Permanent magnet propulsion motor fault data expansion method based on dual adversarial autoencoding

Abstract

The application discloses a permanent magnet propulsion motor fault data expansion method based on double adversarial auto-encoding, and raw material components are as follows: step one: select data groups from a data set to form a sample set, input to an encoder, output a latent space vector, realize mapping of original data in a latent variable space; input the latent space vector and random noise variables conforming to a distribution into a latent space discriminator respectively, realize probability output through a sigmoid activation function, and train a first adversarial network; the permanent magnet propulsion motor fault data expansion method based on double adversarial auto-encoding, according to respective structural advantages of a generative adversarial network and auto-encoding, establishes a learning model, inputs a combination variable formed by a category label and randomly sampled data conforming to a distribution into a decoder to generate new samples through continuous learning and training, meanwhile, the application expands existing few samples to obtain fault data which meets a specific distribution and diversity.

Description

Technical Field

[0001] This invention relates to the fields of fault diagnosis and deep learning technology, specifically to a method for expanding fault data of permanent magnet propulsion motors based on dual adversarial autoencoders. Background Technology

[0002] As complex electrical devices composed of multiple components, permanent magnet propulsion motors often operate in high-pressure, high-intensity environments in mechanical engineering, leading to various faults. These faults can affect the normal operation of the motor, causing irreversible damage and even threatening the safety of workers. Therefore, fault diagnosis of permanent magnet propulsion motors is essential. However, in actual diagnostic processes, online monitoring data exhibits characteristics such as non-stationarity, nonlinearity, multi-source heterogeneity, and low value density. The number of positive and negative samples differs significantly, while the training of diagnostic models requires a large amount of effective fault data. Furthermore, experimental platform data differs considerably from actual operating conditions in terms of data stability and correlation, resulting in insufficient effective samples for diagnostic analysis.

[0003] To address the problems of high training costs, long training times, and low accuracy caused by insufficient sample data, methods for effectively expanding the sample have been extensively studied. Traditional data expansion methods mainly rely on mathematical analysis or interpolation to expand incomplete or imbalanced data, which suffers from drawbacks such as high computational cost, low fit, and unclear data diversity. Deep learning, with its powerful feature extraction capabilities, can extract deep statistical patterns from the original sample dataset and generate new samples that conform to the data distribution patterns through a trained probability distribution model.

[0004] Generative Adversarial Networks (GANs) achieve a Nash equilibrium through a game-like competition between a generator and a discriminator. This equilibrium means the generator can produce data consistent with the original distribution, while the discriminator cannot accurately determine the data source. Variational Autoencoders (VAGs) encode latent variables from the original data and then decode them to obtain reconstructed data, resulting in more stable training compared to GANs. Combining the training stability of autoencoders with the adversarial nature of GANs, this approach builds a learning model based on a small amount of fault data generated during motor operation. Through continuous learning and training, it expands the existing small sample size to obtain fault data that satisfies a specific distribution and exhibits diversity. Training the network model with this expanded data effectively improves diagnostic accuracy.

[0005] In existing methods for intelligent fault diagnosis of permanent magnet propulsion motors, insufficient training data for the network model leads to insufficient diagnostic accuracy. Summary of the Invention

[0006] The purpose of this invention is to provide a method for expanding fault data of permanent magnet propulsion motors based on dual-adversarial self-encoding, so as to solve the problems in the background art mentioned above.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for expanding fault data of permanent magnet propulsion motors based on dual-adversarial self-encoding.

[0008] Preferably, in step one: select data from the dataset to form a sample set, input it into the encoder, and output a latent space vector to realize the mapping of the original data in the latent variable space; input the latent space vector and the random noise variable that conforms to the distribution into the latent space discriminator respectively, and output the probability through the sigmoid activation function to train the first layer of adversarial network;

[0009] Preferably, in step two: a latent space discriminator is designed, which learns the distribution of data through neural network training, selects A-phase current, B-phase current, negative sequence current, and electromagnetic torque to form a combination feature, performs normalization processing, maps it to the [0, 1] interval, and inputs it into the classification model as an input variable;

[0010] Preferably, in step three: one-hot encoding is performed on different types of fault states of the permanent magnet propulsion motor, and the different types of fault samples of the permanent magnet propulsion motor are divided into training set and test set data in a 3:1 ratio. The latent space variables output by the encoder are combined with the corresponding one-hot encoded category labels to obtain combined variables, which are then input into the decoder to generate new reconstructed samples.

