Turn-to-turn short circuit fault diagnosis method based on feature selection and multi-scale residual network

Abstract

The invention discloses a turn-to-turn short-circuit fault diagnosis method based on feature selection and multi-scale residual network, which belongs to the field of motor fault diagnosis and comprises the following steps: collecting the current signal of the permanent magnet synchronous motor under the condition of turn-to-turn short-circuit fault in the internal insulation fault of the permanent magnet synchronous motor, pre-processing the current signal to obtain a training set and a test set; constructing a permanent magnet synchronous motor turn-to-turn short-circuit fault diagnosis model, using the training set to train the permanent magnet synchronous motor turn-to-turn short-circuit fault diagnosis model, and retaining the optimal structure of the permanent magnet synchronous motor turn-to-turn short-circuit fault diagnosis model; putting the test set into the trained permanent magnet synchronous motor turn-to-turn short-circuit fault diagnosis model for feature learning, and then inputting the test set into the trained permanent magnet synchronous motor turn-to-turn short-circuit fault diagnosis model. Softmax The present invention can automatically and efficiently extract weak fault features in current signals, get rid of the dependence on traditional signal processing technology and diagnostic experience, and has good robustness to signals containing noise.

Description

Technical Field

[0001] The present invention relates to the field of motor fault diagnosis, and in particular to a turn-to-turn short circuit fault diagnosis method based on feature selection and a multi-scale residual network. Background Art

[0002] The motor system is one of the important components of various mechanical equipment. During operation, the permanent magnet synchronous motor will be affected by various external factors, such as current, humidity, high temperature, dust, etc. These factors may cause various types of electrical faults in the motor system. If timely measures are not taken, the degree of fault will rapidly increase under extreme working conditions, causing the performance of the permanent magnet synchronous motor to significantly decline, which will lead to the shutdown of the entire system or even more serious accidents.

[0003] The direct cause of the inter-turn short-circuit fault of the permanent magnet synchronous motor is the damage of the insulation layer on the surface of the motor winding coil. When the insulation between two adjacent turns of the coil is damaged, a short-circuit branch will be formed, and the short-circuit current will cause the coil to heat up rapidly, further aggravating the degree of damage. Therefore, accurate and effective fault identification of insulation faults is crucial to ensure the safety and reliability of industrial equipment.

[0004] At present, motor fault diagnosis technology can generally be divided into three categories: model-based, signal processing-based, and deep learning-based. At the beginning of the development of fault diagnosis technology, people have conducted a lot of research on model-based fault diagnosis technology, but it usually requires the construction of a mathematical model that can accurately reflect the actual operating status of the motor. Therefore, intelligent fault diagnosis technology using signal processing and deep learning has good operability and recognition effect, and has been widely used in the field of fault diagnosis. However, intelligent fault diagnosis methods have problems such as complex processes, reliance on relevant professional knowledge, and insufficient learning ability of shallow structural features. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a turn-to-turn short circuit fault diagnosis method based on feature selection and multi-scale residual network, which has simple algorithm, good robustness and high diagnostic accuracy.

[0006] The technical solution of the present invention to solve the above technical problem is: a turn-to-turn short circuit fault diagnosis method based on feature selection and multi-scale residual network, comprising the following steps:

[0007] Step 1: data collection and preprocessing: collecting the current signal of the permanent magnet synchronous motor under the condition of an inter-turn short-circuit fault in the internal insulation fault of the permanent magnet synchronous motor, and preprocessing the inter-turn short-circuit fault current signal of the permanent magnet synchronous motor to obtain a training set and a test set;

[0008] In the step 1, the process of preprocessing the permanent magnet synchronous motor turn-to-turn short-circuit fault current signal is as follows:

[0009] S11: Time-frequency analysis and spectrum analysis are performed through multiple synchronous compression transform (MSST) to obtain signal samples and construct the original feature set;

