Optimized variational mode decomposition-based permanent magnet synchronous motor demagnetization fault diagnosis method

By optimizing variational modal decomposition and deep learning technology, combined with sparrow search algorithm and convolutional neural network, the problem of difficult to identify early demagnetization failures of permanent magnet synchronous motors has been successfully solved, improving the accuracy and efficiency of fault diagnosis, and reducing maintenance cost.

CN118962437APending Publication Date: 2024-11-15EAST CHINA JIAOTONG UNIVERSITY
View PDF 0 Cites 6 Cited by

Patent Information

Application Number
CN202411004087.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The prior art is difficult to effectively identify and diagnose early demagnetization failures of permanent magnet synchronous motors, resulting in the neglect of faults, further aggravating equipment damage and increasing maintenance and replacement costs.

Method used

Using the method based on optimized variational modal decomposition, the parameters of the variational modal algorithm are optimized through the sparrow search algorithm, combined with the convolutional neural network and the long and short-term memory network, time domain features are extracted and fault diagnosis is performed to realize the permanent magnet synchronous motor Efficient identification of demagnetization faults.

Benefits of technology

It improves the accuracy and efficiency of fault identification, reduces the influence of human subjectivity, avoids modal aliasing, enhances the stability and accuracy of diagnosis, and improves the service life of the equipment and the continuous production.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118962437A_ABST
    Figure CN118962437A_ABST
Patent Text Reader

Abstract

The invention discloses a permanent magnet synchronous motor demagnetization fault diagnosis method based on optimized variational mode decomposition, and the method can improve the recognition rate and accuracy of the demagnetization fault of a permanent magnet synchronous motor through the deep research of the demagnetization mechanism and fault features of the permanent magnet synchronous motor and the development of an efficient fault diagnosis technology. Faults can be found and processed in time, and production loss is reduced. The running state of the permanent magnet synchronous motor is monitored and diagnosed by adopting an advanced data detection and processing method, so that quick response can be ensured when a fault occurs, and measures can be taken to improve the continuity and the stability of production.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of motor fault analysis, and in particular to a permanent magnet synchronous motor demagnetization fault diagnosis method based on optimized variational modal decomposition. Background Art

[0002] Permanent magnet synchronous motors are key equipment in modern industrial automation. Their performance directly affects overall production efficiency and product quality. Demagnetization failures can cause motor performance degradation or even failure to operate normally, which seriously threatens production safety. Permanent magnet synchronous motors usually have high costs, and as the core part of the equipment, once a fault occurs, failure to detect it in time and take corresponding measures will directly increase the company's maintenance and replacement costs. With the continuous improvement of the degree of industrial automation, the performance and stability requirements of permanent magnet synchronous motors are gradually increasing. Effective permanent magnet synchronous motor demagnetization fault diagnosis technology can meet market demand and improve the safety, reliability and service life of equipment.

[0003] Although permanent magnet synchronous motors have been widely used in various fields, their fault diagnosis technology is still in the development stage, especially for the early demagnetization fault diagnosis of permanent magnet synchronous motors, which still has certain difficulties and challenges. Due to the complexity and diversity of permanent magnet synchronous motors, traditional fault diagnosis methods are difficult to accurately identify demagnetization faults in a timely manner, which may lead to demagnetization faults being ignored and further aggravate the damage of permanent magnet synchronous motors. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art, and to provide a permanent magnet synchronous motor demagnetization fault diagnosis method based on optimized variational modal decomposition. By deeply studying the demagnetization mechanism and fault characteristics of permanent magnet synchronous motors, the development of efficient fault diagnosis technology can improve the recognition rate and accuracy of permanent magnet synchronous motor demagnetization faults, which helps to timely discover and handle faults and reduce production losses. By adopting advanced data detection and processing methods to monitor and diagnose the operating status of permanent magnet synchronous motors, it can be ensured that a rapid response can be made when a fault occurs and measures can be taken to improve the continuity and stability of production.

[0005] The permanent magnet synchronous motor demagnetization fault diagnosis method based on optimized variational modal decomposition according to an embodiment of the present invention comprises steps S1 to S5:

[0006] S1, fault analysis and signal selection, select the built-in permanent magnet synchronous motor as the verification object, establish a permanent magnet synchronous motor demagnetization fault simulation model, analyze the mechanism of the permanent magnet synchronous motor demagnetization fault, further analyze the change law of the magnetic flux and stator current of the permanent magnet synchronous motor during the irreversible demagnetization process, simulate the demagnetization fault of the permanent magnet synchronous motor, and collect the stator current data obtained by simulation as the fault signal;

[0007] S2, data denoising, using the sparrow search algorithm to optimize the penalty factor and decomposition layer number in the variational modal algorithm. In the variational modal parameter optimization process, the minimum value of the envelope entropy is used as the fitness function to optimize the parameters, and the optimal solution for different fault signals is obtained. The optimal decomposition layer number and penalty factor obtained by optimization are substituted into the variational modal algorithm to perform modal decomposition on the data, and the mode with the minimum envelope entropy is selected for data reconstruction;

[0008] S3, time domain feature extraction, extracting six time domain features including mean, variance, peak, kurtosis, effective value and peak factor from the denoised data, and using the extracted feature vector as input data for the fault diagnosis model;

[0009] S4, diagnostic model construction, which combines convolutional neural network and long short-term memory network, where convolutional neural network extracts key spatial features from input data through convolution and pooling operations, and long short-term memory network is a variant of recurrent neural network, which is used to process sequence data with long-term dependencies in fault diagnosis;

[0010] S5, model training and fault diagnosis, divides the obtained input data into training set samples and test set samples, where the training set samples are used for model learning. The diagnostic model will extract features from the input data, and through training, the diagnostic model can adjust its internal parameters such as weights and biases to minimize the error between the predicted value and the actual value. The test sample set is used to evaluate the performance of the trained diagnostic model on unseen data. The trained model is used to diagnose the demagnetization fault of the permanent magnet synchronous motor to obtain the diagnostic result.