[0011] Preferably, step four involves calculating MSE and MAE to design a sample discriminator, classifying and discriminating reconstructed samples, thereby constraining the decoding network to improve the quality of generated samples. The real samples and reconstructed samples are input into the sample discriminator, and the output is the true / false probability after the sigmoid activation function and the class probability after the softmax function.

[0012] Preferably, in step five: a dual adversarial autoencoder network is trained using fault data of different categories, and the network parameters are updated iteratively by backpropagation using the obtained decoder loss function and sample discriminator loss function, so that the model converges to the global optimum.

[0013] Preferably, step six involves inputting a combined variable consisting of category labels and randomly sampled data from a conforming distribution into the decoder to generate a diverse fault dataset that satisfies a specific distribution.

[0014] Furthermore, considering the large numerical dispersion of the combined features, directly using them as input variables would have an adverse effect on the accuracy of the model, and the neural network training process would encounter the problem of non-convergence. Therefore, the deviation standardization method is used to map them to the [0, 1] interval.

[0015] Furthermore, the encoder consists of five intermediate layers (three 1D convolutional layers and two fully connected layers), with each convolutional kernel having a size of 3 and a stride of 1. Dimensionality reduction is achieved using 1D convolutional layers, and a batch normalization layer is added after each layer. The latent space discriminator consists of four fully connected networks in its intermediate layers, each also followed by a batch normalization layer. All intermediate layers use the LeakyReLU activation function.

[0016] Furthermore, the decoder consists of two fully connected layers and three 1D deconvolutional layers, with four filters in the output layer. The sample discriminator outputs two parallel structures: a softmax output layer that determines the data category, and a fully connected layer that identifies whether the data is real or fake, using the sigmoid activation function.

[0017] Furthermore, the output variable of the model is the state category of the permanent magnet propulsion motor;

[0018] Furthermore, the training iterations were 10,000, the batch size was 64 samples, the latent space data dimension was 8, the learning rate of the encoder and decoder was 0.002, the learning rate of the latent space discriminator and the sample discriminator was 0.001, and the optimization ratio of the encoder, decoder, latent space discriminator and sample discriminator networks in a single training iteration was 2:4:1:1.

[0019] Furthermore, the entropy loss function L of the encoder and latent space discriminator E , The expression is:

[0020]

[0021]

[0022] Decoder D g The overall loss function is in The reconstruction loss function is a combination of MSE and MAE. and Decoder D g The entropy loss function of sample classifier D2 is expressed as follows:

[0023]

[0024]

[0025]

[0026] After the second layer of adversarial interaction, the network parameters are updated, eventually causing the model to converge to the global optimum.

[0027] Compared with existing technologies, the beneficial effects of this invention are as follows: This method for expanding fault data of permanent magnet propulsion motors based on dual adversarial autoencoders establishes a learning model by leveraging the structural advantages of generative adversarial networks and autoencoders. Through continuous learning and training, a combined variable formed by category labels and randomly sampled data from a conformal distribution is input into the decoder to generate new samples. At the same time, this invention expands existing small samples to obtain fault data that satisfies a specific distribution and is diverse. The network model trained with the expanded data can effectively improve the diagnostic accuracy. By integrating with real motor fault datasets, the quality and diversity of fault data are improved, providing reliable data for subsequent fault diagnosis, improving the accuracy of motor fault diagnosis, and enabling early detection and prevention of faults.

[0028] (1) A latent space discriminator was designed and the distribution of data was learned through neural network training. The combined features of phase A current, phase B current, negative sequence current, and electromagnetic torque were selected, normalized, and mapped to the interval [0, 1], and used as input variables into the classification model.

[0029] (2) By performing one-hot encoding on different fault states of permanent magnet propulsion motor, each type of sample is divided into training set and test set data in a 3:1 ratio. The latent space variables and category labels output by the encoder are combined into new variables and input into the decoder to generate new reconstructed samples.