[0010] The principle of synchronous compression transform is: based on the linear time-frequency signal processing algorithm, the time-frequency rearrangement operation is performed to concentrate the energy distribution, but the synchronous compression transform ignores the influence of high-order components, which leads to the loss of some time-varying signal characteristics in non-steady-state signals; therefore, on the basis of MSST, the first-order term of the instantaneous amplitude and the quadratic term of the instantaneous phase are supplemented, and then multiple instantaneous frequency estimation operations are added to enhance the time-frequency aggregation:

[0011] ;

[0012] ;

[0013] Where N is the number of iterations, is the estimated value of the instantaneous frequency, is the estimated value of the instantaneous frequency at the Nth iteration, is the first derivative of the instantaneous phase estimate, is the second derivative of the instantaneous phase estimate, is the Dirac function; is the frequency variable, is the frequency rearrangement operator, is the time-frequency coefficient, is the Nth iteration synchronous compression time-frequency coefficient matrix;

[0014] S12: Feature selection by improving and adjusting the mutual information standard deviation AMISR;

[0015] The specific process of step S12 is as follows:

[0016] S121: Perform FCM cluster analysis on all statistical features:

[0017] ;

[0018] Among them, OF is the original feature set, is the set of the mth feature of the nth signal sample when the fault type is Z;

[0019] S122: using adjusted mutual information AMI as an evaluation index to analyze the clustering results;

[0020] The closer the AMI value of a feature is to 1, the higher its correlation with the fault category is, and the more conducive it is to the subsequent judgment of the fault type. The standard deviation is used to describe the discreteness of the sample data points. For the feature set, the smaller the standard deviation is, the higher the intra-class aggregation is. Therefore, the standard deviation is introduced into AMI to characterize the intra-class aggregation of different statistical features. The row elements in OF represent the same motor state, and the standard deviation of the mth feature in OF is expressed as:

[0021] ;

[0022] ;

[0023] in, is the number of observations, is the standard deviation of the mth feature when the fault type is Z, is the k-th observation value of feature S, is the average value of feature S;

[0024] After calculating the AMI values ​​and standard deviations of all M statistical features, the calculation formula for adjusting the mutual information standard deviation AMISR is defined to obtain the AMISR value sequence:

[0025] ;

[0026] in, To adjust the mutual information standard deviation sequence, To adjust the mutual information, is the standard deviation;

[0027] Using the AMISR feature selection method, various features in the original feature set are analyzed to extract statistical features related to the fault state;

[0028] Step 2: construct a permanent magnet synchronous motor turn-to-turn short-circuit fault diagnosis model based on a multi-scale one-dimensional convolutional residual network, and after training the permanent magnet synchronous motor turn-to-turn short-circuit fault diagnosis model using a training set, retain the optimal structure of the permanent magnet synchronous motor turn-to-turn short-circuit fault diagnosis model;

[0029] Step three: put the test set into the trained permanent magnet synchronous motor turn-to-turn short-circuit fault diagnosis model for feature learning, and then input it into the Softmax classifier for fault diagnosis.

[0030] In the above-mentioned turn-to-turn short-circuit fault diagnosis method based on feature selection and multi-scale residual network, in step 1, for a normal three-phase permanent magnet synchronous motor connected in star shape, its three-phase back electromotive force is expressed as:

[0031] ;

[0032] Where θ is the rotor position angle, , , They are the opposite electromotive forces a, b, and c of a normal three-phase permanent magnet synchronous motor. is a Dirac function. Each phase of a normal three-phase permanent magnet synchronous motor contains the fundamental wave and the vth harmonic term , the amplitude of each phase of a normal three-phase permanent magnet synchronous motor is equal, and the phase difference is 120°; is the phase difference between the vth harmonic and the fundamental wave; when v=3, the 3rd harmonic of the three phases , , They are:

[0033] ;

[0034] The sum of the 3rd harmonics of the three phases for:

[0035] ;

[0036] Similarly, the sum of the 9th, 15th, and 21st harmonics of the three phases are:

[0037] ;

[0038] It can be seen that the 3rd harmonic and its odd-numbered harmonics of the three phases have the same amplitude and phase and cannot be closed when there is no neutral line path, so the zero-sequence component is 0; when a turn-to-turn short circuit occurs in one phase of the stator winding, the three-phase balance relationship of the motor will be destroyed, and a backflow will be generated in the short-circuit branch, which will cause abnormal changes in the zero-sequence component;

[0039] When a permanent magnet synchronous motor has a stator winding fault, the harmonic component in the stator current is significantly enhanced, causing the frequency of the motor current to change. Therefore, the current signal is used as an effective data source that can directly reflect the motor fault status.

[0040] In the above-mentioned turn-to-turn short-circuit fault diagnosis method based on feature selection and multi-scale residual network, in the step 2, the process of constructing a turn-to-turn short-circuit fault diagnosis model based on a multi-scale one-dimensional convolutional residual network is as follows: the original current signal is input into a convolution layer with convolution kernels of various sizes, and a set of feature maps are obtained through an activation function, and then downsampled through a pooling layer to extract short-time features in the original current signal; after repeating the convolution and pooling operation again, it is passed to the residual framework, combined with the multi-scale framework, and a multi-scale one-dimensional convolutional residual network is constructed, thereby obtaining a turn-to-turn short-circuit fault diagnosis model.

[0041] In the above-mentioned turn-to-turn short-circuit fault diagnosis method based on feature selection and multi-scale residual network, in the step 2, the original signal is input into the first convolution layer and a set of feature maps is obtained through the activation function. The operation process is:

[0042] ;

[0043] in, represents the jth input feature map, Indicates The input feature map of the i-th convolution kernel of the layer network, Indicates The jth weight value of the i-th convolution kernel in the layer network, Indicates The additive bias of the jth neuron in the layer network, is the ReLU activation function, , Represents the convolution operation, and the final output is the jth feature map of the lth layer;

[0044] Then the sampling is downsampled through the pooling layer, and the operation process is:

[0045] ;

[0046] in, is the multiplicative bias of the jth neuron in the lth layer of the network, and down(·) represents the sampling function.

[0047] In the above-mentioned turn-to-turn short-circuit fault diagnosis method based on feature selection and multi-scale residual network, in the step 2, each layer of the multi-scale framework is composed of convolution kernels of different sizes. From top to bottom, the convolution kernel sizes are 5×1, 7×1, and 9×1, respectively, and each layer has 3 groups of identical convolution layers and pooling layers for extracting features of the same scale, and then downsampling the features through a 16×1 pooling layer, and finally multi-scale feature fusion is performed through the torch.cat() function. The specific calculation process is as follows:

[0048] ;

[0049] in, , , They represent the outputs of convolutional kernel network layers of different scales. Represents multi-scale fusion features.

[0050] In the above-mentioned turn-to-turn short-circuit fault diagnosis method based on feature selection and multi-scale residual network, in the step 2, the residual network is added to the convolutional neural network to obtain a fault diagnosis model, and the calculation process is expressed as:

[0051] ;

[0052] Among them, x and y represent the input and output of the residual module respectively, represents the residual mapping to be learned.

[0053] In the above-mentioned turn-to-turn short-circuit fault diagnosis method based on feature selection and multi-scale residual network, in step 3, the Softmax function is expressed as:

[0054]

[0055] in, For the h is the ordinal number of the output node, and H is the total number of output nodes, that is, the number of categories.

[0056] The beneficial effects of the present invention are:

[0057] 1. After obtaining signal samples through time-frequency analysis and spectrum analysis by MSST and constructing the original feature set, the present invention proposes a feature selection method based on adjusted mutual information, performs cluster analysis on the original high-dimensional features through the FCM algorithm, and uses the adjusted mutual information and standard deviation ratio value as an indicator to measure the degree of correlation between the feature and the fault state, constructing a new feature evaluation index AMISR, which can extract statistical features related to the fault state and effectively improve the accuracy of turn-to-turn short-circuit fault identification.