[0011] The permanent magnet synchronous motor demagnetization fault diagnosis method based on optimized variational modal decomposition according to an embodiment of the present invention has the following beneficial effects:

[0012] (1) The number of decomposition layers and penalty factors in the variational mode decomposition are optimized through the sparrow search algorithm. Compared with the unoptimized variational mode decomposition, there is no need to manually determine the number of decomposition layers and penalty factors, which realizes the automatic selection of parameters, reduces the influence of human subjectivity on the decomposition effect, effectively avoids the mode aliasing phenomenon, and improves the decomposition performance and stability of the variational mode decomposition.

[0013] (2) The original data is denoised using the optimized variational mode algorithm, which highlights the fault characteristics in the data and is beneficial to the diagnostic effect of the subsequent algorithm.

[0014] (3) The convolutional long short-term memory neural network used in this paper combines the advantages of convolutional neural networks and long short-term memory neural networks. By training, the network can automatically learn the deep features of the input data and represent them as higher-level abstract features. By comparing the diagnostic methods, it can be found that compared with other diagnostic algorithms, it has better generalization ability and higher diagnostic accuracy.

[0015] In addition, the permanent magnet synchronous motor demagnetization fault diagnosis method based on optimized variational modal decomposition according to the above embodiment of the present invention may also have the following additional technical features:

[0016] Furthermore, a simulation model of permanent magnet synchronous motor demagnetization fault is established to analyze the mechanism of permanent magnet synchronous motor demagnetization fault, including:

[0017] S1.1, permanent magnet demagnetization is divided into reversible demagnetization and irreversible demagnetization, irreversible demagnetization is further divided into local demagnetization and uniform demagnetization, the flux equation generated in the three-phase winding of the permanent magnet synchronous motor is PMa , PMb , PMc It is expressed as:

[0018]

[0019] Where k is a positive integer, n p is the number of pole pairs, θ is the angle between the permanent magnet air gap axis and the permanent magnet synchronous motor A phase axis, ψ k ln p is the fundamental wave and harmonic flux amplitude generated by the permanent magnet in the three-phase stator winding of PMSM;

[0020] S1.2, the flux equation of the stator winding of the permanent magnet synchronous motor is expressed as:

[0021]

[0022] Among them, M aa 、M ab 、M ac are the mutual inductance of the three-phase windings; L aa , L bb , L cc are the self-inductance of the three-phase windings; fa , fb , fc They are the flux linkages generated by the permanent magnet in the three-wire winding;

[0023] S1.3, the current I of the permanent magnet synchronous motor stator phase in a healthy state A ,I B ,I C It is expressed as:

[0024] I A =Imax cos(2πft)

[0025] I B =I max cos(2πft+2π / 3)

[0026] I C =I max cos(2πft+2π / 3)

[0027] Where, f is the frequency of the stator current; I max is the maximum value of the stator current amplitude;

[0028] S1.4, when a demagnetization fault occurs, local demagnetization will cause electromagnetic torque pulsation of the permanent magnet synchronous motor. Only the fundamental component needs to be considered. The stator phase current I Af ,I Bf ,I Cf It is expressed as:

[0029]

[0030] Among them, p is the number of pole pairs; J is the rotor inertia; Z is the impedance; is the amplitude of the flux linkage; f1 is the frequency of the local demagnetization fault characteristic; T′ is the electromagnetic torque pulsation amplitude of the permanent magnet synchronous motor under the local demagnetization fault; A(t) is the stator current amplitude of the permanent magnet synchronous motor.

[0031] Furthermore, step S2 specifically includes:

[0032] S2.1, the variational mode decomposition algorithm adaptively matches the optimal center frequency and limited bandwidth of each mode to achieve effective separation of the intrinsic mode components and frequency domain division of the signal. The variational mode decomposition is written as a constrained optimization, expressed as:

[0033]

[0034] Among them, μ k (t) represents the kth mode obtained by decomposition; ω k represents the center frequency of the kth mode; δ(t) is the Dirac function; f(t) is the original signal; t is time; j represents the imaginary unit, represents the partial derivative with respect to t;

[0035] S2.2, introduce the quadratic penalty factor α and the Lagrangian factor λ to transform the constrained variational problem into an unconstrained variational problem, and obtain:

[0036]

[0037] Among them, μ kDenote all modal μ k (t), where λ(t) is the Lagrange multiplier;

[0038] S2.3. Use the alternating direction method of multipliers to alternately update μ k and ω k to search for the saddle point of the augmented Lagrangian expression. The final expressions for the IMF components and the central frequencies are:

[0039]

[0040]

[0041] where, represents the updated value of the k-th mode in the (n + 1)-th iteration, represents the updated value of the central frequency of the k-th mode in the (n + 1)-th iteration, represents the representation of the original signal f(t) in the frequency domain ω, represents the representation of the Lagrange factor in the frequency domain ω;

[0042] S2.4. Obtain K modal components through the following steps:

[0043] Initialize {μ k1}}, {ω k2}}, λ1 and n;

[0044] Let n = n + 1 and execute the outer loop program;

[0045] Let k = 1, k = k + 1. When k < K, execute the inner loop and update u k and ω k and λ according to the following formula:

[0046]

[0047] where, ξ is the discrimination accuracy, τ is the noise tolerance limit. When the iteration condition is satisfied, the loop ends to obtain the corresponding modal output. λ n+1 represents the Lagrange multiplier in the (n + 1)-th iteration, and λ n represents the Lagrange multiplier in the n-th iteration;

[0048] S2.5. In the sparrow algorithm, the sparrow population is divided into three categories: discoverers, followers, and scouting early warning birds. Discoverers: Responsible for finding food and providing the foraging area and direction for the entire population; Followers: Rely on discoverers to obtain food; Scouting early warning birds: Responsible for detecting environmental safety and issuing warnings once danger is detected, enabling the entire population to quickly adjust strategies to ensure safety. At the same time, the identities of the three will also change dynamically. The position update formula for discoverers is as follows:

[0049]

