Motor fault data augmentation method and system

Abstract

The present invention relates to the field of fault diagnosis technology, and discloses a method and system for augmenting motor fault data. The method includes: constructing a real training sample set and a test sample set; constructing a first attribute label set, a first feature label set and a first mixed label set of the real training sample, and constructing an attribute reasoning network; constructing a second attribute label set of the test sample set, obtaining a second feature label set of the test sample set, and determining a second mixed label set of the test sample set; constructing a first preset model, a second preset model and a third preset model of a fault data augmentation model; constructing a codec loss function and a discriminator loss function, and establishing a fault data augmentation model loss function; training the fault data augmentation model, obtaining the optimal model parameters for fault data augmentation; and determining generated data consistent with the test sample attributes. The method is easy to implement, does not require additional hardware equipment, and can augment fault data of multiple operating conditions, multiple types, and multiple degradation levels.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a method and system for augmenting motor fault data. Background Art

[0002] As modern industrial production equipment continues to develop in the direction of structuring, automation and intelligence, motors, as the most important power and drive devices, have been widely used in high-speed trains, subways and intercity rail transit equipment. Data under different operating conditions and different operating states are important prerequisites and foundations for realizing motor operating status monitoring. However, it is difficult and costly to obtain data through real operations and physical tests. The existing data is extremely limited, and the performance degradation, fault and failure data are not comprehensive. The operating conditions involved are rare, the fault types are incomplete, and the degradation degree is single, which makes it difficult to meet the needs of motor operating status monitoring.

[0003] Therefore, how to generate fault data with multiple operating conditions, multiple types, and multiple degradation degrees based on single, incomplete data has become an urgent problem to be solved. Summary of the invention

[0004] The present invention provides a motor fault data augmentation method and system to achieve the purpose of generating multi-operating condition, multi-type, and multi-degradation degree fault data based on single, incomplete data.

[0005] In order to achieve the above object, the present invention is implemented by the following technical solutions:

[0006] In a first aspect, the present invention provides a method for augmenting motor fault data, comprising:

[0007] S1: Construct real training sample sets and test sample sets of motors under different working conditions, different fault types, and different fault severity levels;

[0008] S2: construct a first attribute label set, a first feature label set, and a first mixed label set of real training samples, and construct an attribute reasoning network according to the first attribute label set and the first feature label set;

[0009] S3: construct a second attribute label set for the test sample set, obtain a second feature label set for the test sample set according to the second attribute label set and the attribute inference network, and determine a second mixed label set for the test sample set according to the second attribute label set and the second feature label set;

[0010] S4: constructing a first preset model, a second preset model, and a third preset model of the fault data augmentation model according to the first mixed label set;

[0011] S5: constructing a codec loss function and a discriminator loss function according to the first preset model, the second preset model and the third preset model, and establishing a fault data augmentation model loss function according to the codec loss function and the discriminator loss function;

[0012] S6: Train the fault data augmentation model according to the first preset model, the second preset model, the third preset model and the fault data augmentation model loss function to obtain the fault data augmentation optimal model parameters; determine the generated data consistent with the test sample attributes according to the second mixed label set and the fault data augmentation optimal model parameters.

[0013] In a second aspect, the present application provides a motor fault data augmentation system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the first aspect when executing the computer program.

[0014] Beneficial effects:

[0015] The motor fault data augmentation method provided by the present invention uses the mixed attributes of the generated data to control the fault data augmentation model to generate fault data of corresponding attributes, thereby solving the problems caused by the rare operating conditions, incomplete fault types, and single degradation degree to the operating state monitoring, and laying a foundation for the effective implementation of motor state monitoring. The method is easy to implement, does not require additional hardware equipment, and can augment fault data of multiple operating conditions, multiple types, and multiple degradation degrees. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flow chart of a motor fault data augmentation method according to an embodiment of the present invention;

[0017] Figure 2 This is one of the schematic diagrams of the method for generating motor fault data according to an embodiment of the present invention;

[0018] Figure 3 This is a second schematic diagram of a method for generating motor fault data according to an embodiment of the present invention;

[0019] Figure 4 is a comparison diagram of the time-frequency domain waveforms between the fault data generated by the embodiment of the present invention and the real data under situation 1;

[0020] Figure 5 is a comparison diagram of the time-frequency domain waveforms between the fault data generated by the embodiment of the present invention and the real data under situation 2;

