power spectrum interval from the power spectrum of each power spectrum interval; wherein the filtering bandwidth of the wavelet filter bank is dynamically adjusted according to the frequency of each modal component in each power spectrum interval; The fault feature vector of the fault signal is constructed by using the energy moment of each modal component in each power spectrum interval; the fault feature matrix is ​​constructed by using the fault feature vector of each fault signal; The CatBoost method is used to establish a motor fault diagnosis model, and the motor fault diagnosis model is trained with the improved training set and the improved gradient enhancement method; the trained motor fault diagnosis model performs motor fault diagnosis according to the fault feature matrix.

2. The permanent magnet synchronous motor fault diagnosis method based on improved empirical wavelet transform according to claim 1 is characterized in that: According to the fault signal, the improved Welch method is used to determine the power spectrum of the fault signal, including: The fault signal Divide into non-overlapping Each segment has data, , , the Welch power spectrum obtained by the improved Welch method satisfies the following relationship: （1） In the formula, is the Welch power spectrum of the fault signal, is the number of segments of the fault signal, is the data volume of each fault signal, is the normalization factor, For the Segment fault signal, is the window function, is the sampling number of the fault signal, is the angular frequency.

3. The permanent magnet synchronous motor fault diagnosis method based on improved empirical wavelet transform according to claim 2 is characterized in that: The local maximum method is used to determine the Welch power spectrum of the fault signal. The maximum points and the corresponding The main frequency point, is the number of samples of the fault signal, and The main frequency points are sorted in descending order according to the corresponding extreme value points; The sorted main frequency points are used as the original boundary, and the midpoints of two adjacent main frequency points are used as the segmentation boundary. The original boundary and the segmentation boundary are used to obtain A continuous power spectrum interval.

4. The permanent magnet synchronous motor fault diagnosis method based on improved empirical wavelet transform according to claim 1 is characterized in that: Wavelet filter banks for each power spectrum interval constructed according to the Littlewood-Paley method and the Meyer method, including: empirical scaling function and the empirical wavelet function , respectively satisfying the following relations: ， ， In the formula, , For the The empirical scaling function for each power spectrum interval is For the The empirical wavelet function of the power spectrum interval is is the angular frequency corresponding to the fault signal sampling value, For the The angular frequency of the power spectrum interval, For the The filter bandwidth of the power spectrum interval is is the transition function, is the sampling number of the fault signal; The transition function satisfies the following relationship: ， In the formula, The value range is variables.

5. The permanent magnet synchronous motor fault diagnosis method based on improved empirical wavelet transform according to claim 4 is characterized in that: The wavelet filter bank is used to extract the modal components of each power spectrum interval from the power spectrum of each power spectrum interval, satisfying the following relationship: ， In the formula, is the time domain expression of the global trend of the modal component, For the The time domain expression of the modal components in the power spectrum interval is: is the time domain expression of the approximation coefficient of the global trend of the modal component, is angstrom, and its value is , For the The time domain expression of the approximation coefficients of the modal components in the power spectrum interval is: is the empirical scaling function The time domain expression of For the The empirical wavelet function of the power spectrum interval The time domain expression of is the sampling number of the fault signal.

6. The permanent magnet synchronous motor fault diagnosis method based on improved empirical wavelet transform according to claim 1 is characterized in that: The fault feature vector of the fault signal is constructed by using the energy moments of each modal component in each power spectrum interval, including: No. modal components The energy moment in each power spectrum interval satisfies the following relationship: ， In the formula, For the The energy moment of each modal component in each power spectrum interval, is the sampling number of the fault signal, is the sampling point of the fault signal, is the sampling time interval, For the modal components, ; For The energy moments of the modal components in each power spectrum interval are normalized to the maximum value to satisfy the following relationship: ， In the formula, For the The normalized value of the energy moment of each modal component in each power spectrum interval, , Respectively The minimum and maximum values ​​of the energy moments of the modal components in each power spectrum interval; The fault signal is constructed by the normalized value of the energy moment of each modal component in each power spectrum interval. The fault feature vector satisfies the following relationship: ， In the formula, .