[0050] Among them, i represents the i-th individual in the sparrow population; j represents the total position information of the i-th sparrow in the j-dimensional space; The population summarizes the locations of sparrows at the Tth iteration; The location of the sparrows in the population summary when the iteration number is T+1, where T is the current iteration number; max is the maximum number of iterations, α is a random number in (0,1], ST is a safety value in [0.5,1]; R2 is a warning value in [0,1]; L is a multidimensional all-one matrix with one row; Q is a random number;

[0051] The follower position update formula is as follows:

[0052]

[0053] in, is the optimal position of the discoverer; X worst is the current global worst position; X ij is the explicit information of the i-th sparrow in the j-dimensional space; n represents the population size; A is a row of a multidimensional matrix, each element is randomly assigned a value of 1 or -1, A + =A T (AA T ) -1 ;

[0054] The update formula of the reconnaissance early warning position is as follows:

[0055]

[0056] Among them, β is the step control parameter; X best is the current best position; f g is the current global optimal fitness; f i is the fitness value of the i-th sparrow; f w is the current global worst fitness value; K is a random number; ε is a constant to avoid the denominator being zero;

[0057] S2.6, select the minimum envelope entropy as the optimized fitness function, obtain the envelope signal through Hilbert transform, and extract the signal features by combining the information entropy. When the original signal is converted into a probability distribution sequence after Hilbert mediation, its envelope entropy PE is defined as:

[0058]

[0059] Where: a(j) represents the envelope signal sequence obtained by Hilbert adjustment of signal x(j) (j = 1, 2, ... m); m represents the number of data points; H represents the Hilbert transform of the signal; pj is the normalized form of a(j); according to the above principle, the parameters when the envelope entropy is minimum are used as the optimal solution parameters of VMD decomposition;

[0060] S2.7, the specific steps of optimizing the variational mode based on the sparrow search algorithm are as follows:

[0061] (1) Input the original stator current signal of the permanent magnet synchronous motor;

[0062] (2) Initialize the maximum number of iterations of the sparrow search algorithm and the number of sparrow populations, and set the optimization parameter range;

[0063] (3) Decompose the input signal through variational mode, calculate the minimum envelope entropy of each mode and save it;

[0064] (4) Updating the optimal position and fitness value of the sparrow based on the minimum envelope entropy as the fitness function;

[0065] (5) Determine whether the iteration termination condition is met and output the optimal parameters after the termination condition is met;

[0066] (6) Substitute the obtained optimal parameters into the variational modal algorithm, select the best eigenmode component for data reconstruction, and obtain the optimized fault data.

[0067] Furthermore, the construction of the diagnostic model specifically includes:

[0068] The first layer defines a sequence input, and the input dimension is [6, 1, 1];

[0069] The second layer collapses the sequence so that subsequent layers are applied independently at each time step of the sequence;

[0070] The third layer defines a 2D convolutional layer with the specified filter size and number and stride;

[0071] The fourth layer defines a batch normalization layer to speed up training and prevent overfitting;

[0072] The fifth time, a ReLU layer is defined to introduce nonlinearity. The expression is as follows:

[0073]

[0074] The sixth layer defines a two-dimensional maximum pooling layer to reduce the dimension of the data and introduce translation invariance;

[0075] The seventh layer unfolds the sequence back to its original shape after the folding operation;

[0076] The eighth layer converts the multi-dimensional input into one dimension so that it can be passed to the fully connected layer;

[0077] The ninth layer defines a long short-term memory neural network layer, which is used to process sequence data and only takes the output of the last time step;

[0078] The tenth layer defines a Dropout layer, which is used to randomly discard some neurons during training to prevent overfitting;

[0079] The eleventh layer defines a fully connected layer, and the output dimension is the number of fault types;

[0080] The twelfth layer defines a softmax layer to convert the output of the network into a probability distribution;

[0081] The thirteenth layer defines an output layer for multi-classification problems.

[0082] Furthermore, the training of the diagnostic model includes the following steps:

[0083] S5.1, dividing the obtained time domain feature signal into a training set and a test set in a ratio of 1:4, wherein the training set is used for training the diagnostic model;

[0084] S5.2, use the Adam optimization algorithm to train the network;

[0085] S5.3, set the maximum number of training times to 100;

[0086] S5.4, set the initial learning rate to 0.001, and use the piecewise constant learning rate adjustment strategy to reduce the learning rate after every 50 training times, the new learning rate = current learning rate * 0.5;

[0087] S5.5, uses L2 regularization with a weight decay coefficient of 0.001, which helps prevent overfitting;

[0088] S5.6, the training is divided into the weight initialization stage, the forward propagation stage, and the back propagation stage;

[0089] S5.7, randomly initialize weights during the initialization phase;

[0090] S5.8, in the forward propagation stage, it will pass through the hidden layer and finally reach the output layer and generate a predicted value. During the forward propagation process, the weights and biases of the network remain unchanged, and the input data is nonlinearly transformed through ReLU;

[0091] S5.9 After the forward propagation calculates the predicted value and loss function, the back propagation algorithm calculates the gradient of the loss function relative to the network weights and biases, starting from the output layer and moving forward layer by layer until the input layer. The calculated gradient is then used to update the network weights and biases through Adam, so that the loss function can be smaller in the next forward propagation, and the process is repeated until the preset maximum number of training times is reached.