[0021] Figure 6 is a comparison diagram of the time-frequency domain waveforms between the fault data generated by the embodiment of the present invention and the real data under situation 3;

[0022] Figure 7 is a comparison diagram of the time-frequency domain waveforms between the fault data generated by the embodiment of the present invention and the real data under situation 4;

[0023] Figure 8 It is a comparison diagram of the time-frequency domain waveforms between the fault data generated by the embodiment of the present invention and the real data under situation 5. DETAILED DESCRIPTION

[0024] The technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the usual meanings understood by persons with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "one" or "one" do not indicate quantity restrictions, but indicate the existence of at least one. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the object being described changes, the relative positional relationship also changes accordingly.

[0026] It should be understood that the motor fault data augmentation method and system in the embodiment of the present application can be applied to application scenarios such as motor control systems and rail transit, such as high-speed trains, EMUs, or subways, urban rails, etc. This is only an example and not a limitation.

[0027] See also Figure 1-Figure 3 , the present application provides a motor fault data augmentation method, comprising:

[0028] S1: Construct real training sample sets and test sample sets of motors under different working conditions, different fault types, and different fault severity levels;

[0029] S2: construct a first attribute label set, a first feature label set, and a first mixed label set of real training samples, and construct an attribute reasoning network according to the first attribute label set and the first feature label set;

[0030] S3: construct a second attribute label set for the test sample set, obtain a second feature label set for the test sample set according to the second attribute label set and the attribute reasoning network, and determine a second mixed label set for the test sample set according to the second attribute label set and the second feature label set;

[0031] S4: constructing a first preset model, a second preset model, and a third preset model of the fault data augmentation model according to the first mixed label set;

[0032] S5: constructing a codec loss function and a discriminator loss function according to the first preset model, the second preset model and the third preset model, and establishing a fault data augmentation model loss function according to the codec loss function and the discriminator loss function;

[0033] S6: Train the fault data augmentation model according to the first preset model, the second preset model, the third preset model and the fault data augmentation model loss function to obtain the optimal model parameters for fault data augmentation; determine the generated data consistent with the test sample attributes according to the second mixed label set and the optimal model parameters for fault data augmentation.

[0034] This embodiment takes the augmentation of fault data of a certain type of permanent magnet synchronous motor as an example to further illustrate and verify the method of the present invention. The stator winding inter-turn short circuit fault (hereinafter referred to as "inter-turn short circuit") is a common fault of permanent magnet synchronous motors and is more harmful to the motor. Due to the self-protection mechanism of actual operation, the fault data of this type of fault is scarce, and in physical tests, it can only be set to a single operating condition and a single fault severity. When an inter-turn short circuit fault occurs, the motor current will increase, the temperature will rise, and even sparks will be generated, which is not suitable for long-term operation, resulting in the sparseness of fault data operating conditions and a single degree of degradation, which is difficult to meet the needs of motor operating status monitoring. Therefore, this embodiment takes the augmentation of motor inter-turn short circuit fault data as an example, and uses three-phase current signal (sensor) data to construct the original sensor data set.

[0035] Among them, the first preset model can be an encoder model, the second preset model can be a decoder model, and the third preset model can be a discriminator model.

[0036] The above-mentioned motor fault data augmentation method uses the fault data with rare operating conditions, incomplete fault types, and single degradation degree acquired by sensors to establish a mapping relationship between the existing data and its corresponding mixed attributes, obtain a fault data augmentation model, and use the mixed attributes of the data to be generated and the constructed fault data augmentation model to generate corresponding data, thereby realizing multi-operating conditions, multi-type, and multi-degradation degree fault data augmentation. This method is easy to implement, does not require additional hardware equipment, has a wide range of applications, and lays the foundation for effective real-time health monitoring of motors and trains.

