Motor turn-to-turn short circuit fault detection method based on variational mode decomposition and deep learning

Abstract

The present invention relates to a motor turn-to-turn short-circuit fault detection method based on variational modal decomposition fused with deep learning, which solves the defects of low accuracy and slow efficiency of turn-to-turn short-circuit fault diagnosis compared with the prior art. The present invention includes the following steps: establishing a permanent magnet synchronous motor fault model; extracting fault features by variational modal decomposition algorithm; optimizing variational modal decomposition parameters by sparrow search algorithm; establishing a deep pyramid pooling residual convolutional neural network model; training of SPP-ResCNN network; real-time diagnosis of permanent magnet synchronous motor turn-to-turn short-circuit fault. The present invention uses sparrow search algorithm to optimize variational modal decomposition algorithm, eliminates other harmonic components in fault features, extracts fault features of specific frequencies, and significantly improves the accuracy of fault features. At the same time, pyramid pooling network and residual network are used to optimize convolutional neural network, which improves the accuracy and efficiency of fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of permanent magnet synchronous motor fault diagnosis, and specifically to a motor turn-to-turn short circuit fault detection method based on variational mode decomposition fused with deep learning. Background Art

[0002] Permanent magnet synchronous motor is the core component of the electric drive system of new energy vehicles, and its working reliability directly affects the safety of new energy vehicles. Since the motor runs in a high temperature and high pressure environment for a long time, the coil is very easy to age, which in turn causes inter-turn short circuit faults. When a fault occurs, if the short-circuit current in the fault circuit is not detected in time, the short-circuit current will generate a lot of heat to further damage the coil insulation, causing partial burning of the motor, which will threaten the life and property safety of the driver. Therefore, the research on the diagnosis of inter-turn short circuit faults in new energy vehicle motors is of great significance.

[0003] The fault diagnosis of inter-turn short circuit in new energy vehicle motors is mostly based on the state parameters caused by the fault to determine whether the fault has occurred. However, the fault characteristics of inter-turn short circuit are not obvious in the early stage, and the motor runs unsteadily after the fault occurs and is disturbed by other harmonics such as vibration and noise, resulting in low fault feature accuracy, which seriously affects the accuracy of fault diagnosis.

[0004] The signal processing methods currently used, such as Fourier transform and wavelet transform, have the problem of high bandwidth of extracted signals and inability to remove noise harmonics within the characteristic frequency range, resulting in low accuracy. Traditional convolutional neural networks have defects such as insufficient feature extraction and excessive number of neuron nodes leading to the disappearance of network gradients. Summary of the invention

[0005] The purpose of the present invention is to solve the defects of low accuracy and low efficiency in the diagnosis of turn-to-turn short-circuit faults in the prior art, and to provide a motor turn-to-turn short-circuit fault detection method based on variational mode decomposition fusion deep learning to solve the above problems.

[0006] In order to achieve the above object, the technical solution of the present invention is as follows:

[0007] A motor turn-to-turn short circuit fault detection method based on variational mode decomposition fusion deep learning includes the following steps:

[0008] 11) Establishing a permanent magnet synchronous motor fault model: According to the voltage and current equations in the natural coordinate system of the permanent magnet synchronous motor, the equations of the d-axis and q-axis voltages and the d-axis and q-axis currents in the rotating coordinate system are established, and this is used as a permanent magnet synchronous motor fault model simulating a short circuit between fault phase turns;

[0009] 12) Extract fault features using variational modal decomposition algorithm: According to the simulation results of the permanent magnet synchronous motor fault model, the motor fault phase current and zero-sequence voltage are obtained, and they are respectively input into the variational modal decomposition algorithm for adaptive modal decomposition to obtain the third harmonic of the fault phase current and the fundamental wave of the zero-sequence voltage as the turn-to-turn short-circuit fault features;

[0010] 13) Sparrow search algorithm to optimize variational mode decomposition parameters: The penalty factor α and the number of decomposition levels K in the variational mode decomposition process are optimized within a given range by the sparrow search algorithm to obtain the optimal decomposition results of the third harmonic of the fault phase current and the zero-sequence voltage fundamental wave;

[0011] 14) Establish a deep pyramid pooling residual convolutional neural network model: establish a deep pyramid pooling network and a residual convolutional neural network, use the deep pyramid pooling network to optimize the residual convolutional neural network model, namely the SPP-ResCNN network, and adjust the structure and parameters of the fused SPP-ResCNN network through ablation experiments;

[0012] 15) Training of SPP-ResCNN network: The optimal decomposition results of the third harmonic of the fault phase current and the zero-sequence voltage fundamental obtained by optimizing the variational modal decomposition using the sparrow search algorithm are made into training sets and test sets, which are input into the SPP-ResCNN network for training. The learning rate, maximum number of training rounds, and minimum training batches are adjusted through ablation experiments to minimize the network loss function.

[0013] 16) Real-time diagnosis of permanent magnet synchronous motor inter-turn short-circuit fault: Real-time monitoring of permanent magnet synchronous motor current and zero-sequence voltage, input them into the variational mode decomposition algorithm optimized by the sparrow search algorithm to extract the optimal decomposition results of the third harmonic of the fault phase current and the zero-sequence voltage fundamental, and make them into a data set to be tested, and input them into the trained SPP-ResCNN network to diagnose the degree of inter-turn short-circuit fault.

[0014] The establishment of the permanent magnet synchronous motor fault model comprises the following steps:

[0015] 21) Set the voltage u of phase A, phase B and phase C of the permanent magnet synchronous motor in the natural coordinate system a 、u b 、u c With phase A, phase B, phase C current i a 、i b 、i c and the fault current i f The equation between is as follows:

[0016]

[0017] Among them, u a 、u b 、uc is the voltage of phase A, phase B, and phase C, η is the fault degree, and η is defined as the ratio of the short-circuit turns of the stator winding of phase A to the total turns of its stator winding, R s is the stator resistance, R f is the fault resistance, ψ f is the fundamental amplitude of the flux, θ is the rotor electrical angle, i a 、i b 、i c 、i f are the current and fault current of the stator winding of phase A, phase B, and phase C respectively, L and M are the self-inductance and mutual inductance of the stator winding respectively;

[0018] 22) The voltage equation in the natural coordinate system is transformed into the rotating coordinate system to obtain the permanent magnet synchronous motor fault model, which is convenient for extracting the inter-turn short circuit fault characteristics. The coordinate transformation matrix T is as follows:

[0019]

[0020] The inverse matrix of T is:

[0021]

[0022] Applying the coordinate transformation matrix T to the voltage and current equations in the natural coordinate system of the motor yields:

[0023]

[0024] Where:

[0025] [u dq0f ]=[u d u q -u n 0] T

[0026] [i dq0f ]=[i d i q 0 i f ] T

[0027] [ψ dq0f ]=[ψ f 0 ψ 3h cos3θ ηψ a ] T

[0028] Arranged:

[0029]

[0030] Where:

[0031]

[0032] Among them, u d 、u q 、u n is the d-axis voltage, q-axis voltage and neutral point voltage, i d 、i q is the d and q axis current, u n is the neutral point voltage, ψ d , q , fault are the d-axis, q-axis and fault phase flux, ψ 3h , a , f are the third harmonic, fault phase current flux and fundamental flux respectively, L AA is the inductance of phase A, L abcf , R abcf are the inductance and resistance of the faulty phases A, B and C.