[0092] Furthermore, the method further comprises:

[0093] S6, compares various fault diagnosis algorithms using simulation results, including:

[0094] S6.1, putting the obtained test set data into the convolutional long short-term memory neural network diagnosis model for fault diagnosis, and obtaining the diagnosis result output by the diagnosis model;

[0095] S6.2, the comparison group is divided into three groups. The first group of diagnostic methods uses the traditional variational modal algorithm to perform noise reduction on the collected data, extracts time domain features from the noise reduction data obtained by this method, and then uses the convolutional long short-term memory neural network diagnostic model to perform fault diagnosis; the second group of diagnostic methods does not perform any noise reduction on the collected data, directly extracts time domain features, and uses the convolutional long short-term memory neural network diagnostic model to perform fault diagnosis; the third group of data does not use any noise reduction, and directly uses the convolutional neural network diagnostic model to perform fault diagnosis;

[0096] S6.3, compare and analyze the results of multiple diagnostic methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 A flowchart of a method for diagnosing a turn-to-turn short-circuit fault of a permanent magnet synchronous motor based on deep learning according to the present invention;

[0098] Figure 2 It is a simulation model diagram of the present invention;

[0099] Figure 3 It is a schematic diagram of simulation data of the present invention;

[0100] Figure 4 This is a flow chart of optimizing variational modes based on the sparrow search algorithm of the present invention;

[0101] Figure 5 This is a data graph after noise reduction processing of the present invention;

[0102] Figure 6 It is the feature extraction result diagram of the present invention;

[0103] Figure 7 It is the diagnostic result diagram of the present invention;

[0104] Figure 8 This is the diagnostic result diagram of the first comparison group;

[0105] Fig. 9 This is the diagnostic result diagram of the second comparison group;

[0106] Fig.10 This is the diagnostic result diagram of the third comparison group;

[0107] Fig.11 This is a comparison chart of diagnostic accuracy results. DETAILED DESCRIPTION

[0108] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0109] Please combine Figure 1 The present invention provides a method for diagnosing a turn-to-turn short-circuit fault of a permanent magnet synchronous motor based on deep learning, comprising steps S1 to S5:

[0110] S1, fault analysis and signal selection, select the built-in permanent magnet synchronous motor as the verification object, establish a permanent magnet synchronous motor demagnetization fault simulation model, analyze the mechanism of the permanent magnet synchronous motor demagnetization fault, further analyze the change law of the magnetic flux and stator current of the permanent magnet synchronous motor during the irreversible demagnetization process, simulate the demagnetization fault of the permanent magnet synchronous motor, and collect the stator current data obtained by simulation as the fault signal.

[0111] Step S1 specifically includes:

[0112] S1.1, permanent magnet demagnetization is divided into reversible demagnetization and irreversible demagnetization, irreversible demagnetization is further divided into local demagnetization and uniform demagnetization, the flux equation generated in the three-phase winding of the permanent magnet synchronous motor is PMa , PMb , PMc It is expressed as:

[0113]

[0114]

[0115] Where k is a positive integer, n p is the number of pole pairs, θ is the angle between the permanent magnet air gap axis and the permanent magnet synchronous motor A phase axis, ψ k ln pis the fundamental wave and harmonic flux amplitude generated by the permanent magnet in the three-phase stator winding of PMSM.

[0116] S1.2, the flux equation of the stator winding of the permanent magnet synchronous motor can be expressed as

[0117]

[0118] Among them, M aa 、M ab 、M ac are the mutual inductance of the three-phase windings; L aa , L bb , L cc are the self-inductance of the three-phase windings; fa , fb , fc They are the flux linkage generated by the permanent magnet in the three-wire winding. When a demagnetization fault occurs in the permanent magnet synchronous motor, the applied excitation remains unchanged, the flux linkage generated by the permanent magnet decreases, and the stator current increases.

[0119] S1.3, the current I of the permanent magnet synchronous motor stator phase in a healthy state A ,I B ,I C It is expressed as:

[0120] I A =I max cos(2πft)

[0121] I B =I max cos(2πft-2π / 3)

[0122] I C =I max cos(2πft+2π / 3)

[0123] Where, f is the frequency of the stator current; I max is the maximum value of the stator current amplitude;

[0124] S1.4, when demagnetization fault occurs, local demagnetization will cause electromagnetic torque pulsation of the permanent magnet synchronous motor. Since the amplitude of the higher harmonics of the characteristic frequency of the electromagnetic torque pulsation fault caused by demagnetization is small and negligible during the torque pulsation, only the fundamental component needs to be considered. Therefore, only the fundamental component needs to be considered. The stator phase current I Af ,I Bf ,I Cf It can be expressed as:

[0125]

[0126]

[0127] Among them, p is the number of pole pairs; J is the rotor inertia; Z is the impedance; is the amplitude of the flux linkage; f1 is the frequency of the local demagnetization fault characteristic; T′ is the electromagnetic torque pulsation amplitude of the permanent magnet synchronous motor under the local demagnetization fault; A(t) is the stator current amplitude of the permanent magnet synchronous motor.

[0128] Among them, a permanent magnet synchronous motor simulation model is built, such as Figure 2 The demagnetization fault of the permanent magnet synchronous motor is simulated by simulation software, and six demagnetization fault schemes are set and numbered as D1, D2, D3, D4, D5, and D6, as shown in Table 1:

[0129] Table 1

[0130]

[0131] Add 30db noise to the obtained simulation data, and the final simulation data is as follows Figure 3 shown.

[0132] S2, data denoising, using the sparrow search algorithm to optimize the penalty factor and decomposition layer number in the variational modal algorithm. In the variational modal parameter optimization process, the minimum value of the envelope entropy is used as the fitness function for parameter optimization to obtain the optimal solution for different fault signals. The optimal decomposition layer number and penalty factor obtained by optimization are substituted into the variational modal algorithm to perform modal decomposition on the data, and the mode with the minimum envelope entropy is selected for data reconstruction.

