A method for detecting inter-turn short circuit faults in permanent magnet synchronous motors based on neural network technology

Abstract

The present invention relates to a method for detecting inter-turn short-circuit faults in permanent magnet synchronous motors based on neural network technology. Compared with the prior art, this method overcomes the drawbacks of low fault diagnosis accuracy, such as a small number of inter-turn short-circuit fault diagnosis samples, a single fault feature, and poor effectiveness. The method comprises the following steps: establishing an inter-turn short-circuit fault model for a permanent magnet synchronous motor; acquiring the current and voltage of the permanent magnet synchronous motor under different operating conditions; extracting third harmonics using a wavelet transform; calculating the fault negative-sequence current using a negative-sequence current optimization algorithm; training a convolutional neural network; obtaining data to be tested; and detecting inter-turn short-circuit faults in a permanent magnet synchronous motor. The method utilizes a wavelet transform to decompose the stator current signal under multiple motor operating conditions, thereby obtaining more accurate third harmonic data. A negative-sequence current optimization algorithm is employed to improve the accuracy of the negative-sequence current. The two types of fault features collected are then compiled into a data set and fed into a convolutional neural network for data feature extraction.

Description

Technical Field

[0001] The present invention relates to the technical field of permanent magnet synchronous motors, and in particular to a method for detecting inter-turn short circuit faults of permanent magnet synchronous motors based on neural network technology. Background Art

[0002] As a key device for power conversion or transmission, electric motors are widely used in daily life and production, and their operational reliability is receiving increasing attention. Stator winding interturn short-circuit faults are one of the main causes of permanent magnet synchronous motor failure, with an incidence rate as high as 30% to 40%. When a short-circuit fault is severe, the line protector can immediately cut off the power supply to protect the motor equipment. However, when the number of short-circuited turns is small, the motor can still maintain operation for a period of time. If stator winding interturn short-circuit faults can be detected and eliminated promptly during the fault latency period, the safety of electric vehicles can be guaranteed.

[0003] Current methods for detecting inter-turn short-circuit faults in permanent magnet synchronous motors rely on signal processing and setting fault thresholds. These methods process the current and voltage signals detected by sensors and compare them with pre-set fault thresholds to determine if a fault has occurred. This detection method is significantly affected by the motor's operating environment and relies on expert experience, significantly reducing diagnostic accuracy in complex operating conditions. Neural network-based fault diagnosis eliminates human intervention and addresses the shortcomings of signal processing-based methods. However, this method relies on a large amount of fault data. In real-world operating conditions, motor fault data samples are scarce and difficult to obtain, and the fault characteristics are relatively simple, resulting in low diagnostic accuracy.

[0004] The selection of fault features is the primary factor affecting fault diagnosis accuracy. Fault features should be easily accessible and distinct. Negative-sequence current and the third harmonic of the stator current are prominent and easily accessible features of turn-to-turn short-circuit faults. However, due to the motor's inherent structure, these fault features contain interference items not caused by turn-to-turn short-circuits. Therefore, eliminating these interference items is the primary method for improving fault diagnosis accuracy. Fusion of multiple fault features to create a dataset can enhance the robustness of fault features, offsetting the limited number of fault feature samples, and is another important method for improving fault diagnosis accuracy. Summary of the Invention

[0005] The purpose of the present invention is to solve the defects in the prior art of low fault diagnosis accuracy caused by the small number of turn-to-turn short-circuit diagnosis fault samples, single fault characteristics and poor effectiveness, and to provide a permanent magnet synchronous motor turn-to-turn short-circuit fault detection method based on neural network technology 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 method for detecting inter-turn short-circuit faults in a permanent magnet synchronous motor based on neural network technology comprises the following steps:

[0008] 11) Establish a permanent magnet synchronous motor inter-turn short circuit fault model: Based on the equivalent circuit when the permanent magnet synchronous motor phase A has an inter-turn short circuit, establish the voltage u of phase A, phase B, and phase C. a 、u b 、u c With the A phase, B phase, C phase current i a 、i b 、i c and fault current i f The equation between and establish the torque T e Equation, which is used as the inter-turn short-circuit fault model of permanent magnet synchronous motor;

[0009] 12) Obtain the current and voltage of the permanent magnet synchronous motor under different working conditions: According to the simulation results of the permanent magnet synchronous motor inter-turn short-circuit fault model under different working conditions, obtain the A-phase, B-phase, and C-phase currents i under the inter-turn short-circuit state of the permanent magnet synchronous motor. a 、i b 、i c With the voltage u of phase A, phase B, and phase C a 、u b 、u c ;

[0010] 13) Wavelet transform to extract the third harmonic: The current i of phase A is transformed by wavelet transform. a Decompose and obtain the amplitude of its third harmonic I a_3h As a fault characteristic;

[0011] 14) Negative sequence current optimization algorithm calculates the fault negative sequence current: through the A phase, B phase, C phase current i a 、i b 、i c Calculate the total negative sequence current I - The negative sequence current caused by the inter-turn short circuit is obtained by eliminating the interference term through the negative sequence current optimization algorithm.

[0012] 15) Training of convolutional neural network: The third harmonic amplitude I a_3h and fault negative sequence current At the same time, it is used as the fault feature of inter-turn short circuit and made into a training data set, which is input into the convolutional neural network for training;

[0013] 16) Obtaining the data to be tested: Monitoring the A-phase, B-phase, and C-phase currents i under the actual working conditions of the permanent magnet synchronous motor a ′、i b ′、i c ′, Phase A, Phase B, Phase C voltage ua ′、u b ′、u c ′, and then obtain the third harmonic amplitude I under the actual working condition of the permanent magnet synchronous motor a_3h ′ and fault negative sequence current And make it into a data set to be tested;

[0014] 17) Detection of inter-turn short-circuit faults in permanent magnet synchronous motors: The dataset to be tested is input into the trained convolutional neural network to obtain the detection results of inter-turn short-circuit faults in permanent magnet synchronous motors.