[0033] The variational mode decomposition algorithm for extracting fault features includes the following steps:

[0034] 31) According to the permanent magnet synchronous motor fault model obtained in step 22), the speed is set to 1000 r / min and the torque to 2 N·m, thereby obtaining the third harmonic of the fault phase current and the fundamental wave data of the zero-sequence voltage;

[0035] Setting the variational mode decomposition algorithm to extract fault features includes two processes: constructing a variational problem and solving the variational problem. Based on the variational constraints, iterative optimization is performed to determine the center frequency and bandwidth range of the input signal. The variational constraints are as follows:

[0036]

[0037] st∑ k u k (t) = f(t)

[0038] Among them, {u k}、{w k} represent the modal component and center frequency respectively, {u k}={u1,u2,...u k},{w k}={w1,w2,...w k}; is the first-order partial derivative with respect to time t; δ(t) is the Dirac function; * is the convolution operator; f(t) is the input signal; k is the number of modal components;

[0039] In order to solve the above variational constraint problem, the penalty factor α and the Lagrangian operator λ(t) are introduced to transform it into an unconstrained problem for solution, and the initial value of the variational mode decomposition is obtained, which is expressed as follows:

[0040]

[0041] Using the multiplication alternating direction method, update {u k}、{w k Solve the saddle point of the Lagrangian function, and finally get the modal components and center frequencies:

[0042]

[0043]

[0044] in, is the final mode after Fourier transformation; is the final center frequency;

[0045] 32) Variational mode decomposition decomposes the input signal into K modes, each mode contains a different center frequency. The specific steps of input signal decomposition are as follows:

[0046] Step 1: Initialization λ 1 and n is 0, setting the iteration termination condition;

[0047] Step 2: Let n=n+1 and execute the outer loop variational mode decomposition algorithm;

[0048] Step 3: Let k = k + 1 and execute the inner loop variational mode decomposition algorithm, that is, update {u k}、{w k} and λ, the update rules are as follows:

[0049]

[0050]

[0051]

[0052] Among them, ξ is the discrimination accuracy; τ is the noise tolerance value, which is 1×10 -6 ;

[0053] Step 4: Repeat steps 2 and 3 until the iteration termination condition is met to obtain the modal component of K;

[0054] 33) According to the variational mode decomposition algorithm, the center frequency ω of each modal component is obtained kAfter removing the high-order harmonic noise, the center frequency of the third harmonic of the fault phase current is and the center frequency of the zero-sequence voltage fundamental Based on this, the modal component corresponding to the center frequency is extracted This modal component is the data corresponding to the third harmonic of the fault phase current and the zero-sequence voltage fundamental in the time domain. The specific expression is:

[0055]

[0056]

[0057] The sparrow search algorithm for optimizing variational mode decomposition parameters comprises the following steps:

[0058] 41) Set the finder position update rule in the foraging iteration process as follows:

[0059]

[0060] in, is the position of the i-th sparrow in the j-th dimension; n is the number of iterations; q and L are random numbers and the unit matrix that obey the normal distribution respectively; R2 and S are the alarm value and the safety value respectively, R2=[0,1], S=[0.5,1]; z is a random number between 0 and 1, iter max is the maximum number of iterations, set to 20,

[0061] The follower position update rules are as follows:

[0062]

[0063] in, are the best and worst positions currently occupied by sparrows for foraging; A + For a satisfying + =A T (AA T ) -1 Matrix; c is the number of sparrows, which is 20.

[0064] When danger occurs during predation, the number of vigilant individuals in the sparrow population will be randomly generated at a ratio of 30%, and the mathematical model is as follows:

[0065]

[0066] Among them, β and Q are random numbers between [-1,1]; ε is the minimum random number that ensures that the denominator is not zero; is the global optimal position; f i 、f g 、f wThey are the current fitness value of the population, the global best fitness value, and the global worst fitness value;

[0067] 42) The specific steps of the sparrow search algorithm to optimize the variational mode decomposition algorithm are as follows:

[0068] For step 32), the initial value of variational mode decomposition is obtained, and the sparrow search algorithm is used to optimize the parameters of the variational mode decomposition algorithm. During the optimization, the fitness function is first determined as the iterative evaluation index, and the comprehensive evaluation index is designed as the minimum value of the sample entropy function, the Pearson coefficient, and the relative aggregation algebraic operation. The specific expression is:

[0069] feature=SampEn(data,q,r)=lnB q (r)-lnB q+1 (r)

[0070] pear=corr(data T ,u k T )

[0071] D = lg (size (omega))

[0072] fitness=min((feature / pear)*D)

[0073] Among them, feature, pear, D and fitness are the sample entropy function value, Pearson function value, relative aggregation algebra value and adaptability value; size (omega) is the optimal center frequency ω k The signal length; SampEn and corr are sample entropy function and Pearson function respectively.

[0074] Step 1: Initialize the population, set the decomposition layer number K of the variational mode decomposition algorithm to 1-10, the penalty factor α to 500-3000, randomly initialize [α, K] as the initial position of producers and followers, and set the proportion of producers and followers in the sparrow population;

[0075] Step 2: Perform variational mode decomposition on the producer and follower at each position, calculate the initial fitness value, and update the discoverer position according to the fitness value and warning value;

[0076] Step 3: Update the positions of followers and sentinels, and use them to update the fitness value and the best foraging position;

[0077] Step 4: Repeat steps 2 and 3 until the maximum number of iterations or the loss convergence condition is reached;

[0078] According to the above steps, the optimal decomposition layer number K and penalty factor α are obtained. [α, K] is input into the variational mode decomposition algorithm to decompose the fault phase current and zero-sequence voltage to obtain the center frequency of the third harmonic of the fault phase current and the fundamental wave of the zero-sequence voltage. Based on this, the modal component corresponding to the center frequency is extracted Its specific time domain expression is:

[0079]

[0080]

[0081] in, is the optimal decomposition result of the third harmonic of the fault phase current and the fundamental wave of the zero-sequence voltage; t is the time in seconds.

[0082] The establishment of the deep pyramid pooling residual convolutional neural network model includes the following steps:

[0083] 51) Setting a pyramid pooling network by setting multiple pooling channels and using filters of different sizes to perform pooling operations on the input data, and splicing the pooling results of different scales together;

[0084] The pyramid pooling network structure is set as follows:

[0085] Pyramid pooling channel 1: set the filter size to [3,512], the step size to 1, and use the maximum pooling operation;

[0086] Pyramid pooling channel 2: set the filter size to [64,512], the step size to 1, and use the maximum pooling operation;

[0087] Pyramid pooling channel three: set the filter size to [128,512], the step size to 1, and use the maximum pooling operation;

[0088] The pooling operation process of each channel is expressed as follows:

[0089]

[0090]

[0091]

[0092] in, and is the output of the three channels of pyramid pooling; x l-1 is the input data; same is the padding method to ensure that the output and input data sizes are consistent;

[0093] 52) Set the structure of the residual network as follows:

[0094] The first layer of the residual network: This layer is a convolutional layer. The convolution kernel size is set to 3×3, the number of convolution kernels is set to 64, and the padding method is to keep the output and input sizes consistent;

[0095] The second layer of the residual network: This layer is a batch sample normalization layer to improve the network convergence efficiency;

[0096] The third layer of the residual network: This layer is the activation function layer, and the Relu function is selected;

[0097] The fourth layer of the residual network: This layer is a convolutional layer. The convolution kernel size is set to 5×5, the number of convolution kernels is set to 128, and the padding method is the same;

[0098] The fifth layer of the residual network: This layer is a batch sample normalization layer to improve the network convergence efficiency;

[0099] 53) Set the SPP-ResCNN structure after the fusion of one-dimensional convolutional neural network, pyramid pooling network and residual network, that is, the deep pyramid pooling residual convolutional neural network model is as follows:

[0100] The first layer is the input layer, and the input data size of this layer is set to 2×512;

[0101] The second layer is convolution layer 1. The convolution kernel size is set to 1×3, the number of convolution kernels is set to 64, the step size is set to 1, and the padding method is set to the same. The output of this layer is for:

[0102]

[0103] The third layer is the activation function, which uses the Relu activation function, and its expression is as follows:

[0104]

[0105] The fourth layer is pooling layer 1, which uses maximum pooling to reduce the dimension of input features and prevent network overfitting. The output of this layer for:

[0106]

[0107] The fifth layer is convolution layer 2. The convolution kernel size of this layer is set to 1*3, the number of convolution kernels is set to 128, the step size is set to 1, and the padding method is set to the same. The output of this layer is connected to the input of the residual network.