[0037] Optionally, the S1 includes:

[0038] S11: Collect the original sensor data D of the motor under W different working conditions and C different fault types with S different fault severity 1,1,1 , D 2,1,1 ,…,D w,1,1 ,…,D w,c,s ,…,D W,C,S , where D w,c,s ∈R n The sensor data of the wth working condition, the cth fault type, and the sth fault severity, n represents the number of sampling points of the sensor signal, and the original sensor data set is expressed as D = {D w,c,s ∈R n}, w=1,2,…,W, c=1,2,…,C, s=1,2,…,S;

[0039] S12: Normalize the original data set of the sensor. The normalization description of the i-th data point of the sensor under the w-th working condition, the c-th fault type, and the s-th fault severity satisfies the following formula, where i = 1, 2, ..., n:

[0040]

[0041] In the formula, max(D w,c,s ) and min(D w,c,s ) are the maximum and minimum values ​​of the sensor data collected under the wth working condition, the cth fault type, and the sth fault severity, respectively. is the ith data point, For the corresponding The normalized value of

[0042] For all data D in the sensor raw data D w,c,s Perform normalization processing to obtain normalized sensor data in After normalization, the sensor dataset is represented as

[0043] S13: Normalized sensor data with window size Win and step size Stp Do a time sliding window, and the data after the time sliding window is Where l represents the number of samples, and the calculation formula is as follows:

[0044] l=floor((n-Win) / Stp)+1 (2)

[0045] In the formula, floor means rounding down;

[0046] The normalized sensor data set is processed with a window size of Win and a step size of Stp. All data in Do a time sliding window, the data after the time sliding window is in The sensor data set after the time sliding window can be expressed as

[0047] S14: Arbitrarily select the sensor data set after the time sliding window Construct a test sample set D as a test sample Te , where w1∈1,2,…,W, c1∈1,2,…,C, s1∈1,2,…,S, The test sample set is represented as in represents the i-th test sample;

[0048] The remaining data in the sensor data set after the time sliding window is used as the real training samples, and the real training sample set is constructed by splicing rows in sequence. Among them l Tr Represents the number of real training samples, and its size can be expressed as:

[0049] l Tr =(W+C+S-1)·l (3)

[0050] The real training sample set is represented as in represents the i-th true training sample.

[0051] Optionally, the S2 includes:

[0052] S21: Construct sample attribute labels based on the working conditions, fault types and fault severity of the samples in the real training sample set When the sample When , its attribute label is w=1,2,…,W,c=1,2,…,C,s=1,2,…,S,and w≠w1,c≠c1,s≠s1, construct the attribute labels of all samples in the real training sample set in turn, and obtain the first attribute label set of the real training sample set

[0053] Extract n from the real training sample set f features, and use the extracted features as the feature labels of the corresponding samples to construct the first feature label set of the real training sample set

[0054] Concatenate the first attribute label and the first feature label of the real training sample set by column to construct the first mixed label set of the real training sample set Where n M Indicates the number of mixed labels, its size is:

[0055] n M =n f +3 (4)

[0056] The first mixed label set is composed of l Tr The mixed labels of samples are composed of:

[0057]

[0058] In the formula, is the mixed label of the i-th real training sample;

[0059] S22: According to the first attribute label set and the first feature label set of the real training sample, attribute reasoning network parameters are obtained through learning, and an attribute reasoning network is constructed;

[0060] The first attribute label set of the real training sample set As input, the first feature label set As output, the optimal parameters of the attribute reasoning network are obtained through learning, which is expressed as:

[0061]

[0062] In the formula, θ BPNN are the parameters of the attribute inference network, is the optimal parameter obtained by learning based on the attribute label set and feature label set of the real training samples, and f(·) represents the mapping function of the attribute reasoning network;

[0063] by Construct an attribute reasoning network, the formula is:

[0064]

[0065] Where X is the input of the attribute reasoning network, and Y is the output of the attribute reasoning network.

[0066] Optionally, the S3 includes:

[0067] Construct sample attribute labels based on the working conditions, fault types and fault severity of the samples in the test sample set Dataset The attribute tag is The second attribute label set of the test sample set

[0068] The second attribute label set of the test sample set Input the attribute reasoning network to obtain the second feature label of the test sample set

[0069]

[0070] Concatenate the attribute labels and feature labels of the test sample set by column to construct the second mixed label set of the test sample set

[0071] Optionally, the S4 includes:

[0072] S41: constructing a first preset model of the fault data augmentation model and initializing model parameters;

[0073] The first preset model in the fault data augmentation model is the encoder model. The encoder consists of a convolutional unit and a fully connected layer. The input of the encoder model is a single sample in the real training sample set. Initialize encoder model parameters θ Enc ; The output of the encoder model is the hidden vector under this initialization parameter The mapping relationship between the input and output of the encoder model is expressed as:

[0074]

[0075] Where n Z Represents the hidden vector The dimension of n Z <<Win, Γ(·) represents the encoder mapping function;