[0015] The establishment of the permanent magnet synchronous motor turn-to-turn short circuit fault model comprises the following steps:

[0016] 21) Set the permanent magnet synchronous motor inter-turn short circuit fault model to be composed of phase A, phase B, and phase C voltage u a 、u b 、u c With the A phase, B phase, C phase current i a 、i b 、i c and fault current i f The equation between them and the inter-turn short-circuit electromagnetic torque equation of the permanent magnet synchronous motor;

[0017] 22) Set the voltage u of phase A, phase B and phase C a 、u b 、u c With the A phase, B phase, C phase current i a 、i b 、i c and fault current i f The equation between is as follows:

[0018]

[0019] Among them, u a 、u b 、u c is the voltage of phase A, phase B, and phase C, R s is the stator resistance, u is the fault degree, phase A is the fault phase, then u is the ratio of the short-circuit turns of phase A stator winding to the total number of turns of its stator winding, R f is the fault resistance, i a 、i b 、i c is the current of the stator winding of phase A, phase B, and phase C, i f is the fault current, L is the self-inductance of the stator winding, M is the mutual inductance of the stator winding, ψ f is the fundamental amplitude of the flux linkage, θ is the rotor electrical angle;

[0020] 23) The electromagnetic torque equation of the permanent magnet synchronous motor inter-turn short circuit is set as follows:

[0021]

[0022] Among them, T e is the electromagnetic torque of the permanent magnet synchronous motor, p n is the number of pole pairs of the permanent magnet synchronous motor, i d 、i q is the current of the permanent magnet synchronous motor d-axis and q-axis, L d , L q is the inductance of the d-axis and q-axis of the permanent magnet synchronous motor, ψ f is the fundamental amplitude of the magnetic flux.

[0023] The method of obtaining the current and voltage of the permanent magnet synchronous motor under different working conditions comprises the following steps:

[0024] 31) The linearly varying speed N, the linearly varying torque T, and the four fault levels u are combined to serve as the simulation conditions for the inter-turn short-circuit fault of the permanent magnet synchronous motor. The specific values of the speed N, torque T, and fault level u are as follows:

[0025] N=300n, where n=0, 1, 2…, 10,

[0026] T=[0,1,2,3,4,5,6,7,8,9,10],

[0027] u=[0,0.1,0.2,0.3];

[0028] 32) Perform simulation based on the permanent magnet synchronous motor inter-turn short circuit fault model and in combination with the set simulation working conditions of the permanent magnet synchronous motor inter-turn short circuit fault;

[0029] According to the simulation results, the A-phase current i of the permanent magnet synchronous motor under the inter-turn short-circuit state is obtained. a and B and C phase current i b 、i c , A phase voltage u a and B and C phase voltage u b 、u c , its specific expression is as follows:

[0030] i a (k)=(i a (1),i a (2)...i a (n)),

[0031] i b (k)=(i b (1),i b (2)...i b(n)),

[0032] i c (k)=(i c (1),i c (2)...i c (n)),

[0033] u a (k)=(u a (1),u a (2)...u a (n)),

[0034] u b (k)=(u b (1),u b (2)...u b (n)),

[0035] u c (k)=(u c (1),u c (2)...u c (n)),

[0036] Where k is the sequence number, and its value range is 1, 2, ..., n.

[0037] The wavelet transform to extract the third harmonic comprises the following steps:

[0038] 41) Assume that the specific expression of the wavelet transform algorithm is as follows:

[0039] The basic wavelet ψ(t) is scaled and translated to obtain the wavelet function family ψ a,b (t):

[0040]

[0041] Among them, t is the independent variable, a is the scaling factor, and b is the translation factor;

[0042] Then the continuous wavelet transform of any function x(t) is expressed as:

[0043]

[0044] Considering the phase A current i a Since it is discrete data, the continuous wavelet transform is discretized and the expression is:

[0045]

[0046]

[0047] Using binary discrete form, that is, taking a0 = 2, b0 = 1, the binary discrete wavelet transform is expressed as:

[0048] ψ j,k (t) = 2 -j / 2 ψ(2 -j tk),

[0049] The discrete wavelet transform is expressed as:

[0050]

[0051] in, for The conjugate of , a0 is the discretization scaling factor, b0 is the discretization translation factor, j and m are integers;

[0052] 42) The obtained A phase current i a After discrete wavelet transform, the transformed signal composition has the following relationship:

[0053] i a =A n +D n +D n-1 +...+D1,

[0054] Among them, A n is the approximate signal of the nth layer, D n D n-1 ....D1 is the n-layer detail signal;

[0055] 43) Assume that the fault feature is contained in the detail signal after discrete wavelet transform, and the frequency range of the detail signal is as follows:

[0056]

[0057]

[0058] Among them, f n is the frequency of the signal to be extracted, f is the signal sampling frequency, n is the number of decomposition layers, freq(D j ) is the detail signal D j The frequency of , j is a constant whose value range is [1,n];

[0059] 44) According to the frequency range freq(D j ) and its corresponding number of layers n, we can get the third harmonic amplitude of phase A current I a_3h , and use it as the fault feature, its specific expression is:

[0060] I a_3h (k)=(ia 3h (1),ia3h (2)...ia 3h (n)),

[0061] Where k is the sequence number, and its value range is [1,n].

[0062] The negative sequence current optimization algorithm for calculating the fault negative sequence current comprises the following steps:

[0063] 51) Combine the obtained A phase, B phase, and C phase currents i a 、i b 、i c , the negative sequence current I - The calculation formula is as follows:

[0064]

[0065]

[0066]

[0067] Among them, i a 、i b 、i c is the current of phase A, phase B, and phase C, I + , I - is the amplitude of the positive and negative sequence currents, ω is the fundamental frequency of the current, φ1 and φ2 are the phase angles of the positive and negative sequence currents, α is the operator, α=e j2π / 3 That is 120°;

[0068] 52) Calculated negative sequence current I - Not only does it include the negative sequence current caused by the inter-turn short circuit It also includes the error negative sequence current caused by unstable motor operation and structural asymmetry In order to improve the accuracy of fault diagnosis, the negative sequence current I - For optimization, the optimization algorithm is as follows:

[0069] When no inter-turn short circuit occurs:

[0070]

[0071]

[0072]

[0073]

[0074] When a turn-to-turn short circuit occurs:

[0075]

[0076]

[0077]

[0078]

[0079] Among them, V - is the negative sequence voltage, I - is the negative sequence current, u a 、u b 、u c is the voltage of phase A, phase B, and phase C, i a 、i b 、i c is the current of phase A, phase B, and phase C, Z - is the negative sequence impedance, α is the operator, α=e j2π / 3 That is 120°, is the error negative sequence current, is the fault negative sequence current;

[0080] 53) Obtain the fault negative sequence current according to the negative sequence current optimization algorithm And use it as another fault feature, the fault negative sequence current The specific expression is:

[0081]

[0082] Where k is the sequence number, and its value range is [1,n].

