Synchronous motor rotor broken bar fault state identification method based on stator current

Abstract

The invention provides a synchronous motor rotor broken bar fault state identification method based on stator current, and the method comprises the steps: obtaining the stator current sample data of a synchronous motor when a rotor broken bar fault occurs, and enabling the stator current sample data to be collected through a current sensor and to be generated through a generative adversarial network; performing d-q transformation on the stator current sample data, and then performing Fourier decomposition to obtain feature data; according to feature data of the stator current sample data and a corresponding actual rotor broken bar fault category, a fault diagnosis model based on CNN and Transform is constructed, and hyper-parameters of the fault diagnosis model are optimized by using an optimization algorithm; and using the fault diagnosis model to diagnose the rotor broken bar fault category. According to the invention, the accuracy, efficiency and generalization ability of synchronous motor rotor broken bar fault diagnosis are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of motor fault diagnosis, and particularly to a method for identifying the broken rotor bar fault state of a synchronous motor based on stator current. Background Art

[0002] As the core equipment for power transmission and energy conversion in the modern industrial field, the operating state of a motor directly affects production efficiency and safety. However, during operation, a motor is often prone to failures due to various factors, such as broken rotor bars, rotor eccentricity, inter-turn short circuits, etc. These failures will ultimately lead to equipment damage and thus cause huge economic losses. The incidence of the broken rotor bar fault accounts for about 10% of the total faults. When a broken rotor bar fault occurs, the current cannot pass smoothly through the broken bar, and will concentrate on other unbroken conductors, resulting in local overheating and ultimately causing the motor to stop rotating, affecting industrial production efficiency. Therefore, it is crucial to quickly and accurately identify the broken rotor bars of a synchronous motor.

[0003] Currently, the research methods for motor fault diagnosis can be roughly divided into three categories. The first is to establish a mathematical model of the motor under different working conditions and analyze the operating state of the motor by analyzing the parameter changes during motor operation. This method requires the staff to have a professional knowledge background. The second is to analyze the impact of the signals generated during motor operation on the motor state. However, when the motor operation faces complex and changing environments and states, key characteristic quantities such as current and torque will show dynamic changes, resulting in disadvantages such as low diagnostic efficiency, low accuracy, and difficulty in real-time monitoring for this method. In recent years, artificial intelligence has brought new solutions to various industries, and the adoption of artificial intelligence algorithms has become a current research boom. A large number of methods based on artificial intelligence combined with fault diagnosis have been proposed. Common methods for data preprocessing include Hilbert series transforms, empirical mode decomposition, etc. Hilbert series transforms involve complex integral and convolution operations, and when the signal length is long, the computational complexity will increase significantly. When the frequencies of signals are close, empirical mode decomposition will have a mode mixing problem. Most of the existing literature is based on small samples. Small sample data may not be able to fully represent the distribution and characteristics of the overall data, limiting its generalization ability on new data and resulting in poor performance in practical applications. In addition, whether the data comes from real scenarios or is collected in the laboratory, there is a data imbalance problem, which will seriously affect the accuracy of the diagnostic model. A large number of models have been proposed by scholars, such as Support Vector Machine (SVM), Long Short-Term Memory (LSTM), etc. When the data is unbalanced, the classification results of SVM will be biased, and its performance depends to a large extent on feature engineering. When processing sequence data, due to its cyclic structure characteristics, LSTM consumes a large amount of computing resources. For longer sequences, due to the vanishing gradient problem, the model may have difficulty learning long-distance information.

[0004] Through the above analysis, the problems and defects existing in the prior art are as follows:

[0005] (1) The prior art uses the Hilbert series decomposition algorithm, which involves a large number of convolution operations or Fourier transform operations, and has a high computational complexity. Especially in the case of processing a large amount of data or having high real-time requirements, it cannot meet the needs of rapid diagnosis.

[0006] (2) When the prior art faces the imbalance problem in the data collection process, the classification results will be affected, and the performance is too dependent on the establishment of artificial feature engineering. Summary of the Invention

[0007] In view of the above problems existing in the prior art, the present invention provides a method for identifying the broken rotor bar fault state of a synchronous motor based on stator current, so as to improve the accuracy, efficiency and generalization ability of the broken rotor bar fault diagnosis of the synchronous motor. Its characteristics are simple calculation, high accuracy, less consumption of computing resources and practical feasibility.

[0008] The present invention provides a method for identifying the broken rotor bar fault state of a synchronous motor based on stator current, including:

[0009] Obtain the stator current sample data of the first synchronous motor during the broken rotor bar fault, and the stator current sample data is collected by a current sensor and generated by a generative adversarial network;

[0010] After performing d-q transformation on the stator current sample data, perform Fourier decomposition to obtain feature data;

[0011] Construct a fault diagnosis model based on CNN and Transformer according to the feature data of the stator current sample data and the corresponding actual broken rotor bar fault categories, and optimize the hyperparameters of the fault diagnosis model using an optimization algorithm;

[0012] After performing d-q transformation on the current stator current data of the second synchronous motor, perform Fourier decomposition to obtain feature data, and input the feature data of the current stator current data into the fault diagnosis model to obtain the broken rotor bar fault category output by the fault diagnosis model.