[0108] The sixth layer is the batch sample normalization layer, which standardizes the data input to this layer to accelerate the convergence speed and generalization ability of the network;

[0109] The seventh layer is a shortcut connection layer, which connects the residual network built in 52) and the output of the sixth layer of SPP-ResCNN in 53). This layer consists of a convolution layer 3 and a batch sample normalization layer. The convolution kernel size is set to 1*1, the number of convolutions is 128, the step size is 1, and the padding method is the same;

[0110] The eighth layer is a pyramid pooling layer, and the specific structure of this layer is shown in step 51);

[0111] The ninth layer is the deep connection layer, which is used to connect the outputs of the three channels of the pyramid pooling network to ensure that the data size of each channel output is consistent. The number of channels in this layer is set to 3. After data fusion, the output size of this layer is 2×512×3.

[0112] The tenth layer is a fully connected layer, which is used to integrate the feature information extracted by the previous layers and map the sample distribution features to the sample label space. The number of classification categories in this layer is set to 6. The output of the fully connected layer is for:

[0113]

[0114] Among them, W l and b l are weight coefficients and biases; f() is the activation function, x l-1 is the input of this layer;

[0115] The eleventh layer is the Softmax layer, which normalizes the output of the fully connected layer so that the value of each element is between 0 and 1, and the sum of all elements is 1. The output of this layer is the probability that the element belongs to a certain category;

[0116] The twelfth layer is the Classification layer, which is used to output the classification results of the inter-turn short-circuit fault and calculate the loss function for back propagation.

[0117] The training of the SPP-ResCNN network includes the following steps:

[0118] 61) Create a training set and a test set for the SPP-ResCNN network, specifically including the following steps:

[0119] The optimal decomposition results of the third harmonic of the fault phase current and the zero-sequence voltage fundamental obtained by the sparrow search algorithm optimized variational modal decomposition in step 4 of step 42) are made into a labeled training set XTrain and a test set YTest. Two fault features are set, and the size of each training sample is 2×512, with a total of 1698 samples. The size of each training sample in the test set is 2×512, with a total of 702 samples. Different fault degrees η are set and used as classification labels Label. The specific expressions of the training set, test set, fault degree, and classification label are:

[0120]

[0121]

[0122] η=[0,0.05,0.1,0.15,0.2,0.25]

[0123] Label = [1,2,3,4,5,6];

[0124] 62) Input the labeled training set XTrain and test set YTest in 61) into the SPP-ResCNN network for training and verification. The specific steps are as follows:

[0125] Set the maximum number of training rounds to 10 and the minimum number of training batches to 64, then the number of iterations per round is 26, the initial learning rate is 0.01, and the loss function is the cross entropy function. The specific expression is as follows:

[0126]

[0127] Where n is the number of samples in the training set, m(i) is the true sample distribution, w(i) is the predicted distribution, y and f(x) are the true value and predicted value, and L(y,f(x)) is the cross entropy function loss error value;

[0128] The optimizer selects Adam to update the network parameters to minimize the loss function. The specific network training steps are as follows:

[0129] Step 1: Weight initialization process, using Gaussian distribution with a mean of 0 and a variance of 0.01 to initialize the weight coefficients of the network neuron nodes;

[0130] Step 2: Input the training set XTrain prepared in 61) into SPP-ResCNN for forward propagation, that is, the training set is propagated layer by layer through the network constructed in 53), and the deep feature extraction of the training set is completed through convolution operation, pooling operation, residual operation and pyramid pooling in sequence, and finally input into the fully connected layer, which integrates and classifies the features, and obtains the classification results by the Softmax layer and the Classification layer;

[0131] Step 3: When the output result of the SPP-ResCNN network does not match the expected value, the SPP-ResCNN network performs back propagation to find the error between the output result and the expected value, and then back propagates the error to the fully connected layer, pyramid pooling layer, residual layer, pooling layer, and convolution layer to update the weight coefficient of each layer;

[0132] Step 4: Repeat steps 2 and 3 until the training termination condition is reached to complete the SPP-ResCNN training;

[0133] Step 5: Input the test set YTest prepared in step 61) into the SPP-ResCNN trained in step 4 of step 62) for testing, calculate the classification accuracy based on the network output results, and use it as an indicator for network evaluation, and use the network convergence speed as a network performance evaluation indicator.

[0134] The real-time diagnosis of the inter-turn short circuit fault of the permanent magnet synchronous motor comprises the following steps:

[0135] 71) Real-time detection of the three-phase current and zero-sequence voltage of the permanent magnet synchronous motor. When it is detected that the current and zero-sequence voltage of a certain phase are greater than the set threshold, it indicates that the phase is a fault phase. The current and zero-sequence voltage of the phase are extracted, which specifically includes the following steps:

[0136] The detected fault phase current and zero-sequence voltage are input into the variational mode decomposition algorithm optimized by the sparrow search algorithm described in 42) to obtain the real-time third harmonic of the fault phase current and zero sequence voltage fundamental Its specific expression is:

[0137]

[0138]

[0139] Where, t is time in seconds;

[0140] 72) The third harmonic of the fault phase current collected in real time and zero sequence voltage fundamental As two features of turn-to-turn short-circuit fault, and make a training set X time Train and test set Y time Test, whose expression is as follows:

[0141]

[0142]

[0143] 73) The prepared training set X time Train, test set Ytime Test is input into the SPP-ResCNN network and the steps in 62) are repeated to obtain the turn-to-turn short circuit fault diagnosis result.

[0144] Beneficial Effects

[0145] Compared with the prior art, the motor inter-turn short-circuit fault detection method based on variational modal decomposition fusion deep learning of the present invention adopts the sparrow search algorithm to optimize the variational modal decomposition algorithm, eliminates other harmonic components in the fault characteristics, and extracts specific frequency fault characteristics, which significantly improves the accuracy of the fault characteristics. At the same time, the pyramid pooling network and the residual network are used to optimize the convolutional neural network, which improves the accuracy and efficiency of fault diagnosis.

[0146] 1. The present invention uses a variational mode decomposition algorithm to adaptively decompose the input signal according to the characteristics of the fault characteristic signal, thereby avoiding the shortcomings of traditional signal decomposition methods, such as excessive bandwidth and difficulty in completely eliminating harmonics.

[0147] 2. The present invention optimizes the parameters in the variational mode decomposition process through the sparrow search algorithm, avoids the error caused by the artificial experience in selecting parameters, adopts multiple adaptive functions as comprehensive evaluation indicators, and significantly improves the accuracy of the variational mode decomposition algorithm.