[0076] S42: constructing a second preset model of the fault data augmentation model and initializing model parameters;

[0077] The second preset model in the fault data augmentation model is the decoder model. The decoder consists of transposed convolution units. The input of the decoder model is the hidden vector Mixed label with the i-th true training sample Vector after column concatenation Initialize decoder model parameters θ Dec ; The output of the decoder model is the generated sample under the initialization parameters The mapping relationship between the input and output of the decoder model is expressed as:

[0078]

[0079] Where Ψ(·) represents the decoder mapping function;

[0080] S43: constructing a third preset model of the fault data augmentation model and initializing model parameters;

[0081] The third preset model in the fault data augmentation model is the discriminator model. The discriminator consists of a convolutional unit and a fully connected layer. Mixed label with the i-th true training sample After column concatenation, we get a vector Will generate samples Mixed label with the i-th true training sample After column concatenation, we get a vector Vector and Input decoder model; initialize discriminator model parameters θ Dis ; The output of the discriminator model is the discriminant function value under the initialization parameters The mapping relationship between the input and output of the discriminator model is expressed as:

[0082]

[0083] Where Φ(·) represents the discriminator mapping function, x i represents the input of the discriminator, when When , it means that the sample comes from the real training sample. If the corresponding It indicates that the discriminator determines that the sample comes from the generated sample, indicating that the judgment is wrong. It indicates that the discriminator determines that the sample comes from the real training sample, indicating that the judgment is correct. When , it means that the sample comes from the generated sample. If the corresponding It indicates that the discriminator determines that the sample comes from the generated sample, indicating that the judgment is correct. It indicates that the discriminator determines that the sample comes from the real training sample, indicating that the judgment is wrong;

[0084] The fault data augmentation model parameters are composed of three parts: encoder, decoder, and discriminator model parameters, which can be expressed as:

[0085] θ DA =[θ Enc ,θ Dec ,θ Dis ] (12)

[0086] Optionally, the S5 includes:

[0087] S51: construct encoder-decoder loss function;

[0088] Set the batch size to m, that is, randomly select m samples from the real training samples as a batch Use the first preset model to Mapping is performed to obtain the corresponding latent variables Then the hidden variables are input into the second preset model to obtain the corresponding generated samples The encoder-decoder loss function is constructed as:

[0089]

[0090] Where D KL represents the KL divergence, q(z Tr |x Tr ) represents the approximate posterior distribution, P(z Tr |x Tr ) represents the true posterior distribution, express and The cross entropy between is used to characterize the similarity between the real training samples and the generated samples;

[0091] S52: Construct discriminator loss function;

[0092] The vectors obtained by concatenating the real training samples with the corresponding mixed labels and the generated samples with the corresponding mixed labels are Input the third preset model, and obtain the corresponding discrimination result 0 or 1 through the discriminator. The expression of the discriminator loss function is:

[0093]

[0094] The expected goal of the discriminator loss function is to make the discriminator distinguish whether the sample comes from a real training sample or a generated sample;

[0095] S53: Constructing the loss function of the fault data augmentation model;

[0096] The loss function of the fault data augmentation model is expressed as:

[0097] L=λ vae L VAE +λ d L D (15)

[0098] In the formula, λ vae , d Represent the loss function L VAE , L D Corresponding to the coefficient in the total loss.

[0099] Optionally, the S6 includes:

[0100] S61: According to the set number of iterations N, batch size m, and learning rate lr, the real training sample set D TrInput the fault data augmentation model in batches, and train the fault data augmentation model according to the first preset model, the second preset model, the third preset model and the fault data augmentation model loss function, with the goal of minimizing the loss function, to obtain the optimal model parameters:

[0101]

[0102] In the formula, They are the optimal fault data augmentation model parameters, the optimal encoder model parameters, the optimal decoder model parameters, and the optimal discriminator model parameters respectively;

[0103] S62: Randomly generate and hide the vector z i The Gaussian vector matrix Z′ has the same dimension and the same number of samples as the test sample set. Concatenate the Gaussian vector matrix Z′ with the second mixed label set of the test sample to obtain the concatenated matrix The concatenated matrix is ​​input into the decoder model constructed by the optimal decoder model parameters to obtain the generated sample D′ that is consistent with the test sample attributes Te ∈R l×Win :

[0104]