7. The permanent magnet synchronous motor fault diagnosis method based on improved empirical wavelet transform according to claim 6 is characterized in that: The fault feature matrix is ​​constructed using the fault feature vector of each fault signal, including: use Fault signal The fault feature vector of the fault is used to construct the fault feature matrix, which satisfies the following relationship: ， In the formula, is the fault feature matrix, For the The fault feature vector of the fault signal, , is the number of fault signals.

8. The permanent magnet synchronous motor fault diagnosis method based on improved empirical wavelet transform according to claim 1 is characterized in that: The improvements to the training set are as follows: Define the label value corresponding to each row element in the fault feature matrix; cluster each element in the fault feature matrix to obtain a cluster sequence , take the average value of the label values ​​corresponding to each element in the cluster as the new label value of all elements in the cluster; Set the prior probability corresponding to the new label value of each element in the cluster and the weight corresponding to the prior probability, and determine the numerical characteristics of each element in the cluster according to the following relationship: ， In the formula, For the Cluster Neidi The numerical characteristics of the elements, , is the number of fault signals, , Respectively Cluster Neidi elements and Cluster Neidi The fault characteristics of each element, For the Cluster The label value is Cluster The new tag values ​​for all elements in is the prior probability corresponding to the new label value of each element in the cluster, is the weight corresponding to the prior probability, Brackets for Iverson.

9. The permanent magnet synchronous motor fault diagnosis method based on improved empirical wavelet transform according to claim 8 is characterized in that: The improvements to the gradient enhancement method are as follows: In the In the iteration, the numerical features of the improved training set are randomly sorted and a numerical feature is randomly removed from the sorting result. After that, a new training set is obtained, and the motor fault diagnosis model is trained using the new training set to establish an underfitting tree model. , calculate the arrangement in The previous numerical features were in the underfitting tree model The first-order gradient of the loss function on is used to train the first The first iteration Model trees to build a fitted tree model , with an underfit tree model and fit tree model The cumulative result is taken as the model tree and enter the +1 iteration.

10. The permanent magnet synchronous motor fault diagnosis method based on improved empirical wavelet transform according to claim 1, characterized in that: The AOA optimization algorithm is used to determine the hyperparameters of the motor fault diagnosis model, including: The initial random solution is obtained by initializing the AOA parameters; when When the fitness function is calculated to determine the current best solution, the value of the mathematical accelerator optimization function is updated based on the current best solution. and the value of the probability function ;when Execute the exploration phase operations, including: Then perform the division operation, if Then perform the multiplication operation; when Execute the development phase operations at the same time, including: Then perform the subtraction operation, if Then run the addition operation; after executing the exploration phase operation or the development phase operation, execute ;when When , the optimal solution of the hyperparameters of the motor fault diagnosis model is returned; where , , for A random number in the range, , are the current number of iterations and the maximum number of iterations respectively.

11. A permanent magnet synchronous motor fault diagnosis system based on improved empirical wavelet transform, characterized in that: include: Improve the empirical wavelet transform decomposition module to obtain the fault signal, and use the improved Welch method to determine the power spectrum of the fault signal; The power spectrum of the fault signal is divided into multiple continuous spectrum intervals; the Littlewood-Paley method and the Meyer method are used to construct the wavelet filter group of each power spectrum interval, and the wavelet filter group is used to extract each modal component of each power spectrum interval from the power spectrum of each power spectrum interval; wherein the filtering bandwidth of the wavelet filter group is dynamically adjusted according to the frequency of each modal component in each power spectrum interval; A feature extraction module is used to construct a fault feature vector of a fault signal using the energy moment of each modal component in each power spectrum interval; and to construct a fault feature matrix using the fault feature vectors of each fault signal; The Catboost classifier module is used to establish a motor fault diagnosis model using the CatBoost method, and train the motor fault diagnosis model using an improved training set and an improved gradient enhancement method; the trained motor fault diagnosis model performs motor fault diagnosis according to a fault feature matrix.

12. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 10.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.

Images