[0013] According to the method for identifying the broken rotor bar fault state of a synchronous motor based on stator current provided by the present invention, obtaining the stator current sample data of the first synchronous motor during the broken rotor bar fault includes:

[0014] Respectively obtain the stator current sample data of the first synchronous motor when the broken rotor bar fault categories are 1 broken bar, 2 broken bars and 3 broken bars.

[0015] According to the method for identifying the broken rotor bar fault state of a synchronous motor based on stator current provided by the present invention, after performing d-q transformation on the stator current sample data, perform Fourier decomposition to obtain feature data, including:

[0016] Perform two-phase shifts on the stator current sample data i1 of each frequency component, and the stator current sample data after the two-phase shifts are i2 and i3 respectively;

[0017] For I = [i1, i2, i3] T Perform d-q transformation to obtain the d-axis component i d and the q-axis component i q ;

[0018] Extract the d-axis component current i through Fourier decomposition d and the DC component current i da in d q and the q-axis component current i qa ;

[0019] Calculate the sum of squares of the DC component current i da and i qa to obtain the amplitude of each frequency component as the characteristic data.

[0020] According to a method for identifying the rotor broken bar fault state of a synchronous motor based on stator current provided by the present invention, when the first synchronous motor has a rotor broken bar fault, current harmonic components are generated, and the formula for the stator current sample data is:

[0021]

[0022] where n is a positive integer; A1, ω1, and θ1 are respectively the fundamental wave amplitude, angular frequency, and initial phase angle of the stator current sample data; A2,..., A n are the amplitudes of the harmonic components of the stator current sample data; ω2,..., ω n are the angular frequencies of the harmonic components of the stator current sample data; θ2,..., θ n are the initial phase angles of the harmonic components of the stator current sample data.

[0023] According to a method for identifying the rotor broken bar fault state of a synchronous motor based on stator current provided by the present invention, after performing d-q transformation on the stator current sample data and then performing Fourier decomposition to obtain characteristic data, including:

[0024] Shift the stator current sample data i1 of the frequency component ω2 120° to the left in the time domain to obtain the current i2, and shift it 240° to the left to obtain the current i3. i2 and i3 are expressed as:

[0025]

[0026] where θ′ and θ″ are respectively the phase differences of the fundamental wave components in i2 and i3 relative to i1; θ′ n and θ″ n are respectively the phase differences of the harmonic components in i2 and i3 relative to i1;

[0027] The formula for performing d-q transformation on I = [i1, i2, i3] T is as follows, P 3 / 2 is the d-q transformation matrix:

[0028]

[0029] Extract the d-axis component i of the stator current by Fourier decomposition d The DC component i in da and the q-axis component i of the stator current q The DC component i in qa The formulas are as follows:

[0030]

[0031] Calculate the DC component i through the following formula da and i qa The sum of squares to obtain the amplitude of the frequency component ω2 as the characteristic data:

[0032]

[0033] According to a method for identifying the broken rotor bar fault state of a synchronous motor based on stator current provided by the present invention, the generative adversarial network includes a generator and a discriminator;

[0034] The generator is used to generate new data according to the input random noise;

[0035] The discriminator is used to determine the possibility that the new data belongs to real data;

[0036] The structure of the generator includes a fully connected layer, an activation function LeakyReLU, a first deconvolution layer, a second deconvolution layer, and a third deconvolution layer connected in sequence;

[0037] The structure of the discriminator includes a first convolution layer, a second convolution layer, a Dropout layer, and a fully connected layer connected in sequence.

[0038] According to a method for identifying the broken rotor bar fault state of a synchronous motor based on stator current provided by the present invention, the fault diagnosis model based on CNN and Transformer includes a CNN, a Transformer encoder, and a classifier connected in sequence;

[0039] The CNN is used to extract the local features of the characteristic data of the current stator current data;

[0040] The Transformer encoder is used to capture the global dependencies of the characteristic data of the current stator current data;

[0041] The classifier is used to output the broken rotor bar fault category of the second synchronous motor.

[0042] A method for identifying the rotor broken bar fault state of a synchronous motor based on stator current provided by the present invention constructs a fault diagnosis model based on CNN and Transformer according to the characteristic data of the stator current sample data and the corresponding actual rotor broken bar fault categories. The hyperparameters of the fault diagnosis model are optimized using an optimization algorithm, including:

[0043] Divide the characteristic data of the stator current sample data and the corresponding actual rotor broken bar fault categories as samples into a training set and a validation set;

[0044] Under the current hyperparameters of the fault diagnosis model, use the training set to train the fault diagnosis model, and evaluate the trained fault diagnosis model on the validation set to obtain the fitness value corresponding to the current hyperparameters;

[0045] If the fitness value corresponding to the current hyperparameters is better than the global optimal fitness value, update the global optimal fitness value to the fitness value corresponding to the current hyperparameters, and update the global optimal parameters corresponding to the global optimal fitness value to the current hyperparameters;

[0046] After updating the current hyperparameters of the fault diagnosis model using the optimization algorithm, continue to train, evaluate, and update the global optimal parameters of the fault diagnosis model until the maximum number of iterations is reached, and output the global optimal parameters;

[0047] Use the model obtained by training the fault diagnosis model with the training set under the global optimal parameters as the final fault diagnosis model.