[0148] 3. The present invention improves the problems of insufficient feature extraction, gradient vanishing and gradient exploding in the training process of traditional convolutional neural networks by effectively integrating pyramid pooling networks, residual networks and convolutional neural networks, extracts the features of input data to the maximum extent, improves the network diagnosis accuracy, and greatly speeds up the efficiency of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0149] Figure 1 is a method sequence diagram of the present invention;

[0150] Figure 2 is the Fourier transform diagram of the original signal;

[0151] Figure 3 Optimize the adaptive function graph for the sparrow search algorithm;

[0152] Figure 4 This is a comparison chart of variational mode decomposition and wavelet transform results after the sparrow search algorithm is optimized;

[0153] Figure 5 Comparison chart of SPP-ResCNN and CNN loss functions;

[0154] Figure 6 This is the result diagram of the permanent magnet synchronous motor turn-to-turn short circuit fault diagnosis. DETAILED DESCRIPTION

[0155] In order to have a further understanding and recognition of the structural features and the effects achieved by the present invention, a preferred embodiment and accompanying drawings are used for detailed description as follows:

[0156] like Figure 1 As shown, the method for detecting inter-turn short-circuit fault of a permanent magnet synchronous motor based on neural network technology of the present invention comprises the following steps:

[0157] The first step is to establish a permanent magnet synchronous motor fault model: According to the voltage and current equations in the natural coordinate system of the permanent magnet synchronous motor, the equations of the d-axis and q-axis voltages and the d-axis and q-axis currents in the rotating coordinate system are established, which is used as the permanent magnet synchronous motor fault model to simulate the short circuit between the fault phase turns.

[0158] (1) Set the voltage u of phase A, phase B, and phase C of the permanent magnet synchronous motor in the natural coordinate system a 、u b 、u c With phase A, phase B, phase C current i a 、i b 、i c and the fault current i f The equation between is as follows:

[0159]

[0160] Among them, u a 、u b 、u c is the voltage of phase A, phase B, and phase C, η is the fault degree, and η is defined as the ratio of the short-circuit turns of the stator winding of phase A to the total turns of its stator winding, R s is the stator resistance, R f is the fault resistance, ψ f is the fundamental amplitude of the flux, θ is the rotor electrical angle, i a 、i b 、i c 、i f are the current and fault current of the stator winding of phase A, phase B and phase C respectively. L and M are the self-inductance and mutual inductance of the stator winding respectively.

[0161] (2) The voltage equation in the natural coordinate system is transformed into the rotating coordinate system to obtain the permanent magnet synchronous motor fault model, which is convenient for extracting the inter-turn short circuit fault characteristics. The coordinate transformation matrix T is as follows:

[0162]

[0163] The inverse matrix of T is:

[0164]

[0165] Applying the coordinate transformation matrix T to the voltage and current equations in the natural coordinate system of the motor yields:

[0166]

[0167] Where:

[0168] [u dq0f ]=[u d u q -u n 0] T

[0169] [i dq0f ]=[i d i q 0i f ] T

[0170] [ψ dq0f ]=[ψ f 0ψ 3h cos3θηψ a ] T

[0171] Arranged:

[0172]

[0173] Where:

[0174]

[0175] Among them, u d 、u q 、u n is the d-axis voltage, q-axis voltage and neutral point voltage, i d 、i q is the d and q axis current, u n is the neutral point voltage, ψ d , q , fault are the d-axis, q-axis and fault phase flux, ψ 3h , a , f are the third harmonic, fault phase current flux and fundamental flux respectively, L AA is the inductance of phase A, L abcf , R abcf are the inductance and resistance of the faulty phases A, B and C.

[0176] In the second step, the variational modal decomposition algorithm is used to extract fault characteristics: the motor fault phase current and zero-sequence voltage are obtained according to the simulation results of the permanent magnet synchronous motor fault model, and are respectively input into the variational modal decomposition algorithm for adaptive modal decomposition to obtain the third harmonic of the fault phase current and the fundamental wave of the zero-sequence voltage as the turn-to-turn short-circuit fault characteristics.

[0177] The traditional signal decomposition method has the defects of too high frequency bandwidth, which makes it impossible to completely eliminate the noise harmonics near the signal center frequency and low fault feature accuracy. The present invention applies the variational mode decomposition algorithm to fault feature extraction. The algorithm adaptively adjusts the frequency bandwidth according to the signal characteristics and extracts the third harmonic of the fault phase current and the zero-sequence voltage fundamental with higher accuracy.

[0178] (1) According to the obtained permanent magnet synchronous motor fault model, the speed is set to 1000 r / min and the torque is set to 2 N·m to obtain the third harmonic of the fault phase current and the fundamental wave data of the zero-sequence voltage;

[0179] Setting the variational mode decomposition algorithm to extract fault features includes two processes: constructing a variational problem and solving the variational problem. Based on the variational constraints, iterative optimization is performed to determine the center frequency and bandwidth range of the input signal. The variational constraints are as follows:

[0180]

[0181] st∑ k u k (t) = f(t)

[0182] Among them, {u k}、{w k} represent the modal component and center frequency respectively, {u k}={u1,u2,...u k},{w k}={w1,w2,...w k}; is the first-order partial derivative with respect to time t; δ(t) is the Dirac function; * is the convolution operator; f(t) is the input signal; k is the number of modal components;

[0183] In order to solve the above variational constraint problem, the penalty factor α and the Lagrangian operator λ(t) are introduced to transform it into an unconstrained problem for solution, and the initial value of the variational mode decomposition is obtained, which is expressed as follows:

[0184]

[0185] Using the multiplication alternating direction method, update {u k}、{w k Solve the saddle point of the Lagrangian function, and finally get the modal components and center frequencies:

[0186]

[0187]

[0188] in, is the final mode after Fourier transformation; is the final center frequency.

[0189] (2) Variational mode decomposition decomposes the input signal into K modes, each of which contains a different center frequency. The specific steps of input signal decomposition are as follows:

[0190] Step 1: Initialization λ 1 and n is 0, setting the iteration termination condition;

[0191] Step 2: Let n=n+1 and execute the outer loop variational mode decomposition algorithm;

[0192] Step 3: Let k = k + 1 and execute the inner loop variational mode decomposition algorithm, that is, update {u k}、{w k} and λ, the update rules are as follows:

[0193]

[0194]

[0195]

[0196] Among them, ξ is the discrimination accuracy; τ is the noise tolerance value, which is 1×10 -6 .

[0197] Step 4: Repeat steps 2 and 3 until the iteration termination condition is met to obtain the modal component of K;

[0198] (3) According to the variational mode decomposition algorithm, the center frequency ω of each modal component is obtained k After removing the high-order harmonic noise, the center frequency of the third harmonic of the fault phase current is and the center frequency of the zero-sequence voltage fundamental Based on this, the modal component corresponding to the center frequency is extracted This modal component is the data corresponding to the third harmonic of the fault phase current and the zero-sequence voltage fundamental in the time domain. The specific expression is:

[0199]

[0200]

[0201] like Figure 2 As shown in the figure, there are multiple harmonic components in the frequency domain image of the original signal after Fourier transform, but only one harmonic can reflect the fault characteristics. Therefore, the noise harmonic interference in the original signal is large, which affects the accuracy and efficiency of fault diagnosis.

[0202] The third step is to optimize the variational modal decomposition parameters using the sparrow search algorithm: the penalty factor α and the number of decomposition levels K in the variational modal decomposition process are optimized within a given range using the sparrow search algorithm to obtain the optimal decomposition results of the third harmonic of the fault phase current and the zero-sequence voltage fundamental.

[0203] It can be seen from the process of extracting fault features by the variational mode decomposition algorithm that the decomposition result is greatly affected by the penalty factor α and the number of decomposition layers K. At present, the penalty factor α and the number of decomposition layers K are mainly set based on human experience. However, this method usually causes the phenomenon of over-decomposition and under-decomposition of the signal, resulting in problems such as the inconvenience of promoting the variational mode decomposition algorithm. In response to this problem, the present invention effectively combines the sparrow search algorithm with the variational mode decomposition algorithm, and uses the sparrow search algorithm to iterate layer by layer to find the optimal penalty factor α and the number of decomposition layers K, which significantly improves the accuracy of fault feature extraction.