[0030] (3) A sample discriminator is designed by calculating the mean squared error (MSE) and average error (MAE) to classify and discriminate the reconstructed samples, thereby constraining the decoding network and improving the quality of the generated samples. The two models compete with each other and eventually reach the Nash mean, resulting in a well-trained generative model that generates fault data that meets a specific distribution and is diverse. This data, together with real data, constitutes motor fault samples, achieving the purpose of data expansion. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the overall process structure of the present invention;

[0032] Figure 2 This is a schematic diagram of the first layer of adversarial network structure of the dual adversarial autoencoder of the present invention;

[0033] Figure 3 This is a schematic diagram of the second layer of adversarial network structure of the dual adversarial autoencoder of the present invention. Detailed Implementation

[0034] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0035] Example 1:

[0036] 1. A method for intelligent diagnosis of permanent magnet motor faults, characterized in that the method is based on an intelligent diagnostic device for permanent magnet motor faults, comprising a permanent magnet motor, an inverter, an open-circuit fault diagnostic device, and a processor for preventing misjudgment under no-load or light-load conditions, wherein:

[0037] The permanent magnet motor includes three-phase winding currents iA, iB, and iC. These three-phase winding currents iA, iB, and iC are transformed using an abc / dq coordinate system to obtain the actual dq-axis currents id and iq. The actual speed ωr and the given speed ωr* are then PI-regulated to output the given q-axis current iq*. The given q-axis current iq* and the actual q-axis current iq are then PI-regulated to output the given q-axis voltage Vq*. The given d-axis current id* and the actual d-axis current id are then PI-regulated to output the given d-axis voltage Vd*. The given q-axis and d-axis voltages Vq* and Vd* are then transformed using an abc / dq coordinate system to obtain the αβ-axis reference voltages Vα* and Vβ*. The αβ-axis reference voltages Vα* and Vβ* are then subjected to voltage space vector pulse width modulation to obtain a three-phase PWM wave, which is then output to the inverter.

[0038] The inverter receives a three-phase PWM wave after voltage space vector pulse width modulation and drives the permanent magnet motor to run;

[0039] To prevent the system from misjudging under no-load or light-load conditions, the processor determines the value of the given d-axis current id* based on the actual q-axis current value iq of the system.

[0040] The open-circuit fault diagnostic device receives the three-phase winding currents iA, iB, and iC of the permanent magnet motor (ABC), as well as the normalized average current diagnostic constant I0, the maximum value M of the normalized absolute value of the average current, and the minimum value m of the normalized absolute value of the average current, and performs intelligent diagnosis of open-circuit faults, thereby realizing intelligent diagnosis of faults in the permanent magnet motor drive system.

[0041] The method includes the following steps:

[0042] Step 1: The collected three-phase winding currents iA, iB, and iC of the permanent magnet motor (PMM) are transformed using an abc / dq coordinate system to obtain the actual dq-axis currents id and iq. The actual speed ωr and the given speed ωr* are PI-regulated to output the given q-axis current iq*. The given q-axis current iq* and the actual q-axis current iq are PI-regulated to output the given q-axis voltage Vq*. The given d-axis current id* and the actual d-axis current id are PI-regulated to output the given d-axis voltage Vd*. The given q-axis and d-axis voltages Vq* and Vd* are transformed using a dq / αβ coordinate system to obtain the αβ-axis reference voltages Vα* and Vβ*. The αβ-axis reference voltages Vα* and Vβ* are then modulated using voltage space vector pulse width modulation (PWM) to obtain a three-phase PWM wave. The inverter receives the three-phase PWM wave modulated by voltage space vector pulse width modulation and drives the PMM to run.

[0043] Step 2: Use a processor to prevent the system from misjudging no-load or light-load conditions, and determine the value of the given d-axis current id* based on the actual q-axis current value iq of the system.

[0044] Step 3: Substitute the collected three-phase winding currents iA, iB, and iC of the permanent magnet motor (ABC), as well as the normalized average current diagnostic constant I0, the maximum value M of the average current value after normalization of absolute values, and the minimum value m of the average current value after normalization of absolute values, into the open circuit fault diagnostic tool to perform intelligent diagnosis of open circuit faults, thereby realizing intelligent diagnosis of faults in the permanent magnet motor drive system.