[0048] A method for identifying the rotor broken bar fault state of a synchronous motor based on stator current provided by the present invention, the hyperparameters of the fault diagnosis model include the number of CNN layers, the attention dimension, the number of Transformer encoder layers, and the number of multi-head attention heads.

[0049] A method for identifying the rotor broken bar fault state of a synchronous motor based on stator current provided by the present invention, evaluating the trained fault diagnosis model on the validation set to obtain the fitness value corresponding to the current hyperparameters, including:

[0050] Use the cross-entropy loss of the validation set as the fitness value corresponding to the current hyperparameters.

[0051] A method for identifying the broken rotor bar fault state of a synchronous motor based on stator current provided by the present invention. By collecting the stator current of the motor rotor bar fault state, after performing d-q transformation on the stator current, and then performing Fourier decomposition on the transformation result to obtain characteristic data, it can more directly and clearly reflect the current characteristics of the motor operating state. At the same time, the calculation process is relatively simple, improving the diagnostic efficiency while ensuring the diagnostic accuracy; using GAN to generate synthetic data highly similar to the real motor fault data, expanding the scale of the data set and balancing the data categories. Based on GAN data augmentation, it can effectively alleviate the problem of data imbalance, and at the same time generate high-quality data that meets the model requirements. By deeply learning the internal characteristics of the data, the generalization ability of the model is improved; CNN-Transformer, as an end-to-end model, has complementary advantages of local feature extraction and global dependence capture in processing one-dimensional time series data, can adapt to various complex data patterns, reduce the dependence on manual feature engineering, and its parallel structure has higher efficiency and accuracy when processing large-scale data sets; using an optimization algorithm to globally optimize the hyperparameters of the CNN-Transformer network model to find the optimal combination of model parameters, improving the diagnostic performance and efficiency of the model. The augmented data generated by GAN and the original data are used together to train the CNN-Transformer model, which is beneficial to enhancing the sensitivity of the model to fault characteristics and improving the accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0053] Figure 1 is one of the flow schematic diagrams of the method for identifying the broken rotor bar fault state of a synchronous motor based on stator current provided by the present invention;

[0054] Figure 2 is the second flow schematic diagram of the method for identifying the broken rotor bar fault state of a synchronous motor based on stator current provided by the present invention;

[0055] Figure 3 is the structural schematic diagram of the fault diagnosis model based on CNN and Transformer in the method for identifying the broken rotor bar fault state of a synchronous motor based on stator current provided by the present invention;

[0056] Figure 4 is the schematic diagram of iteratively optimizing the parameters in the method for identifying the broken rotor bar fault state of a synchronous motor based on stator current provided by the present invention;

[0057] Figure 5 It is a schematic diagram of the training effect of the CNN-Transformer model in the method for identifying the rotor broken bar fault state of a synchronous motor based on stator current provided by the present invention. Specific embodiments

[0058] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Obviously, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without making creative efforts fall within the protection scope of the present invention.

[0059] The prior art generally uses Hilbert series transforms to extract fault feature frequency components and uses an SVM model for fault diagnosis. The Hilbert series decomposition algorithm involves a large number of convolution operations or Fourier transform operations, and the computational complexity is relatively high. When the signal length is long, the computational complexity will increase significantly, which is not suitable for real-time diagnosis; the performance of the SVM depends too much on the establishment of artificial feature engineering, and when the data is unbalanced, there will be deviations in the SVM classification results.

[0060] The following combines Figure 1 Describe a method for identifying the rotor broken bar fault state of a synchronous motor based on stator current of the present invention, including:

[0061] Step 101, obtain the stator current sample data of the first synchronous motor during the rotor broken bar fault, and the stator current sample data is collected by a current sensor and generated by a generative adversarial network;

[0062] Step 102, after performing d-q transformation on the stator current sample data, perform Fourier decomposition to obtain feature data;

[0063] Step 103, construct a fault diagnosis model based on CNN and Transformer according to the feature data of the stator current sample data and the corresponding actual rotor broken bar fault category, and optimize the hyperparameters of the fault diagnosis model using an optimization algorithm;

[0064] Step 104, after performing d-q transformation on the current stator current data of the second synchronous motor, perform Fourier decomposition to obtain feature data, and input the feature data of the current stator current data into the fault diagnosis model to obtain the rotor broken bar fault category output by the fault diagnosis model.

[0065] In the data acquisition and processing step, the stator current in the broken rotor bar fault state of the synchronous motor is collected. After performing d-q transformation on the stator current, Fourier decomposition is performed on the transformation result to obtain characteristic data. A Hall current sensor can be used to collect the stator current in the broken rotor bar fault state. The d-q transformation has good effects in electrical signal applications. As a linear transformation, the d-q transformation does not change the key information in the original current signal, and can more directly and clearly reflect the current characteristics under the motor operating state. In addition, its calculation is simple, mainly involving algebraic and trigonometric calculations, which is easy to implement and is conducive to real-time diagnosis.