[0105] In one example, when the permanent magnet synchronous motor runs at a speed of 1400 r / min, the data of the A-phase current signal under five conditions, namely normal and turn-to-turn short circuit (the severity of the turn-to-turn short circuit fault is 0.05, 0.1, 0.15, and 0.2 respectively), are collected. The system sampling frequency is 2.5×10 3 , the number of sampling points is 1.01×10 5 After normalization and time sliding window preprocessing, the window size of the time sliding window Win = 1024, the step size Stp = 100, and a total of 1000 samples of data in each case are obtained. Four of the five cases are randomly selected as the real training sample set D Tr ∈R 4000×1024 , and the remaining one is used as the test sample set D Te ∈R 1000 ×1024 The data collection information and attribute labels of the five cases are shown in Table 1. Attribute label set of the real training sample set Attribute label set of the test sample set

[0106] Table 1 Data collection information and attribute labels in five cases

[0107]

[0108] A convolutional neural network (CNN) is used to extract the features of the real training sample set, with a feature dimension of 32, and to construct a feature label set of the real training sample set. Concatenate the attribute label set and feature label set of the real training sample set by column to obtain the mixed label set of the real training sample set The attribute labels and feature labels of the training sample set are used to train the attribute inference network to obtain the optimal model parameters. Based on the obtained attribute inference network and the attribute labels of the test sample set, the mixed label set of the test sample set is obtained.

[0109] Set the batch size to m = 32, the number of iterations N = 125, and the initial learning rate Lr = 0.001. Input the samples in the training sample set into the constructed fault data augmentation model in batches. The structure diagram of the fault data augmentation model is shown in Figure 4-Figure 8 As shown in the figure, the Adam optimization algorithm is used to minimize the loss function, train the fault data augmentation model, and obtain the optimal fault data augmentation model parameters (including the optimal encoder model, decoder model, and discriminator model parameters).

[0110] Randomly generate Gaussian vector matrix Z′∈R 1000×20 , concatenate the Gaussian vector matrix Z′ with the test sample mixed label set constructed in step S23 to obtain the concatenated matrix Input the concatenated matrix into the optimal decoder model to obtain the generated sample D′ that is consistent with the test sample attributes Te ∈R 1000×1024 :

[0111] Generate samples and test samples such as Figure 4 As shown, Figure 4 The time domain and frequency domain comparison results of the test samples and generated samples under 5 test conditions (as shown in Table 2) are shown. From the time domain waveform, the waveform of the generated data is highly similar to that of the real data. From the frequency domain, the fundamental frequencies of the generated data and the real data are basically consistent.

[0112] Furthermore, the similarity of the generated data and the test data is evaluated, and the Euclidean distance, KL divergence, cosine similarity, and Pearson correlation coefficient are selected to calculate the similarity of the two data. The similarity index results are shown in Table 2. As can be seen from Table 2, the distance between the generated data and the test data is small and the similarity value is high.

[0113] comprehensive Figure 4 The results in Table 2 show that this method can effectively generate motor fault data and realize motor fault data augmentation.

[0114] Table 2 Similarity index results

[0115]

[0116] An embodiment of the present application also provides a motor fault data augmentation system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the corresponding steps of the above method when executing the computer program.

[0117] The above-mentioned motor fault data augmentation system can implement various embodiments of the above-mentioned motor fault data augmentation method and can achieve the same beneficial effects, which will not be described in detail here.

[0118] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for augmenting motor fault data, characterized in that: include: S1: Construct real training sample sets and test sample sets of motors under different working conditions, different fault types, and different fault severity levels; S2: construct a first attribute label set, a first feature label set, and a first mixed label set of real training samples, and construct an attribute reasoning network according to the first attribute label set and the first feature label set; S3: construct a second attribute label set for the test sample set, obtain a second feature label set for the test sample set according to the second attribute label set and the attribute inference network, and determine a second mixed label set for the test sample set according to the second attribute label set and the second feature label set; S4: constructing a first preset model, a second preset model, and a third preset model of the fault data augmentation model according to the first mixed label set; S5: constructing a codec loss function and a discriminator loss function according to the first preset model, the second preset model and the third preset model, and establishing a fault data augmentation model loss function according to the codec loss function and the discriminator loss function; S6: Train the fault data augmentation model according to the first preset model, the second preset model, the third preset model and the fault data augmentation model loss function to obtain the fault data augmentation optimal model parameters; determine the generated data consistent with the test sample attributes according to the second mixed label set and the fault data augmentation optimal model parameters.

