Permanent magnet synchronous motor fault diagnosis method based on deep learning

Abstract

The application discloses a kind of permanent magnet synchronous motor fault diagnosis methods based on deep learning, utilizes the expansion of fault data of modified generative adversarial network, after joining damage noise in dataset, input deep sparse auto-encoding network combines softmax classifier to realize the classification of motor turn-to-turn short circuit fault.The method comprises the following steps: 1) collect the three-phase stator current signal and zero sequence voltage signal of motor different degree turn-to-turn short circuit, and establish dataset;2) build auto-encoder model;3) the auto-encoder is used as the generation model of generative adversarial network, and the normal data collected is used as the input of generation model, to complete the expansion of fault data;4) the expanded dataset is divided into training set and test set according to the set proportion;5) unsupervised learning and supervised fine-tuning are organically combined to train deep sparse auto-encoding network, and the classifier is combined to realize the classification of turn-to-turn short circuit fault.Compared with similar methods, the application has higher applicability and feasibility.

Description

Technical Field

[0001] This invention belongs to the field of permanent magnet synchronous motor fault diagnosis technology, and relates to a fault diagnosis method for permanent magnet synchronous motors based on deep learning. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs) mainly consist of a stator, rotor, and end covers. They offer advantages such as high efficiency, low noise, small size, light weight, and low losses. However, due to the complex operating environment of PMSMs, they are highly susceptible to inter-turn short-circuit faults. If minor inter-turn short-circuit faults are ignored, the continuously rising short-circuit circulating current generates high temperatures, leading to demagnetization and ultimately causing the motor to malfunction or even burn out. Therefore, research on inter-turn short-circuit faults in PMSMs and the optimization of datasets are of great significance for efficient, accurate, low-consumption, and time-saving fault detection.

[0003] Deep learning-based fault diagnosis technology for permanent magnet synchronous motors has become a current research hotspot. To address the challenges of scarce fault samples and independent, singular features in real-world scenarios, a data expansion model is introduced to perform unsupervised learning on the samples. Pseudo-data is generated to simulate real samples. A sparse autoencoder network is then used to classify and diagnose fault data features, improving the accuracy and stability of fault diagnosis.

[0004] Previous research on deep learning-based fault diagnosis methods typically employs nonlinear mapping functions in both the generative and discriminative models of generative adversarial networks (GANs). These models are implemented using multilayer perceptrons or deep neural networks, inputting Gaussian noise into the generative model to synthesize data of the corresponding dimension. The generative model then learns features from real samples to achieve data expansion. However, this approach is not only slow to train but also requires a sufficient number of fault samples to meet the training needs of the GAN model. In reality, faults occur infrequently, and the data obtained is often only from the normal operation of the motor, making it difficult to acquire enough fault data. This makes training the GAN challenging and increases the training time.

[0005] Sparse autoencoders have a large number of neurons and a large number of network parameters, which can easily lead to overfitting. In addition, sparse autoencoders are shallow networks, containing only three layers: input, hidden, and output. Their learning ability is limited, which will affect their ability to learn features from the input data and thus reduce the accuracy of different fault classifications. Summary of the Invention

[0006] The purpose of this invention is to provide a method for diagnosing inter-turn short-circuit faults in permanent magnet synchronous motors based on deep learning, thereby overcoming the shortcomings mentioned in the background art.

[0007] The technical solution to achieve the purpose of this invention is as follows:

[0008] A deep learning-based method for diagnosing inter-turn short-circuit faults in permanent magnet synchronous motors includes the following steps:

[0009] Step 1: Collect the three-phase stator current signal and zero-sequence voltage signal of the permanent magnet synchronous motor under normal operation and various inter-turn short circuit faults of different degrees as the combined feature quantity for inter-turn short circuit fault diagnosis. The collected data sets are processed in batches to obtain the corresponding feature components and form a dataset, which includes a normal sample set and a sample set of inter-turn short circuit faults of different degrees.

[0010] Step 2: Build the autoencoder model by using the normal sample set from Step 1 as input to the autoencoder network to complete the training of the autoencoder network.

