CN117572300A - Motor turn-to-turn short circuit fault detection method based on variational modal decomposition fusion deep learning - Google Patents

Abstract

The invention relates to a motor turn-to-turn short circuit fault detection method based on variation modal decomposition fusion deep learning, which solves the defects of low turn-to-turn short circuit fault diagnosis precision and low efficiency compared with the prior art. The invention comprises the following steps: establishing a permanent magnet synchronous motor fault model; extracting fault characteristics by a variational modal decomposition algorithm; optimizing variation modal decomposition parameters by using a sparrow search algorithm; establishing a depth pyramid pooling residual convolution neural network model; training of SPP-ResCNN network; and (5) real-time diagnosis of turn-to-turn short circuit faults of the permanent magnet synchronous motor. The invention adopts the sparrow search algorithm to optimize the variational modal decomposition algorithm, eliminates other harmonic components in fault characteristics, extracts the fault characteristics with specific frequency, obviously improves the precision of the fault characteristics, adopts the pyramid pooling network and the residual error network to optimize the convolutional neural network, and improves the precision and the efficiency of fault diagnosis.

Description

Motor turn-to-turn short circuit fault detection method based on variational modal decomposition fusion deep learning

Technical Field

The invention relates to the technical field of permanent magnet synchronous motor fault diagnosis, in particular to a motor turn-to-turn short circuit fault detection method based on variation modal decomposition fusion deep learning.

Background

The permanent magnet synchronous motor is a core part of the electric drive system of the new energy automobile, and the working reliability of the permanent magnet synchronous motor directly influences the safety of the new energy automobile. Because the motor runs in a high-temperature and high-pressure environment for a long time, the coil is extremely easy to age, and turn-to-turn short circuit faults are further caused. When a fault occurs, if the short-circuit current in the fault loop is not detected in time, a large amount of heat is generated by the short-circuit current to further damage the insulation of the coil, so that the motor is partially burnt out, and the life and property safety of a driver is threatened. Therefore, the method has important significance for research on turn-to-turn short circuit fault diagnosis of the motor of the new energy automobile.

The turn-to-turn short circuit fault diagnosis of the new energy automobile motor mostly judges whether the fault occurs or not through state parameters caused by the fault. However, the initial fault characteristics of the turn-to-turn short circuit are not obvious, and after the motor fails, the motor does not operate stably and is interfered by other harmonics such as vibration noise, so that the precision of the fault characteristics is low, and the accuracy of fault diagnosis is seriously affected.

The signal processing method used at present, such as Fourier transform, wavelet transform and the like, has the problems that the bandwidth of extracted signals is high, noise harmonics in the characteristic frequency range can not be removed, and the precision is low. The traditional convolutional neural network has the defects of network gradient disappearance and the like caused by insufficient feature extraction and excessive number of neuron nodes.

Disclosure of Invention

The invention aims to solve the defects of low accuracy and low efficiency of turn-to-turn short circuit fault diagnosis in the prior art, and provides a motor turn-to-turn short circuit fault detection method based on variation modal decomposition fusion deep learning to solve the problems.

In order to achieve the above object, the technical scheme of the present invention is as follows:

a motor turn-to-turn short circuit fault detection method based on variation modal decomposition fusion deep learning comprises the following steps:

11 A permanent magnet synchronous motor fault model is established: establishing equations of d-axis and q-axis voltages and d-axis and q-axis currents of a rotating coordinate system according to a voltage-current equation under a natural coordinate system of the permanent magnet synchronous motor, and using the equations as a permanent magnet synchronous motor fault model for simulating fault phase turn-to-turn short circuits;

12 A variational modal decomposition algorithm extracts fault features: according to the simulation result of the permanent magnet synchronous motor fault model, obtaining motor fault phase current and zero sequence voltage, respectively inputting the motor fault phase current and zero sequence voltage into a variation modal decomposition algorithm to perform self-adaptive modal decomposition, and obtaining fault phase current third harmonic and zero sequence voltage fundamental wave as turn-to-turn short circuit fault characteristics;

13 Sparrow search algorithm optimizing variation modal decomposition parameters: optimizing a punishment factor alpha and a decomposition layer number K in a variation modal decomposition process in a set range by a sparrow search algorithm to obtain an optimal decomposition result of a fault phase current third harmonic and a zero sequence voltage fundamental wave;

14 A depth pyramid pooling residual convolution neural network model is established: establishing a deep pyramid pooling network and a residual convolution neural network, optimizing a residual convolution neural network model, namely an SPP-ResCNN network by adopting the deep pyramid pooling network, and adjusting the structure and parameters of the fused SPP-ResCNN network through an ablation test;

15 Training of SPP-ResCNN network: the optimal decomposition result of the third harmonic of the fault phase current and the zero sequence voltage fundamental wave, which are obtained by optimizing the variational modal decomposition through a sparrow search algorithm, is manufactured into a training set and a testing set, the training set and the testing set are input into an SPP-ResCNN network for training, and the learning rate, the maximum training round number and the minimum training batch are adjusted through an ablation test, so that the network loss function is reduced to the minimum;

16 Real-time diagnosis of turn-to-turn short circuit faults of the permanent magnet synchronous motor: the method comprises the steps of monitoring the current and zero sequence voltage of a permanent magnet synchronous motor in real time, inputting the current and zero sequence voltage of the permanent magnet synchronous motor into a variational modal decomposition algorithm optimized by a sparrow search algorithm, extracting the optimal decomposition result of the third harmonic wave and zero sequence voltage fundamental wave of the fault phase current, manufacturing the optimal decomposition result into a data set to be detected, and inputting the data set into a trained SPP-ResCNN network to diagnose the turn-to-turn short circuit fault degree.

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

21 Setting the voltage u of A phase, B phase and C phase of the permanent magnet synchronous motor under a natural coordinate system _a 、u _b 、u _c With current i of phase A, phase B and phase C _a 、i _b 、i _c And fault current i _f The equation between them is as follows:

wherein u is _a 、u _b 、u _c For the voltages of A phase, B phase and C phase, eta is the fault degree, eta is defined as the ratio of the short-circuit turns of the A phase stator winding to the total turns of the stator winding, R _s R is the stator resistance _f Is a fault resistance, ψ _f The fundamental wave amplitude of flux linkage, theta is the rotor electrical angle, i _a 、i _b 、i _c 、i _f The current and fault current of the stator windings of the phase A, the phase B and the phase C are respectively, and L, M is the self inductance and mutual inductance of the stator windings respectively;

22 The voltage equation under the natural coordinate system is transformed to the rotating coordinate system to obtain a permanent magnet synchronous motor fault model, the inter-turn short circuit fault feature extraction is facilitated, and the coordinate transformation matrix T is as follows:

the inverse of T is:

the coordinate transformation matrix T is applied to a voltage-current equation under a motor natural coordinate system to obtain the following steps:

wherein:

[u _dq0f ]＝[u _d u _q -u _n 0] ^T

[i _dq0f ]＝[i _d i _q 0 i _f ] ^T

[ψ _dq0f ]＝[ψ _f 0 ψ _3h cos3θ ηψ _a ] ^T

and (3) finishing to obtain:

wherein:

wherein u is _d 、u _q 、u _n I is the d, q axis voltage and neutral point voltage _d 、i _q For d, q-axis current, u _n Is neutral point voltage, ψ _d 、ψ _q 、ψ _fault Are d, q axes and fault phase flux linkage respectively, ψ _3h 、ψ _a 、ψ _f Respectively third harmonic, fault phase current flux linkage and fundamental flux linkage, L _AA Is A phase inductance, L _abcf 、R _abcf The inductance and resistance of A, B, C phase fault phase.