[0204] (1) The rules for updating the finder's position during the foraging iteration are set as follows:

[0205]

[0206] in, is the position of the i-th sparrow in the j-th dimension; n is the number of iterations; q and L are random numbers and the unit matrix that obey the normal distribution respectively; R2 and S are the alarm value and the safety value respectively, R2=[0,1], S=[0.5,1]; z is a random number between 0 and 1, iter max is the maximum number of iterations, set to 20,

[0207] The follower position update rules are as follows:

[0208]

[0209] in, are the best and worst positions currently occupied by sparrows for foraging; A + For a satisfying + =A T (AA T ) -1 Matrix; c is the number of sparrows, which is 20.

[0210] When danger occurs during predation, the number of vigilant individuals in the sparrow population will be randomly generated at a ratio of 30%, and the mathematical model is as follows:

[0211]

[0212] Among them, β and Q are random numbers between [-1,1]; ε is the minimum random number that ensures that the denominator is not zero; is the global optimal position; f i 、f g 、f w are the current fitness value of the population, the global best fitness value, and the global worst fitness value, respectively.

[0213] (2) The specific steps of optimizing the variational mode decomposition algorithm using the sparrow search algorithm are as follows:

[0214] In order to obtain the initial value of variational mode decomposition, the sparrow search algorithm is used to optimize the parameters of the variational mode decomposition algorithm. During the optimization, the fitness function is first determined as the iterative evaluation index, and the comprehensive evaluation index is designed as the minimum value of the sample entropy function, the Pearson coefficient, and the relative aggregation algebraic operation. The specific expression is:

[0215] feature=SampEn(data,q,r)=lnB q (r)-lnB q+1 (r)

[0216]

[0217] D = lg (size (omega))

[0218] fitness=min((feature / pear)*D)

[0219] Among them, feature, pear, D and fitness are the sample entropy function value, Pearson function value, relative aggregation algebra value and adaptability value; size (omega) is the optimal center frequency ω k The signal length; SampEn and corr are sample entropy function and Pearson function respectively.

[0220] Step 1: Initialize the population, set the decomposition layer number K of the variational mode decomposition algorithm to 1-10, the penalty factor α to 500-3000, randomly initialize [α, K] as the initial position of producers and followers, and set the proportion of producers and followers in the sparrow population;

[0221] Step 2: Perform variational mode decomposition on the producer and follower at each position, calculate the initial fitness value, and update the discoverer position according to the fitness value and warning value;

[0222] Step 3: Update the positions of followers and sentinels, and use them to update the fitness value and the best foraging position;

[0223] Step 4: Repeat steps 2 and 3 until the maximum number of iterations or the loss convergence condition is reached;

[0224] According to the above steps, the optimal parameters [α, K] of the variational mode decomposition algorithm are [2951, 3], as shown in Figure 3 As shown in Figure 2, the adaptive function converges after 5 iterations, and the fitness value is about 7.59×10 -3 , input [α, K] into the variational mode decomposition algorithm, decompose the fault phase current and zero-sequence voltage, and obtain the center frequency of the third harmonic of the fault phase current and the fundamental wave of the zero-sequence voltage Based on this, the modal component corresponding to the center frequency is extracted Its specific time domain expression is:

[0225]

[0226]

[0227] in, is the optimal decomposition result of the third harmonic of the fault phase current and the fundamental wave of the zero-sequence voltage; tzai is the time in seconds.

[0228] like Figure 4 As shown, the sparrow search algorithm proposed in the present invention is used to optimize the variational modal decomposition algorithm. Under the same number of layers as the wavelet transform decomposition, the modal components decomposed by the method of the present invention only contain fault characteristic harmonics, and the extraction accuracy is higher.

[0229] The fourth step is to establish a deep pyramid pooling residual convolutional neural network model: establish a deep pyramid pooling network and a residual convolutional neural network, use the deep pyramid pooling network to optimize the residual convolutional neural network, and adjust the structure and parameters of the fused SPP-ResCNN network through ablation tests.

[0230] In order to improve classification accuracy, traditional convolutional neural networks will continuously increase network depth, causing network redundancy and easy overfitting. The present invention integrates the pyramid pooling network into the traditional convolutional neural network, establishes a multi-scale pooling channel, and deeply extracts fault semantic features, thereby effectively reducing the number of network neurons. While ensuring the accuracy of fault diagnosis, the efficiency of fault diagnosis is improved. At the same time, a residual network is introduced to solve the overfitting problem that occurs during network training.

[0231] (1) Setting the pyramid pooling network By setting multiple pooling channels and using filters of different sizes to perform pooling operations on the input data, and splicing the pooling results of different scales together, the fault feature extraction is more sufficient. The pyramid pooling network structure is set as follows:

[0232] Pyramid pooling channel 1: set the filter size to [3,512], the step size to 1, and use the maximum pooling operation;

[0233] Pyramid pooling channel 2: set the filter size to [64,512], the step size to 1, and use the maximum pooling operation;

[0234] Pyramid pooling channel three: set the filter size to [128,512], the step size to 1, and use the maximum pooling operation;

[0235] The pooling operation process of each channel is expressed as follows:

[0236]

[0237]

[0238]

[0239] in, and is the output of the three channels of pyramid pooling; x l-1 is the input data; same is the padding method to ensure that the output and input data sizes are consistent.

[0240] (2) Set the structure of the residual network as follows:

[0241] The first layer of the residual network: This layer is a convolutional layer. The convolution kernel size is set to 3×3, the number of convolution kernels is set to 64, and the padding method is to keep the output and input sizes consistent;

[0242] The second layer of the residual network: This layer is a batch sample normalization layer to improve the network convergence efficiency;

[0243] The third layer of the residual network: This layer is the activation function layer, and the Relu function is selected;

[0244] The fourth layer of the residual network: This layer is a convolutional layer. The convolution kernel size is set to 5×5, the number of convolution kernels is set to 128, and the padding method is the same;

[0245] The fifth layer of the residual network: This layer is a batch sample normalization layer to improve the network convergence efficiency.

[0246] (3) The SPP-ResCNN structure after the fusion of one-dimensional convolutional neural network, pyramid pooling network and residual network is set as follows:

[0247] The first layer is the input layer, and the input data size of this layer is set to 2×512;

[0248] The second layer is convolution layer 1. The convolution kernel size is set to 1×3, the number of convolution kernels is set to 64, the step size is set to 1, and the padding method is set to the same. The output of this layer is for:

[0249]

[0250] The third layer is the activation function, which uses the Relu activation function, and its expression is as follows:

[0251]

[0252] The fourth layer is pooling layer 1, which uses maximum pooling to reduce the dimension of input features and prevent network overfitting. The output of this layer for:

[0253]

[0254] The fifth layer is convolution layer 2. The convolution kernel size of this layer is set to 1*3, the number of convolution kernels is set to 128, the step size is set to 1, and the padding method is set to the same. The output of this layer is connected to the input of the residual network.

[0255] The sixth layer is the batch sample normalization layer, which standardizes the data input to this layer to accelerate the convergence speed and generalization ability of the network;

[0256] The seventh layer is the shortcut connection layer, which connects the residual network built in this layer with the output of the sixth layer of SPP-ResCNN. This layer consists of convolution layer 3 and batch sample normalization layer. The convolution kernel size is set to 1*1, the number of convolutions is 128, the step size is 1, and the padding method is the same;

[0257] The eighth layer is the pyramid pooling layer;

[0258] The ninth layer is the deep connection layer, which is used to connect the outputs of the three channels of the pyramid pooling network to ensure that the data size of each channel output is consistent. The number of channels in this layer is set to 3. After data fusion, the output size of this layer is 2×512×3.