[0011] Step 3: Build a generative adversarial network (GAN) model. Use the autoencoder network trained in Step 2 as the generative model of the GAN, and a deep neural network as the discriminant model. Use normal samples as input to the generative model, and use the output of the generative model and samples of inter-turn short-circuit faults of different degrees as input to the discriminant model. When the GAN obtains the optimal objective function, the training of the network model is completed. At this time, the generative model and the discriminant model reach Nash equilibrium. The expansion of samples of inter-turn short-circuit faults of different degrees is completed through the trained GAN.

[0012] Step 4: Mix the fault samples generated in Step 3 with the dataset formed in Step 1 to form an expanded dataset. Divide the expanded dataset into an unlabeled dataset and a labeled dataset with type labels added according to the severity of the inter-turn short circuit fault according to a set ratio. Divide the unlabeled dataset into a training set and a test set according to a set ratio.

[0013] Step 5: After adding damage noise to the training set divided in Step 4, input it into the deep sparse autoencoder network for feature learning. Add a softmax classification layer after the last hidden layer of the deep sparse autoencoder network. The number of neurons in the classification layer is the number of inter-turn short circuit fault categories of different degrees in the motor. Complete the network pre-training. Use a labeled dataset to perform supervised fine-tuning on the deep sparse autoencoder network. Determine whether the model has been trained by calculating whether the overall cost function has converged. The output is the label probability value of inter-turn short circuit faults of different degrees. The label with the highest probability value is the corresponding fault type, thereby determining the severity of the inter-turn short circuit fault in the motor. Use a test set to verify the effectiveness of the model. Finally, obtain the trained inter-turn short circuit fault diagnosis model for the motor.

[0014] The significant advantages of this invention compared to existing technologies are:

[0015] (1) This invention uses deep learning to use a trained autoencoder as the generative model of a generative adversarial network, and uses the collected normal motor three-phase stator current data and zero-sequence voltage data as the input of the generative model, thereby generating inter-turn short-circuit fault data of different degrees. This greatly reduces the dependence of the generative adversarial network model on the collected real fault data during training, and also reduces the training batches of the generative adversarial network.

[0016] (2) Adding certain impairment noise to the input layer of a deep sparse autoencoder network can effectively suppress the overfitting problem of the network.

[0017] (3) Using deep sparse autoencoder networks, compared with general sparse autoencoder networks, it is able to learn better feature representations of input data.

[0018] (4) By combining multi-layer unsupervised learning and supervised fine-tuning, the performance of deep sparse autoencoders can be improved. Attached Figure Description

[0019] Figure 1 A flowchart illustrating a specific implementation method for diagnosing inter-turn short-circuit faults in a permanent magnet synchronous motor;

[0020] Figure 2 This is a schematic diagram of the adversarial network structure generated in this invention;

[0021] Figure 3 This is a schematic diagram of the deep sparse autoencoder network structure in this invention; Detailed Implementation

[0022] Specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be noted that in the following description, detailed descriptions of known functions and designs that might obscure the main points of the invention will be omitted here. References Figure 1 , Figure 2 as well as Figure 3 The specific steps of the deep learning-based fault diagnosis method for permanent magnet synchronous motors are as follows:

[0023] 1) Using a sampling frequency of 10kHz, the three-phase stator current signal and zero-sequence voltage signal of the permanent magnet synchronous motor under normal operating conditions and under five different levels of inter-turn short circuit fault conditions (1%, 2%, 5%, 10%, and 15%) were collected as combined feature quantities for inter-turn short circuit fault diagnosis. 3500 sets of three-phase stator current and zero-sequence voltage data sets of the motor under different operating conditions were collected, including 1000 sets under normal motor conditions and 500 sets under each of the five different levels of inter-turn short circuit fault conditions. The corresponding feature components were obtained through batch calculation. The 3500 combined features were formed into a dataset, which includes a normal sample set and a fault sample set for the five different levels of inter-turn short circuit. The dataset was divided into a training set and a test set in a 3:1 ratio.