[0045] Step 2 describes the use of a processor to prevent erroneous judgments about system no-load or light-load conditions. The value of the given d-axis current id* is determined based on the actual q-axis current value iq. The specific processing method is as follows:

[0046] Where id* is the given d-axis current, iq is the actual q-axis current, F is the threshold for judging the system load condition, and F is 20% of the rated current value of the permanent magnet motor, and L is the d-axis current injection value when the system is under no-load or light-load conditions.

[0047] Step 3 involves substituting the collected three-phase winding currents iA, iB, and iC of the permanent magnet motor (ABC), along with the normalized average current diagnostic constant I0, the maximum value M of the normalized absolute average current value, and the minimum value m of the normalized absolute average current value, into the open-circuit fault diagnostic tool for intelligent open-circuit fault diagnosis. This enables intelligent diagnosis of faults in the permanent magnet motor drive system, as detailed below:

[0048] Step 3.1: Perform Clark / Park transformation on the collected three-phase winding currents iA, iB, and iC of the permanent magnet motor (ABC) to obtain the actual dq-axis current values ​​id and iq of the ABC windings in the dq-axis coordinate system. The Clark / Park transformation method is as follows:

[0049] Based on the Clark / Park transformed values ​​id and iq, after filtering by a low-pass filter and processing by a Park vector processor, the processed three-phase current Park vector reference value |is| is obtained as follows:

[0050] Step 3.2: Normalize the currents of the three-phase windings ABC. The normalized phase current inN is expressed as:

[0051] Where n = A, B, C;

[0052] Step 3.3: Average the obtained normalized phase current over one period to obtain the normalized average current value:

[0053] Where fPMSM is the operating frequency of the permanent magnet motor;

[0054] The normalized average current diagnostic variable In is:

[0055] In= / I0

[0056] Where I0 is the normalized average current diagnostic constant;

[0057] Based on the normalized average current diagnostic variable In, the fault diagnosis signal Mn for normalized average current judgment is obtained as follows:

[0058] Wherein, Mn is the fault diagnosis signal for normalized average current judgment. When the system is in the open-circuit fault state of the lower power switch, Mn = HM; when the system is in the normal state or the winding open-circuit fault state, Mn = NM; when the system is in the open-circuit fault state of the upper power switch, Mn = LM.

[0059] Step 3.4: Take the absolute value of the obtained normalized phase current inN and then average it over one period to obtain the normalized absolute average current value <|inN|>. Based on the obtained <|inN|> and the maximum value M and minimum value m of the normalized absolute average current value, the diagnostic variable dn of the extreme value difference of the normalized absolute average current is obtained as follows:

[0060] The fault diagnosis signal Dn, obtained by determining the fault through the average current extreme difference of the absolute value, is obtained. The fault diagnosis criterion for the average current extreme difference of the normalized absolute value is expressed as follows:

[0061] Where Dn is the fault diagnosis signal of the average current extreme value difference after normalization of absolute value. When the system is in normal state, Dn = ND; when the system is in power switch open circuit fault state, Dn = HD; when the system is in winding open circuit fault state, Dn = LD.

[0062] Step 3.5: Process the fault diagnosis signal Mn obtained by normalizing the average current and the fault diagnosis signal Dn of the extreme difference of the average current after normalization of the absolute value to obtain the specific fault of the permanent magnet motor drive system.

[0063] Step 3.5 involves processing the fault diagnosis signal Mn obtained from the normalized average current and the fault diagnosis signal Dn from the difference between the extreme values ​​of the normalized absolute values ​​of the average current to obtain the specific fault of the permanent magnet motor drive system. The specific fault diagnosis and location criteria are as follows:

[0064] When DA, DB, and DC are HD, ND, and ND respectively, and MA is LM, the diagnostic result is that the power transistor driving phase A has an open circuit fault.

[0065] When DA, DB, DC are HD, ND, ND respectively, and MA is HM, the diagnostic result is that the lower power transistor driving phase A has an open circuit fault.

[0066] When DA is LD and MA is NM, the diagnosis result is an open circuit fault in the winding driving phase A.

[0067] When DA, DB, and DC are ND, HD, and ND respectively, and MB is LM, the diagnostic result is that the power transistor driving phase B has an open circuit fault.