[0058] 2. The present invention can automatically and efficiently extract weak fault features in the current signal of the permanent magnet synchronous motor, get rid of the dependence on traditional signal processing technology and diagnostic experience, and has good robustness to signals containing noise. It can not only be used for insulation fault diagnosis in the internal system of the permanent magnet synchronous motor, but also can be widely used in a series of complex systems such as electric vehicles, aviation, and high-speed railways. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a flow chart of the present invention.

[0060] Figure 2 This is the time domain waveform of the stator current of the permanent magnet synchronous motor when it is healthy.

[0061] Figure 3 This is the frequency domain harmonic amplitude-frequency diagram of the stator current of the permanent magnet synchronous motor under healthy conditions.

[0062] Figure 4 This is the time domain waveform of the stator current in the case of a permanent magnet synchronous motor inter-turn short circuit fault.

[0063] Figure 5This is the frequency domain harmonic amplitude-frequency diagram of the stator current in the case of a permanent magnet synchronous motor inter-turn short circuit fault.

[0064] Figure 6 This is the algorithm structure diagram.

[0065] Figure 7 Schematic diagram of experimental results under the inter-turn short-circuit fault current data set. DETAILED DESCRIPTION

[0066] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0067] like Figure 1 As shown, a turn-to-turn short circuit fault diagnosis method based on feature selection and multi-scale residual network includes the following steps:

[0068] Step 1: data collection and preprocessing: collect the current signal of the permanent magnet synchronous motor under the condition of inter-turn short circuit fault in the internal insulation fault of the permanent magnet synchronous motor, and preprocess the current signal of the permanent magnet synchronous motor to obtain the training set and the test set.

[0069] First of all, the direct cause of the inter-turn short circuit fault is the damage of the insulation layer on the surface of the permanent magnet synchronous motor winding coil. Insulation materials generally have a very high resistivity to block current, so they can withstand strong electric field strength to prevent breakdown. When the insulation between two adjacent turns of the coil is damaged, a short-circuit branch will be formed, and the short-circuit current will cause the coil to heat up rapidly, further aggravating the degree of damage. For a normal three-phase permanent magnet synchronous motor with star connection, its three-phase back electromotive force is expressed as:

[0070] ;

[0071] Where θ is the rotor position angle, , , They are the opposite electromotive forces a, b, and c of a normal three-phase permanent magnet synchronous motor. is a Dirac function. Each phase of a normal three-phase permanent magnet synchronous motor contains the fundamental wave and the vth harmonic term , the amplitude of each phase of a normal three-phase permanent magnet synchronous motor is equal, and the phase difference is 120°; is the phase difference between the vth harmonic and the fundamental wave; when v=3, the 3rd harmonic of the three phases , , They are:

[0072] ;

[0073] The sum of the 3rd harmonics of the three phases for:

[0074] ;

[0075] Similarly, the sum of the 9th, 15th, and 21st harmonics of the three phases are:

[0076] ;

[0077] It can be seen from this that the third and odd-order harmonics of the three-phase have the same amplitude and phase, and cannot be closed when there is no path in the neutral line, so the zero-sequence component is 0; when a turn-to-turn short circuit occurs in one phase of the stator winding, the three-phase balance relationship of the permanent magnet synchronous motor will be destroyed, and backflow will be generated in the short-circuit branch, which will cause abnormal changes in the zero-sequence component.

[0078] When the permanent magnet synchronous motor is in normal condition, such as Figure 2 and Figure 3 As shown in the figure, the stator current is smooth and there is no current surge. In the frequency domain, no 3rd, 5th, 7th, and 9th harmonics appear, indicating that the motor operates normally without fault. When a permanent magnet synchronous motor has a stator winding fault, such as Figure 4 and Figure 5 As shown in the figure, a turn-to-turn short circuit fault occurs in the stator current at the 25th second, causing the current to surge and become unsmooth. The harmonic components in the stator current are significantly enhanced, and the 3rd, 5th, 7th, and 9th harmonics appear, causing the frequency of the permanent magnet synchronous motor current to change accordingly. Therefore, the current signal is used as an effective data source that can directly reflect the fault state of the permanent magnet synchronous motor.