The fault feature extraction method by the variational modal decomposition algorithm comprises the following steps of:

31 According to the permanent magnet synchronous motor fault model obtained in the step 22), setting the rotating speed of 1000r/min and the torque of 2 N.m, so as to obtain the third harmonic wave of fault phase current and the fundamental wave data of zero sequence voltage;

setting a variation modal decomposition algorithm to extract fault characteristics comprises two processes of constructing a variation problem and solving the variation problem, carrying out iterative optimization on the basis of variation constraint, and determining the central frequency and bandwidth range of an input signal, wherein the variation constraint conditions are as follows:

s.t.∑ _k u _k (t)＝f(t)

wherein { u } _k }、{w _k Respectively represent modal components and center frequencies, { u } _k }＝{u ₁ ,u ₂ ,...u _k }，{w _k }＝{w ₁ ,w ₂ ,...w _k }；Is a first order bias derivative for time t; delta (t) is a dirac function; * Is a convolution operator; f (t) is an input signal; k is the number of modal components;

in order to solve the variation constraint problem, a penalty factor alpha and a Lagrange operator lambda (t) are introduced to convert the variation constraint problem into an unconstrained problem to be solved, and an initial value of variation modal decomposition is obtained, wherein the expression is as follows:

updating { u } by using multiplication alternate direction method _k }、{w _k Solving saddle points of the Lagrangian function, and finally obtaining modal components and center frequencies as follows:

wherein, The mode after Fourier transformation is finally obtained; />The center frequency is finally obtained;

32 The method comprises the following specific steps of) decomposing an input signal into K modes by using a variational mode decomposition, wherein each mode comprises different center frequencies, and the input signal decomposition comprises the following specific steps of:

step one: initialization ofλ ¹ And n is 0, set up the stackA substitute termination condition;

step two: let n=n+1, execute the outer loop variation modal decomposition algorithm;

step three: let k=k+1, perform the inner loop variation modal decomposition algorithm, i.e. update { u) according to the multiplicative alternating direction method in step 31) _k }、{w _k And λ, the updated rule is as follows:

wherein ζ is the discrimination accuracy; τ is the noise margin value, 1×10 ^-6 ；

Step four: repeating the second step and the third step until the iteration termination condition is met, so as to obtain a mode component of K;

33 Obtaining the center frequency omega of each modal component according to the variational modal decomposition algorithm _k After eliminating the higher harmonic noise, the center frequency of the third harmonic of the fault phase current is obtained asAnd the center frequency of the zero sequence voltage fundamental wave->From this the modal component corresponding to the center frequency is extracted +.>The modal component is data corresponding to a fault phase current third harmonic and a zero sequence voltage fundamental wave in a time domain, and the specific expression is:

The sparrow search algorithm optimizing variation modal decomposition parameters comprises the following steps:

41 Setting the finder position update rule in the foraging iteration process as follows:

wherein,the position of the ith sparrow in the j-th dimension; n is the iteration number; q and L are random numbers and identity matrixes obeying normal distribution respectively; r is R ₂ S is an alarm value and a safety value respectively, R ₂ ＝[0,1]，S＝[0.5,1]The method comprises the steps of carrying out a first treatment on the surface of the z is a random number between 0 and 1, iter _max For the maximum number of iterations, 20 is taken,

the follower position update rule is as follows:

wherein,the best and worst currently occupied positions for sparrow foraging are respectively; a is that ⁺ Is a meeting A ⁺ ＝A ^T (AA ^T ) ^-1 Is a matrix of (a); c is the number of sparrows, 20 is taken,

when a danger occurs during predation, the number of alertors in the sparrow population is randomly generated at a rate of 30%, and the mathematical model is as follows:

wherein beta and Q are [ -1,1]Random numbers in between; epsilon is the minimum random number for ensuring that the denominator is not zero;is a global optimal position; f (f) _i 、f _g 、f _w The current fitness value, the global optimal fitness value and the global worst fitness value of the population are respectively;

42 The detailed steps of the sparrow search algorithm optimization variation modal decomposition algorithm are as follows:

aiming at the initial value of the variation modal decomposition obtained in the step 32), the sparrow search algorithm is adopted to optimize the parameters of the variation modal decomposition algorithm, the fitness function is firstly determined as an iteration evaluation index during optimization, the comprehensive evaluation index is designed to be the minimum value of the sample entropy function, the Pelson coefficient and the relative aggregation algebra operation, and the specific expression is as follows:

feature＝SampEn(data，q,r)＝lnB ^q (r)-lnB ^q+1 (r)

pear＝corr(data ^T ,u _k ^T )

D＝lg(size(omega))

fitness＝min((feature/pear)*D)

Wherein feature, pear, D and fitness are sample entropy function values, pearson function values, relative aggregate values and adaptation values; size (omega) is the optimum center frequency ω for extraction _k Is a signal length of (a); sampEn and corr are the sample entropy function and pearson function respectively,

step one: initializing a population, setting the decomposition layer number K of a variation modal decomposition algorithm to be 1-10, setting the penalty factor alpha to be 500-3000, randomly initializing [ alpha, K ] as initial positions of producers and followers, and setting the proportion of the sparrow population producers and followers;

step two: carrying out variation modal decomposition operation on producers and followers at each position, calculating an initial fitness value, and updating the position of a finder according to the self-fitness value and the early warning value;

step three: updating positions of the follower and the alerter, and updating the fitness value and the optimal foraging position according to the updated positions;

step four: repeating the second step and the third step until the maximum iteration times or the loss convergence condition is reached;

obtaining the optimal decomposition layer number K and penalty factor alpha according to the steps, and obtaining [ alpha, K ]]In the input variation modal decomposition algorithm, the fault phase current and the zero sequence voltage are decomposed to obtain the center frequency of the third harmonic of the fault phase current and the zero sequence voltage fundamental wave From this the modal component corresponding to the center frequency is extracted +.> The specific time domain expression is as follows:

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

The establishing the depth pyramid pooling residual convolution neural network model comprises the following steps:

51 Setting a pyramid pooling network, pooling input data by setting a plurality of pooling channels and adopting filters with different sizes, and splicing pooling results with different scales together;

the pyramid pooling network structure is set as follows:

pyramid pooling channel one: setting the size of a filter as [3,512], setting the step length as 1, and adopting maximum pooling operation;

pyramid pooling channel two: setting the size of a filter as [64,512], setting the step length as 1, and adopting maximum pooling operation;

pyramid pooling channel three: setting the size of a filter as [128,512], setting the step length as 1, and adopting maximum pooling operation;

the pooling operation process of each channel is expressed as follows:

wherein,and->The output of the three channels is pooled for the pyramid; x is x ^l-1 Is input data; the same is a filling mode, so that the output data size is consistent with the input data size;

52 The structure of the set residual network is as follows:

Residual network first layer: the layer is a convolution layer, the size of convolution kernel is set to be 3 multiplied by 3, the number of convolution kernels is 64, and the filling mode is that the output size is consistent with the input size;

residual network second layer: the layer is a batch sample normalization layer, so that the network convergence efficiency is improved;

residual network third layer: the layer is an activation function layer, and a Relu function is selected;

residual network fourth layer: the layer is a convolution layer, the size of convolution kernels is set to be 5 multiplied by 5, the number of the convolution kernels is 128, and the filling mode is same;

residual network fifth layer: the layer is a batch sample normalization layer, so that the network convergence efficiency is improved;