[0259] The tenth layer is a fully connected layer, which is used to integrate the feature information extracted by the previous layers and map the sample distribution features to the sample label space. The number of classification categories in this layer is set to 6. The output x of the fully connected layer is l full for:

[0260]

[0261] Among them, W l and b l are weight coefficients and biases; f() is the activation function, x l-1 is the input of this layer;

[0262] The eleventh layer is the Softmax layer, which normalizes the output of the fully connected layer so that the value of each element is between 0 and 1, and the sum of all elements is 1. The output of this layer is the probability that the element belongs to a certain category;

[0263] The twelfth layer is the Classification layer, which is used to output the classification results of the inter-turn short-circuit fault and calculate the loss function for back propagation.

[0264] The fifth step is the training of the SPP-ResCNN network: the optimal decomposition results of the third harmonic of the fault phase current and the zero-sequence voltage fundamental obtained by optimizing the variational modal decomposition by the sparrow search algorithm are made into training sets and test sets, which are input into the SPP-ResCNN for training. The learning rate, maximum number of training rounds, and minimum training batches are adjusted through ablation experiments to minimize the network loss function.

[0265] (1) Prepare the training set and test set of the SPP-ResCNN network, which specifically includes the following steps:

[0266] The best decomposition results of the third harmonic of the fault phase current and the zero-sequence voltage fundamental obtained by optimizing the variational modal decomposition by the sparrow search algorithm are made into a labeled training set XTrain and a test set YTest. Two fault features are set, and the size of each training sample is 2×512, with a total of 1698 samples. The size of each training sample in the test set is 2×512, with a total of 702 samples. Different fault degrees η are set and used as classification labels Label. The specific expressions of the training set, test set, fault degree, and classification label are:

[0267]

[0268]

[0269] η=[0,0.05,0.1,0.15,0.2,0.25]

[0270] Label = [1,2,3,4,5,6];

[0271] (2) Input the labeled training set XTrain and test set YTest into the SPP-ResCNN network for training and verification. The specific steps are as follows:

[0272] Set the maximum number of training rounds to 10 and the minimum number of training batches to 64, then the number of iterations per round is 26, the initial learning rate is 0.01, and the loss function is the cross entropy function. The specific expression is as follows:

[0273]

[0274] Where n is the number of samples in the training set, m(i) is the true sample distribution, w(i) is the predicted distribution, y and f(x) are the true value and predicted value, and L(y,f(x)) is the cross entropy function loss error value;

[0275] The optimizer selects Adam to update the network parameters to minimize the loss function. The specific network training steps are as follows:

[0276] Step 1: Weight initialization process, using Gaussian distribution with a mean of 0 and a variance of 0.01 to initialize the weight coefficients of the network neuron nodes;

[0277] Step 2: Input the prepared training set XTrain into SPP-ResCNN for forward propagation, that is, the training set is propagated layer by layer through the constructed network, and the deep feature extraction of the training set is completed through convolution operation, pooling operation, residual operation and pyramid pooling in sequence, and finally input into the fully connected layer, which integrates and classifies the features, and obtains the classification results by the Softmax layer and the Classification layer;

[0278] Step 3: When the output result of the SPP-ResCNN network does not match the expected value, the SPP-ResCNN network performs back propagation to find the error between the output result and the expected value, and then back propagates the error to the fully connected layer, pyramid pooling layer, residual layer, pooling layer, and convolution layer to update the weight coefficient of each layer;

[0279] Step 4: Repeat steps 2 and 3 until the training termination condition is reached to complete the SPP-ResCNN training;

[0280] Step 5: Input the prepared test set YTest into the SPP-ResCNN trained in step 4 for testing. Calculate the classification accuracy based on the network output results and use it as an indicator for network evaluation. Use the network convergence speed as a network performance evaluation indicator.

[0281] like Figure 5 As shown in Figure 2, the loss function of SPP-ResCNN decreases significantly faster than that of CNN. The CNN network converges in about 150 rounds, while SPP-ResCNN converges in about 50 rounds, and the convergence speed is increased by 200%. Figure 6 As shown in Figure 2, the SPP-ResCNN network training result has an accuracy of 100%, and both the training set and the test set are correctly classified.

[0282] The sixth step is real-time diagnosis of permanent magnet synchronous motor inter-turn short-circuit fault: real-time monitoring of permanent magnet synchronous motor current and zero-sequence voltage, input them into the variational mode decomposition algorithm optimized by the sparrow search algorithm to extract the optimal decomposition results of the third harmonic of the fault phase current and the zero-sequence voltage fundamental, and make it into a data set to be tested, and input it into the trained SPP-ResCNN network to diagnose the degree of inter-turn short-circuit fault.

[0283] (1) Real-time detection of the three-phase current and zero-sequence voltage of the permanent magnet synchronous motor. When it is detected that the current and zero-sequence voltage of a certain phase are greater than a set threshold, it indicates that the phase is a fault phase. The current and zero-sequence voltage of the phase are extracted, which specifically includes the following steps:

[0284] The detected fault phase current and zero-sequence voltage are input into the variational mode decomposition algorithm optimized by the sparrow search algorithm described in 3b) to obtain the real-time third harmonic of the fault phase current and zero sequence voltage fundamental Its specific expression is:

[0285]

[0286]

[0287] Where, t is time in seconds;

[0288] (2) The third harmonic of the fault phase current collected in real time and zero sequence voltage fundamental As two features of turn-to-turn short-circuit fault, and make a training set X time Train and test set Y time Test, whose expression is as follows:

[0289]

[0290]

[0291] (3) The prepared training set X time Train, test set Y time Test is input into the SPP-ResCNN network to obtain the turn-to-turn short-circuit fault diagnosis result.

[0292] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A motor turn-to-turn short-circuit fault detection method based on variational mode decomposition fusion deep learning, characterized in that: The following steps are involved: 11) Establishing a permanent magnet synchronous motor fault model: According to the voltage and current equations in the natural coordinate system of the permanent magnet synchronous motor, the equations of the d-axis and q-axis voltages and the d-axis and q-axis currents in the rotating coordinate system are established, and this is used as a permanent magnet synchronous motor fault model simulating a short circuit between fault phase turns; 12) Extract fault features using variational modal decomposition algorithm: According to the simulation results of the permanent magnet synchronous motor fault model, the motor fault phase current and zero-sequence voltage are obtained, and they are respectively input into the variational modal decomposition algorithm for adaptive modal decomposition to obtain the third harmonic of the fault phase current and the fundamental wave of the zero-sequence voltage as the turn-to-turn short-circuit fault features; 13) Sparrow search algorithm to optimize variational mode decomposition parameters: The penalty factor α and the number of decomposition levels K in the variational mode decomposition process are optimized within a given range by the sparrow search algorithm to obtain the optimal decomposition results of the third harmonic of the fault phase current and the zero-sequence voltage fundamental wave; 14) Establish a deep pyramid pooling residual convolutional neural network model: establish a deep pyramid pooling network and a residual convolutional neural network, use the deep pyramid pooling network to optimize the residual convolutional neural network model, namely the SPP-ResCNN network, and adjust the structure and parameters of the fused SPP-ResCNN network through ablation experiments; 15) Training of SPP-ResCNN network: The optimal decomposition results of the third harmonic of the fault phase current and the zero-sequence voltage fundamental obtained by optimizing the variational modal decomposition using the sparrow search algorithm are made into training sets and test sets, which are input into the SPP-ResCNN network for training. The learning rate, maximum number of training rounds, and minimum training batches are adjusted through ablation experiments to minimize the network loss function. 16) Real-time diagnosis of permanent magnet synchronous motor inter-turn short-circuit fault: Real-time monitoring of permanent magnet synchronous motor current and zero-sequence voltage, input them into the variational mode decomposition algorithm optimized by the sparrow search algorithm to extract the optimal decomposition results of the third harmonic of the fault phase current and the zero-sequence voltage fundamental, and make them into a data set to be tested, and input them into the trained SPP-ResCNN network to diagnose the degree of inter-turn short-circuit fault.