[0079] The process of preprocessing the current signal is as follows:

[0080] S11: Time-frequency analysis and spectrum analysis are performed through multiple synchronous compression transform (MSST) to obtain signal samples and construct the original feature set.

[0081] The principle of synchronous compression transform is: based on the linear time-frequency signal processing algorithm, the time-frequency rearrangement operation is performed to concentrate the energy distribution, but the synchronous compression transform ignores the influence of high-order components, which leads to the loss of some time-varying signal characteristics in non-steady-state signals; therefore, on the basis of MSST, the first-order term of the instantaneous amplitude and the quadratic term of the instantaneous phase are supplemented, and then multiple instantaneous frequency estimation operations are added to enhance the time-frequency aggregation:

[0082] ;

[0083] ;

[0084] Where N is the number of iterations, is the estimated value of the instantaneous frequency, is the estimated value of the instantaneous frequency at the Nth iteration, is the first derivative of the instantaneous phase estimate, is the second derivative of the instantaneous phase estimate, is the Dirac function; is the frequency variable, is the frequency rearrangement operator, is the time-frequency coefficient, is the Nth iteration synchronous compression time-frequency coefficient matrix;

[0085] S12: Feature selection is performed by improving and adjusting the mutual information standard deviation AMISR. The specific process of step S12 is:

[0086] S121: Perform FCM cluster analysis on all statistical features:

[0087] ;

[0088] Among them, OF is the original feature set, is the set of the mth feature of the nth signal sample when the fault type is Z;

[0089] S122: Use the adjusted mutual information AMI as an evaluation index to analyze the clustering results.

[0090] The closer the AMI value of a feature is to 1, the higher its correlation with the fault category is, and the more conducive it is to the subsequent judgment of the fault type. The standard deviation is used to describe the discreteness of the sample data points. For the feature set, the smaller the standard deviation is, the higher the intra-class aggregation is. Therefore, the standard deviation is introduced into AMI to characterize the intra-class aggregation of different statistical features. The row elements in OF represent the same motor state, and the standard deviation of the mth feature in OF is expressed as:

[0091] ;

[0092] ;

[0093] in, is the number of observations, is the standard deviation of the mth feature when the fault type is Z, is the k-th observation value of feature S, is the average value of feature S;

[0094] After calculating the AMI values ​​and standard deviations of all M statistical features, the calculation formula for adjusting the mutual information standard deviation AMISR is defined to obtain the AMISR value sequence:

[0095] ;

[0096] in, To adjust the mutual information standard deviation sequence, To adjust the mutual information, is the standard deviation;

[0097] It can be seen that the larger the AMISR value is, the higher the correlation with the current motor fault type is. The AMISR feature selection method is used to analyze the various features in the original feature set and extract the statistical features related to the fault state in order to better improve the diagnostic performance.

[0098] Step 2: construct a permanent magnet synchronous motor turn-to-turn short-circuit fault diagnosis model based on a multi-scale one-dimensional convolutional residual network. After training the permanent magnet synchronous motor turn-to-turn short-circuit fault diagnosis model using a training set, the optimal structure of the permanent magnet synchronous motor turn-to-turn short-circuit fault diagnosis model is retained.

[0099] like Figure 6 As shown in the figure, the process of constructing a turn-to-turn short-circuit fault diagnosis model based on a multi-scale one-dimensional convolutional residual network (MS-1DCNN-ResNet) is as follows: the original current signal is input into a convolution layer with convolution kernels of various sizes, and a set of feature maps are obtained through an activation function. Then, the signal is downsampled through a pooling layer to extract short-time features from the original current signal. After repeating the convolution and pooling operation, the signal is passed to the residual framework, and combined with the multi-scale framework, a multi-scale one-dimensional convolutional residual network is constructed to obtain a turn-to-turn short-circuit fault diagnosis model.