[0133] Step S2 specifically includes:

[0134] S2.1, the variational mode decomposition algorithm adaptively matches the optimal center frequency and limited bandwidth of each mode to achieve effective separation of the intrinsic mode components and frequency domain division of the signal. The variational mode decomposition can be written as a constrained optimization, which can be expressed as:

[0135]

[0136] Among them, μ k (t) represents the kth mode obtained by decomposition; ω k represents the center frequency of the kth mode; δ(t) is the Dirac function; f(t) is the original signal; t is time; j represents the imaginary unit, represents the partial derivative with respect to t;

[0137] S2.2, introduce the quadratic penalty factor α and the Lagrangian factor λ to transform the constrained variational problem into an unconstrained variational problem, and obtain:

[0138]

[0139] Among them, μ k represents the set of all modes μ k (t), and λ(t) is the Lagrange multiplier;

[0140] S2.3. Use the alternating direction method of multipliers to alternately update μ k and ω k to search for the saddle point of the augmented Lagrangian expression. The final obtained IMF components and the expressions of the center frequencies are:

[0141]

[0142] Among them, represents the updated value of the k-th mode in the (n + 1)-th iteration, represents the updated value of the center frequency of the k-th mode in the (n + 1)-th iteration, represents the representation of the original signal f(t) in the frequency domain ω, represents the representation of the Lagrange factor in the frequency domain ω;

[0143] S2.4. Obtain K mode components through the following steps:

[0144] Initialize {μ k1}, {ω k2}, λ1 and n;

[0145] Let n = n + 1 and execute the outer loop program;

[0146] Let k = 1, k = k + 1. When k < K, execute the inner loop, and update u k and ω k and λ according to the following formula:

[0147]

[0148] Among them, ξ is the discrimination accuracy, τ is the noise tolerance limit. When the iteration condition is satisfied, the loop ends to obtain the corresponding mode output. λ n+1 represents the Lagrange multiplier in the (n + 1)-th iteration, and λ n represents the Lagrange multiplier in the n-th iteration;

[0149] S2.5. In the sparrow algorithm, the sparrow population is divided into three categories: discoverers, followers, and scouting early warning birds. Discoverers: Responsible for finding food and providing the foraging area and direction for the entire population; Followers: Rely on discoverers to obtain food; Scouting early warning birds: Responsible for detecting environmental safety and will issue a warning once danger is found, enabling the entire population to quickly adjust strategies to ensure safety. At the same time, their identities will also change dynamically. The position update formula for discoverers is as follows:

[0150]

[0151] Among them, i represents the i-th individual in the sparrow population; j represents the total position information of the i-th sparrow in the j-dimensional space; The population summarizes the locations of sparrows at the Tth iteration; The location of the sparrows in the population summary when the iteration number is T+1, where T is the current iteration number; max is the maximum number of iterations, α is a random number in (0,1], ST is a safety value in [0.5,1]; R2 is a warning value in [0,1]; L is a multidimensional all-one matrix with one row; Q is a random number;

[0152] The follower position update formula is as follows:

[0153]

[0154] in, is the optimal position of the finder; X worst is the current global worst position; X ij is the explicit information of the i-th sparrow in the j-dimensional space; n represents the population size; A is a row of a multidimensional matrix, each element is randomly assigned a value of 1 or -1, A + =A T (AA T ) -1 ;

[0155] The update formula of the reconnaissance early warning position is as follows:

[0156]

[0157] Among them, β is the step control parameter; X best is the current best position; f g is the current global optimal fitness; f i is the fitness value of the i-th sparrow; f w is the current global worst fitness value; K is a random number; ε is a constant to avoid the denominator being zero.

[0158] S2.6, select the minimum envelope entropy as the optimized fitness function, obtain the envelope signal through Hilbert transform, and extract the signal features by combining the information entropy. When the original signal is converted into a probability distribution sequence after Hilbert mediation, its envelope entropy PE is defined as:

[0159]

[0160] Where: a(j) represents the envelope signal sequence obtained by Hilbert adjustment of signal x(j) (j = 1, 2, ... m); m represents the number of data points; H represents the Hilbert transform of the signal; pj is the normalized form of a(j); according to the above principle, the parameters when the envelope entropy is minimum are used as the optimal solution parameters of VMD decomposition;

[0161] S2.7, please combine Figure 4 ,The specific steps of optimizing variational modes based on the sparrow search algorithm are as follows:

[0162] (1) Input the original stator current signal of the permanent magnet synchronous motor;

[0163] (2) Initialize the maximum number of iterations of the sparrow search algorithm and the number of sparrow populations, and set the optimization parameter range. The initialization parameters of the sparrow search algorithm are: population size 30, maximum number of iterations 20, dimension 2, the range of decomposition layers is [3, 10], and the range of penalty factors is [500, 2500];

[0164] (3) Decompose the input signal through variational mode, calculate the minimum envelope entropy of each mode and save it;

[0165] (4) Updating the optimal position and fitness value of the sparrow based on the minimum envelope entropy as the fitness function;

[0166] (5) Determine whether the iteration termination condition is met and output the optimal parameters after the termination condition is met;

[0167] (6) Substitute the obtained optimal parameters into the variational modal algorithm, select the best eigenmode component for data reconstruction, and obtain the optimized fault data.

[0168] The simulation data is modally decomposed using the optimized variational modal algorithm.

[0169] After noise reduction, the data is Figure 5 shown.

[0170] S3, time domain feature extraction, extracts 6 time domain features including mean, variance, peak, kurtosis, effective value and peak factor from the denoised data, and uses the extracted feature vector as the input data of the fault diagnosis model.

[0171] Step S3 further extracts time domain features from the denoised data, including the following steps:

[0172] The collected stator current data are subjected to six time domain feature extractions, including mean, variance, peak value, kurtosis, effective value, and peak factor.

[0173] The feature extraction results of the six different demagnetization fault data obtained by simulation are as follows Figure 6 shown.

[0174] S4, diagnostic model construction, which combines convolutional neural network and long short-term memory network. The convolutional neural network extracts key spatial features from the input data through convolution and pooling operations. The long short-term memory network is a variant of the recurrent neural network and is used to process sequence data with long-term dependencies in fault diagnosis.