[0068] When DA, DB, and DC are ND, HD, and ND respectively, and MB is HM, the diagnostic result is that the lower power transistor driving phase B has an open circuit fault.

[0069] When DB is LD and MB is NM, the diagnostic result is an open circuit fault in the winding of phase B of the drive.

[0070] When DA, DB, and DC are ND, ND, and HD respectively, and MC is LM, the diagnostic result is that the power transistor driving phase C has an open circuit fault.

[0071] When DA, DB, and DC are ND, ND, and HD respectively, and MC is HM, the diagnostic result is that the lower power transistor driving phase C has an open circuit fault.

[0072] When DC is LD and MC is NM, the diagnostic result is an open circuit fault in the winding of the C phase drive.

[0073] Example 2:

[0074] A permanent magnet motor structure capable of detecting permanent magnet demagnetization and rotor eccentricity faults, characterized in that each phase of the motor stator armature winding consists of a fault detection winding and a main armature winding arranged in a cross pattern.

[0075] The pitch and number of turns of each coil in the fault detection winding and the main armature winding are the same;

[0076] Each phase's main armature winding consists of one or more coils connected in series; the main armature windings of each phase are symmetrically distributed in space.

[0077] Each phase's fault detection winding consists of a coil; the fault detection windings of each phase are spatially symmetrically distributed.

[0078] When the permanent magnet motor is running normally, each phase fault detection winding is connected in series with the main armature winding to form an m-phase stator armature winding, and then the power supply is carried out normally according to the connection method of the m-phase winding.

[0079] When detecting a fault in a permanent magnet motor, the fault detection winding and the main armature winding of each phase are disconnected, and the motor rotor is rotated at a constant speed. The fault condition of the permanent magnet motor is determined by the back electromotive force waveform induced in the m-phase fault detection winding.

[0080] 2. The permanent magnet motor structure for detecting permanent magnet demagnetization and rotor eccentricity faults according to claim 1, characterized in that the structure further includes a stator core and a rotating shaft;

[0081] A permanent magnet pole and a rotor core are sequentially arranged between the stator core and the shaft.

[0082] 3. The permanent magnet motor structure for detecting permanent magnet demagnetization and rotor eccentricity faults according to claim 2, characterized in that a housing is provided outside the stator core, a junction box is also provided on the housing, and a base is fixedly connected to the housing.

[0083] 4. A method for diagnosing faults in a permanent magnet motor, characterized in that the permanent magnet motor includes the permanent magnet motor structure described in any one of claims 1 to 3, capable of detecting permanent magnet demagnetization and rotor eccentricity faults, and the method includes:

[0084] If the waveforms of the back electromotive force are the same for each period, and the waveforms of the positive half-cycle and the negative half-cycle are symmetrical, then there is no permanent magnet demagnetization or eccentricity fault.

[0085] If the waveforms of the back electromotive force are the same, but the positive and negative half-cycles are asymmetrical and there is a dip, then there is a local demagnetization fault in the permanent magnet.

[0086] If the positive and negative half-cycles of the back EMF waveform of each phase are symmetrical, but the amplitude of the back EMF waveform of each phase is different, then there is a rotor eccentricity fault.

[0087] A method for fault data expansion of permanent magnet propulsion motors based on dual adversarial autoencoders:

[0088] Step 1: Select data from the dataset to form a sample set, input it into the encoder, and output the latent space vector to realize the mapping of the original data in the latent variable space; input the latent space vector and the random noise variable that conforms to the distribution into the latent space discriminator respectively, and output the probability through the sigmoid activation function to train the first layer of adversarial network;

[0089] Step 2: A latent space discriminator was designed. The distribution of data was learned through neural network training. A-phase current, B-phase current, negative sequence current, and electromagnetic torque were selected to form a combination feature. The feature was normalized and mapped to the [0, 1] interval, and used as input variables into the classification model.

[0090] Step 3: One-hot encoding is performed on different types of fault states of permanent magnet propulsion motor. The fault samples of different types of permanent magnet propulsion motor are divided into training set and test set data in a 3:1 ratio. The latent space variables output by the encoder are combined with the corresponding one-hot encoded category labels to obtain combined variables, which are input into the decoder to generate new reconstructed samples.