[0100] For one-dimensional current signals, based on convolutional neural networks, a one-dimensional convolution structure is proposed to construct a permanent magnet synchronous motor turn-to-turn short-circuit fault diagnosis model for current signals. This method avoids the increase in parameter calculation caused by stacking multiple convolution layers with smaller convolution kernels in two-dimensional convolutional neural networks. The original signal is input into the first convolution layer and a set of feature maps is obtained through the activation function. The operation process is:

[0101] ;

[0102] in, represents the jth input feature map, Indicates The input feature map of the i-th convolution kernel of the layer network, Indicates The jth weight value of the i-th convolution kernel in the layer network, Indicates The additive bias of the jth neuron in the layer network, is the ReLU activation function, , Represents the convolution operation, and the final output is the j-th feature map of the l-th layer.

[0103] Then the sampling is downsampled through the pooling layer, and the operation process is:

[0104] ;

[0105] in, is the multiplicative bias of the jth neuron in the lth layer of the network, and down(·) represents the sampling function.

[0106] The concept of multi-scale feature fusion is introduced to fuse feature kernels of different sizes to enhance more diverse feature information, achieve information complementarity, and comprehensively enhance the model's feature expression capabilities. Each layer of the multi-scale framework is composed of convolution kernels of different sizes. From top to bottom, the convolution kernel sizes are 5×1, 7×1, and 9×1, respectively. Each layer has 3 sets of identical convolution layers and pooling layers to extract features of the same scale, and then downsample the features through a 16×1 pooling layer. Finally, multi-scale feature fusion is performed through the torch.cat() function. The specific calculation process is as follows:

[0107] ;

[0108] in, , , They represent the outputs of convolutional kernel network layers of different scales. Represents multi-scale fusion features.

[0109] As the number of neural network layers increases, the activation function will also experience a gradient decrease, the network will become difficult to train, and the network training accuracy will reach saturation, resulting in network degradation. Therefore, the residual network is added to the convolutional neural network to obtain a fault diagnosis model. The calculation process is expressed as:

[0110] ;

[0111] Among them, x and y represent the input and output of the residual module respectively, represents the residual mapping to be learned.

[0112] Step three: put the test set into the trained permanent magnet synchronous motor turn-to-turn short-circuit fault diagnosis model for feature learning, and then input it into the Softmax classifier for fault diagnosis.

[0113] The Softmax function is expressed as:

[0114]

[0115] in, For the h is the ordinal number of the output node, and H is the total number of output nodes, that is, the number of categories.

[0116] Advantages of the model proposed by the present invention:

[0117] a) Compared with the traditional CNN method, the traditional CNN is more sensitive to data and requires a large amount of labeled data for effective training, which leads to large fluctuations in the calculation results. At the same time, the computational complexity is relatively high, resulting in a long calculation time, which is not conducive to practical application. The method proposed in the present invention, on the basis of a one-dimensional convolutional neural network, constructs a multi-scale framework with different convolution kernel sizes, which can have better feature extraction capabilities, and introduces a residual network framework to avoid problems such as gradient descent caused by the deepening of the network depth, effectively improving the recognition accuracy of permanent magnet synchronous motor fault diagnosis, and has good adaptability to variable working conditions and anti-noise performance.

[0118] b) Compared with machine learning methods that require manual feature extraction, deep learning methods have better diagnostic performance. This is because deep learning methods can adaptively learn representative features from input data, thereby reducing expert experience and manual intervention, and have a high degree of generalization.

[0119] The diagnosis results of the inter-turn short-circuit fault of the permanent magnet synchronous motor are as follows: Figure 7 The performance of the model proposed in the present invention is good, and the average accuracy rate reaches 97.12%.