2. The motor turn-to-turn short circuit fault detection method based on variational mode decomposition fusion deep learning according to claim 1 is characterized in that: The establishment of the permanent magnet synchronous motor fault model comprises the following steps: 21) Set the voltage u of phase A, phase B and phase C of the permanent magnet synchronous motor in the natural coordinate system a 、u b 、u c With phase A, phase B, phase C current i a 、i b 、i c and fault current i f The equation between is as follows: Among them, u a 、u b 、u c is the voltage of phase A, phase B, and phase C, η is the fault degree, and η is defined as the ratio of the short-circuit turns of the stator winding of phase A to the total turns of its stator winding, R s is the stator resistance, R f is the fault resistance, ψ f is the fundamental amplitude of the flux, θ is the rotor electrical angle, i a 、i b 、i c 、i f are the current and fault current of the stator winding of phase A, phase B, and phase C respectively, L and M are the self-inductance and mutual inductance of the stator winding respectively; 22) The voltage equation in the natural coordinate system is transformed into the rotating coordinate system to obtain the permanent magnet synchronous motor fault model, which is convenient for extracting the inter-turn short circuit fault characteristics. The coordinate transformation matrix T is as follows: The inverse matrix of T is: Applying the coordinate transformation matrix T to the voltage and current equations in the natural coordinate system of the motor yields: Where: [in dq0f ]=[in d in q -in n 0] T [i dq0f ]=[i d i q 0 i f ] T [ψ dq0f ]=[ψ f 0 ψ 3h cos3θ εψ a ] T Arranged: Where: Among them, u d 、u q 、u n is the d-axis voltage, q-axis voltage and neutral point voltage, i d 、i q is the d and q axis current, u n is the neutral point voltage, ψ d , q , fault are the d-axis, q-axis and fault phase flux, ψ 3h , a , f are the third harmonic, fault phase current flux and fundamental flux respectively, L AA is the inductance of phase A, L abcf , R abcf are the inductance and resistance of the faulty phases A, B and C.

3. The motor turn-to-turn short circuit fault detection method based on variational mode decomposition fusion deep learning according to claim 1 is characterized in that: The variational mode decomposition algorithm for extracting fault features includes the following steps: 31) According to the permanent magnet synchronous motor fault model obtained in step 22), the speed is set to 1000 r / min and the torque to 2 N·m, thereby obtaining the third harmonic of the fault phase current and the fundamental wave data of the zero-sequence voltage; Setting the variational mode decomposition algorithm to extract fault features includes two processes: constructing a variational problem and solving the variational problem. Based on the variational constraints, iterative optimization is performed to determine the center frequency and bandwidth range of the input signal. The variational constraints are as follows: s.t.∑ k u k (t)=f(t) Among them, {u k }、{w k } represent the modal component and center frequency respectively, {u k }={u1,u2,...u k },{w k }={w1,w2,...w k }; is the first-order partial derivative with respect to time t; δ(t) is the Dirac function; * is the convolution operator; f(t) is the input signal; k is the number of modal components; In order to solve the above variational constraint problem, the penalty factor α and the Lagrangian operator λ(t) are introduced to transform it into an unconstrained problem for solution, and the initial value of the variational mode decomposition is obtained, which is expressed as follows: Using the multiplication alternating direction method, update {u k }、{w k Solve the saddle point of the Lagrangian function, and finally get the modal components and center frequencies: in, is the final mode after Fourier transformation; is the final center frequency; 32) Variational mode decomposition decomposes the input signal into K modes, each mode contains a different center frequency. The specific steps of input signal decomposition are as follows: Step 1: Initialization λ 1 and n is 0, setting the iteration termination condition; Step 2: Let n=n+1 and execute the outer loop variational mode decomposition algorithm; Step 3: Let k = k + 1 and execute the inner loop variational mode decomposition algorithm, that is, update {u k }、{w k } and λ, the update rules are as follows: Among them, ξ is the discrimination accuracy; τ is the noise tolerance value, which is 1×10 -6 ; Step 4: Repeat steps 2 and 3 until the iteration termination condition is met to obtain the modal component of K; 33) According to the variational mode decomposition algorithm, the center frequency ω of each modal component is obtained k After removing the high-order harmonic noise, the center frequency of the third harmonic of the fault phase current is and the center frequency of the zero-sequence voltage fundamental Based on this, the modal component corresponding to the center frequency is extracted This modal component is the data corresponding to the third harmonic of the fault phase current and the zero-sequence voltage fundamental in the time domain. The specific expression is:

4. The motor turn-to-turn short circuit fault detection method based on variational mode decomposition fusion deep learning according to claim 1 is characterized in that: The sparrow search algorithm for optimizing variational mode decomposition parameters comprises the following steps: 41) Set the finder position update rule in the foraging iteration process as follows: in, is the position of the i-th sparrow in the j-th dimension; n is the number of iterations; q and L are random numbers and the unit matrix that obey the normal distribution respectively; R2 and S are the alarm value and the safety value respectively, R2=[0,1], S=[0.5,1]; z is a random number between 0 and 1, iter max is the maximum number of iterations, set to 20, The follower position update rules are as follows: in, are the best and worst positions currently occupied by sparrows for foraging; A + For a satisfying + =A T (AA T ) -1 Matrix; c is the number of sparrows, which is 20. When danger occurs during predation, the number of vigilant individuals in the sparrow population will be randomly generated at a ratio of 30%, and the mathematical model is as follows: Among them, β and Q are random numbers between [-1,1]; ε is the minimum random number that ensures that the denominator is not zero; is the global optimal position; f i 、f g 、f w They are the current fitness value of the population, the global best fitness value, and the global worst fitness value; 42) The specific steps of the sparrow search algorithm to optimize the variational mode decomposition algorithm are as follows: For step 32), the initial value of variational mode decomposition is obtained, and the sparrow search algorithm is used to optimize the parameters of the variational mode decomposition algorithm. During the optimization, the fitness function is first determined as the iterative evaluation index, and the comprehensive evaluation index is designed as the minimum value of the sample entropy function, the Pearson coefficient, and the relative aggregation algebraic operation. The specific expression is: featureSampEn(data,q,r)ln B q (r)-ln B q+1 (r) D = lg (size (omega)) fitness=min((feature / pear)*D) Among them, feature, pear, D and fitness are the sample entropy function value, Pearson function value, relative aggregation algebra value and adaptability value; size (omega) is the optimal center frequency ω k The signal length; SampEn and corr are sample entropy function and Pearson function respectively. Step 1: Initialize the population, set the decomposition layer number K of the variational mode decomposition algorithm to 1-10, the penalty factor α to 500-3000, randomly initialize [α, K] as the initial position of producers and followers, and set the proportion of producers and followers in the sparrow population; Step 2: Perform variational mode decomposition on the producer and follower at each position, calculate the initial fitness value, and update the discoverer position according to the fitness value and warning value; Step 3: Update the positions of followers and sentinels, and use them to update the fitness value and the best foraging position; Step 4: Repeat steps 2 and 3 until the maximum number of iterations or the loss convergence condition is reached; According to the above steps, the optimal decomposition layer number K and penalty factor α are obtained. [α, K] is input into the variational mode decomposition algorithm to decompose the fault phase current and zero-sequence voltage to obtain the center frequency of the third harmonic of the fault phase current and the fundamental wave of the zero-sequence voltage. Based on this, the modal component corresponding to the center frequency is extracted Its specific time domain expression is: in, is the optimal decomposition result of the third harmonic of the fault phase current and the fundamental wave of the zero-sequence voltage; t is the time in seconds.