53 Setting an SPP-ResCNN structure after the one-dimensional convolutional neural network, the pyramid pooling network and the residual error network are fused, namely a depth pyramid pooling residual error convolutional neural network model is as follows:

the first layer is an input layer, and the input data size of the layer is set to be 2 multiplied by 512;

the second layer is convolution layer 1, the convolution kernel size of the layer is set to be 1 multiplied by 3, the number of convolution kernels is set to be 64, the step length is set to be 1, the filling mode is the same, and the output of the layerThe method comprises the following steps:

the third layer is an activation function, and a Relu activation function is adopted, and the expression is as follows:

the fourth layer is a pooling layer 1, and the largest pooling is adopted to reduce the dimension of the input features and prevent the network from being over fitted, and the output of the layer The method comprises the following steps:

the fifth layer is a convolution layer 2, the convolution kernel size of the layer is 1*3, the number of convolution kernels is 128, the step length is 1, the filling mode is same, and the output of the layer is connected with the input of a residual error network;

the sixth layer is a batch sample normalization layer, and the data input into the layer is subjected to standardization processing, so that the convergence speed and generalization capability of the network are accelerated;

the seventh layer is a shortcut connection layer, the layer connects the residual network built in 52) and the output of the SPP-ResCNN sixth layer in 53), the layer is composed of a convolution layer 3 and a batch sample normalization layer, the convolution kernel size is 1*1, the convolution sum and the number are 128, the step length is 1, and the filling mode is same;

the eighth layer is a pyramid pooling layer, and the specific structure of the layer is shown in the step 51);

the ninth layer is a depth connection layer, and the layer is used for connecting the output of three channels of the pyramid pooling network, ensuring the consistent data size output by each channel, setting the number of channels of the layer to be 3, and the output size of the layer after data fusion to be 2 multiplied by 512 multiplied by 3;

the tenth layer is a fully connected layer for integrating the feature information extracted from the previous layers, mapping the sample distribution features to a sample mark space, setting the classification category number of the layer as 6, and outputting the fully connected layer The method comprises the following steps:

wherein W is ^l And b ^l The weight coefficient and the bias are adopted; f () is an activation function, x ^l-1 An input for the layer;

the eleventh layer is a Softmax layer that normalizes the output of the fully connected layer such that the value of each element is between 0 and 1 and the sum of all elements is 1, the layer output being the probability that the element is of a certain class;

the twelfth layer is a classification layer, which is used for outputting turn-to-turn short circuit fault classification results and calculating a loss function for back propagation.

The training of the SPP-ResCNN network comprises the following steps:

61 Manufacturing a training set and a testing set of the SPP-ResCNN network, which specifically comprises the following steps:

the method comprises the steps of preparing a training set XTrain and a test set YTEST with labels from fault phase current third harmonic and zero sequence voltage fundamental wave optimal decomposition results obtained by optimizing variational modal decomposition through a sparrow search algorithm, wherein the training set is 2X 512 in size, 1698 samples are taken as a total, the test set is 2X 512 in size, 702 samples are taken as a total, different fault degrees eta are set, the fault degrees are used as classification labels Label, and the training set, the test set, the fault degrees and classification labels are respectively expressed as follows:

η＝[0,0.05,0.1,0.15,0.2,0.25]

Label＝[1,2,3,4,5,6]；

62 Training and validating the training set XTrain and test set YTest with labels in 61) to the SPP-ResCNN network, the specific steps are as follows:

setting the maximum training round number as 10, the minimum training batch as 64, and the iteration number of each round as 26, initializing the learning rate as 0.01, and selecting a cross entropy function by a loss function, wherein the specific expression is as follows:

wherein n is the number of samples of the training set, m (i) is the real sample distribution, w (i) is the prediction distribution, y, f (x) are the real value and the prediction value, and L (y, f (x)) is the cross entropy function error difference;

the optimizer selects Adam for updating network parameters to minimize the loss function, and the network specific training steps are as follows:

step one: initializing a network neuron node weight coefficient by adopting Gaussian distribution with a mean value of 0 and a variance of 0.01 in the weight initialization process;

step two: inputting the training set XTrain manufactured in 61) into SPP-ResCNN for forward propagation, namely, propagating the training set layer by layer through a network built in 53), completing training set depth feature extraction through convolution operation, pooling operation, residual error operation and pyramid pooling in sequence, and finally inputting the training set XTrain into a full-connection layer, wherein the full-connection layer integrates and classifies the features, and classification results are obtained by a Softmax layer and a classification layer;

Step three: when the output result of the SPP-ResCNN network is not consistent with the expected value, the SPP-ResCNN network carries out back propagation to calculate the error of the output result and the expected value, and then the error is reversely propagated to the full-connection layer, the pyramid pooling layer, the residual layer, the pooling layer and the convolution layer, so that the weight coefficient of each layer is updated;

step four: repeating the second step and the third step until reaching the training termination condition, and completing SPP-ResCNN training;

step five: and (3) inputting the test set YTEST manufactured in 61) into the SPP-ResCNN trained in the step four for testing, calculating the accuracy of classification according to the result output by the network, taking the accuracy as an index of network evaluation, and taking the network convergence speed as an index of network performance evaluation.

The real-time diagnosis of the turn-to-turn short circuit fault of the permanent magnet synchronous motor comprises the following steps:

71 Detecting three-phase current and zero sequence voltage of the permanent magnet synchronous motor in real time, and when detecting that the current and the zero sequence voltage of a certain phase are larger than a set threshold value, indicating that the phase is a fault phase, extracting the current and the zero sequence voltage of the phase, wherein the method specifically comprises the following steps:

the detected fault phase current and the zero sequence voltage are input 42) in a variational modal decomposition algorithm optimized by the sparrow search algorithm to obtain the real-time fault phase current third harmonic And zero sequence voltage fundamental->The specific expression is as follows:

wherein t is time in seconds;

72 Fault phase current third harmonic to be collected in real time)And zero sequence voltage fundamental->As two characteristics of turn-to-turn short circuit fault and making training set X _time Train and test set Y _time Test, its expression is as follows:

73 To make the training set X _time Train, test set Y _time And (4) repeating the step 62) in the SPP-ResCNN network by inputting Test to obtain the turn-to-turn short circuit fault diagnosis result.

Advantageous effects

Compared with the prior art, the motor turn-to-turn short circuit fault detection method based on variation modal decomposition fusion deep learning adopts a sparrow search algorithm to optimize variation modal decomposition algorithm, eliminates other harmonic components in fault characteristics, extracts specific frequency fault characteristics, remarkably improves the precision of the fault characteristics, and adopts a pyramid pooling network and a residual error network to optimize a convolutional neural network, thereby improving the precision and efficiency of fault diagnosis.

1. According to the invention, the input signal is adaptively decomposed according to the characteristics of the fault characteristic signal by the variation modal decomposition algorithm, so that the defect that the bandwidth is too large and the harmonic wave is difficult to completely remove in the traditional signal decomposition means is avoided.

2. According to the invention, parameters in the variational modal decomposition process are optimized through the sparrow search algorithm, errors caused by artificial experience parameter selection are avoided, and the precision of the variational modal decomposition algorithm is remarkably improved by adopting a multi-adaptive function as a comprehensive evaluation index.

3. According to the invention, the pyramid pooling network, the residual network and the convolutional neural network are effectively fused, so that the problems of insufficient feature extraction, gradient disappearance and gradient explosion in the traditional convolutional neural network training process are solved, the features of input data are extracted to the maximum extent, the network diagnosis precision is improved, and the fault diagnosis efficiency is greatly accelerated.

Drawings

FIG. 1 is a process sequence diagram of the present invention;

FIG. 2 is a Fourier transform plot of the original signal;

FIG. 3 is a graph of a sparrow search algorithm optimization fitness function;

FIG. 4 is a graph comparing the decomposition of the variation mode after optimization of the sparrow search algorithm with the wavelet transformation result;

FIG. 5 is a graph comparing SPP-ResCNN with CNN loss functions;

fig. 6 is a diagram of the diagnosis result of the turn-to-turn short circuit fault of the permanent magnet synchronous motor.