[0083] The training of the convolutional neural network includes the following steps:

[0084] 61) Create a training data set, specifically including the following steps:

[0085] The third harmonic amplitude of the A phase current I a_3h and fault negative sequence current As two characteristics of the inter-turn short circuit fault of the permanent magnet synchronous motor, the training data set data1 is a two-dimensional array, and its expression is as follows:

[0086]

[0087] 62) To build a convolutional neural network, the steps are as follows:

[0088] The convolutional neural network architecture is set to have 11 layers, consisting of an input layer, a convolutional layer, an activation function, a pooling layer, a fully connected layer, and an output layer; among them:

[0089] The first layer is the input layer, and the input feature dimension of this layer is set to 2;

[0090] The second layer is the convolution layer, and the convolution kernel size is set to 3*3 and the number of convolution kernels is set to 64;

[0091] The third layer is the activation function. The constructed convolutional neural network uses the Relu activation function, which is expressed as follows:

[0092]

[0093] The fourth layer is the normalization layer, which is used to prevent overfitting and gradient explosion and improve the training efficiency of convolutional neural networks;

[0094] The fifth layer is the pooling layer. The convolutional neural network constructed uses maximum pooling. Maximum pooling refers to selecting the maximum value of the feature area as the pooled value of the area.

[0095] The sixth layer is the convolution layer, and the convolution kernel size is set to 3*3 and the number of convolution kernels is set to 128;

[0096] The seventh layer is the activation function layer, which uses the Relu activation function;

[0097] The eighth layer is the pooling layer, which uses maximum pooling;

[0098] The ninth layer is a fully connected layer, and the feature dimension of this layer is set to 4;

[0099] The tenth layer is the Softmax layer. The Softmax layer outputs the probability of each type in the classification case. The sum of the output probabilities is 1. The size of the probability value can be used to determine whether the data feature belongs to a certain type.

[0100] The eleventh layer is the output layer. The convolutional neural network is used to classify the inter-turn short-circuit fault of the permanent magnet synchronous motor, so classification is used.

[0101] 63) Input the training data set into the constructed convolutional neural network for training. The specific training steps are as follows:

[0102] The convolutional neural network training sets the minimum batch size for each round of training to 15, with a total of 1000 rounds of training. The training consists of the following three stages:

[0103] The first stage is the weight initialization process, which uses a Gaussian distribution with a mean of 0 and a variance of 0.01 to randomly initialize the weights;

[0104] The second stage is the forward propagation process. The input training set is continuously reduced in feature dimension through the convolution layer and pooling layer, and feature extraction is finally input to the fully connected layer. The specific process is as follows:

[0105] The convolution operation is expressed as follows:

[0106] x l=f(x l-1 *W l +b l )

[0107] The pooling operation is expressed as follows:

[0108] x l =pooling(x l-1 )

[0109] The output of the fully connected layer is:

[0110] x l =f(W l x l-1 +b l )

[0111] Among them, x l is the output of the lth layer, W l is the weight of the lth layer, b l is the bias of the lth layer, f() is the activation function, and pooling() is the pooling operation;

[0112] The third stage is the back propagation process, which compares the output value with the true value and uses the loss function to represent the error. The error function used in training is the cross entropy function, and its specific expression is:

[0113]

[0114] Where n is the number of samples in the training set, m(i) is the true sample distribution, w(i) is the predicted distribution, and L(y,f(x)) is the cross entropy loss error value;

[0115] The cross entropy function loss error value L(y,f(x)) is back-propagated to the fully connected layer, pooling layer, and convolution layer to update the weights and biases, and then the forward propagation process is continued. The three stages are iteratively trained to minimize the loss function. At this time, it indicates that the convolutional neural network has reached the convergence condition. Further training is performed until the convolutional neural network reaches the iterative termination condition and the training is terminated; the trained convolutional neural network is obtained.

[0116] The acquisition of the data to be detected comprises the following steps:

[0117] 71) Obtain the third harmonic amplitude I under actual working conditions a_3h ', specifically comprising the following steps:

[0118] The A-phase current i under the actual working condition of the permanent magnet synchronous motor obtained by monitoring a 'Perform discrete wavelet transform in steps 41), 42), and 43) to obtain the A phase current i under actual working conditions a The third harmonic amplitude of ′ is Ia_3h ′, its specific expression is:

[0119] I a_3h ′(k)=(ia′ 3h (1),ia′ 3h (2)...ia′ 3h (n)),

[0120] Where k is the sequence number, and its value range is [1,n];

[0121] 72) Obtain the negative sequence current under actual working conditions The specific steps include:

[0122] The A-phase, B-phase, and C-phase currents i under the actual working conditions of the permanent magnet synchronous motor obtained by monitoring a ′、i b ′、i c ′, Phase A, Phase B, Phase C voltage u a ′、u b ′、u c ', perform the negative sequence current calculation and optimization in steps 51) and 52) to obtain the negative sequence current under actual working conditions The specific expression is:

[0123]

[0124] Where k is the sequence number, and its value range is [1,n];

[0125] 74) The third harmonic amplitude of the A phase current under the actual working condition is obtained a_3h ′ and fault negative sequence current As two characteristics of the inter-turn short-circuit fault of the permanent magnet synchronous motor under actual working conditions, the prepared data set data1′ to be tested is a two-dimensional array, and its expression is as follows:

[0126]

[0127] The permanent magnet synchronous motor inter-turn short circuit fault detection comprises the following steps:

[0128] 81) Input the data set data1′ to be tested into the trained convolutional neural network.

[0129] 82) The convolutional neural network performs the training process in step 63) until the convolutional neural network reaches the iteration termination condition.

[0130] 83) Output the classification result of the data set data1′ to be tested, and complete the diagnosis of the permanent magnet synchronous motor inter-turn short circuit fault.

[0131] Beneficial effects

[0132] The present invention provides a permanent magnet synchronous motor inter-turn short-circuit fault detection method based on neural network technology. Compared with the existing technology, the stator current signal under multiple working conditions of the motor is decomposed by wavelet transform to obtain more accurate third harmonic data. The negative sequence current optimization algorithm is used to improve the negative sequence current accuracy. At the same time, the two types of fault features collected are made into a data set to enhance the robustness of the fault features. The data are then transmitted to a convolutional neural network for data feature extraction, thereby achieving higher detection accuracy.

[0133] The wavelet transform involved in the present invention can effectively decompose the stator current signal of the motor under multiple working conditions in the time domain and frequency domain, thereby obtaining more accurate third harmonic data and improving the fault diagnosis accuracy; the negative sequence current optimization algorithm can effectively eliminate the asymmetric factors of the motor body and improve the fault diagnosis accuracy; the convolutional neural network has a powerful data feature extraction capability and can effectively classify inter-turn short circuit faults of different degrees. Compared with traditional fault detection methods, its diagnostic accuracy is higher.