[0066] In the data augmentation step based on GAN (Generative Adversarial Networks), GAN is used for data augmentation and balancing data categories.

[0067] A fault diagnosis model based on a hybrid model of CNN and Transformer is constructed. By automatically extracting data features in an end-to-end manner, it can overcome the disadvantages of weak generalization ability, poor robustness, and difficult maintenance. And an optimization algorithm is used to optimize hyperparameters such as the number of convolutional layers, the number of channels, and the number of attention heads of the model. To achieve accurate classification of motor faults, it has the advantages of high accuracy, fast diagnosis speed, and strong robustness. The first synchronous motor and the second synchronous motor can be the same motor or different motors, which is not limited in this embodiment. The complete flowchart is as Figure 2 shown.

[0068] In this embodiment, by collecting the stator current in the broken rotor bar fault state of the motor, performing d-q transformation on the stator current, and then performing Fourier decomposition on the transformation result to obtain characteristic data, it can more directly and clearly reflect the current characteristics under the motor operating state. At the same time, the calculation process is relatively simple, which improves the diagnosis efficiency while ensuring the diagnosis accuracy; using GAN to generate synthetic data highly similar to the real motor fault data, expanding the scale of the dataset and balancing data categories. Data augmentation based on GAN can effectively alleviate the problem of data imbalance, and at the same time generate high-quality data that meets the model requirements. By deeply learning the internal characteristics of the data, the generalization ability of the model is improved; CNN-Transformer, as an end-to-end model, enables it to process one-dimensional time series data with complementary advantages of local feature extraction and global dependence capture, can adapt to various complex data patterns, reduce the dependence on artificial feature engineering, and its parallel structure has higher efficiency and accuracy when processing large-scale datasets; using an optimization algorithm to globally optimize the hyperparameters of the CNN-Transformer network model, finding the optimal combination of model parameters, improving the diagnostic performance and efficiency of the model. The augmented data generated by GAN and the original data are used together to train the CNN-Transformer model, which is beneficial to enhancing the sensitivity of the model to fault features and improving the accuracy.

[0069] Based on the above embodiments, in this embodiment, obtaining the stator current sample data of the first synchronous motor during the rotor bar breaking fault includes:

[0070] Obtaining the stator current sample data of the first synchronous motor when the rotor bar breaking fault categories are 1 broken bar, 2 broken bars, and 3 broken bars respectively.

[0071] Collect the stator current under three working conditions of 1 broken bar, 2 broken bars, and 3 broken bars of the motor through a current sensor. After performing d-q transformation on the stator current, perform Fourier decomposition on the transformation result, extract the DC component to highlight the fault characteristics, and finally obtain the amplitude of the frequency component as the characteristic data by calculating the sum of squares of the DC components.

[0072] Based on the above embodiments, in this embodiment, after performing d-q transformation on the stator current sample data and then performing Fourier decomposition to obtain the characteristic data, it includes:

[0073] Perform two-phase shifts on the stator current sample data i1 of each frequency component. The stator current sample data after the two-phase shifts are i2 and i3 respectively;

[0074] For I = [i1, i2, i3] T Perform d-q transformation to obtain the d-axis component i d and the q-axis component i q ;

[0075] Extract the DC component i d in the d-axis component i da and the DC component i q in the q-axis component i qa ;

[0076] Calculate the sum of squares of the DC components i da and i qa to obtain the amplitude of each frequency component as the characteristic data.

[0077] Based on the above embodiments, in this embodiment, when the first synchronous motor has a rotor bar breaking fault, current harmonic components are generated. The formula for the stator current sample data is:

[0078]

[0079] where n is a positive integer; A1, ω1, and θ1 are respectively the fundamental wave amplitude, angular frequency, and initial phase angle of the stator current sample data; A2,..., A n are the amplitudes of the respective harmonic components of the stator current sample data; ω2,..., ω n are the angular frequencies of the respective harmonic components of the stator current sample data; θ2,..., θn is the initial phase angle of each harmonic component of the stator current sample data.

[0080] On the basis of the above embodiments, in this embodiment, after performing d-q transformation on the stator current sample data, Fourier decomposition is performed to obtain characteristic data, including:

[0081] The stator current sample data i1 of the frequency component ω2 is shifted 120° to the left in the time domain to obtain the current i2, and shifted 240° to the left to obtain the current i3. i2 and i3 are expressed as:

[0082]

[0083] where θ′ and θ″ are the phase differences of the fundamental wave components in i2 and i3 relative to i1; θ′ n and θ″ n are the phase differences of each harmonic component in i2 and i3 relative to i1, respectively;

[0084] For I = [i1, i2, i3] T The formula for performing d-q transformation is as follows, P 3 / 2 is the d-q transformation matrix:

[0085]

[0086] Extract the DC component i d in the d-axis component through Fourier decomposition da and the DC component i q in the q-axis component qa The formula is as follows:

[0087]

[0088] Calculate the sum of squares of the DC components i da and i qa through the following formula to obtain the amplitude of the frequency component ω2 as the characteristic data:

[0089]

[0090] Calculate the sum of squares of the DC components obtained by Fourier decomposition of the d-q transformation result to obtain the amplitude of the frequency ω2 component, and use this as the fault recognition characteristic spectrum data. Through d-q transformation, any frequency component of the stator current i1 can be transformed into a DC component to highlight the fault characteristics.