Detailed Description

For a further understanding and appreciation of the structural features and advantages achieved by the present invention, the following description is provided in connection with the accompanying drawings, which are presently preferred embodiments and are incorporated in the accompanying drawings, in which:

As shown in fig. 1, the method for detecting the turn-to-turn short circuit fault of the permanent magnet synchronous motor based on the neural network technology comprises the following steps:

firstly, establishing a permanent magnet synchronous motor fault model: and establishing equations of d-axis and q-axis voltages and d-axis and q-axis currents of a rotating coordinate system according to a voltage and current equation under a natural coordinate system of the permanent magnet synchronous motor, and using the equations as a permanent magnet synchronous motor fault model for simulating fault phase turn-to-turn short circuits.

(1) Setting the voltage u of A phase, B phase and C phase of the permanent magnet synchronous motor under a natural coordinate system _a 、u _b 、u _c With current i of phase A, phase B and phase C _a 、i _b 、i _c And fault current i _f The equation between them is as follows:

wherein u is _a 、u _b 、u _c For the voltages of A phase, B phase and C phase, eta is the fault degree, eta is defined as the ratio of the short-circuit turns of the A phase stator winding to the total turns of the stator winding, R _s R is the stator resistance _f Is a fault resistance, ψ _f The fundamental wave amplitude of flux linkage, theta is the rotor electrical angle, i _a 、i _b 、i _c 、i _f The current and fault current of the stator windings of the phase A, the phase B and the phase C are respectively, and L, M is the self inductance and the mutual inductance of the stator windings.

(2) And transforming a voltage equation under a natural coordinate system to a rotating coordinate system to obtain a permanent magnet synchronous motor fault model, wherein the characteristic extraction of turn-to-turn short circuit faults is facilitated, and a coordinate transformation matrix T is as follows:

The inverse of T is:

the coordinate transformation matrix T is applied to a voltage-current equation under a motor natural coordinate system to obtain the following steps:

wherein:

[u _dq0f ]＝[u _d u _q -u _n 0] ^T

[i _dq0f ]＝[i _d i _q 0i _f ] ^T

[ψ _dq0f ]＝[ψ _f 0ψ _3h cos3θηψ _a ] ^T

and (3) finishing to obtain:

wherein:

wherein u is _d 、u _q 、u _n I is the d, q axis voltage and neutral point voltage _d 、i _q For d, q-axis current, u _n Is neutral point voltage, ψ _d 、ψ _q 、ψ _fault Are d, q axes and fault phase flux linkage respectively, ψ _3h 、ψ _a 、ψ _f Respectively third harmonic, fault phase current flux linkage and fundamental flux linkage, L _AA Is A phase inductance, L _abcf 、R _abcf The inductance and resistance of A, B, C phase fault phase.

Secondly, extracting fault characteristics by a variation modal decomposition algorithm: and obtaining motor fault phase current and zero sequence voltage according to a simulation result of the permanent magnet synchronous motor fault model, respectively inputting the motor fault phase current and zero sequence voltage into a variation modal decomposition algorithm to perform self-adaptive modal decomposition, and obtaining a fault phase current third harmonic and zero sequence voltage fundamental wave as turn-to-turn short circuit fault characteristics.

The traditional signal decomposition mode has the defects that noise harmonics near the center frequency of a signal cannot be thoroughly removed due to the fact that the frequency bandwidth is too high, and the fault feature accuracy is low.

(1) Setting the rotating speed of 1000r/min and the torque of 2 N.m according to the obtained fault model of the permanent magnet synchronous motor, so as to obtain the third harmonic wave of fault phase current and zero sequence voltage fundamental wave data;

setting a variation modal decomposition algorithm to extract fault characteristics comprises two processes of constructing a variation problem and solving the variation problem, carrying out iterative optimization on the basis of variation constraint, and determining the central frequency and bandwidth range of an input signal, wherein the variation constraint conditions are as follows:

s.t.∑ _k u _k (t)＝f(t)

wherein { u } _k }、{w _k Respectively represent modal components and center frequencies, { u } _k }＝{u ₁ ,u ₂ ,...u _k }，{w _k }＝{w ₁ ,w ₂ ,...w _k }；Is a first order bias derivative for time t; delta (t) is a dirac function; * Is a convolution operator; f (t) is an input signal; k is the number of modal components; />

In order to solve the variation constraint problem, a penalty factor alpha and a Lagrange operator lambda (t) are introduced to convert the variation constraint problem into an unconstrained problem to be solved, and an initial value of variation modal decomposition is obtained, wherein the expression is as follows:

updating { u } by using multiplication alternate direction method _k }、{w _k Solving saddle points of the Lagrangian function, and finally obtaining modal components and center frequencies as follows:

wherein,the mode after Fourier transformation is finally obtained; />The resulting center frequency.

(2) The variation mode decomposition decomposes the input signal into K modes, each mode containing a different center frequency, the input signal decomposition comprises the following specific steps:

Step one: initialization ofλ ¹ And n is 0, and setting an iteration termination condition;

step two: let n=n+1, execute the outer loop variation modal decomposition algorithm;

step three: let k=k+1, perform the inner loop variation modal decomposition algorithm, i.e. update { u) according to the multiplicative alternating direction method in step 31) _k }、{w _k And λ, the updated rule is as follows:

wherein ζ is the discrimination accuracy; τ is the noise margin value, 1×10 ^-6 。

Step four: repeating the second step and the third step until the iteration termination condition is met, so as to obtain a mode component of K;

(3) Obtaining the center frequency omega of each modal component according to the variational modal decomposition algorithm _k After eliminating the higher harmonic noise, the center frequency of the third harmonic of the fault phase current is obtained asAnd the center frequency of the zero sequence voltage fundamental wave->From this the modal component corresponding to the center frequency is extracted +.>The modal component is data corresponding to a fault phase current third harmonic and a zero sequence voltage fundamental wave in a time domain, and the specific expression is:

/>

as shown in fig. 2, the original signal has a plurality of harmonic components in the frequency domain image after fourier transformation, and only one harmonic of the fault characteristic can be represented, so that noise harmonic interference in the original signal is large, and the accuracy and efficiency of fault diagnosis are affected.

Thirdly, optimizing variation modal decomposition parameters by using a sparrow search algorithm: and optimizing the punishment factor alpha and the decomposition layer number K in the decomposition process of the variation mode through a sparrow search algorithm in a given range to obtain the optimal decomposition result of the third harmonic of the fault phase current and the zero sequence voltage fundamental wave.

The process of extracting fault characteristics by the variation modal decomposition algorithm can be known that the decomposition result is greatly influenced by the punishment factor alpha and the decomposition layer number K, and the punishment factor alpha and the decomposition layer number K are mainly set according to human experience at present. However, the mode generally causes the phenomena of over-decomposition and under-decomposition of signals, so that the problem that a variation modal decomposition algorithm is inconvenient to popularize and the like is caused. Aiming at the problem, the invention effectively combines the sparrow search algorithm with the variation modal decomposition algorithm, and the optimal punishment factor alpha and the decomposition layer number K are searched for in a layer-by-layer iteration mode through the sparrow search algorithm, so that the fault feature extraction precision is obviously improved.

(1) Setting the position updating rule of the discoverer in the foraging iterative process as follows:

wherein,the position of the ith sparrow in the j-th dimension; n is the iteration number; q and L are random numbers and identity matrixes obeying normal distribution respectively; r is R ₂ S is an alarm value and a safety value respectively, R ₂ ＝[0,1]，S＝[0.5,1]The method comprises the steps of carrying out a first treatment on the surface of the z is a random number between 0 and 1, iter _max For the maximum number of iterations, 20 is taken,

the follower position update rule is as follows:

wherein,the best and worst currently occupied positions for sparrow foraging are respectively; a is that ⁺ Is a meeting A ⁺ ＝A ^T (AA ^T ) ^-1 Is a matrix of (a); c is the number of sparrows, 20 is taken,

when a danger occurs during predation, the number of alertors in the sparrow population is randomly generated at a rate of 30%, and the mathematical model is as follows:

wherein beta and Q are [ -1,1]Random numbers in between; epsilon is the minimum random number for ensuring that the denominator is not zero;is a global optimal position; f (f) _i 、f _g 、f _w The current fitness value, the global best fitness value and the global worst fitness value of the population are respectively.