[0120] In summary, the present invention can automatically and efficiently extract weak fault features in the current signal of the permanent magnet synchronous motor, and realize high-precision fault diagnosis of the rotating machinery.

Claims

1. A turn-to-turn short circuit fault diagnosis method based on feature selection and multi-scale residual network, characterized in that: The following steps are involved: Step 1: data collection and preprocessing: collecting the current signal of the permanent magnet synchronous motor under the condition of an inter-turn short-circuit fault in the internal insulation fault of the permanent magnet synchronous motor, and preprocessing the inter-turn short-circuit fault current signal of the permanent magnet synchronous motor to obtain a training set and a test set; In the step 1, the process of preprocessing the permanent magnet synchronous motor turn-to-turn short-circuit fault current signal is as follows: S11: Time-frequency analysis and spectrum analysis are performed through multiple synchronous compression transform (MSST) to obtain signal samples and construct the original feature set; The principle of synchronous compression transform is: based on the linear time-frequency signal processing algorithm, the time-frequency rearrangement operation is performed to concentrate the energy distribution, but the synchronous compression transform ignores the influence of high-order components, which leads to the loss of some time-varying signal characteristics in non-steady-state signals; therefore, on the basis of MSST, the first-order term of the instantaneous amplitude and the second-order term of the instantaneous phase are supplemented, and then multiple instantaneous frequency estimation operations are added to enhance the time-frequency aggregation: ； ； Where N is the number of iterations, is the estimated value of the instantaneous frequency, is the estimated value of the instantaneous frequency at the Nth iteration, is the first derivative of the instantaneous phase estimate, is the second derivative of the instantaneous phase estimate, is the Dirac function; is the frequency variable, is the frequency rearrangement operator, is the time-frequency coefficient, is the Nth iteration synchronous compression time-frequency coefficient matrix; S12: Feature selection by improving and adjusting the mutual information standard deviation AMISR; The specific process of step S12 is as follows: S121: Perform FCM cluster analysis on all statistical features: ； Among them, OF is the original feature set, is the set of the mth feature of the nth signal sample when the fault type is Z; S122: using adjusted mutual information AMI as an evaluation index to analyze the clustering results; The closer the AMI value of a feature is to 1, the higher its correlation with the fault category is, and the more conducive it is to the subsequent judgment of the fault type. The standard deviation is used to describe the discreteness of the sample data points. For the feature set, the smaller the standard deviation is, the higher the intra-class aggregation is. Therefore, the standard deviation is introduced into AMI to characterize the intra-class aggregation of different statistical features. The row elements in OF represent the same motor state, and the standard deviation of the mth feature in OF is expressed as: ； ； in, is the number of observations, is the standard deviation of the mth feature when the fault type is Z, is the k-th observation value of feature S, is the average value of feature S; After calculating the AMI values ​​and standard deviations of all M statistical features, the calculation formula for adjusting the mutual information standard deviation AMISR is defined to obtain the AMISR value sequence: ； in, To adjust the mutual information standard deviation sequence, To adjust the mutual information, is the standard deviation; Using the AMISR feature selection method, various features in the original feature set are analyzed to extract statistical features related to the fault state; Step 2: construct a permanent magnet synchronous motor turn-to-turn short-circuit fault diagnosis model based on a multi-scale one-dimensional convolutional residual network, and after training the permanent magnet synchronous motor turn-to-turn short-circuit fault diagnosis model using a training set, retain the optimal structure of the permanent magnet synchronous motor turn-to-turn short-circuit fault diagnosis model; Step three: put the test set into the trained permanent magnet synchronous motor turn-to-turn short-circuit fault diagnosis model for feature learning, and then input it into the Softmax classifier for fault diagnosis.