[0024] 2) Build an autoencoder model, setting hyperparameters such as initial learning rate, number of hidden layers, and stride, and selecting the cross-entropy loss function and Adam optimization algorithm. Use normal samples from the training set as input to the autoencoder network, and update the neuron weights layer by layer using gradient descent to make the network output approximately equal to the network input. When the autoencoder network loss function is less than a set threshold, the training of the autoencoder network is complete. Input normal samples from the test set into the trained autoencoder network to verify its effectiveness. Save the trained autoencoder model.

[0025] 3) Construct a generative adversarial network model;

[0026] The result diagram of the generative adversarial network model in this invention is shown below. Figure 2 As shown, the specific steps of the setup process are as follows:

[0027] (1) The autoencoder trained in step 2 is used as the generator of the generative adversarial network (GAN), and the discriminator is composed of a deep neural network. The input of the GAN model is divided into two parts: the first is the input of the generator, which is the collected normal samples, and the second is the input of the discriminator, which is the output of the generator and fault samples with different degrees of inter-turn short circuits. The output of the GAN model is [1,0], which means that the probability of the discriminator classifying the fault sample as true is 1, and the probability of classifying the data output by the generator as true is 0.

[0028] Define data parameters:

[0029]

[0030] Where: x t and z t These represent the fault data and normal data of the t-th motor, respectively; y t [1,0] represents the output of the generative adversarial network model, and T represents the number of times the GAN is trained.

[0031] (2) Initialize the discriminator parameters of the generative adversarial network (GAN). The generator parameters are determined by the autoencoder model parameters from step 2. During training, the Adam optimization algorithm is used to backpropagate the error to each layer node. By modifying the weights between nodes and the biases of each layer, the error is mitigated. The parameters are continuously updated in repeated iterations to obtain the state with the lowest error. The lowest error is evaluated by the optimization objective function of the GAN. The final optimization objective function of the GAN is described as follows:

[0032]

[0033] Where: ∑ a～p(x) log(D(a, θ) D )) represents the minimum loss function of the corresponding discriminative model obtained by fixing the generative model. This indicates that the generative model maximizes the objective function, θ. D To determine the parameters to be optimized in the model, θ G To generate the parameters to be optimized in the model, 'a' represents a fault sample, and 'a' ~ 'p(a)' follows the sampling of motor fault samples. This represents the output data of the generative model. Sampling of generated data following the distribution of the generative model, D(α, θ) D ) is the mapping function of the discriminant model. The mapping function for generating the model, D(α, θ) D () represents the probability value that 'a' will be judged as true. Indicates will The probability value of being identified as fake.

[0034] When the generative adversarial network obtains the optimal objective function, the network model training is complete. At this point, the generative model and the discriminative model reach Nash equilibrium. The conditions for Nash equilibrium are:

[0035] 1. The pseudo-fault data generated by the generator is consistent with the distribution of fault samples in the dataset;

[0036] 2. The discriminator can only randomly guess whether a specific sample is true or false, that is, the probability of being true or false is 50%.

[0037] (3) Each time the experiment is conducted, 60 normal samples and 60 fault samples with certain inter-turn short circuits are randomly sampled from the training set. The experiment is iterated 60 times. When D(G(z)) = 1, the generated model output data and fault data reach the most ideal state. The network parameters at this time are recorded as the final data expansion network model.

[0038] (4) The effectiveness of the generative adversarial network in this invention is verified using a test set, and finally a well-trained generative adversarial network fault data generation model is obtained.

[0039] (5) The inter-turn short-circuit fault data of the motor are expanded using conventional generative adversarial networks and the generative adversarial network model in the method of this invention. The number of samples and training time required to complete the model training are as follows:

[0040]

[0041] As can be seen from the above experimental process and results, this invention addresses the issue of insufficient inter-turn short-circuit fault data in real-world situations by using a modified generative adversarial network to expand the data, while also shortening the training batches and training time of the generative adversarial network model. It has broad application prospects in expanding fault data for motors and other mechanical devices.