(2) The detailed steps of the sparrow search algorithm optimization variation modal decomposition algorithm are as follows:

aiming at obtaining an initial value of the variation modal decomposition, optimizing parameters of the variation modal decomposition algorithm by adopting a sparrow search algorithm, firstly determining a fitness function as an iteration evaluation index during optimization, and designing a comprehensive evaluation index as a minimum value of a sample entropy function, a Pelson coefficient and a relative aggregation algebra operation, wherein the specific expression is as follows:

feature＝SampEn(data，q,r)＝lnB ^q (r)-lnB ^q+1 (r)

D＝lg(size(omega))

fitness＝min((feature/pear)*D)

wherein feature, pear, D and fitness are sample entropy function values, pearson function values, relative aggregate values and adaptation values; size (omega) is the optimum center frequency ω for extraction _k Is a signal length of (a); sampEn and corr are the sample entropy function and pearson function respectively,

step one: initializing a population, setting the decomposition layer number K of a variation modal decomposition algorithm to be 1-10, setting the penalty factor alpha to be 500-3000, randomly initializing [ alpha, K ] as initial positions of producers and followers, and setting the proportion of the sparrow population producers and followers;

step two: carrying out variation modal decomposition operation on producers and followers at each position, calculating an initial fitness value, and updating the position of a finder according to the self-fitness value and the early warning value;

step three: updating positions of the follower and the alerter, and updating the fitness value and the optimal foraging position according to the updated positions;

step four: repeating the second step and the third step until the maximum iteration times or the loss convergence condition is reached;

obtaining optimal parameters [ alpha, K ] of the variation modal decomposition algorithm according to the steps]Is [2951,3 ]]As shown in FIG. 3, the fitness function converges after 5 iterations, with a fitness value of about 7.59X10 ^-3 Will [ alpha, K ]]In the input variation modal decomposition algorithm, the fault phase current and the zero sequence voltage are decomposed to obtain the center frequency of the third harmonic of the fault phase current and the zero sequence voltage fundamental wave From this the modal component corresponding to the center frequency is extracted +.>The specific time domain expression is as follows:

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

As shown in FIG. 4, the sparrow search algorithm provided by the invention is adopted to optimize the variational modal decomposition algorithm, and the modal components decomposed by the method only contain fault characteristic harmonic waves under the same layer number as wavelet transformation decomposition, so that the extraction precision is higher.

Fourth, establishing a depth pyramid pooling residual convolution neural network model: establishing a deep pyramid pooling network and a residual convolution neural network, optimizing the residual convolution neural network by adopting the deep pyramid pooling network, and adjusting the structure and parameters of the fused SPP-ResCNN network through an ablation test.

In order to improve the classification precision, the traditional convolutional neural network can continuously increase the network depth, so that network redundancy is caused and fitting is easy to carry out.

(1) Setting pyramid pooling network to pool input data by setting multiple pooling channels and adopting filters of different sizes, and splicing pooling results of different scales together, so as to extract fault characteristics more fully, wherein the structure of the pyramid pooling network is as follows:

pyramid pooling channel one: setting the size of a filter as [3,512], setting the step length as 1, and adopting maximum pooling operation;

pyramid pooling channel two: setting the size of a filter as [64,512], setting the step length as 1, and adopting maximum pooling operation;

pyramid pooling channel three: setting the size of a filter as [128,512], setting the step length as 1, and adopting maximum pooling operation;

the pooling operation process of each channel is expressed as follows:

wherein,and->The output of the three channels is pooled for the pyramid; x is x ^l-1 Is input data; the same is a filling mode, and the output and input data sizes are ensured to be consistent.

(2) The structure of the set residual network is as follows:

residual network first layer: the layer is a convolution layer, the size of convolution kernel is set to be 3 multiplied by 3, the number of convolution kernels is 64, and the filling mode is that the output size is consistent with the input size;

residual network second layer: the layer is a batch sample normalization layer, so that the network convergence efficiency is improved;

Residual network third layer: the layer is an activation function layer, and a Relu function is selected;

residual network fourth layer: the layer is a convolution layer, the size of convolution kernels is set to be 5 multiplied by 5, the number of the convolution kernels is 128, and the filling mode is same;

residual network fifth layer: the layer is a batch sample normalization layer, and network convergence efficiency is improved.

(3) The SPP-ResCNN structure after the one-dimensional convolutional neural network, the pyramid pooling network and the residual network are fused is set as follows:

the first layer is an input layer, and the input data size of the layer is set to be 2 multiplied by 512;

the second layer is convolution layer 1, the convolution kernel size of the layer is set to be 1 multiplied by 3, the number of convolution kernels is set to be 64, the step length is set to be 1, the filling mode is the same, and the output of the layerThe method comprises the following steps:

the third layer is an activation function, and a Relu activation function is adopted, and the expression is as follows:

the fourth layer is a pooling layer 1, and the largest pooling is adopted to reduce the dimension of the input features and prevent the network from being over fitted, and the output of the layerThe method comprises the following steps:

the fifth layer is a convolution layer 2, the convolution kernel size of the layer is 1*3, the number of convolution kernels is 128, the step length is 1, the filling mode is same, and the output of the layer is connected with the input of a residual error network;

the sixth layer is a batch sample normalization layer, and the data input into the layer is subjected to standardization processing, so that the convergence speed and generalization capability of the network are accelerated;

The seventh layer is a shortcut connection layer, the output of the residual network and the SPP-ResCNN sixth layer built in the connection of the layers is composed of a convolution layer 3 and a batch sample normalization layer, the convolution kernel size is 1*1, the convolution sum number is 128, the step length is 1, and the filling mode is same;

the eighth layer is a pyramid pooling layer;

the ninth layer is a depth connection layer, and the layer is used for connecting the output of three channels of the pyramid pooling network, ensuring the consistent data size output by each channel, setting the number of channels of the layer to be 3, and the output size of the layer after data fusion to be 2 multiplied by 512 multiplied by 3;

the tenth layer is a fully connected layer for integrating the feature information extracted from the previous layers, mapping the sample distribution features to a sample mark space, setting the classification category number of the layer as 6, and outputting x of the fully connected layer ^l _full The method comprises the following steps:

wherein W is ^l And b ^l The weight coefficient and the bias are adopted; f () is an activation function, x ^l-1 An input for the layer;

the eleventh layer is a Softmax layer that normalizes the output of the fully connected layer such that the value of each element is between 0 and 1 and the sum of all elements is 1, the layer output being the probability that the element is of a certain class;

the twelfth layer is a classification layer, which is used for outputting turn-to-turn short circuit fault classification results and calculating a loss function for back propagation.

Fifth step, training SPP-ResCNN network: and (3) manufacturing a training set and a testing set from the optimal decomposition result of the third harmonic of the fault phase current and the zero sequence voltage fundamental wave, which are obtained by optimizing variation modal decomposition through a sparrow search algorithm, inputting the training set and the testing set into the SPP-ResCNN for training, and adjusting the learning rate, the maximum training round number and the minimum training batch through an ablation test so as to minimize the network loss function.