[0175] The construction of the diagnostic model specifically includes:

[0176] The first layer defines a sequence input, and the input dimension is [6, 1, 1];

[0177] The second layer collapses the sequence so that subsequent layers are applied independently at each time step of the sequence;

[0178] The third layer defines a 2D convolutional layer with the specified filter size and number and stride;

[0179] The fourth layer defines a batch normalization layer to speed up training and prevent overfitting;

[0180] The fifth time, a ReLU layer is defined to introduce nonlinearity. The expression is as follows:

[0181]

[0182] The sixth layer defines a two-dimensional maximum pooling layer to reduce the dimension of the data and introduce translation invariance;

[0183] The seventh layer unfolds the sequence back to its original shape after the folding operation;

[0184] The eighth layer converts the multi-dimensional input into one dimension so that it can be passed to the fully connected layer;

[0185] The ninth layer defines a long short-term memory neural network layer, which is used to process sequence data and only takes the output of the last time step;

[0186] The tenth layer defines a Dropout layer, which is used to randomly discard some neurons during training to prevent overfitting;

[0187] The eleventh layer defines a fully connected layer, and the output dimension is the number of fault types;

[0188] The twelfth layer defines a softmax layer to convert the output of the network into a probability distribution;

[0189] The thirteenth layer defines an output layer for multi-classification problems.

[0190] S5, model training and fault diagnosis, divides the obtained input data into training set samples and test set samples, where the training set samples are used for model learning. The diagnostic model will extract features from the input data, and through training, the diagnostic model can adjust its internal parameters such as weights and biases to minimize the error between the predicted value and the actual value. The test sample set is used to evaluate the performance of the trained diagnostic model on unseen data. The trained model is used to diagnose the demagnetization fault of the permanent magnet synchronous motor to obtain the diagnostic result.

[0191] Among them, the diagnostic model training includes the following steps:

[0192] S5.1, dividing the obtained time domain feature signal into a training set and a test set in a ratio of 1:4, wherein the training set is used for training the diagnostic model;

[0193] S5.2, use the Adam optimization algorithm to train the network;

[0194] S5.3, set the maximum number of training times to 100;

[0195] S5.4, set the initial learning rate to 0.001, and use the piecewise constant learning rate adjustment strategy to reduce the learning rate after every 50 training times, the new learning rate = current learning rate * 0.5;

[0196] S5.5, uses L2 regularization with a weight decay coefficient of 0.001, which helps prevent overfitting;

[0197] S5.6, the training is divided into the weight initialization stage, the forward propagation stage, and the back propagation stage;

[0198] S5.7, randomly initialize weights during the initialization phase;

[0199] S5.8, in the forward propagation stage, it will pass through the hidden layer and finally reach the output layer and generate a predicted value. During the forward propagation process, the weights and biases of the network remain unchanged, and the input data is nonlinearly transformed through ReLU;

[0200] S5.9 After the forward propagation calculates the predicted value and loss function, the back propagation algorithm calculates the gradient of the loss function relative to the network weights and biases, starting from the output layer and moving forward layer by layer until the input layer. The calculated gradient is then used to update the network weights and biases through Adam, so that the loss function can be smaller in the next forward propagation, and the process is repeated until the preset maximum number of training times is reached.

[0201] The diagnostic results of the present invention are as follows Figure 7 shown.

[0202] In addition, in this embodiment, the method further includes:

[0203] S6, compares various fault diagnosis algorithms using simulation results, including:

[0204] Step 6: Result verification and comparative analysis includes the following steps:

[0205] S6.1, putting the obtained test set data into the convolutional long short-term memory neural network diagnosis model for fault diagnosis, and obtaining the diagnosis result output by the diagnosis model;

[0206] S6.2, the comparison group is divided into three groups. The diagnostic method of the first comparison group is to use the traditional variational mode algorithm to perform noise reduction processing on the collected data, extract time domain features from the noise reduction data obtained by this method, and then perform fault diagnosis through the convolutional long short-term memory neural network diagnostic model; the diagnostic method of the second comparison group is to directly perform time domain feature extraction on the collected data without any noise reduction processing, and perform fault diagnosis through the convolutional long short-term memory neural network diagnostic model; the third comparison group does not take any noise reduction processing, and directly uses the convolutional neural network diagnostic model for fault diagnosis;

[0207] S6.3, compare and analyze the results of various diagnostic methods. Figures 8 to 11 As shown, it can be seen from the combination that the accuracy of the diagnostic method of the present invention is significantly better than other methods.

[0208] In summary, the permanent magnet synchronous motor demagnetization fault diagnosis method based on optimized variational mode decomposition according to the embodiment of the present invention has the following beneficial effects:

[0209] (1) The number of decomposition layers and penalty factors in the variational mode decomposition are optimized through the sparrow search algorithm. Compared with the unoptimized variational mode decomposition, there is no need to manually determine the number of decomposition layers and penalty factors, which realizes the automatic selection of parameters, reduces the influence of human subjectivity on the decomposition effect, effectively avoids the mode aliasing phenomenon, and improves the decomposition performance and stability of the variational mode decomposition.

[0210] (2) The original data is denoised using the optimized variational mode algorithm, which highlights the fault characteristics in the data and is beneficial to the diagnostic effect of the subsequent algorithm.

[0211] (3) The convolutional long short-term memory neural network used in this paper combines the advantages of convolutional neural networks and long short-term memory neural networks. By training, the network can automatically learn the deep features of the input data and represent them as higher-level abstract features. By comparing the diagnostic methods, it can be found that compared with other diagnostic algorithms, it has better generalization ability and higher diagnostic accuracy.

[0212] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0213] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A permanent magnet synchronous motor demagnetization fault diagnosis method based on optimized variational mode decomposition, characterized in that: The method comprises steps S1 to S5: S1, fault analysis and signal selection, select the built-in permanent magnet synchronous motor as the verification object, establish a permanent magnet synchronous motor demagnetization fault simulation model, analyze the mechanism of the permanent magnet synchronous motor demagnetization fault, further analyze the change law of the magnetic flux and stator current of the permanent magnet synchronous motor during the irreversible demagnetization process, simulate the demagnetization fault of the permanent magnet synchronous motor, and collect the stator current data obtained by simulation as the fault signal; S2, data denoising, using the sparrow search algorithm to optimize the penalty factor and decomposition layer number in the variational modal algorithm. In the variational modal parameter optimization process, the minimum value of the envelope entropy is used as the fitness function to optimize the parameters, and the optimal solution for different fault signals is obtained. The optimal decomposition layer number and penalty factor obtained by optimization are substituted into the variational modal algorithm to perform modal decomposition on the data, and the mode with the minimum envelope entropy is selected for data reconstruction; S3, time domain feature extraction, extracting six time domain features including mean, variance, peak, kurtosis, effective value and peak factor from the denoised data, and using the extracted feature vector as input data for the fault diagnosis model; S4, diagnostic model construction, which combines convolutional neural network and long short-term memory network, where convolutional neural network extracts key spatial features from input data through convolution and pooling operations, and long short-term memory network is a variant of recurrent neural network, which is used to process sequence data with long-term dependencies in fault diagnosis; S5, model training and fault diagnosis, divides the obtained input data into training set samples and test set samples, where the training set samples are used for model learning. The diagnostic model will extract features from the input data, and through training, the diagnostic model can adjust its internal parameters such as weights and biases to minimize the error between the predicted value and the actual value. The test sample set is used to evaluate the performance of the trained diagnostic model on unseen data. The trained model is used to diagnose the demagnetization fault of the permanent magnet synchronous motor to obtain the diagnostic result.