2. The turn-to-turn short circuit fault diagnosis method based on feature selection and multi-scale residual network according to claim 1 is characterized in that: In the step 1, for a normal three-phase permanent magnet synchronous motor connected in star shape, its three-phase back electromotive force is expressed as: ； Where θ is the rotor position angle, , , They are the opposite electromotive forces a, b, and c of a normal three-phase permanent magnet synchronous motor. is a Dirac function. Each phase of a normal three-phase permanent magnet synchronous motor contains the fundamental wave and the vth harmonic term , the amplitude of each phase of a normal three-phase permanent magnet synchronous motor is equal, and the phase difference is 120°; is the phase difference between the vth harmonic and the fundamental wave; when v=3, the 3rd harmonic of the three phases , , They are: ； The sum of the 3rd harmonics of the three phases for: ； Similarly, the sum of the 9th, 15th, and 21st harmonics of the three phases are: ； It can be seen that the 3rd harmonic and its odd-numbered harmonics of the three phases have the same amplitude and phase and cannot be closed when there is no neutral line path, so the zero-sequence component is 0; when a turn-to-turn short circuit occurs in one phase of the stator winding, the three-phase balance relationship of the motor will be destroyed, and a backflow will be generated in the short-circuit branch, which will cause abnormal changes in the zero-sequence component; When a permanent magnet synchronous motor has a stator winding fault, the harmonic component in the stator current is significantly enhanced, causing the frequency of the motor current to change. Therefore, the current signal is used as an effective data source that can directly reflect the motor fault status.

3. The turn-to-turn short circuit fault diagnosis method based on feature selection and multi-scale residual network according to claim 2 is characterized in that: In the step 2, the process of constructing the turn-to-turn short-circuit fault diagnosis model based on the multi-scale one-dimensional convolutional residual network is as follows: the original current signal is input into a convolution layer with convolution kernels of various sizes, and then a set of feature maps are obtained through an activation function, and then downsampled through a pooling layer to extract short-time features in the original current signal; After repeating the convolution and pooling operation again, it is passed to the residual framework and combined with the multi-scale framework to construct a multi-scale one-dimensional convolutional residual network to obtain the turn-to-turn short circuit fault diagnosis model.

4. The turn-to-turn short circuit fault diagnosis method based on feature selection and multi-scale residual network according to claim 3 is characterized in that: In step 2, the original signal is input into the first convolution layer and a set of feature maps is obtained through the activation function. The operation process is: ； in, represents the jth input feature map, Indicates The input feature map of the i-th convolution kernel of the layer network, Indicates The jth weight value of the i-th convolution kernel in the layer network, Indicates The additive bias of the jth neuron in the layer network, is the ReLU activation function, , Represents the convolution operation, and the final output is the jth feature map of the lth layer; Then the sampling is downsampled through the pooling layer, and the operation process is: ； in, is the multiplicative bias of the jth neuron in the lth layer of the network, and down(·) represents the sampling function.

5. The turn-to-turn short circuit fault diagnosis method based on feature selection and multi-scale residual network according to claim 4 is characterized in that: In step 2, each layer of the multi-scale framework is composed of convolution kernels of different sizes. From top to bottom, the convolution kernel sizes are 5×1, 7×1, and 9×1, respectively. Each layer has 3 sets of identical convolution layers and pooling layers to extract features of the same scale. The features are then downsampled through a 16×1 pooling layer, and finally multi-scale feature fusion is performed through the torch.cat() function. The specific calculation process is as follows: ； in, , , They represent the outputs of convolutional kernel network layers of different scales. Represents multi-scale fusion features.

6. The turn-to-turn short circuit fault diagnosis method based on feature selection and multi-scale residual network according to claim 5 is characterized in that: In step 2, the residual network is added to the convolutional neural network to obtain a fault diagnosis model. The calculation process is expressed as: ； Among them, x and y represent the input and output of the residual module respectively, represents the residual mapping to be learned.

7. The turn-to-turn short circuit fault diagnosis method based on feature selection and multi-scale residual network according to claim 6 is characterized in that: In step 3, the Softmax function is expressed as: in, For the h is the ordinal number of the output node, and H is the total number of output nodes, that is, the number of categories.

Images