[0091] Use a Hall current sensor to collect the stator current in the rotor broken bar fault state; perform phase shift on the stator current to obtain I; perform d-q transformation on I to obtain i d and i q; The DC components \(i\) in formula (4) are obtained through Fourier decomposition da and \(i\) qa ; The amplitude of this frequency component is obtained through formula (5);

[0092] Select the characteristic frequency amplitudes from 0 to 200 Hz as the characteristic data set and normalize it. In this embodiment, the rotor broken bar fault is classified according to the number of broken bars in the rotor. One broken bar is classified as an early minor fault, marked as 0; two broken bars are classified as an early moderate fault, marked as 1; three broken bars are classified as an early severe fault, marked as 2. Table 1 shows some samples of the data set, \(f\) n (n = 1, 2, …, 200) represents the characteristic frequency.

[0093] Table 1 Some data samples

[0094]

[0095] Based on the above embodiments, in this embodiment, the generative adversarial network (GAN) includes a generator and a discriminator;

[0096] The generator is used to generate new data according to the input random noise;

[0097] The discriminator is used to determine the possibility that the new data belongs to the real data;

[0098] The structure of the generator includes a fully connected layer, an activation function LeakyReLU, a first deconvolution layer, a second deconvolution layer, and a third deconvolution layer connected in sequence;

[0099] The structure of the discriminator includes a first convolution layer, a second convolution layer, a Dropout layer, and a fully connected layer connected in sequence.

[0100] Construct a GAN model, which consists of two parts: a generator and a discriminator. The role of the generator is to generate new data similar to the original data. The input is random noise, and the output is the generated new data. The task of the discriminator is to distinguish the data generated by the generator from the original data. Its input is the original data and the new data generated by the generator, and the output is 0 or 1, indicating the possibility that the data belongs to the real data.

[0101] The specific structure of the generator model can be: the first layer is a fully connected layer, and the activation function is selected as LeakyReLU. The second layer is a deconvolution layer with a stride of 1, and the third and fourth layers are both deconvolution layers with a stride of 2. The discriminator model structure can be: first, two consecutive convolution layers with a convolution kernel of 5×5 and a stride of 2. The third layer introduces a Dropout layer, and finally, there is a fully connected layer.

[0102] Determine the training strategy and train the GAN model. In this case, the cross-entropy function is selected as the loss function, the Adam algorithm is used as the optimizer, the initial learning rate is 0.0001, and the number of training epochs is 500.

[0103] After training is completed, call the generator model with the best effect. Input random noise to the generator to obtain new data samples. For example, there are 2000 samples of each of the three types of data collected. A total of 10500 samples are obtained through this step.

[0104] This embodiment utilizes the Generative Adversarial Network (GAN) technology to generate synthetic data highly similar to real motor fault data, expand the scale of the dataset, and fully exert the learning ability of the model.

[0105] Based on the above embodiments, the fault diagnosis model based on CNN and Transformer in this embodiment includes a CNN, a Transformer encoder, and a classifier connected in sequence;

[0106] The CNN is used to efficiently extract the local features of the feature data of the current stator current data;

[0107] The Transformer encoder is used to capture the global dependencies of the feature data of the current stator current data;

[0108] The classifier is used to output the rotor bar breakage fault category of the second synchronous motor, realizing accurate identification and diagnosis of motor fault features.

[0109] The structure of the fault diagnosis model based on CNN and Transformer is as Figure 3 shown. This model consists of three parts in total. The first part is a convolutional neural network. The size of each convolutional kernel in the convolutional layer is 3×3, the padding is 1, and finally, a max-pooling layer with a convolutional kernel of 2×2 and a stride of 2 is used for downsampling operation to reduce the computational amount.

[0110] The second part consists of a Transformer encoder structure (Encoding). The self-attention mechanism considers all positions in the sequence simultaneously when processing sequence data, improving the computational efficiency. Its encoder layer consists of two sub-layers, namely the multi-head self-attention mechanism and the feed-forward neural network.

[0111] The third part is the classifier. The average pooling layer is used to reduce the data dimension, and then the feature is mapped to the classification result through the fully connected layer to obtain the final prediction output. There are three categories in the dataset of this case, and the output parameters of the fully connected layer are set to 3.