2. The permanent magnet synchronous motor demagnetization fault diagnosis method based on optimized variational modal decomposition according to claim 1 is characterized in that: A simulation model of permanent magnet synchronous motor demagnetization fault is established to analyze the mechanism of permanent magnet synchronous motor demagnetization fault, including: S1.1, permanent magnet demagnetization is divided into reversible demagnetization and irreversible demagnetization, irreversible demagnetization is further divided into local demagnetization and uniform demagnetization, the flux equation generated in the three-phase winding of the permanent magnet synchronous motor is PMa , PMb , PMc It is expressed as: Where k is a positive integer, n p is the number of pole pairs, θ is the angle between the permanent magnet air gap axis and the permanent magnet synchronous motor A phase axis, ψ k ln p is the fundamental wave and harmonic flux amplitude generated by the permanent magnet in the three-phase stator winding of PMSM; S1.2, the flux equation of the stator winding of the permanent magnet synchronous motor is expressed as: Among them, M aa 、M ab 、M ac are the mutual inductance of the three-phase windings; L aa , L bb , L cc are the self-inductance of the three-phase windings; fa , fb , fc They are the flux linkages generated by the permanent magnet in the three-wire winding; S1.3, the current I of the permanent magnet synchronous motor stator phase in a healthy state A ,I B ,I C It is expressed as: I A =I max cos(2πft) I B =I max cos(2πft-2π / 3) I C =I max cos(2πft+2π / 3) Where, f is the frequency of the stator current; I max is the maximum value of the stator current amplitude; S1.4, when a demagnetization fault occurs, local demagnetization will cause electromagnetic torque pulsation of the permanent magnet synchronous motor. Only the fundamental component needs to be considered. The stator phase current I Af ,I Bf ,I Cf It is expressed as: Among them, p is the number of pole pairs; J is the rotor inertia; Z is the impedance; is the amplitude of the flux linkage; f1 is the frequency of the local demagnetization fault characteristic; T′ is the electromagnetic torque pulsation amplitude of the permanent magnet synchronous motor under the local demagnetization fault; A(t) is the stator current amplitude of the permanent magnet synchronous motor.

3. The permanent magnet synchronous motor demagnetization fault diagnosis method based on optimized variational modal decomposition according to claim 2 is characterized in that: Step S2 specifically includes: S2.1, the variational mode decomposition algorithm adaptively matches the optimal center frequency and limited bandwidth of each mode to achieve effective separation of the intrinsic mode components and frequency domain division of the signal. The variational mode decomposition is written as a constrained optimization, expressed as: Among them, μ k (t) represents the kth mode obtained by decomposition; ω k represents the center frequency of the kth mode; δ(t) is the Dirac function; f(t) is the original signal; t is time; j represents the imaginary unit, represents the partial derivative with respect to t; S2.2, introduce the quadratic penalty factor α and the Lagrangian factor λ to transform the constrained variational problem into an unconstrained variational problem, and obtain: Among them, μ k Represents all modes μ k (t), λ(t) is the Lagrange multiplier; S2.3, using the alternating direction multiplier method to alternately update μ k and ω k To search for the saddle point of the augmented Lagrangian expression, the expressions of the IMF components and center frequencies are finally obtained: in, represents the updated value of the kth mode in the n+1th iteration, represents the updated value of the kth modal center frequency in the n+1th iteration, represents the representation of the original signal f(t) in the frequency domain ω, represents the representation of the Lagrangian factor in the frequency domain ω; S2.4, obtain K modal components by the following steps: Initialize {μ k1 },{ω k2 }, λ1 and n; Set n=n+1 and execute the outer loop program; Set \(k = 1\), \(k=k + 1\), execute the inner loop when \(k < K\), and update \(u\) according to the following formula k and \(\omega\) k and \(\lambda\): Among them, ξ is the discrimination accuracy, τ is the noise tolerance limit, and when the iteration condition is met, the loop ends and the corresponding modal output is obtained. n+1 represents the Lagrange multiplier in the n+1th iteration, λ n represents the Lagrange multiplier in the nth iteration; S2.5, in the sparrow algorithm, the sparrow population is divided into three categories: discoverers, followers, and scouts and early warnings; discoverers: responsible for finding food and providing foraging areas and directions for the entire population; followers: rely on discoverers to obtain food; scouts and early warnings: responsible for detecting environmental safety, once danger is found, they will issue a warning so that the entire population can quickly adjust strategies to ensure safety; at the same time, the identities of the three will also change dynamically, and the formula for updating the discoverer's position is as follows: Among them, i represents the i-th individual in the sparrow population; j represents the total position information of the i-th sparrow in the j-dimensional space; The population summarizes the locations of sparrows at the Tth iteration; The location of the sparrows in the population summary when the iteration number is T+1, where T is the current iteration number; max is the maximum number of iterations, α is a random number in (0,1], ST is a safety value in [0.5,1]; R2 is a warning value in [0,1]; L is a multidimensional all-one matrix with one row; Q is a random number; The follower position update formula is as follows: in, is the optimal position of the finder; X worst is the current global worst position; X ij is the explicit information of the i-th sparrow in the j-dimensional space; n represents the population size; A is a row of a multidimensional matrix, each element is randomly assigned a value of 1 or -1, A + =A T (AA T ) -1 ; The update formula of the reconnaissance early warning position is as follows: Among them, β is the step control parameter; X best is the current best position; f g is the current global optimal fitness; f i is the fitness value of the i-th sparrow; f w is the current global worst fitness value; K is a random number; ε is a constant to avoid the denominator being zero; S2.6, select the minimum envelope entropy as the optimized fitness function, obtain the envelope signal through Hilbert transform, and extract the signal features by combining the information entropy. When the original signal is converted into a probability distribution sequence after Hilbert mediation, its envelope entropy PE is defined as: Where: a(j) represents the envelope signal sequence obtained by Hilbert adjustment of signal x(j) (j = 1, 2, ... m); m represents the number of data points; H represents the Hilbert transform of the signal; p j is the normalized form of a(j); according to the above principle, the parameters when the envelope entropy is minimum are used as the optimal solution parameters of VMD decomposition; S2.7, the specific steps of optimizing the variational mode based on the sparrow search algorithm are as follows: (1) Input the original stator current signal of the permanent magnet synchronous motor; (2) Initialize the maximum number of iterations of the sparrow search algorithm and the number of sparrow populations, and set the optimization parameter range; (3) Decompose the input signal through variational mode, calculate the minimum envelope entropy of each mode and save it; (4) Updating the optimal position and fitness value of the sparrow based on the minimum envelope entropy as the fitness function; (5) Determine whether the iteration termination condition is met and output the optimal parameters after the termination condition is met; (6) Substitute the obtained optimal parameters into the variational modal algorithm, select the best eigenmode component for data reconstruction, and obtain the optimized fault data.