2. The motor fault data augmentation method according to claim 1, characterized in that: The S1 includes: S11: Collect the original sensor data D of the motor under W different working conditions and C different fault types with S different fault severity 1,1,1 , D 2,1,1 ,…,D w,1,1 ,…,D w,c,s ,…,D W,C,S , where D w,c,s ∈R n The sensor data of the wth working condition, the cth fault type, and the sth fault severity, n represents the number of sampling points of the sensor signal, and the original sensor data set is expressed as D = {D w,c,s ∈R n }, w=1,2,…,W, c=1,2,…,C, s=1,2,…,S; S12: Normalize the original data set of the sensor. The normalization description of the i-th data point of the sensor under the w-th working condition, the c-th fault type, and the s-th fault severity satisfies the following formula, where i = 1, 2, ..., n: In the formula, max(D w,c,s ) and min(D w,c,s ) are the maximum and minimum values ​​of the sensor data collected under the wth working condition, the cth fault type, and the sth fault severity, respectively. is the ith data point, For the corresponding The normalized value of For all data D in the sensor raw data D w,c,s Perform normalization processing to obtain normalized sensor data in After normalization, the sensor dataset is represented as S13: Normalized sensor data with window size Win and step size Stp Do a time sliding window, and the data after the time sliding window is Where l represents the number of samples, and the calculation formula is as follows: l=floor((n-Win) / Stp)+1 (2) In the formula, floor means rounding down; The normalized sensor data set is processed with a window size of Win and a step size of Stp. All data in Do a time sliding window, the data after the time sliding window is in The sensor data set after the time sliding window can be expressed as S14: Arbitrarily select the sensor data set after the time sliding window Construct a test sample set D as a test sample Te , where w1∈1,2,…,W, c1∈1,2,…,C, s1∈1,2,…,S, The test sample set is represented as in represents the i-th test sample; The remaining data in the sensor data set after the time sliding window is used as the real training samples, and the real training sample set is constructed by splicing rows in sequence. Among them l Tr Represents the number of real training samples, and its size can be expressed as: l Tr =(W+C+S-1)·l (3) The real training sample set is represented as in represents the i-th true training sample.

3. The motor fault data augmentation method according to claim 1, characterized in that: The S2 includes: S21: Construct sample attribute labels based on the working conditions, fault types and fault severity of the samples in the real training sample set when When , its attribute label is And w≠w1,c≠c1,s≠s1, construct the attribute labels of all samples in the real training sample set in turn, and obtain the first attribute label set of the real training sample set Extract n from the real training sample set f features, and use the extracted features as the feature labels of the corresponding samples to construct the first feature label set of the real training sample set Concatenate the first attribute label and the first feature label of the real training sample set by column to construct the first mixed label set of the real training sample set where n M Indicates the number of mixed labels, its size is: n M =n f +3 (4) The first mixed label set is composed of l Tr The mixed labels of samples are composed of: In the formula, is the mixed label of the i-th real training sample; S22: According to the first attribute label set and the first feature label set of the real training sample, attribute reasoning network parameters are obtained through learning, and an attribute reasoning network is constructed; The first attribute label set of the real training sample set As input, the first feature label set As output, the optimal parameters of the attribute reasoning network are obtained through learning, which is expressed as: In the formula, θ BPNN are the parameters of the attribute inference network, is the optimal parameter obtained by learning based on the attribute label set and feature label set of the real training samples, and f(·) represents the mapping function of the attribute reasoning network; by Construct an attribute reasoning network, the formula is: Where X is the input of the attribute reasoning network, and Y is the output of the attribute reasoning network.