[0134] The present invention also has the following advantages:

[0135] 1. The stator current of the motor under multiple working conditions is extracted by wavelet transform, which greatly improves the accuracy of the third harmonic amplitude obtained under various complex working conditions;

[0136] 2. The negative sequence current optimization algorithm eliminates the error negative sequence current caused by the motor itself, thereby improving the calculation accuracy of the negative sequence current;

[0137] 3. Using the two fault features obtained under multiple working conditions as the feature quantity for fault detection greatly improves the defects of fault detection such as poor robustness and insufficient data caused by a single fault feature;

[0138] 4. Apply convolutional neural networks to fault detection. The convolutional neural network is used to identify fault features of the original fault data. Compared with the traditional threshold detection method, the accuracy of fault detection is significantly improved, and the classification of different degrees of inter-turn short circuit faults is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0139] Figure 1 is a method sequence diagram of the present invention;

[0140] Figure 2 This is a comparison chart between wavelet transform and Fourier transform;

[0141] Figure 3 This is a comparison chart of negative sequence current optimization algorithms;

[0142] Figure 4 This is the convolutional neural network classification result diagram. DETAILED DESCRIPTION

[0143] In order to provide a further understanding and appreciation of the structural features and effects achieved by the present invention, a detailed description is provided with reference to preferred embodiments and accompanying drawings as follows:

[0144] like Figure 1 As shown, the method for detecting inter-turn short-circuit faults of permanent magnet synchronous motors based on neural network technology of the present invention includes the following steps:

[0145] The first step is to establish a permanent magnet synchronous motor inter-turn short circuit fault model. Based on the equivalent circuit when the permanent magnet synchronous motor has an inter-turn short circuit, the voltages u of phase A, phase B, and phase C are established. a 、u b 、u c With the A phase, B phase, C phase current i a 、i b 、i c and fault current i f The equation between and establish the torque T e The equation is used as the permanent magnet synchronous motor turn-to-turn short circuit fault model. The specific steps are as follows:

[0146] (1) The permanent magnet synchronous motor inter-turn short circuit fault model is set by the voltage u of phase A, phase B and phase C. a 、u b 、u c With the A phase, B phase, C phase current i a 、i b 、i c and fault current i f The equation between them and the inter-turn short-circuit electromagnetic torque equation of the permanent magnet synchronous motor.

[0147] (2) Set the voltage u of phase A, phase B, and phase C a 、u b 、u c With the A phase, B phase, C phase current i a 、i b 、i c and fault current i f The equation between is as follows:

[0148]

[0149] Among them, u a 、u b 、u c is the voltage of phase A, phase B, and phase C, R s is the stator resistance, u is the fault degree, phase A is the fault phase, then u is the ratio of the short-circuit turns of phase A stator winding to the total number of turns of its stator winding, R f is the fault resistance, i a 、ib 、i c is the current of the stator winding of phase A, phase B, and phase C, i f is the fault current, L is the self-inductance of the stator winding, M is the mutual inductance of the stator winding, ψ f is the fundamental amplitude of the flux linkage, θ is the rotor electrical angle;

[0150] (3) The electromagnetic torque equation of the permanent magnet synchronous motor inter-turn short-circuit is set as follows:

[0151]

[0152] Among them, T e is the electromagnetic torque of the permanent magnet synchronous motor, p n is the number of pole pairs of the permanent magnet synchronous motor, ψ f is the fundamental amplitude of the magnetic flux i d 、i q is the current of the permanent magnet synchronous motor d-axis and q-axis, L d , L q is the inductance of the d-axis and q-axis of the permanent magnet synchronous motor.

[0153] The second step is to obtain the current and voltage of the permanent magnet synchronous motor under different working conditions: According to the simulation results of the permanent magnet synchronous motor interturn short circuit fault model under different working conditions, the A-phase, B-phase, and C-phase voltages u of the permanent magnet synchronous motor under the interturn short circuit state are obtained. a 、u b 、u c With the A phase, B phase, C phase current i a 、i b 、i c .

[0154] The specific steps are as follows:

[0155] (1) The linearly varying speed N, the linearly varying torque T, and the four fault levels u are combined to serve as the simulation conditions for the inter-turn short-circuit fault of the permanent magnet synchronous motor. The specific values of speed N, torque T, and fault level u are as follows:

[0156] N=300n, where n=0, 1, 2…, 10,

[0157] T=[0,1,2,3,4,5,6,7,8,9,10],

[0158] u=[0,0.1,0.2,0.3].

[0159] (2) performing simulation based on the permanent magnet synchronous motor inter-turn short circuit fault model and in combination with the set simulation working conditions of the permanent magnet synchronous motor inter-turn short circuit fault;

[0160] According to the simulation results, the A-phase current i of the permanent magnet synchronous motor under the inter-turn short-circuit state is obtained. a and B and C phase current i b 、i c and A phase voltage u a and B and C phase voltage u b 、u c , its specific expression is as follows:

[0161] i a (k)=(i a (1),i a (2)...i a (n))

[0162] i b (k)=(i b (1),i b (2)...i b (n))

[0163] i c (k)=(i c (1),i c (2)...i c (n))

[0164] u a (k)=(u a (1),u a (2)...u a (n))

[0165] u b (k)=(u b (1),u b (2)...u b (n))

[0166] u c (k)=(u c (1),u c (2)...u c (n))

[0167] Where k is the sequence number, and its value range is 1, 2, ..., n.

[0168] The third step is to extract the third harmonic by wavelet transform: the current i of phase A is transformed by wavelet transform. a Decompose and obtain the amplitude of its third harmonic I a_3h As a fault feature. Compared with traditional signal processing methods, wavelet transform requires the selection of wavelets suitable for the signal to be analyzed, and the order of the wavelet and the number of decomposition layers need to be continuously debugged to obtain the correct time-frequency signal.

[0169] The specific steps of wavelet transform to extract the third harmonic are as follows:

[0170] (1) Assume that the specific expression of the wavelet transform algorithm is as follows:

[0171] The basic wavelet ψ(t) is scaled and translated to obtain the wavelet function family ψ a,b (t):

[0172]

[0173] Among them, t is the independent variable, a is the scaling factor, and b is the translation factor;

[0174] Then the continuous wavelet transform of any function x(t) is expressed as:

[0175]

[0176] Considering the phase A current i a Since it is discrete data, the continuous wavelet transform is discretized and the expression is:

[0177]

[0178]

[0179] Using binary discrete form, that is, taking a0 = 2, b0 = 1, the binary discrete wavelet transform is expressed as:

[0180] ψ j,k (t) = 2 -j / 2 ψ(2 -j tk),

[0181] The discrete wavelet transform is expressed as:

[0182]

[0183] in, for The conjugate of , a0 is the discretization scaling factor, b0 is the discretization translation factor, j and m are integers.

[0184] (2) The obtained A phase current i a After discrete wavelet transform, the transformed signal composition has the following relationship:

[0185] i a =A n +D n +D n-1 +...+D1,

[0186] Among them, A n is the approximate signal of the nth layer, D n Dn-1 ....D1 is the n-layer detail signal.