[0042] 4) Based on step 3, perform data augmentation on the real fault samples, generating 500 pseudo-samples for each of the five different degrees of inter-turn short-circuit faults. After mixing the real dataset and the augmented dataset, 6000 samples are obtained. These samples are then divided into an unlabeled dataset and a labeled dataset with type labels added according to the severity of the inter-turn short-circuit fault, according to a set ratio. The unlabeled dataset is then divided into a training set and a test set according to a set ratio.

[0043] 5) Construct a deep sparse autoencoder network model;

[0044] The result diagram of the deep sparse autoencoder network model in this invention is shown below. Figure 3 As shown, the specific steps of the model training process are as follows:

[0045] (1) Randomly set some elements of the input matrix sample z in the training set to zero according to the qD distribution to transform it into a matrix sample Z containing damaged noise. Input the matrix sample X into the input layer of the deep sparse autoencoder network, and Z satisfies:

[0046] Z = qD(Z|z)

[0047] Where: qD is a form of noise distribution, that is, the impairment noise is added according to the qD distribution. z is the input matrix sample without impairment noise.

[0048] (2) The first layer of the sparse autoencoder is trained in an unsupervised manner. The cost function of the sparse autoencoder network is minimized as follows:

[0049]

[0050] Where w and b represent the weight matrix connecting the input layer and the hidden layer, and the bias matrix between the two layers, respectively. h is the mean square error reconstruction term. w，b x (i) y is the output value of the output layer of the neural network for the i-th sample group. (i)Here, is the input value of the neural network input layer for the i-th sample group, and m is the number of training samples, KL(ρ||ρ j ) represents the KL divergence. For sparsity penalty, β is the coefficient of the sparsity penalty term, d′ is the encoding vector, and ρ is the sparsity parameter, which is usually taken as a small value (close to 0). j denoted as the average activation level of neuron j in the hidden layer.

[0051] (3) Use the previous sparse autoencoder hidden layer as the input of the next sparse autoencoder;

[0052] (4) Repeat (3) to complete the training of all sparse autoencoder networks;

[0053] (5) The output of the last hidden layer is used as the input of the softmax classifier. The number of neurons in the classification layer is the number of different types of inter-turn short circuit faults in the motor.

[0054] (6) After pre-training the deep sparse autoencoder network through the above five steps, the network is then fine-tuned in a supervised manner using a labeled dataset to improve its performance. The overall cost function of the deep sparse autoencoder network model is:

[0055]

[0056] Where: w and b represent the weight matrix connecting the input layer and the hidden layer, and the bias matrix between the two layers, respectively; m is the number of training samples; J(w, b; q) e s e ) represents the sparse autoencoder network of each layer, q e Let s represent the e-th input sample. e This represents the category label for group e.

[0057] (7) The output of the last hidden layer is used as the input of the softmax classification layer. The output is the label probability value of the inter-turn short circuit fault of different degrees. The label with the largest probability value is the corresponding fault type, thereby determining the severity of the inter-turn short circuit fault of the motor. The test set verifies the effectiveness of the network model, and finally the trained inter-turn short circuit fault diagnosis model of the motor is obtained.

[0058] (8) The test set was compared with the test set using a sparse autoencoder network and a shallow neural network, and the deep sparse autoencoder network in the method of this invention. The classification accuracy is as follows:

[0059]

[0060] The experimental results above demonstrate that this invention achieves higher accuracy in motor fault diagnosis, effectively diagnosing inter-turn short-circuit faults in motors. It has broad application prospects in the field of motor and other simulateable mechanical fault diagnosis, and also provides a new method for deep learning-based fault diagnosis technology.