4. The motor fault data augmentation method according to claim 3, characterized in that: The S3 includes: Construct sample attribute labels based on the working conditions, fault types and fault severity of the samples in the test sample set Dataset The attribute tag is The second attribute label set of the test sample set The second attribute label set of the test sample set Input the attribute reasoning network to obtain the second feature label of the test sample set Concatenate the attribute labels and feature labels of the test sample set by column to construct the second mixed label set of the test sample set 5. The motor fault data augmentation method according to claim 1, characterized in that: The S4 includes: S41: constructing a first preset model of the fault data augmentation model and initializing model parameters; The first preset model in the fault data augmentation model is the encoder model. The encoder consists of a convolutional unit and a fully connected layer. The input of the encoder model is a single sample in the real training sample set. Initialize encoder model parameters θ Enc ; The output of the encoder model is the hidden vector under this initialization parameter The mapping relationship between the input and output of the encoder model is expressed as: Where n Z Represents the hidden vector The dimension of n Z <<Win, Γ(·) represents the encoder mapping function; S42: constructing a second preset model of the fault data augmentation model and initializing model parameters; The second preset model in the fault data augmentation model is the decoder model. The decoder consists of transposed convolution units. The input of the decoder model is the hidden vector Mixed label with the i-th true training sample Vector after column concatenation Initialize decoder model parameters θ Dec ; The output of the decoder model is the generated sample under the initialization parameters The mapping relationship between the input and output of the decoder model is expressed as: Where Ψ(·) represents the encoder mapping function; S43: constructing a third preset model of the fault data augmentation model and initializing model parameters; The third preset model in the fault data augmentation model is the discriminator model. The discriminator consists of a convolutional unit and a fully connected layer. Mixed label with the i-th true training sample After column concatenation, we get a vector Will generate samples Mixed label with the i-th true training sample After column concatenation, we get a vector Vector and Input decoder model; initialize discriminator model parameters θ Dis ; The output of the discriminator model is the discriminant function value under the initialization parameters The mapping relationship between the input and output of the discriminator model is expressed as: In the formula, Φ(·) represents the discriminator mapping function, x i represents the input of the discriminator, when When , it means that the sample comes from the real training sample. If the corresponding It indicates that the discriminator determines that the sample comes from the generated sample, indicating that the judgment is wrong. It indicates that the discriminator determines that the sample comes from the real training sample, indicating that the judgment is correct. When , it means that the sample comes from the generated sample. If the corresponding It indicates that the discriminator determines that the sample comes from the generated sample, indicating that the judgment is correct. It indicates that the discriminator determines that the sample comes from the real training sample, indicating that the judgment is wrong; The fault data augmentation model parameters are composed of three parts: encoder, decoder, and discriminator model parameters, which can be expressed as: i DA =[θ Enc ,i Dec ,i Dis ] (12).

6. The motor fault data augmentation method according to claim 1, characterized in that: The S5 includes: S51: construct encoder-decoder loss function; Set the batch size to m, that is, randomly select m samples from the real training samples as a batch Use the first preset model to Mapping is performed to obtain the corresponding latent variables Then the hidden variables are input into the second preset model to obtain the corresponding generated samples The encoder-decoder loss function is constructed as: Where D KL represents the KL divergence, q(z Tr |x Tr ) represents the approximate posterior distribution, P(z Tr |x Tr ) represents the true posterior distribution, express and The cross entropy between is used to characterize the similarity between the real training samples and the generated samples; S52: Construct discriminator loss function; The vectors obtained by concatenating the real training samples with the corresponding mixed labels and the generated samples with the corresponding mixed labels are Input the third preset model, and obtain the corresponding discrimination result 0 or 1 through the discriminator. The expression of the discriminator loss function is: The expected goal of the discriminator loss function is to make the discriminator distinguish whether the sample comes from a real training sample or a generated sample; S53: Constructing the loss function of the fault data augmentation model; The loss function of the fault data augmentation model is expressed as: L=λ vae L VAE +λ d L D (15) In the formula, λ vae , d Represent the loss function L VAE , L D Corresponding to the coefficient in the total loss.

7. The motor fault data augmentation method according to claim 1, characterized in that: The S6 includes: S61: According to the set number of iterations N, batch size m, and learning rate lr, the real training sample set D Tr Input the fault data augmentation model in batches, and train the fault data augmentation model according to the first preset model, the second preset model, the third preset model and the fault data augmentation model loss function, with the goal of minimizing the loss function, to obtain the optimal model parameters: In the formula, They are the optimal fault data augmentation model parameters, the optimal encoder model parameters, the optimal decoder model parameters, and the optimal discriminator model parameters respectively; S62: Randomly generate and hide vector z i The Gaussian vector matrix Z′ has the same dimension and the same number of samples as the test sample set. Concatenate the Gaussian vector matrix Z′ with the second mixed label set of the test sample to obtain the concatenated matrix The concatenated matrix is ​​input into the decoder model constructed by the optimal decoder model parameters to obtain the generated sample D′ that is consistent with the test sample attributes Te ∈R l×Win :

8. A motor fault data augmentation system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method described in any one of claims 1 to 7 are implemented.

Images