5. The motor turn-to-turn short circuit fault detection method based on variational mode decomposition fusion deep learning according to claim 1 is characterized in that: The establishment of the deep pyramid pooling residual convolutional neural network model includes the following steps: 51) Setting a pyramid pooling network by setting multiple pooling channels and using filters of different sizes to perform pooling operations on the input data, and splicing the pooling results of different scales together; The pyramid pooling network structure is set as follows: Pyramid pooling channel 1: set the filter size to [3,512], the step size to 1, and use the maximum pooling operation; Pyramid pooling channel 2: set the filter size to [64,512], the step size to 1, and use the maximum pooling operation; Pyramid pooling channel three: set the filter size to [128,512], the step size to 1, and use the maximum pooling operation; The pooling operation process of each channel is expressed as follows: in, and is the output of the three channels of pyramid pooling; x l-1 is the input data; same is the padding method to ensure that the output and input data sizes are consistent; 52) Set the structure of the residual network as follows: The first layer of the residual network: This layer is a convolutional layer. The convolution kernel size is set to 3×3, the number of convolution kernels is set to 64, and the padding method is to keep the output and input sizes consistent; The second layer of the residual network: This layer is a batch sample normalization layer to improve the network convergence efficiency; The third layer of the residual network: This layer is the activation function layer, and the Relu function is selected; The fourth layer of the residual network: This layer is a convolutional layer. The convolution kernel size is set to 5×5, the number of convolution kernels is set to 128, and the padding method is the same; The fifth layer of the residual network: This layer is a batch sample normalization layer to improve the network convergence efficiency; 53) Set the SPP-ResCNN structure after the fusion of one-dimensional convolutional neural network, pyramid pooling network and residual network, that is, the deep pyramid pooling residual convolutional neural network model is as follows: The first layer is the input layer, and the input data size of this layer is set to 2×512; The second layer is convolution layer 1. The convolution kernel size is set to 1×3, the number of convolution kernels is set to 64, the step size is set to 1, and the padding method is set to the same. The output of this layer is for: The third layer is the activation function, which uses the Relu activation function, and its expression is as follows: The fourth layer is pooling layer 1, which uses maximum pooling to reduce the dimension of input features and prevent network overfitting. The output of this layer for: The fifth layer is convolution layer 2. The convolution kernel size of this layer is set to 1*3, the number of convolution kernels is set to 128, the step size is set to 1, and the padding method is set to the same. The output of this layer is connected to the input of the residual network. The sixth layer is the batch sample normalization layer, which standardizes the data input to this layer to accelerate the convergence speed and generalization ability of the network; The seventh layer is a shortcut connection layer, which connects the residual network built in 52) and the output of the sixth layer of SPP-ResCNN in 53). This layer consists of a convolution layer 3 and a batch sample normalization layer. The convolution kernel size is set to 1*1, the number of convolutions is 128, the step size is 1, and the padding method is the same; The eighth layer is a pyramid pooling layer, and the specific structure of this layer is shown in step 51); The ninth layer is the deep connection layer, which is used to connect the outputs of the three channels of the pyramid pooling network to ensure that the data size of each channel output is consistent. The number of channels in this layer is set to 3. After data fusion, the output size of this layer is 2×512×3. The tenth layer is a fully connected layer, which is used to integrate the feature information extracted by the previous layers and map the sample distribution features to the sample label space. The number of classification categories in this layer is set to 6. The output of the fully connected layer is for: Among them, W l and b l are weight coefficients and biases; f() is the activation function, x l-1 is the input of this layer; The eleventh layer is the Softmax layer, which normalizes the output of the fully connected layer so that the value of each element is between 0 and 1, and the sum of all elements is 1. The output of this layer is the probability that the element belongs to a certain category; The twelfth layer is the Classification layer, which is used to output the classification results of the inter-turn short-circuit fault and calculate the loss function for back propagation.

6. The motor turn-to-turn short circuit fault detection method based on variational mode decomposition fusion deep learning according to claim 1 is characterized in that: The training of the SPP-ResCNN network includes the following steps: 61) Create a training set and a test set for the SPP-ResCNN network, specifically including the following steps: The optimal decomposition results of the third harmonic of the fault phase current and the zero-sequence voltage fundamental obtained by the sparrow search algorithm optimized variational modal decomposition in step 4 of step 42) are made into a labeled training set XTrain and a test set YTest. Two fault features are set, and the size of each training sample is 2×512, with a total of 1698 samples. The size of each training sample in the test set is 2×512, with a total of 702 samples. Different fault degrees η are set and used as classification labels Label. The specific expressions of the training set, test set, fault degree, and classification label are: η=[0,0.05,0.1,0.15,0.2,0.25] Label = [1,2,3,4,5,6]; 62) Input the labeled training set XTrain and test set YTest in 61) into the SPP-ResCNN network for training and verification. The specific steps are as follows: Set the maximum number of training rounds to 10 and the minimum number of training batches to 64, then the number of iterations per round is 26, the initial learning rate is 0.01, and the loss function is the cross entropy function. The specific expression is as follows: Where n is the number of samples in the training set, m(i) is the true sample distribution, w(i) is the predicted distribution, y and f(x) are the true value and predicted value, and L(y,f(x)) is the cross entropy function loss error value; The optimizer selects Adam to update the network parameters to minimize the loss function. The specific network training steps are as follows: Step 1: Weight initialization process, using Gaussian distribution with a mean of 0 and a variance of 0.01 to initialize the weight coefficients of the network neuron nodes; Step 2: Input the training set XTrain prepared in 61) into SPP-ResCNN for forward propagation, that is, the training set is propagated layer by layer through the network constructed in 53), and the deep feature extraction of the training set is completed through convolution operation, pooling operation, residual operation and pyramid pooling in sequence, and finally input into the fully connected layer, which integrates and classifies the features, and obtains the classification results by the Softmax layer and the Classification layer; Step 3: When the output result of the SPP-ResCNN network does not match the expected value, the SPP-ResCNN network performs back propagation to find the error between the output result and the expected value, and then back propagates the error to the fully connected layer, pyramid pooling layer, residual layer, pooling layer, and convolution layer to update the weight coefficient of each layer; Step 4: Repeat steps 2 and 3 until the training termination condition is reached to complete the SPP-ResCNN training; Step 5: Input the test set YTest prepared in step 61) into the SPP-ResCNN trained in step 4 of step 62) for testing, calculate the classification accuracy based on the network output results, and use it as an indicator for network evaluation, and use the network convergence speed as a network performance evaluation indicator.

7. The motor turn-to-turn short circuit fault detection method based on variational mode decomposition fusion deep learning according to claim 1 is characterized in that: The real-time diagnosis of the inter-turn short circuit fault of the permanent magnet synchronous motor comprises the following steps: 71) Real-time detection of the three-phase current and zero-sequence voltage of the permanent magnet synchronous motor. When it is detected that the current and zero-sequence voltage of a certain phase are greater than the set threshold, it indicates that the phase is a fault phase. The current and zero-sequence voltage of the phase are extracted, which specifically includes the following steps: The detected fault phase current and zero-sequence voltage are input into the variational mode decomposition algorithm optimized by the sparrow search algorithm described in 42) to obtain the real-time third harmonic of the fault phase current and zero sequence voltage fundamental Its specific expression is: Where, t is time in seconds; 72) The third harmonic of the fault phase current collected in real time and zero sequence voltage fundamental As two features of turn-to-turn short-circuit fault, and make a training set X time Train and test set Y time Test, whose expression is as follows: 73) The prepared training set X time Train, test set Y time Test is input into the SPP-ResCNN network and the steps in 62) are repeated to obtain the turn-to-turn short circuit fault diagnosis result.

Images