Claims

1. A deep learning-based inter-turn short circuit fault diagnosis method for a permanent magnet synchronous motor, characterized by, Comprising the following steps: Step 1, collecting three-phase stator current signals and zero sequence voltage signals of permanent magnet synchronous motor under normal working and various degrees of turn-to-turn short circuit fault as the combined characteristic quantity of motor turn-to-turn short circuit fault diagnosis, obtaining the corresponding characteristic component through batch calculation of the collected data set to form a data set containing a normal sample set and a variety of different degrees of turn-to-turn short circuit fault sample set; Step 2, building an autoencoder model, taking the normal sample set in step 1 as the input of the autoencoding network, and completing the training of the autoencoding network; Step 3, building a generative adversarial network model, taking the trained autoencoding network in step 2 as the generative model of the generative adversarial network, and taking the deep neural network as the discriminant model; Using the normal sample set as the input of the generative model, the output of the generative model and the different degrees of turn-to-turn short circuit fault sample set as the input of the discriminant model, when the optimal objective function of the generative adversarial network is obtained, the training of the network model is completed, at this time the generative model and the discriminant model reach Nash equilibrium, and the expansion of different degrees of turn-to-turn short circuit fault sample is completed through the trained generative adversarial network; The optimal objective function is: ； wherein: represents the minimum loss function of the corresponding discriminative model obtained by the fixed generative model, represents the maximum objective function of the generative model, is the parameter to be optimized of the discriminative model, is the parameter to be optimized of the generative model, represents a fault sample, obeys the sampling under the motor fault sample, represents the output data of the generative model, obeys the sampling under the data distribution generated by the generative model, is the mapping function of the discriminative model, is the mapping function of the generative model, represents that will be the probability value of discriminating as true, represents that will be the probability value of discriminating as false; Step 4, mixing the fault samples generated in step 3 with the data set in step 1 to form an expanded data set, and dividing the expanded data set into training set and test set according to the set proportion; Step 5, after adding damage noise to the training set divided in step 4, input the deep sparse autoencoding network for feature learning, add a softmax classification layer after the last hidden layer of the deep sparse autoencoding network, the number of neurons of the classification layer is the number of different degrees of turn-to-turn short circuit fault classes of the motor, complete the network pre-training, add type labels to the deep sparse autoencoding network according to the severity of the motor turn-to-turn short circuit fault in the sample expansion data set in step 4, and perform supervised fine tuning, whether the model is trained is judged by calculating whether the overall cost function converges, and the output is the label probability value of different degrees of turn-to-turn short circuit fault, the maximum probability value is the corresponding fault type, so as to determine the severity of the motor turn-to-turn short circuit fault, and finally obtain the trained motor turn-to-turn short circuit fault diagnosis model; The overall cost function of the deep sparse autoencoding network model is: ； in: , These represent the weight matrix connecting the input layer and the hidden layer, and the bias matrix between the two layers, respectively. The number of training samples, This represents the sparse autoencoder network of each layer. Indicates the first Group input samples, Indicates the first Group category labels.

2. The method according to claim 1, wherein, The Nash equilibrium in step 3 satisfies the following conditions: (1) the pseudo-fault data generated by the generator is consistent with the distribution of the fault samples in the data set; (2) the discriminator can only randomly guess whether a specific sample is real or fake, i.e. the probabilities of true and false are both 50%.

3. The method according to claim 1, wherein the method further comprises: determining the fault type of the PMSM based on the fault feature vector. The impairment noise described in step 5 is to randomly zero the part of elements in the input matrix sample in the training set according to the distribution and become the matrix sample containing impairment noise , satisfies: ； where: is a form of noise distribution, i.e. impairment noise is added according to ; and is the input matrix sample without impairment noise.

4. The method of claim 1, wherein the method further comprises: The pre-training of the deep sparse autoencoding network model in step 5 is to complete the training of each layer of sparse autoencoding network in turn until all sparse autoencoding networks are trained; the sign of completion of training of each layer of sparse autoencoding network is whether the cost function converges or not, and its formula is: ； wherein: , W and b represent the weight matrix connecting the input layer and the hidden layer and the bias value matrix between the two layers, respectively, is the mean square error reconstruction term, is the output value of the output layer of the neural network for the i-th sample, represents the input layer sample, is the output of the hidden layer for the i-th sample, is the number of training samples, represents the divergence, is the sparsity penalty term, is the coefficient of the sparsity penalty term, is the encoding vector, is the sparsity parameter, is the average activation of the hidden layer neuron.​​​

Images