(1) The method for manufacturing the training set and the testing set of the SPP-ResCNN network specifically comprises the following steps:

the obtained fault phase current third harmonic and zero sequence voltage fundamental wave optimal decomposition result obtained through the sparrow search algorithm optimization variation modal decomposition is manufactured into a training set XTrain and a testing set YTest with labels, two fault characteristics are set, the size of each training sample is 2×512, the total size of each training sample of the testing set is 1698 samples, the size of each training sample of the testing set is 2×512, 702 samples are all, different fault degrees eta are set, the fault degrees eta are used as classification labels Label, and the training set, the testing set, the fault degree and the classification labels are respectively expressed as follows:

η＝[0,0.05,0.1,0.15,0.2,0.25]

Label＝[1,2,3,4,5,6]；

(2) Inputting a training set XTrain and a test set YTEST with labels into an SPP-ResCNN network for training and verification, wherein the specific steps are as follows:

setting the maximum training round number as 10, the minimum training batch as 64, and the iteration number of each round as 26, initializing the learning rate as 0.01, and selecting a cross entropy function by a loss function, wherein the specific expression is as follows:

Wherein n is the number of samples of the training set, m (i) is the real sample distribution, w (i) is the prediction distribution, y, f (x) are the real value and the prediction value, and L (y, f (x)) is the cross entropy function error difference;

the optimizer selects Adam for updating network parameters to minimize the loss function, and the network specific training steps are as follows:

step one: initializing a network neuron node weight coefficient by adopting Gaussian distribution with a mean value of 0 and a variance of 0.01 in the weight initialization process;

step two: inputting the prepared training set XTrain into the SPP-ResCNN for forward propagation, namely propagating the training set layer by layer through a built network, completing training set depth feature extraction through convolution operation, pooling operation, residual error operation and pyramid pooling in sequence, finally inputting the training set XTrain into a full-connection layer, carrying out integrated classification on the features by the full-connection layer, and obtaining classification results by a Softmax layer and a classification layer;

step three: when the output result of the SPP-ResCNN network is not consistent with the expected value, the SPP-ResCNN network carries out back propagation to calculate the error of the output result and the expected value, and then the error is reversely propagated to the full-connection layer, the pyramid pooling layer, the residual layer, the pooling layer and the convolution layer, so that the weight coefficient of each layer is updated;

Step four: repeating the second step and the third step until reaching the training termination condition, and completing SPP-ResCNN training;

step five: and (3) inputting the manufactured test set YTEST into the SPP-ResCNN trained in the step four for testing, calculating the accuracy of classification according to the result output by the network, taking the accuracy as an index of network evaluation, and taking the network convergence speed as an index of network performance evaluation.

As shown in FIG. 5, compared with the CNN network, the loss function decreasing rate of the SPP-ResCNN is obviously accelerated, the CNN network achieves convergence about 150 rounds, the SPP-ResCNN achieves convergence about 50 rounds, and the convergence speed is increased by 200%; as shown in FIG. 6, the training result of the SPP-ResCNN network is 100% accurate, and the training set and the test set are correctly classified.

Sixth, real-time diagnosis of turn-to-turn short circuit faults of the permanent magnet synchronous motor: the method comprises the steps of monitoring the current and zero sequence voltage of a permanent magnet synchronous motor in real time, inputting the current and zero sequence voltage of the permanent magnet synchronous motor into a variational modal decomposition algorithm optimized by a sparrow search algorithm, extracting the optimal decomposition result of the third harmonic wave and zero sequence voltage fundamental wave of the fault phase current, manufacturing the optimal decomposition result into a data set to be detected, and inputting the data set into a trained SPP-ResCNN network to diagnose the turn-to-turn short circuit fault degree.

(1) Detecting three-phase current and zero sequence voltage of a permanent magnet synchronous motor in real time, and when detecting that the current and the zero sequence voltage of a certain phase are larger than a set threshold value, indicating that the phase is a fault phase, extracting the current and the zero sequence voltage of the phase, wherein the method specifically comprises the following steps:

inputting the detected fault phase current and the zero sequence voltage into a variational modal decomposition algorithm optimized by the sparrow search algorithm 3 b) to obtain real-time fault phase current third harmonicAnd zero sequence voltage fundamental->The specific expression is as follows:

wherein t is time in seconds;

(2) Fault phase current third harmonic wave collected in real timeAnd zero sequence voltage fundamental->As two characteristics of turn-to-turn short circuit fault and making training set X _time Train and test set Y _time Test, its expression is as follows:

(3) The training set X is manufactured _time Train, test set Y _time And (5) inputting the Test into an SPP-ResCNN network to obtain a turn-to-turn short circuit fault diagnosis result.

The foregoing has shown and described the basic principles, principal features and advantages of the invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, and that the above embodiments and descriptions are merely illustrative of the principles of the present invention, and various changes and modifications may be made therein without departing from the spirit and scope of the invention, which is defined by the appended claims. The scope of the invention is defined by the appended claims and equivalents thereof.

Claims (7)

1. The motor turn-to-turn short circuit fault detection method based on variation modal decomposition fusion deep learning is characterized by comprising the following steps of:

11 A permanent magnet synchronous motor fault model is established: establishing equations of d-axis and q-axis voltages and d-axis and q-axis currents of a rotating coordinate system according to a voltage-current equation under a natural coordinate system of the permanent magnet synchronous motor, and using the equations as a permanent magnet synchronous motor fault model for simulating fault phase turn-to-turn short circuits;

12 A variational modal decomposition algorithm extracts fault features: according to the simulation result of the permanent magnet synchronous motor fault model, obtaining motor fault phase current and zero sequence voltage, respectively inputting the motor fault phase current and zero sequence voltage into a variation modal decomposition algorithm to perform self-adaptive modal decomposition, and obtaining fault phase current third harmonic and zero sequence voltage fundamental wave as turn-to-turn short circuit fault characteristics;

13 Sparrow search algorithm optimizing variation modal decomposition parameters: optimizing a punishment factor alpha and a decomposition layer number K in a variation modal decomposition process in a set range by a sparrow search algorithm to obtain an optimal decomposition result of a fault phase current third harmonic and a zero sequence voltage fundamental wave;

14 A depth pyramid pooling residual convolution neural network model is established: establishing a deep pyramid pooling network and a residual convolution neural network, optimizing a residual convolution neural network model, namely an SPP-ResCNN network by adopting the deep pyramid pooling network, and adjusting the structure and parameters of the fused SPP-ResCNN network through an ablation test;

15 Training of SPP-ResCNN network: the optimal decomposition result of the third harmonic of the fault phase current and the zero sequence voltage fundamental wave, which are obtained by optimizing the variational modal decomposition through a sparrow search algorithm, is manufactured into a training set and a testing set, the training set and the testing set are input into an SPP-ResCNN network for training, and the learning rate, the maximum training round number and the minimum training batch are adjusted through an ablation test, so that the network loss function is reduced to the minimum;

16 Real-time diagnosis of turn-to-turn short circuit faults of the permanent magnet synchronous motor: the method comprises the steps of monitoring the current and zero sequence voltage of a permanent magnet synchronous motor in real time, inputting the current and zero sequence voltage of the permanent magnet synchronous motor into a variational modal decomposition algorithm optimized by a sparrow search algorithm, extracting the optimal decomposition result of the third harmonic wave and zero sequence voltage fundamental wave of the fault phase current, manufacturing the optimal decomposition result into a data set to be detected, and inputting the data set into a trained SPP-ResCNN network to diagnose the turn-to-turn short circuit fault degree.