[0187] (3) Assume that the fault feature is contained in the detail signal after discrete wavelet transform, and the frequency range of the detail signal is as follows:

[0188]

[0189]

[0190] Among them, f n is the frequency of the signal to be extracted, f is the signal sampling frequency, n is the number of decomposition layers, freq(D j ) is the detail signal D j The frequency of j is a constant whose value range is [1,n].

[0191] (4) According to the frequency range freq(D j ) and its corresponding number of layers n, we can get the third harmonic amplitude of phase A current I a_3h , its specific expression is:

[0192] I a_3h (k)=(ia 3h (1),ia 3h (2)...ia 3h (n)),

[0193] Where k is the sequence number, and its value range is [1,n].

[0194] like Figure 2 As shown in the figure, the third harmonics extracted by wavelet transform and Fourier transform are compared when the speed is between 2800rpm and 3000rpm. It can be seen from the image that the wavelet transform is more stable for the extracted third harmonic amplitude during the dynamic change of the working condition, and it is more efficient in removing the interference term, and the extracted third harmonic amplitude has a higher accuracy.

[0195] The fourth step is to calculate the negative sequence current of the fault by the negative sequence current optimization algorithm: a 、i b 、i c Calculate the total negative sequence current I - The negative sequence current caused by the inter-turn short circuit is obtained by eliminating the interference term through the negative sequence current optimization algorithm. The error negative-sequence current is mainly caused by the asymmetric structure of the motor itself, which can be calculated through the negative-sequence impedance. The negative-sequence impedance is usually not affected by the motor operating conditions. Therefore, the negative-sequence current optimization algorithm cleverly uses this characteristic of negative-sequence impedance to eliminate the error negative-sequence current and obtain a more accurate fault negative-sequence current.

[0196] The negative sequence current optimization algorithm for calculating the fault negative sequence current comprises the following steps:

[0197] (1) Combined with the obtained A phase, B phase, and C phase current i a 、i b 、i c , the negative sequence current I - The calculation formula is as follows:

[0198]

[0199]

[0200]

[0201] Among them, i a 、i b 、i c is the current of phase A, phase B, and phase C, I + , I - is the amplitude of the positive and negative sequence currents, ω is the fundamental frequency of the current, φ1 and φ2 are the phase angles of the positive and negative sequence currents, α is the operator, α=e j2π / 3 That is 120°.

[0202] (2) Calculated negative sequence current I - Not only does it include the negative sequence current caused by the inter-turn short circuit It also includes the error negative sequence current caused by unstable motor operation and structural asymmetry In order to improve the accuracy of fault diagnosis, the negative sequence current I - For optimization, the optimization algorithm is as follows:

[0203] When no inter-turn short circuit occurs:

[0204]

[0205]

[0206]

[0207]

[0208] When a turn-to-turn short circuit occurs:

[0209]

[0210]

[0211]

[0212]

[0213] Among them, V - is the negative sequence voltage, I - is the negative sequence current, u a 、u b 、u c is the voltage of phase A, phase B, and phase C, Z - is the negative sequence impedance, α is the operator, α=e j2π / 3 That is 120°, is the error negative sequence current, is the fault negative sequence current.

[0214] (3) Obtain the fault negative sequence current according to the negative sequence current optimization algorithm This is used as the fault feature input for convolutional neural network training, and the fault negative sequence current The specific expression is:

[0215]

[0216] Where k is the sequence number, and its value range is [1,n].

[0217] like Figure 3 As shown, the horizontal axis is the nine working conditions set, and the vertical axis is the amplitude of the negative sequence current. Figure 3 From the comparison in , it can be concluded that the negative-sequence current optimization algorithm can effectively eliminate the negative-sequence current caused by the asymmetry of the motor itself under different working conditions, thereby improving the accuracy of the negative-sequence current.

[0218] Step 5: Training of convolutional neural network: transform the third harmonic amplitude I a_3h and fault negative sequence current This fault signature, representing a turn-to-turn short circuit, is then generated into a training dataset and fed into a convolutional neural network for training. Convolutional neural networks have powerful data feature extraction capabilities, eliminating the need for preprocessing raw data and enabling direct extraction of fault signatures. The network architecture, kernel size, number of kernels, and number of training rounds require continuous tuning to minimize the loss function.

[0219] The training of the convolutional neural network includes the following steps:

[0220] (1) Creating a training data set, specifically including the following steps:

[0221] The third harmonic amplitude of the A phase current I a_3h The fault negative sequence current obtained from As two characteristics of the inter-turn short circuit fault of the permanent magnet synchronous motor, the training data set data1 is a two-dimensional array, and its expression is as follows:

[0222]

[0223] (2) Build a convolutional neural network. The steps are as follows:

[0224] The convolutional neural network architecture is set to have 11 layers, consisting of an input layer, a convolutional layer, an activation function, a pooling layer, a fully connected layer, and an output layer; among them:

[0225] The first layer is the input layer, and the input feature dimension of this layer is set to 2;

[0226] The second layer is the convolution layer, and the convolution kernel size is set to 3*3 and the number of convolution kernels is set to 64;

[0227] The third layer is the activation function. The constructed convolutional neural network uses the Relu activation function, which is expressed as follows:

[0228] The fourth layer is the normalization layer, which is used to prevent overfitting and gradient explosion and improve the training efficiency of convolutional neural networks;

[0229] The fifth layer is the pooling layer. The convolutional neural network constructed uses maximum pooling. Maximum pooling refers to selecting the maximum value of the feature area as the pooled value of the area.

[0230] The sixth layer is the convolution layer, and the convolution kernel size is set to 3*3 and the number of convolution kernels is set to 128;

[0231] The seventh layer is the activation function layer, which uses the Relu activation function;

[0232] The eighth layer is the pooling layer, which uses maximum pooling;

[0233] The ninth layer is a fully connected layer, and the feature dimension of this layer is set to 4;

[0234] The tenth layer is the Softmax layer. The Softmax layer outputs the probability of each type in the classification case. The sum of the output probabilities is 1. The size of the probability value can be used to determine whether the data feature belongs to a certain type.

[0235] The eleventh layer is the output layer. The convolutional neural network constructed is used to realize the classification of inter-turn short-circuit faults of permanent magnet synchronous motors, so classification is adopted.