4. The permanent magnet synchronous motor demagnetization fault diagnosis method based on optimized variational modal decomposition according to claim 3 is characterized in that: The construction of the diagnostic model specifically includes: The first layer defines a sequence input, and the input dimension is [6, 1, 1]; The second layer collapses the sequence so that subsequent layers are applied independently at each time step of the sequence; The third layer defines a 2D convolutional layer with the specified filter size and number and stride; The fourth layer defines a batch normalization layer to speed up training and prevent overfitting; The fifth time, a ReLU layer is defined to introduce nonlinearity. The expression is as follows: The sixth layer defines a two-dimensional maximum pooling layer to reduce the dimension of the data and introduce translation invariance; The seventh layer unfolds the sequence back to its original shape after the folding operation; The eighth layer converts the multi-dimensional input into one dimension so that it can be passed to the fully connected layer; The ninth layer defines a long short-term memory neural network layer, which is used to process sequence data and only takes the output of the last time step; The tenth layer defines a Dropout layer, which is used to randomly discard some neurons during training to prevent overfitting; The eleventh layer defines a fully connected layer, and the output dimension is the number of fault types; The twelfth layer defines a softmax layer to convert the output of the network into a probability distribution; The thirteenth layer defines an output layer for multi-classification problems.

5. The permanent magnet synchronous motor demagnetization fault diagnosis method based on optimized variational modal decomposition according to claim 4 is characterized in that: The training of the diagnostic model consists of the following steps: S5.1, dividing the obtained time domain feature signal into a training set and a test set in a ratio of 1:4, wherein the training set is used for training the diagnostic model; S5.2, use the Adam optimization algorithm to train the network; S5.3, set the maximum number of training times to 100; S5.4, set the initial learning rate to 0.001, and use the piecewise constant learning rate adjustment strategy to reduce the learning rate after every 50 training times, the new learning rate = current learning rate * 0.5; S5.5, uses L2 regularization with a weight decay coefficient of 0.001, which helps prevent overfitting; S5.6, the training is divided into the weight initialization stage, the forward propagation stage, and the back propagation stage; S5.7, randomly initialize weights during the initialization phase; S5.8, in the forward propagation stage, it will pass through the hidden layer and finally reach the output layer and generate a predicted value. During the forward propagation process, the weights and biases of the network remain unchanged, and the input data is nonlinearly transformed through ReLU; S5.9 After the forward propagation calculates the predicted value and loss function, the back propagation algorithm calculates the gradient of the loss function relative to the network weights and biases, starting from the output layer and moving forward layer by layer until the input layer. The calculated gradient is then used to update the network weights and biases through Adam, so that the loss function can be smaller in the next forward propagation, and the process is repeated until the preset maximum number of training times is reached.

6. The method for diagnosing demagnetization fault of a permanent magnet synchronous motor based on optimized variational modal decomposition according to claim 5, characterized in that: The method further comprises: S6, compares various fault diagnosis algorithms using simulation results, including: S6.1, putting the obtained test set data into the convolutional long short-term memory neural network diagnosis model for fault diagnosis, and obtaining the diagnosis result output by the diagnosis model; S6.2, the comparison group is divided into three groups. The first group of diagnostic methods uses the traditional variational modal algorithm to perform noise reduction on the collected data, extracts time domain features from the noise reduction data obtained by this method, and then uses the convolutional long short-term memory neural network diagnostic model to perform fault diagnosis; the second group of diagnostic methods does not perform any noise reduction on the collected data, directly extracts time domain features, and uses the convolutional long short-term memory neural network diagnostic model to perform fault diagnosis; the third group of data does not use any noise reduction, and directly uses the convolutional neural network diagnostic model to perform fault diagnosis; S6.3, compare and analyze the results of multiple diagnostic methods.

Citation Information

Cited By

  • Motor fault prediction method and device, electronic equipment and storage medium

    CN118566718A

  • Motor fault prediction method and device, electronic equipment and storage medium

    CN118566718B

  • Fault feature extraction method of flexible hand for deep submersible vehicle

    CN119337063A

  • CNN-based power measuring instrument short circuit fault automatic diagnosis method

    CN119577586A

  • Demagnetization fault detection method and system based on high-precision torque distribution

    CN120949042A