2. The motor turn-to-turn short circuit fault detection method based on variational modal decomposition fusion deep learning as claimed in claim 1, wherein said establishing a permanent magnet synchronous motor fault model comprises the steps of:

21 Setting the voltage u of A phase, B phase and C phase of the permanent magnet synchronous motor under a natural coordinate system _a 、u _b 、u _c With current i of phase A, phase B and phase C _a 、i _b 、i _c And fault current i _f The equation between them is as follows:

wherein u is _a 、u _b 、u _c For the voltages of A phase, B phase and C phase, eta is the fault degree, eta is defined as the ratio of the short-circuit turns of the A phase stator winding to the total turns of the stator winding, R _s R is the stator resistance _f Is a fault resistance, ψ _f The fundamental wave amplitude of flux linkage, theta is the rotor electrical angle, i _a 、i _b 、i _c 、i _f The current and fault current of the stator windings of the phase A, the phase B and the phase C are respectively, and L, M is the self inductance and mutual inductance of the stator windings respectively;

22 The voltage equation under the natural coordinate system is transformed to the rotating coordinate system to obtain a permanent magnet synchronous motor fault model, the inter-turn short circuit fault feature extraction is facilitated, and the coordinate transformation matrix T is as follows:

the inverse of T is:

the coordinate transformation matrix T is applied to a voltage-current equation under a motor natural coordinate system to obtain the following steps:

wherein:

[u _dq0f ]＝[u _d u _q -u _n 0] ^T

[i _dq0f ]＝[i _d i _q 0 i _f ] ^T

[ψ _dq0f ]＝[ψ _f 0 ψ _3h cos3θ ηψ _a ] ^T

and (3) finishing to obtain:

wherein:

wherein u is _d 、u _q 、u _n I is the d, q axis voltage and neutral point voltage _d 、i _q For d, q-axis current, u _n Is neutral point voltage, ψ _d 、ψ _q 、ψ _fault Are d, q axes and fault phase flux linkage respectively, ψ _3h 、ψ _a 、ψ _f Respectively third harmonic, fault phase current flux linkage and fundamental flux linkage, L _AA Is A phase inductance, L _abcf 、R _abcf The inductance and resistance of A, B, C phase fault phase.

3. The motor turn-to-turn short circuit fault detection method based on variational modal decomposition fusion deep learning as claimed in claim 1, wherein the variational modal decomposition algorithm extracting fault features comprises the following steps:

31 According to the permanent magnet synchronous motor fault model obtained in the step 22), setting the rotating speed of 1000r/min and the torque of 2 N.m, so as to obtain the third harmonic wave of fault phase current and the fundamental wave data of zero sequence voltage;

setting a variation modal decomposition algorithm to extract fault characteristics comprises two processes of constructing a variation problem and solving the variation problem, carrying out iterative optimization on the basis of variation constraint, and determining the central frequency and bandwidth range of an input signal, wherein the variation constraint conditions are as follows:

s.t.∑ _k u _k (t)＝f(t)

wherein { u } _k }、{w _k Respectively represent modal components and center frequencies, { u } _k }＝{u ₁ ,u ₂ ,...u _k }，{w _k }＝{w ₁ ,w ₂ ,...w _k }；Is a first order bias derivative for time t; delta (t) is a dirac function; * Is a convolution operator; f (t) is an input signal; k is the number of modal components;

in order to solve the variation constraint problem, a penalty factor alpha and a Lagrange operator lambda (t) are introduced to convert the variation constraint problem into an unconstrained problem to be solved, and an initial value of variation modal decomposition is obtained, wherein the expression is as follows:

updating { u } by using multiplication alternate direction method _k }、{w _k Solution of saddle points of Lagrangian function, the resulting modal component andthe center frequency is:

wherein,the mode after Fourier transformation is finally obtained; />The center frequency is finally obtained;

32 The method comprises the following specific steps of) decomposing an input signal into K modes by using a variational mode decomposition, wherein each mode comprises different center frequencies, and the input signal decomposition comprises the following specific steps of:

Step one: initialization ofλ ¹ And n is 0, and setting an iteration termination condition;

step two: let n=n+1, execute the outer loop variation modal decomposition algorithm;

step three: let k=k+1, perform the inner loop variation modal decomposition algorithm, i.e. update { u) according to the multiplicative alternating direction method in step 31) _k }、{w _k And λ, the updated rule is as follows:

wherein ζ is the discrimination accuracy; τ is the noise margin value, 1×10 ^-6 ；

Step four: repeating the second step and the third step until the iteration termination condition is met, so as to obtain a mode component of K;

33 Obtaining the center frequency omega of each modal component according to the variational modal decomposition algorithm _k After eliminating the higher harmonic noise, the center frequency of the third harmonic of the fault phase current is obtained asAnd the center frequency of the zero sequence voltage fundamental wave->From this the modal component corresponding to the center frequency is extracted +.>The modal component is data corresponding to a fault phase current third harmonic and a zero sequence voltage fundamental wave in a time domain, and the specific expression is:

4. the motor turn-to-turn short circuit fault detection method based on variational modal decomposition fusion deep learning according to claim 1, wherein the sparrow search algorithm optimizing variational modal decomposition parameters comprises the following steps:

41 Setting the finder position update rule in the foraging iteration process as follows:

Wherein,the position of the ith sparrow in the j-th dimension; n is the iteration number; q and L are random numbers and identity matrixes obeying normal distribution respectively; r is R ₂ S is an alarm value and a safety value respectively, R ₂ ＝[0,1]，S＝[0.5,1]The method comprises the steps of carrying out a first treatment on the surface of the z is a random number between 0 and 1, iter _max For the maximum number of iterations, 20 is taken,

the follower position update rule is as follows:

wherein,the best and worst currently occupied positions for sparrow foraging are respectively; a is that ⁺ Is a meeting A ⁺ ＝A ^T (AA ^T ) ^-1 Is a matrix of (a); c is the number of sparrows, 20 is taken,

when a danger occurs during predation, the number of alertors in the sparrow population is randomly generated at a rate of 30%, and the mathematical model is as follows:

wherein beta and Q are [ -1,1]Random numbers in between; epsilon is the minimum random number for ensuring that the denominator is not zero;is a global optimal position; f (f) _i 、f _g 、f _w The current fitness value, the global optimal fitness value and the global worst fitness value of the population are respectively;

42 The detailed steps of the sparrow search algorithm optimization variation modal decomposition algorithm are as follows:

aiming at the initial value of the variation modal decomposition obtained in the step 32), the sparrow search algorithm is adopted to optimize the parameters of the variation modal decomposition algorithm, the fitness function is firstly determined as an iteration evaluation index during optimization, the comprehensive evaluation index is designed to be the minimum value of the sample entropy function, the Pelson coefficient and the relative aggregation algebra operation, and the specific expression is as follows:

feature＝SampEn(data，q,r)＝ln B ^q (r)-ln B ^q+1 (r)

D＝lg(size(omega))

fitness＝min((feature/pear)*D)

Wherein feature, pear, D and fitness are sample entropy function values, pearson function values, relative aggregate values and adaptation values; size (omega) is the optimum center frequency ω for extraction _k Is a signal length of (a); sampEn and corr are the sample entropy function and pearson function respectively,

step one: initializing a population, setting the decomposition layer number K of a variation modal decomposition algorithm to be 1-10, setting the penalty factor alpha to be 500-3000, randomly initializing [ alpha, K ] as initial positions of producers and followers, and setting the proportion of the sparrow population producers and followers;

step two: carrying out variation modal decomposition operation on producers and followers at each position, calculating an initial fitness value, and updating the position of a finder according to the self-fitness value and the early warning value;

step three: updating positions of the follower and the alerter, and updating the fitness value and the optimal foraging position according to the updated positions;

step four: repeating the second step and the third step until the maximum iteration times or the loss convergence condition is reached;

obtaining the optimal decomposition layer number K and penalty factor alpha according to the steps, and obtaining [ alpha, K ]]In the input variation modal decomposition algorithm, the fault phase current and the zero sequence voltage are decomposed to obtain the center frequency of the third harmonic of the fault phase current and the zero sequence voltage fundamental wave From this the modal component corresponding to the center frequency is extracted +.>The specific time domain expression is as follows:

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

5. The motor turn-to-turn short circuit fault detection method based on variational modal decomposition fusion deep learning of claim 1, wherein the establishing a deep pyramid pooling residual convolution neural network model comprises the following steps:

51 Setting a pyramid pooling network, pooling input data by setting a plurality of pooling channels and adopting filters with different sizes, and splicing pooling results with different scales together;

the pyramid pooling network structure is set as follows:

pyramid pooling channel one: setting the size of a filter as [3,512], setting the step length as 1, and adopting maximum pooling operation;

pyramid pooling channel two: setting the size of a filter as [64,512], setting the step length as 1, and adopting maximum pooling operation;

pyramid pooling channel three: setting the size of a filter as [128,512], setting the step length as 1, and adopting maximum pooling operation;

the pooling operation process of each channel is expressed as follows:

wherein,and->The output of the three channels is pooled for the pyramid; x is x ^l-1 Is input data; the same is a filling mode, so that the output data size is consistent with the input data size;

52 The structure of the set residual network is as follows:

residual network first layer: the layer is a convolution layer, the size of convolution kernel is set to be 3 multiplied by 3, the number of convolution kernels is 64, and the filling mode is that the output size is consistent with the input size;

residual network second layer: the layer is a batch sample normalization layer, so that the network convergence efficiency is improved;

residual network third layer: the layer is an activation function layer, and a Relu function is selected;

residual network fourth layer: the layer is a convolution layer, the size of convolution kernels is set to be 5 multiplied by 5, the number of the convolution kernels is 128, and the filling mode is same;

residual network fifth layer: the layer is a batch sample normalization layer, so that the network convergence efficiency is improved;

53 Setting an SPP-ResCNN structure after the one-dimensional convolutional neural network, the pyramid pooling network and the residual error network are fused, namely a depth pyramid pooling residual error convolutional neural network model is as follows:

the first layer is an input layer, and the input data size of the layer is set to be 2 multiplied by 512;

the second layer is convolution layer 1, the convolution kernel size of the layer is set to be 1 multiplied by 3, the number of convolution kernels is set to be 64, the step length is set to be 1, the filling mode is the same, and the output of the layerThe method comprises the following steps:

the third layer is an activation function, and a Relu activation function is adopted, and the expression is as follows:

The fourth layer is a pooling layer 1, and the largest pooling is adopted to reduce the dimension of the input features and prevent the network from being over fitted, and the output of the layerThe method comprises the following steps:

the fifth layer is a convolution layer 2, the convolution kernel size of the layer is 1*3, the number of convolution kernels is 128, the step length is 1, the filling mode is same, and the output of the layer is connected with the input of a residual error network;

the sixth layer is a batch sample normalization layer, and the data input into the layer is subjected to standardization processing, so that the convergence speed and generalization capability of the network are accelerated;

the seventh layer is a shortcut connection layer, the layer connects the residual network built in 52) and the output of the SPP-ResCNN sixth layer in 53), the layer is composed of a convolution layer 3 and a batch sample normalization layer, the convolution kernel size is 1*1, the convolution sum and the number are 128, the step length is 1, and the filling mode is same;

the eighth layer is a pyramid pooling layer, and the specific structure of the layer is shown in the step 51);

the ninth layer is a depth connection layer, and the layer is used for connecting the output of three channels of the pyramid pooling network, ensuring the consistent data size output by each channel, setting the number of channels of the layer to be 3, and the output size of the layer after data fusion to be 2 multiplied by 512 multiplied by 3;

the tenth layer is a fully connected layer for integrating the feature information extracted from the previous layers, mapping the sample distribution features to a sample mark space, setting the classification category number of the layer as 6, and outputting the fully connected layer The method comprises the following steps:

wherein W is ^l And b ^l The weight coefficient and the bias are adopted; f () is an activation function, x ^l-1 An input for the layer;

the eleventh layer is a Softmax layer that normalizes the output of the fully connected layer such that the value of each element is between 0 and 1 and the sum of all elements is 1, the layer output being the probability that the element is of a certain class;

the twelfth layer is a classification layer, which is used for outputting turn-to-turn short circuit fault classification results and calculating a loss function for back propagation.

6. The motor turn-to-turn short circuit fault detection method based on variational modal decomposition fusion deep learning as claimed in claim 1, wherein the training of the SPP-ResCNN network comprises the following steps:

61 Manufacturing a training set and a testing set of the SPP-ResCNN network, which specifically comprises the following steps:

the method comprises the steps of preparing a training set XTrain and a test set YTEST with labels from fault phase current third harmonic and zero sequence voltage fundamental wave optimal decomposition results obtained by optimizing variational modal decomposition through a sparrow search algorithm, wherein the training set is 2X 512 in size, 1698 samples are taken as a total, the test set is 2X 512 in size, 702 samples are taken as a total, different fault degrees eta are set, the fault degrees are used as classification labels Label, and the training set, the test set, the fault degrees and classification labels are respectively expressed as follows:

η＝[0,0.05,0.1,0.15,0.2,0.25]

Label＝[1,2,3,4,5,6]；

62 Training and validating the training set XTrain and test set YTest with labels in 61) to the SPP-ResCNN network, the specific steps are as follows:

setting the maximum training round number as 10, the minimum training batch as 64, and the iteration number of each round as 26, initializing the learning rate as 0.01, and selecting a cross entropy function by a loss function, wherein the specific expression is as follows:

wherein n is the number of samples of the training set, m (i) is the real sample distribution, w (i) is the prediction distribution, y, f (x) are the real value and the prediction value, and L (y, f (x)) is the cross entropy function error difference;

the optimizer selects Adam for updating network parameters to minimize the loss function, and the network specific training steps are as follows:

step one: initializing a network neuron node weight coefficient by adopting Gaussian distribution with a mean value of 0 and a variance of 0.01 in the weight initialization process;

step two: inputting the training set XTrain manufactured in 61) into SPP-ResCNN for forward propagation, namely, propagating the training set layer by layer through a network built in 53), completing training set depth feature extraction through convolution operation, pooling operation, residual error operation and pyramid pooling in sequence, and finally inputting the training set XTrain into a full-connection layer, wherein the full-connection layer integrates and classifies the features, and classification results are obtained by a Softmax layer and a classification layer;

Step three: when the output result of the SPP-ResCNN network is not consistent with the expected value, the SPP-ResCNN network carries out back propagation to calculate the error of the output result and the expected value, and then the error is reversely propagated to the full-connection layer, the pyramid pooling layer, the residual layer, the pooling layer and the convolution layer, so that the weight coefficient of each layer is updated;

step four: repeating the second step and the third step until reaching the training termination condition, and completing SPP-ResCNN training;

step five: and (3) inputting the test set YTEST manufactured in 61) into the SPP-ResCNN trained in the step four for testing, calculating the accuracy of classification according to the result output by the network, taking the accuracy as an index of network evaluation, and taking the network convergence speed as an index of network performance evaluation.

7. The motor turn-to-turn short circuit fault detection method based on variant modal decomposition fusion deep learning of claim 1, wherein the real-time diagnosis of the permanent magnet synchronous motor turn-to-turn short circuit fault comprises the following steps:

71 Detecting three-phase current and zero sequence voltage of the permanent magnet synchronous motor in real time, and when detecting that the current and the zero sequence voltage of a certain phase are larger than a set threshold value, indicating that the phase is a fault phase, extracting the current and the zero sequence voltage of the phase, wherein the method specifically comprises the following steps:

The detected fault phase current and the zero sequence voltage are input 42) in a variational modal decomposition algorithm optimized by the sparrow search algorithm to obtain the real-time fault phase current third harmonicAnd zero sequence voltage fundamental->The specific expression is as follows:

wherein t is time in seconds;

72 Fault phase current third harmonic to be collected in real time)And zero sequence voltage fundamental->As two characteristics of turn-to-turn short circuit fault and making training set X _time Train and test set Y _time Test, its expression is as follows:

73 To make the training set X _time Train, test set Y _time And (4) repeating the step 62) in the SPP-ResCNN network by inputting Test to obtain the turn-to-turn short circuit fault diagnosis result.