[0236] (3) Input the training data set into the constructed convolutional neural network for training. The specific training steps are as follows:

[0237] The convolutional neural network training sets the minimum batch size for each round of training to 15, with a total of 1000 rounds of training. The training consists of the following three stages:

[0238] The first stage is the weight initialization process, which uses a Gaussian distribution with a mean of 0 and a variance of 0.01 to randomly initialize the weights;

[0239] The second stage is the forward propagation process. The input training set is continuously reduced in feature dimension through the convolution layer and pooling layer, and feature extraction is finally input to the fully connected layer. The specific process is as follows:

[0240] The convolution operation is expressed as follows:

[0241] x l =f(x l-1 *W l +b l )

[0242] The pooling operation is expressed as follows:

[0243] x l =pooling(x l-1 )

[0244] The output of the fully connected layer is:

[0245] x l =f(W l x l-1 +b l )

[0246] Among them, x l is the output of the lth layer, W l is the weight of the lth layer, b l is the bias of the lth layer, f() is the activation function, and pooling() is the pooling operation;

[0247] The third stage is the back propagation process, which compares the output value with the true value and uses the loss function to represent the error. The error function used in training is the cross entropy function, and its specific expression is:

[0248]

[0249] Where n is the number of samples in the training set, m(i) is the true sample distribution, w(i) is the predicted distribution, and L(y,f(x)) is the cross entropy loss error value;

[0250] The cross entropy function loss error value L(y,f(x)) is back-propagated to the fully connected layer, pooling layer, and convolution layer to update the weights and biases, and then the forward propagation process is continued. The three stages are repeatedly iterated to minimize the loss function, indicating that the convolutional neural network has reached the convergence condition. Further training is performed until the convolutional neural network reaches the iterative termination condition, and the training is terminated; the trained convolutional neural network is obtained.

[0251] Step 6: Obtaining the data to be tested: Monitoring the A-phase, B-phase, and C-phase currents i under the actual working conditions of the permanent magnet synchronous motor a ′、i b ′、i c ′, Phase A, Phase B, Phase C voltage u a ′、u b ′、u c ′, and then obtain the third harmonic amplitude I under the actual working condition of the permanent magnet synchronous motor a_3h ′ and fault negative sequence current And make it into a data set to be tested.

[0252] (1) Obtain the third harmonic amplitude I under actual working conditions a_3h ', specifically comprising the following steps:

[0253] The A-phase current i under the actual working condition of the permanent magnet synchronous motor obtained by monitoring a 'Perform the discrete wavelet transform in the above steps to obtain the third harmonic amplitude I of the A phase current under actual working conditions a_3h ′, its specific expression is:

[0254] I a_3h ′(k)=(ia′ 3h (1),ia′ 3h (2)...ia′ 3h (n)),

[0255] Where k is the sequence number, and its value range is [1,n].

[0256] (2) Obtaining the negative sequence current under actual operating conditions The specific steps include:

[0257] The A-phase, B-phase, and C-phase currents i under the actual working conditions of the permanent magnet synchronous motor obtained by monitoring a ′、i b ′、i c ′, Phase A, Phase B, Phase C voltage u a ′、u b ′、u c ', perform the negative sequence current calculation and optimization in the above steps to obtain the negative sequence current under actual working conditions The specific expression is:

[0258]

[0259] Where k is the sequence number, and its value range is [1,n].

[0260] (3) The third harmonic amplitude I of the A-phase current under the actual working condition is obtained a_3h ′ and fault negative sequence current As two characteristics of the inter-turn short-circuit fault of the permanent magnet synchronous motor under actual working conditions, the prepared data set data1′ to be tested is a two-dimensional array, and its expression is as follows:

[0261]

[0262] Step 7: Detection of inter-turn short-circuit faults in permanent magnet synchronous motors: Input the dataset to be detected into the trained convolutional neural network to obtain the detection results of inter-turn short-circuit faults in permanent magnet synchronous motors.

[0263] (1) Input the data set data1′ to be tested into the trained convolutional neural network.

[0264] (2) The convolutional neural network performs the training process in the above steps until the convolutional neural network reaches the iteration termination condition.

[0265] (3) Output the classification results of the data set data1′ to be tested and complete the diagnosis of the inter-turn short circuit fault of the permanent magnet synchronous motor.

[0266] In practical applications, the prepared data set data1 can also be divided into 90% as a training set and 10% as a test set, and labeled with category labels 1, 2, 3, and 4. The training set is input into the convolutional neural network and trained for 1000 rounds to obtain a trained convolutional neural network.

[0267] The test set is input into the trained convolutional neural network to realize the classification of permanent magnet synchronous motor inter-turn short circuit fault diagnosis, and the root mean square value RMSE is used to evaluate the network accuracy, which is expressed as follows:

[0268]

[0269] The third harmonic amplitude of phase A current I is obtained respectively a_3h and the obtained fault negative sequence current Create data sets data2 and data3, whose expressions are as follows:

[0270] data2=[I a_3h (1),I a_3h (2)...I a_3h (n)]

[0271]

[0272] Repeat the above steps and use RMSE to evaluate the accuracy of the convolutional neural network trained on the three data sets.

[0273] Table 1 Convolutional neural network accuracy table

[0274]

[0275] As shown in Table 1, the RMSE values of the convolutional neural network evaluation indicators are obtained after the data data1, data2, and data3 are input into the convolutional neural network for training. From the data comparison in Table 1, it can be seen that the method proposed in the present invention of simultaneously inputting two types of fault features into the convolutional neural network for training can effectively improve the diagnosis accuracy of the permanent magnet synchronous motor inter-turn short circuit fault.

[0276] like Figure 4 As shown, the vertical axis is the actual classification of the test set, which has four categories in total, and the horizontal axis is the classification obtained after the convolutional neural network recognizes the fault features of the test set. The test set has a total of 80 sample data. Figure 4 It can be seen that correct classification is achieved, which reflects the effectiveness of the convolutional neural network proposed in this invention for detecting inter-turn short-circuit faults in permanent magnet synchronous motors.