[0112] Based on the above embodiments, in this embodiment, a fault diagnosis model based on CNN and Transformer is constructed according to the characteristic data of the stator current sample data and the corresponding actual rotor broken bar fault category. The hyperparameters of the fault diagnosis model are optimized using an optimization algorithm, including:

[0113] Divide the characteristic data of the stator current sample data and the corresponding actual rotor broken bar fault category as samples into a training set and a validation set, and initialize the optimization algorithm parameters;

[0114] Under the current hyperparameters of the fault diagnosis model, use the training set to train the fault diagnosis model, and evaluate the trained fault diagnosis model on the validation set to obtain the fitness value corresponding to the current hyperparameters;

[0115] If the fitness value corresponding to the current hyperparameters is better than the global optimal fitness value, update the global optimal fitness value to the fitness value corresponding to the current hyperparameters, and update the global optimal parameters corresponding to the global optimal fitness value to the current hyperparameters;

[0116] After updating the current hyperparameters of the fault diagnosis model using the optimization algorithm, continue to train, evaluate, and update the global optimal parameters of the fault diagnosis model;

[0117] Judge whether the iteration reaches the termination condition. If the condition is met, stop the iteration and output the global optimal parameters. If the number of iterations has not exceeded the maximum number of iterations, continue to execute the next step and increase the number of iterations by 1;

[0118] Use the model obtained by training the fault diagnosis model with the training set under the global optimal parameters as the final fault diagnosis model.

[0119] Determine the best parameters of the fault diagnosis model based on the optimization search algorithm. Use global optimization algorithms such as the sparrow search algorithm to optimize the hyperparameters of the model, such as the attention dimension, the number of network layers, the number of attention heads, etc., to find the optimal model parameter combination.

[0120] Configure the model according to the optimal parameter combination obtained from the optimization algorithm, and use the augmented data generated by GAN and the original data together to train the fault diagnosis model, enhance the sensitivity of the model to fault features, and improve the diagnostic accuracy.

[0121] The steps of configuring the model according to the optimal parameter combination obtained from the optimization algorithm include:

[0122] Determine the fitness function and the model hyperparameters to be optimized. The cross-entropy loss of the validation set can be selected as the objective function. The optimized model parameters include the number of CNN layers, the attention dimension, the number of encoder layers, and the number of multi-head attention heads.

[0123] Data partitioning and initialization of the parameters of the sparrow optimization algorithm. The obtained dataset is partitioned into a training set and a validation set according to a certain ratio. The training set is used to train the model, and the validation set is used to evaluate the model performance under the optimized algorithm parameter combinations. Initialize the sparrow positions and number, the maximum number of iterations, the number of training epochs, and the upper and lower bounds. Each sparrow position represents a model parameter configuration, and the position dimension of the sparrow individual is the parameter dimension to be optimized in the model. In this case, the dataset is partitioned at a ratio of 8:2, and the values of the sparrow number, the maximum number of iterations, and the number of training epochs are set to 10, 15, and 15 respectively. The upper bound lb = [1, 32, 1, 2]; the lower bound ub = [3, 128, 3, 3].

[0124] Calculate the fitness of the current sparrow. Input the current position of each sparrow (i.e., the parameter configuration of the network model) into the network model, train and calculate the error of the model under this configuration as the fitness value. In each iteration, record the sparrow with the optimal fitness in the sparrow population as the current global optimal solution;

[0125] Update the current global optimal parameters. Compare the fitness of the current sparrow with the global optimal fitness. If the fitness of the current sparrow is better, update the global optimal solution and the corresponding parameters;

[0126] Update the sparrow positions. There are three roles in the sparrow algorithm, namely discoverers, followers, and vigilant ones. The new position of the discoverer is obtained by the following formula;

[0127]

[0128] where, represents the position of the i-th sparrow in the j-th dimension at the t-th iteration, α is a random number within (0, 1], ST represents the safety value, Q is a random number from a normal distribution, L is a 1×d unit vector, R2 represents the warning value, and iter max is the maximum number of iterations.

[0129] The new position of the follower is calculated by the following formula:

[0130]

[0131] where, represents the worst individual at the t-th iteration, X P represents the current optimal parameters, A + = A T (AA T ) -1 ;

[0132] The new position of the vigilant is calculated by the following formula:

[0133]

[0134] where β is a random number following a normal distribution with a mean of 0 and a variance of 1, represents the worst individual in the t-th iteration, K is a random number in [-1, 1], f i represents the fitness value of the current sparrow, f g and f w represent the current optimal fitness value and the current worst fitness value respectively.

[0135] Evaluation and update of the fitness function. After the position of the sparrow is updated, the fitness of the new position of each sparrow is re-evaluated. By training the network model, the loss function value is calculated and updated as the new fitness value;

[0136] Judge whether the iteration reaches the termination condition. If the condition is met, stop the iteration and output the optimal network model parameters. If the number of iterations has not exceeded the maximum number of iterations, continue to execute the next step;

[0137] Update the sparrow position information and the number of iterations. After each iteration, the optimal solution and fitness of the sparrow population are updated; the number of iterations is increased by 1, and then go to the step of calculating the fitness of the current sparrow. Figure 4 This is a schematic diagram of the optimal parameters of the algorithm iteration for this case.

[0138] In this case, the fitness value converges to the lowest at the 9th iteration. At this time, the fitness value is 0.1294, and the corresponding hyperparameter combination is: the number of CNN layers is 3, the attention dimension is 78, the number of encoder layers is 2, and the number of multi-head attention heads is 2.