[0091] Step 4: Calculate MSE and MAE to design a sample discriminator to classify and discriminate reconstructed samples, thereby constraining the decoding network and improving the quality of generated samples. Input real samples and reconstructed samples into the sample discriminator respectively, and output the true and false probabilities after the sigmoid activation function and the class probability after the softmax function.

[0092] Step 5: Train a dual adversarial autoencoder network using fault data of different categories. Use the obtained decoder loss function and sample discriminator loss function to perform backpropagation and iteratively update the network parameters of each network until the model converges to the global optimum.

[0093] Step 6: Input the combined variables formed by the category labels and randomly sampled data from the conforming distribution into the decoder to generate a diverse fault dataset that conforms to a specific distribution.

[0094] Considering the large numerical dispersion of the combined features, directly using them as input variables would have an adverse effect on the accuracy of the model, and the training process of the neural network would not converge. Therefore, the deviation standardization method is used to map them to the interval [0, 1].

[0095] Furthermore, the encoder consists of five intermediate layers (three 1D convolutional layers and two fully connected layers), with each convolutional kernel having a size of 3 and a stride of 1. Dimensionality reduction is achieved using 1D convolutional layers, and a batch normalization layer is added after each layer. The latent space discriminator consists of four fully connected networks in its intermediate layers, each also followed by a batch normalization layer. All intermediate layers use the LeakyReLU activation function.

[0096] Furthermore, the decoder consists of two fully connected layers and three 1D deconvolutional layers, with four filters in the output layer. The sample discriminator outputs two parallel structures: a softmax output layer that determines the data category, and a fully connected layer that identifies whether the data is real or fake, using the sigmoid activation function.

[0097] Furthermore, the output variable of the model is the state category of the permanent magnet propulsion motor;

[0098] Furthermore, the training iterations were 10,000, the batch size was 64 samples, the latent space data dimension was 8, the learning rate of the encoder and decoder was 0.002, the learning rate of the latent space discriminator and the sample discriminator was 0.001, and the optimization ratio of the encoder, decoder, latent space discriminator and sample discriminator networks in a single training iteration was 2:4:1:1.

[0099] Furthermore, the entropy loss function L of the encoder and latent space discriminator E , The expression is:

[0100]

[0101]

[0102] Decoder D g The overall loss function is in The reconstruction loss function is a combination of MSE and MAE. and Decoder D g The entropy loss function of sample classifier D2 is expressed as follows:

[0103]

[0104]

[0105]

[0106] After the second layer of adversarial interaction, the network parameters are updated, eventually causing the model to converge to the global optimum.

[0107] First, users need to preprocess the real data, select combined features, perform one-hot encoding on the device's operating status, divide the dataset into training and test sets, and normalize the samples. Based on the VAE, a discriminator and data category information are added. The dual-adversarial autoencoder model framework consists of an encoder, decoder, latent space discriminator, and sample discriminator. To ensure that the variables mapped to the latent space satisfy a specific distribution, a latent space discriminator is designed. This discriminator learns the data distribution through neural network training, forming the first layer of adversarial interaction in the model. The latent space variables output by the encoder are combined with the category labels to form new variables, which are then input into the decoder to generate new reconstructed samples. MSE and MAE are calculated to design the sample discriminator, which classifies and discriminates the reconstructed samples, constraining the decoding network to improve the quality of the generated samples. Real samples and reconstructed samples are input into the sample discriminator, which outputs the true / false probabilities after a sigmoid activation function and the class probabilities after a softmax function, thus forming the second layer of adversarial interaction. A dual adversarial autoencoder network is trained using fault data of different categories. The resulting decoder loss function and discriminator loss function are then used for backpropagation to iteratively update their respective network parameters, ultimately leading to model convergence to the global optimum. Through iterative training of the entire network, only a combination of class labels and randomly sampled data conforming to a distribution needs to be input into the decoder to obtain a high-quality dataset containing samples from multiple classes.