[0277] 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 merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting inter-turn short circuit faults in a permanent magnet synchronous motor based on neural network technology, characterized in that: The following steps are involved: 11) Establish a permanent magnet synchronous motor inter-turn short circuit fault model: Based on the equivalent circuit when the permanent magnet synchronous motor phase A has an inter-turn short circuit, establish the voltage u of phase A, phase B, and phase C. a 、u b 、u c With the A phase, B phase, C phase current i a 、i b 、i c and fault current i f The equation between and establish the torque T e Equation, which is used as the inter-turn short-circuit fault model of permanent magnet synchronous motor; 12) Obtain the current and voltage of the permanent magnet synchronous motor under different working conditions: According to the simulation results of the permanent magnet synchronous motor inter-turn short-circuit fault model under different working conditions, obtain the A-phase, B-phase, and C-phase currents i under the inter-turn short-circuit state of the permanent magnet synchronous motor. a 、i b 、i c With phase A, phase B, phase C voltage u a 、u b 、u c ; 13) Wavelet transform to extract the third harmonic: The current i of phase A is transformed by wavelet transform. a Decompose and obtain the amplitude of its third harmonic I a_3h As a fault characteristic; 14) Negative sequence current optimization algorithm calculates the fault negative sequence current: through the A phase, B phase, C phase current i a 、i b 、i c Calculate the total negative sequence current I - The negative sequence current caused by the inter-turn short circuit is obtained by eliminating the interference term through the negative sequence current optimization algorithm. 15) Training of convolutional neural network: The third harmonic amplitude I a_3h and fault negative sequence current At the same time, it is used as the fault feature of inter-turn short circuit and made into a training data set, which is input into the convolutional neural network for training; 16) Obtaining the data to be tested: Monitoring the A-phase, B-phase, and C-phase currents i under the actual working conditions of the permanent magnet synchronous motor a ′、i b ′、i c ′, Phase A, Phase B, Phase C voltage u a ′、u b ′、u c ′, and then obtain the third harmonic amplitude I under the actual working condition of the permanent magnet synchronous motor a_3h ′ and fault negative sequence current And make it into a data set to be tested; 17) Detection of inter-turn short-circuit faults in permanent magnet synchronous motors: The dataset to be tested is input into the trained convolutional neural network to obtain the detection results of inter-turn short-circuit faults in permanent magnet synchronous motors.

2. The method for detecting inter-turn short circuit fault of a permanent magnet synchronous motor based on neural network technology according to claim 1 is characterized in that: The establishment of the permanent magnet synchronous motor turn-to-turn short circuit fault model comprises the following steps: 21) Set the permanent magnet synchronous motor inter-turn short circuit fault model to be composed of phase A, phase B, and phase C voltage u a 、u b 、u c With the A phase, B phase, C phase current i a 、i b 、i c and fault current i f The equation between them and the inter-turn short-circuit electromagnetic torque equation of the permanent magnet synchronous motor; 22) Set the voltage u of phase A, phase B and phase C a 、u b 、u c With the A phase, B phase, C phase 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, R s is the stator resistance, u is the fault degree, phase A is the fault phase, then u is the ratio of the short-circuit turns of phase A stator winding to the total number of turns of its stator winding, R f is the fault resistance, i a 、i b 、i c is the current of the stator winding of phase A, phase B, and phase C, i f is the fault current, L is the self-inductance of the stator winding, M is the mutual inductance of the stator winding, ψ f is the fundamental amplitude of the flux linkage, θ is the rotor electrical angle; 23) The electromagnetic torque equation of the permanent magnet synchronous motor inter-turn short circuit is set as follows: Among them, T e is the electromagnetic torque of the permanent magnet synchronous motor, p n is the number of pole pairs of the permanent magnet synchronous motor, i d 、i q is the current of the permanent magnet synchronous motor d-axis and q-axis, L d , L q is the inductance of the d-axis and q-axis of the permanent magnet synchronous motor, ψ f is the fundamental amplitude of the magnetic flux.

3. The method for detecting inter-turn short circuit fault of a permanent magnet synchronous motor based on neural network technology according to claim 1 is characterized in that: The method of obtaining the current and voltage of the permanent magnet synchronous motor under different working conditions comprises the following steps: 31) The linearly varying speed N, the linearly varying torque T, and the four fault levels u are combined to serve as the simulation conditions for the inter-turn short-circuit fault of the permanent magnet synchronous motor. The specific values of the speed N, torque T, and fault level u are as follows: N=300n, where n=0, 1, 2…, 10, T＝[0,1,2,3,4,5,6,7,8,9,10]， u＝[0,0.1,0.2,0.3]； 32) Perform simulation based on the permanent magnet synchronous motor inter-turn short circuit fault model and in combination with the set simulation working conditions of the permanent magnet synchronous motor inter-turn short circuit fault; According to the simulation results, the A-phase current i of the permanent magnet synchronous motor under the inter-turn short-circuit state is obtained. a and B and C phase current i b 、i c , A phase voltage u a and B and C phase voltage u b 、u c , its specific expression is as follows: i a (k)=(i a (1),i a (2)...i a (n)), i b (k)=(i b (1),i b (2)...i b (n)), i c (k)=(i c (1),i c (2)...i c (n)), u a (k)=(u a (1),u a (2)...u a (n)), u b (k)=(u b (1),u b (2)...u b (n)), u c (k)=(u c (1),u c (2)...u c (n)), Where k is the sequence number, and its value range is 1, 2, ..., n.

4. The method for detecting inter-turn short circuit fault of a permanent magnet synchronous motor based on neural network technology according to claim 1, characterized in that: The wavelet transform to extract the third harmonic comprises the following steps: 41) Assume that the specific expression of the wavelet transform algorithm is as follows: The basic wavelet ψ(t) is scaled and translated to obtain the wavelet function family ψ a,b (t): Among them, t is the independent variable, a is the scaling factor, and b is the translation factor; Then the continuous wavelet transform of any function x(t) is expressed as: Considering the phase A current i a Since it is discrete data, the continuous wavelet transform is discretized and the expression is: Using binary discrete form, that is, taking a0 = 2, b0 = 1, the binary discrete wavelet transform is expressed as: ψ j,k (t)=2 -j / 2 ψ(2 -j tk), The discrete wavelet transform is expressed as: in, for The conjugate of , a0 is the discretization scaling factor, b0 is the discretization translation factor, j and m are integers; 42) The obtained A phase current i a After discrete wavelet transform, the transformed signal composition has the following relationship: i a =A n +D n +D n-1 +...+D1, Among them, A n is the approximate signal of the nth layer, D n D n-1 ....D1 is the n-layer detail signal; 43) Assume that the fault feature is contained in the detail signal after discrete wavelet transform, and the frequency range of the detail signal is as follows: Among them, f n is the frequency of the signal to be extracted, f is the signal sampling frequency, n is the number of decomposition layers, freq(D j ) is the detail signal D j The frequency of , j is a constant whose value range is [1,n]; 44) According to the frequency range freq(D j ) and its corresponding number of layers n, we can get the third harmonic amplitude of phase A current I a_3h , and use it as the fault feature, its specific expression is: I a_3h (k)=(it) 3h (1),it 3h (2)...it 3h (n)), Where k is the sequence number, and its value range is [1,n].