[0139] The model configured according to the optimal parameter combination obtained by the sparrow search algorithm is: the CNN module has three convolutional layers, the number of encoder layers is 2, the number of attention heads is set to 2, the attention dimension is set to 78, and the augmented data generated by the GAN and the original data are used together to train the fault diagnosis model. The cross-entropy loss is selected as the loss function, and the formula is as follows:

[0140]

[0141] where m is the number of samples, p(x i ) is the true probability distribution of the samples, and q(x i ) is the probability distribution predicted by the model.

[0142] Based on the above embodiments, the hyperparameters of the fault diagnosis model in this embodiment include the number of CNN layers, the attention dimension, the number of Transformer encoder layers, and the number of multi-head attention heads.

[0143] Based on the above embodiments, in this embodiment, the trained fault diagnosis model is evaluated on the validation set to obtain the fitness value corresponding to the current hyperparameters, including:

[0144] Taking the cross-entropy loss of the validation set as the fitness value corresponding to the current hyperparameters.

[0145] The effect diagram of the training of the CNN and Transformer hybrid model based on the sparrow search algorithm constructed in this embodiment is as Figure 5 shown. Experimental verification confirms that this method has high accuracy, reaching 99.2%, and a short diagnosis time. The recognition speed of each sample is 0.35 s on average, proving that this method is practical and feasible. Table 2 shows the recognition results of some sample models.

[0146] Table 2 Some samples

[0147]

[0148]

[0149] The expected benefits and commercial values after the transformation of the technical solution of the present invention are:

[0150] 1. Improve the efficiency and real-time performance of synchronous motor fault diagnosis: This method extracts fault features based on the d-q transformation of the stator current. Compared with means such as Hilbert series transformation, wavelet transformation, and EMD (Empirical Mode Decomposition), as a linear transformation, the d-q transformation will not change the key information in the original current signal and can more directly and clearly reflect the current characteristics under the motor operating state. At the same time, the calculation process of the d-q transformation is relatively simple. This computational simplicity shortens the time required for fault diagnosis, enabling this method to improve the diagnosis efficiency while ensuring the diagnosis accuracy and meeting the industrial application scenarios with high real-time requirements.

[0151] 2. Enhanced the robustness of the motor fault diagnosis model: By using GAN for data augmentation, it ensures that the model can learn features from a large amount of data, which helps to reduce the overfitting phenomenon of the model to specific fault data. In addition, factors such as noise and sensor errors during the operation of the motor may lead to the instability of fault data. By generating fault samples with noise and variations through GAN, the model can learn how to extract useful fault information from these interferences during the training process. In practical applications, even when facing inputs with high noise or low data quality, the model can maintain a high diagnostic accuracy.

[0152] 3. Market competition advantage: Enterprises adopting the method for identifying the broken rotor bar fault state of synchronous motors based on the characteristic spectrum of stator current will gain obvious technical advantages in the market. Timely and accurate fault diagnosis not only improves the quality and reliability of motor products but also enables enterprises to lead their peers in product upgrading and after-sales service, attracting more customers and expanding the market share.

[0153] The technical solution of the present invention fills the domestic and foreign industry gaps: From the perspective of existing technical literature, although there are some synchronous motor fault diagnosis methods, most of them are limited to traditional signal processing techniques or are based on small-sample learning, which to a certain extent limits the diagnostic accuracy and generalization ability. The present invention combines various technical means such as d-q transformation of stator current, GAN data augmentation, and deep learning to form a comprehensive fault diagnosis system. This method not only has fast diagnosis and high fault identification accuracy but also has strong system robustness, effectively coping with complex and changeable motor fault conditions. Moreover, this method is applicable to various fault types of different motors, and for each specific motor, this method can automatically construct the most optimized model structure that best matches its own characteristics.

[0154] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying a broken rotor bar fault state of a synchronous motor based on stator current, characterized in that: include: Acquire stator current sample data of the first synchronous motor when a rotor bar failure occurs, wherein the stator current sample data is acquired by current sensor collection and generated by a generative adversarial network; After performing dq transformation on the stator current sample data, Fourier decomposition is performed to obtain characteristic data; According to the characteristic data of the stator current sample data and the corresponding actual rotor bar broken fault category, a fault diagnosis model based on CNN and Transformer is constructed, and the hyperparameters of the fault diagnosis model are optimized using an optimization algorithm; The current stator current data of the second synchronous motor is subjected to dq transformation and then Fourier decomposition to obtain characteristic data, and the characteristic data of the current stator current data is input into the fault diagnosis model to obtain the rotor broken bar fault category output by the fault diagnosis model.

2. The method for identifying a synchronous motor rotor broken bar fault state based on stator current according to claim 1, characterized in that: Obtaining stator current sample data of the first synchronous motor when a rotor bar failure occurs, including: The stator current sample data of the first synchronous motor when the rotor broken bar fault categories are 1 broken bar, 2 broken bars and 3 broken bars are respectively obtained.