[0108] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0109] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for expanding fault data of permanent magnet propulsion motors based on dual adversarial self-encoding, characterized in that: The manufacturing steps are as follows: Step 1: Select data from the dataset to form a sample set, input it into the encoder, and output the latent space vector to realize the mapping of the original data in the latent variable space; input the latent space vector and the random noise variable that conforms to the distribution into the latent space discriminator respectively, and output the probability through the sigmoid activation function to train the first layer of adversarial network; Step 2: A latent space discriminator was designed. The distribution of data was learned through neural network training. A-phase current, B-phase current, negative sequence current, and electromagnetic torque were selected to form a combination feature. The feature was normalized and mapped to the [0, 1] interval, and used as input variables into the classification model. Step 3: One-hot encoding is performed on different types of fault states of permanent magnet propulsion motor. The fault samples of different types of permanent magnet propulsion motor are divided into training set and test set data in a 3:1 ratio. The latent space variables output by the encoder are combined with the corresponding one-hot encoded category labels to obtain combined variables, which are input into the decoder to generate new reconstructed samples. Step 4: Calculate MSE and MAE to design a sample discriminator to classify and discriminate reconstructed samples, thereby constraining the decoding network and improving the quality of generated samples. Input real samples and reconstructed samples into the sample discriminator respectively, and output the true and false probabilities after the sigmoid activation function and the class probability after the softmax function. Step 5: Train a dual adversarial autoencoder network using fault data of different categories. Use the obtained decoder loss function and sample discriminator loss function to perform backpropagation and iteratively update the network parameters of each network until the model converges to the global optimum. Step 6: Input the combined variables formed by the category labels and randomly sampled data from the conforming distribution into the decoder to generate a diverse fault dataset that conforms to a specific distribution.

2. The method for expanding fault data of permanent magnet propulsion motors based on dual adversarial autoencoders according to claim 1, characterized in that: In step (a), considering the large numerical dispersion of the combined features, directly using them as input variables would adversely affect the accuracy of the model, and the neural network training process would encounter non-convergence problems. Therefore, deviation normalization is used to map them to the [0,1] interval. The encoder in step (a) consists of 5 intermediate layers, including 3 1D convolutional layers and 2 fully connected layers. The kernel size of each convolutional layer is 3, the stride is 1, 1D convolutional layers are used for dimensionality reduction, and a batch normalization layer is added after each layer. The intermediate layers all use the LeakyReLU activation function.

3. The method for expanding fault data of permanent magnet propulsion motors based on dual adversarial autoencoders according to claim 1, characterized in that: The latent space discriminator in step (ii) is used to receive the latent space vector output by the encoder and the random noise variable that conforms to the distribution. Its network structure consists of four fully connected layers, with a batch normalization layer added after each layer. The intermediate layers use the Leaky Relu activation function, and the output layer uses the sigmoid activation function to output the probability value, thereby realizing the discrimination between true and false latent space vectors.

4. The method for expanding fault data of permanent magnet propulsion motors based on dual adversarial autoencoders according to claim 1, characterized in that: The decoder in step (iii) consists of two fully connected layers and three 1D deconvolutional layers. The output layer is set with a filter that determines the number of fault types. The sample discriminator output has two parallel structures: one is a softmax output layer that can determine the data category, and the other is a fully connected layer that can distinguish between true and false data, using the sigmoid activation function X.

5. The method for expanding fault data of permanent magnet propulsion motors based on dual adversarial autoencoders according to claim 1, characterized in that: The output variable of the model in step (iv) is the state category of the permanent magnet propulsion motor.

6. The method for expanding fault data of permanent magnet propulsion motor based on dual adversarial autoencoder according to claim 1, characterized in that: In step (v), the training iterations are 10,000, the sample size in each batch is 64, the data dimension of the latent space is 8, the learning rate of the encoder and decoder is 0.002, the learning rate of the latent space discriminator and the sample discriminator is 0.001, and the optimization ratio of the encoder, decoder, latent space discriminator and sample discriminator networks in a single training iteration is 2:4:1:

1.

7. The method for expanding fault data of permanent magnet propulsion motor based on dual adversarial autoencoder according to claim 1, characterized in that: Entropy loss function of encoder and latent space discriminator , The expression is: ； ； decoder The overall loss function is ,in The reconstruction loss function is a combination of MSE and MAE. and Decoders and sample classifier The entropy loss function is expressed as follows: ； ； ； After the second layer of adversarial interaction, the network parameters are updated, eventually causing the model to converge to the global optimum.

Images