5. The method for detecting inter-turn short circuit fault of a permanent magnet synchronous motor based on neural network technology according to claim 1, characterized in that: The negative sequence current optimization algorithm for calculating the fault negative sequence current comprises the following steps: 51) Combine the obtained A phase, B phase, and C phase currents i a 、i b 、i c , the negative sequence current I - The calculation formula is as follows: Among them, i a 、i b 、i c is the current of phase A, phase B, and phase C, I + , I - is the amplitude of the positive and negative sequence currents, ω is the fundamental frequency of the current, φ1 and φ2 are the phase angles of the positive and negative sequence currents, α is the operator, α=e j2π / 3 That is 120°; 52) Calculated negative sequence current I - Not only does it include the negative sequence current caused by the inter-turn short circuit It also includes the error negative sequence current caused by unstable motor operation and structural asymmetry In order to improve the accuracy of fault diagnosis, the negative sequence current I - For optimization, the optimization algorithm is as follows: When no inter-turn short circuit occurs: When a turn-to-turn short circuit occurs: Among them, V - is the negative sequence voltage, I - is the negative sequence current, u a 、u b 、u c is the voltage of phase A, phase B, and phase C, i a 、i b 、i c is the current of phase A, phase B, and phase C, Z - is the negative sequence impedance, α is the operator, α=e j2π / 3 That is 120°, is the error negative sequence current, is the fault negative sequence current; 53) Obtain the fault negative sequence current according to the negative sequence current optimization algorithm And use it as another fault feature, the fault negative sequence current The specific expression is: Where k is the sequence number, and its value range is [1,n].

6. The method for detecting inter-turn short circuit fault of a permanent magnet synchronous motor based on neural network technology according to claim 1, characterized in that: The training of the convolutional neural network includes the following steps: 61) Create a training data set, specifically including the following steps: The third harmonic amplitude of the A phase current I a_3h and fault negative sequence current As two characteristics of the inter-turn short circuit fault of the permanent magnet synchronous motor, the training data set data1 is a two-dimensional array, and its expression is as follows: 62) To build a convolutional neural network, the steps are as follows: The convolutional neural network architecture is set to have 11 layers, consisting of an input layer, a convolutional layer, an activation function, a pooling layer, a fully connected layer, and an output layer; among them: The first layer is the input layer, and the input feature dimension of this layer is set to 2; The second layer is the convolution layer, and the convolution kernel size is set to 3*3 and the number of convolution kernels is set to 64; The third layer is the activation function. The constructed convolutional neural network uses the Relu activation function, which is expressed as follows: The fourth layer is the normalization layer, which is used to prevent overfitting and gradient explosion and improve the training efficiency of convolutional neural networks; The fifth layer is the pooling layer. The convolutional neural network constructed uses maximum pooling. Maximum pooling refers to selecting the maximum value of the feature area as the pooled value of the area. The sixth layer is the convolution layer, and the convolution kernel size is set to 3*3 and the number of convolution kernels is set to 128; The seventh layer is the activation function layer, which uses the Relu activation function; The eighth layer is the pooling layer, which uses maximum pooling; The ninth layer is a fully connected layer, and the feature dimension of this layer is set to 4; The tenth layer is the Softmax layer. The Softmax layer outputs the probability of each type in the classification case. The sum of the output probabilities is 1. The size of the probability value is used to determine whether the data feature belongs to a certain type. The eleventh layer is the output layer. The convolutional neural network is used to classify the inter-turn short-circuit fault of the permanent magnet synchronous motor, so classification is used. 63) Input the training data set into the constructed convolutional neural network for training. The specific training steps are as follows: The convolutional neural network training sets the minimum batch size for each round of training to 15, with a total of 1000 rounds of training. The training consists of the following three stages: The first stage is the weight initialization process, which uses a Gaussian distribution with a mean of 0 and a variance of 0.01 to randomly initialize the weights; The second stage is the forward propagation process. The input training set is continuously reduced in feature dimension through the convolution layer and pooling layer, and feature extraction is finally input to the fully connected layer. The specific process is as follows: The convolution operation is expressed as follows: x l =f(x l-1 *W l +b l ) The pooling operation is expressed as follows: x l =pooling(x l-1 ) The output of the fully connected layer is: x l =f(W l x l-1 +b l ) Among them, x l is the output of the lth layer, W l is the weight of the lth layer, b l is the bias of the lth layer, f() is the activation function, and pooling() is the pooling operation; The third stage is the back propagation process, which compares the output value with the true value and uses the loss function to represent the error. The error function used in training is the cross entropy function, and its specific expression is: Where n is the number of samples in the training set, m(i) is the true sample distribution, w(i) is the predicted distribution, and L(y,f(x)) is the cross entropy loss error value; The cross entropy function loss error value L(y,f(x)) is back-propagated to the fully connected layer, pooling layer, and convolution layer to update the weights and biases, and then the forward propagation process is continued. The three stages are iteratively trained to minimize the loss function. At this time, it indicates that the convolutional neural network has reached the convergence condition. Further training is performed until the convolutional neural network reaches the iterative termination condition and the training is terminated; the trained convolutional neural network is obtained.

7. The method for detecting inter-turn short circuit fault of a permanent magnet synchronous motor based on neural network technology according to claim 1, characterized in that: The acquisition of the data to be detected comprises the following steps: 71) Obtain the third harmonic amplitude I under actual working conditions a_3h ', specifically comprising the following steps: The A-phase current i under the actual working condition of the permanent magnet synchronous motor obtained by monitoring a 'Perform discrete wavelet transform in steps 41), 42), and 43) to obtain the A phase current i under actual working conditions a The third harmonic amplitude of ′ is I a_3h ′, its specific expression is: I a_3h ′(k)=(it′ 3h (1), it' 3h (2)...it' 3h (n)), Where k is the sequence number, and its value range is [1,n]; 72) Obtain the negative sequence current under actual working conditions The specific steps include: The A-phase, B-phase, and C-phase currents i under the actual working conditions of the permanent magnet synchronous motor obtained by monitoring a ′、i b ′、i c ′, Phase A, Phase B, Phase C voltage u a ′、u b ′、u c ', perform the negative sequence current calculation and optimization in steps 51) and 52) to obtain the negative sequence current under actual working conditions The specific expression is: Where k is the sequence number, and its value range is [1,n]; 73) The third harmonic amplitude of the A phase current under the actual working condition is obtained a_3h ′ and fault negative sequence current As two characteristics of the inter-turn short-circuit fault of the permanent magnet synchronous motor under actual working conditions, the prepared data set data1′ to be tested is a two-dimensional array, and its expression is as follows:

8. The method for detecting inter-turn short circuit fault of a permanent magnet synchronous motor based on neural network technology according to claim 1, characterized in that: The permanent magnet synchronous motor inter-turn short circuit fault detection comprises the following steps: 81) Input the data set data1′ to be tested into the trained convolutional neural network; 82) The convolutional neural network performs the training process in step 63) until the convolutional neural network reaches the iteration termination condition; 83) Output the classification result of the data set data1′ to be tested, and complete the diagnosis of the permanent magnet synchronous motor inter-turn short circuit fault.

Images