3. The method for identifying a synchronous motor rotor broken bar fault state based on stator current according to claim 1, characterized in that: After the stator current sample data is subjected to dq transformation, Fourier decomposition is performed to obtain characteristic data, including: Performing two phase shifts on the stator current sample data i1 of each frequency component, the stator current sample data after the two phase shifts are i2 and i3 respectively; For I = [i1,i2,i3] T Perform dq transformation to obtain the d-axis component i d and the q-axis component i q ; Extract the d-axis component i by Fourier decomposition d The DC component i da and the q-axis component i q The DC component i qa ; Calculate the DC component i da and i qa The square sum is calculated to obtain the amplitude of each frequency component as the characteristic data.

4. The method for identifying a synchronous motor rotor broken bar fault state based on stator current according to claim 3, characterized in that: The first synchronous motor generates current harmonic components when a rotor bar fails, and the formula of the stator current sample data is: Where n is a positive integer; A1, ω1 and θ1 are the fundamental amplitude, angular frequency and initial phase angle of the stator current sample data respectively; A2,…,A n is the amplitude of each harmonic component of the stator current sample data; ω2,…,ω n is the angular frequency of each harmonic component of the stator current sample data; θ2,…,θ n is the initial phase angle of each harmonic component of the stator current sample data.

5. The method for identifying a broken rotor bar fault state of a synchronous motor based on stator current according to claim 4, characterized in that: After the stator current sample data is subjected to dq transformation, Fourier decomposition is performed to obtain characteristic data, including: The stator current sample data i1 of the frequency component ω2 is shifted 120° to the left in the time domain to obtain the current i2, and is shifted 240° to the left to obtain the current i3. i2 and i3 are expressed as: Where θ′ and θ″ are the phase differences of the fundamental components in i2 and i3 relative to i1; θ′ n and θ″ n are the phase differences of each harmonic component in i2 and i3 relative to i1; For I = [i1,i2,i3] T The formula for dq transformation is as follows, P 3 / 2 is the dq transformation matrix: Extract the d-axis component i by Fourier decomposition d The DC component i da and the q-axis component i q The DC component i qa The formula is as follows: The DC component i is calculated by the following formula da and i qa The square sum is used to obtain the amplitude of the frequency component ω2 as the characteristic data:

6. The method for identifying a synchronous motor rotor broken bar fault state based on stator current according to any one of claims 1 to 5, characterized in that: The generative adversarial network includes a generator and a discriminator; The generator is used to generate new data according to the input random noise; The discriminator is used to determine the possibility that the new data belongs to real data; The structure of the generator includes a fully connected layer, an activation function LeakyReLU, a first deconvolution layer, a second deconvolution layer, and a third deconvolution layer connected in sequence; The structure of the discriminator includes a first convolutional layer, a second convolutional layer, a Dropout layer and a fully connected layer which are connected in sequence.

7. The method for identifying a broken rotor bar fault state of a synchronous motor based on stator current according to any one of claims 1 to 5, characterized in that: The fault diagnosis model based on CNN and Transformer includes CNN, Transformer encoder and classifier connected in sequence; The CNN is used to extract local features of feature data of the current stator current data; The Transformer encoder is used to capture the global dependency of the feature data of the current stator current data; The classifier is used to output a broken rotor bar fault category of the second synchronous motor.

8. The method for identifying a synchronous motor rotor broken bar fault state based on stator current according to any one of claims 1 to 5, characterized in that: According to the characteristic data of the stator current sample data and the corresponding actual rotor broken bar fault category, a fault diagnosis model based on CNN and Transformer is constructed, and the hyperparameters of the fault diagnosis model are optimized using an optimization algorithm, including: Dividing the characteristic data of the stator current sample data and the corresponding actual rotor broken bar fault category as samples into a training set and a validation set; Under the current hyperparameters of the fault diagnosis model, the fault diagnosis model is trained using the training set, and the trained fault diagnosis model is evaluated on the validation set to obtain a fitness value corresponding to the current hyperparameters; If the fitness value corresponding to the current hyperparameter is better than the global optimal fitness value, the global optimal fitness value is updated to the fitness value corresponding to the current hyperparameter, and the global optimal parameter corresponding to the global optimal fitness value is updated to the current hyperparameter; After the current hyperparameters of the fault diagnosis model are updated using the optimization algorithm, the fault diagnosis model is continuously trained, evaluated, and globally optimal parameters are updated until a maximum number of iterations is reached, and the globally optimal parameters are output; The model obtained by training the fault diagnosis model using the training set under the global optimal parameters is used as the final fault diagnosis model.

9. The method for identifying a synchronous motor rotor broken bar fault state based on stator current according to claim 8, characterized in that: The hyperparameters of the fault diagnosis model include the number of CNN layers, attention dimension, the number of Transformer encoder layers, and the number of multi-head attention heads.

10. The method for identifying a synchronous motor rotor broken bar fault state based on stator current according to claim 8, characterized in that: Evaluating the trained fault diagnosis model on the validation set to obtain a fitness value corresponding to the current hyperparameter includes: The cross entropy loss of the validation set is used as the fitness value corresponding to the current hyperparameter.

Images