A current sensor fault diagnosis method based on a multi-channel convolutional neural network

Abstract

The application discloses a current sensor fault diagnosis method based on a multi-channel convolutional neural network, and specifically comprises the following steps: a permanent magnet motor mathematical model and mathematical models under four typical fault modes of a current sensor are established; original three-phase current data under normal mode and the four typical fault modes are obtained to form a data set; the data set is subjected to normalization processing and random overlapping sampling, and new data samples are divided into a training set, a cross-validation set and a test set; a three-channel global pooling one-dimensional convolutional neural network is constructed, three-phase current data are taken as an input layer, prediction classification of input data is taken as output, a cross-entropy cost function is taken as an evaluation index to adjust a network structure and hyperparameters, and an optimal network is selected according to the index; data under different motor speeds and load working conditions are obtained and put into the trained model for testing, and high-precision fault diagnosis is realized. The application realizes efficient and accurate diagnosis of the four typical current sensor faults under different working conditions of the permanent magnet motor.

Description

Technical Field

[0001] This invention belongs to the field of current sensor fault diagnosis, specifically relating to a current sensor fault diagnosis method based on a multi-channel convolutional neural network. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs) have advantages such as simple structure, high power density, and high efficiency, and are widely used in AC speed control drives. In various control strategies of PMSMs, such as field-oriented control (FOC), model predictive control (MPC), and direct torque control (DTC), current sensors provide feedback signals to the closed-loop control system. However, current sensors are susceptible to the influence of external environment and electrothermal stress, leading to signal loss or deviation, affecting motor control performance, and even causing system shutdown. Furthermore, different types of current sensor failures directly impact the efficiency of subsequent fault tolerance and maintenance. Therefore, to ensure the stable and safe operation of the motor drive system and provide a basis for fault-tolerant control and maintenance decisions, current sensor fault detection and diagnosis (FDD) has become a major research focus.

[0003] Currently, mainstream fault diagnosis methods for current sensors can be broadly categorized into model-based methods, signal-based methods, and data-driven methods. Model-based methods focus on establishing an equivalent mathematical model of the drive system to generate predictive signals and detect faults by comparing residuals with predefined diagnostic thresholds. However, this method is highly dependent on model parameters and has poor model adaptability. As high-speed train traction drive systems become increasingly complex, it is difficult to establish accurate mathematical models. Signal-based methods do not require mathematical models; instead, they process the measured signals, extract features that can identify faults, and then make diagnostic decisions based on prior knowledge. This method is typically only effective for specific fault types and relies on prior knowledge, resulting in poor robustness. Therefore, there is a need to develop a simpler, more accurate, and more universal fault detection and diagnosis algorithm.

[0004] With the development of artificial intelligence, Convolutional Neural Networks (CNNs), as one of the representatives of deep learning, have been widely used in fields such as natural language processing, image recognition, and biotechnology. They have powerful end-to-end feature extraction capabilities, good automatic learning capabilities, and plasticity. CNNs have great application advantages in fault diagnosis of current sensors due to their superior performance. Summary of the Invention

[0005] To address the above shortcomings, this invention provides a current sensor fault diagnosis method based on a multi-channel convolutional neural network.

[0006] The present invention provides a current sensor fault diagnosis method based on a multi-channel convolutional neural network, comprising the following steps:

[0007] Step 1: Establish mathematical models for the permanent magnet motor and four typical fault modes: current sensor disconnection, jamming, gain, and zero offset. Build a model of a permanent magnet motor drive system powered by a three-level inverter based on vector control. Obtain raw three-phase current data for normal mode and the four typical fault modes through MATLAB / Simulink simulation. a i b i c Next, the data is labeled to form a dataset, which is then used for subsequent network training and testing.

[0008] Step 2: Preprocess the dataset. First, normalize the data within the range of [-1, 1]. Then, use random overlapping sampling to augment the data and generate a new data sample. Divide the new data sample into training set, cross-validation set, and test set according to the proportion, which are used for network training, network parameter design, and network performance evaluation, respectively.

[0009] Step 3: Construct a three-channel global pooling one-dimensional convolutional neural network. Take the labeled three-phase current data as input and the probability of correctly predicting the classification of each fault category as output. Train the network based on the dataset and the Adam algorithm. Adjust the network structure and hyperparameters using the cross-entropy cost function as the evaluation index, and select the network with the optimal index.

[0010] Step 4: Obtain data under different motor speeds and load conditions on the experimental platform, process it according to Step 2, put the divided dataset into the trained model for testing, realize high-precision fault diagnosis under different operating conditions, and verify the accuracy and robustness under different operating conditions.

[0011] Furthermore, the establishment of the mathematical model of the permanent magnet motor and the mathematical models of the four typical faults in step 1 is specifically as follows:

[0012] Taking a permanent magnet motor fed by a three-level active midpoint clamped inverter as the controlled object, a mathematical model of the permanent magnet motor in a synchronous rotating reference coordinate system (dq) is established, which is expressed as:

[0013]

[0014]

[0015] Among them, u d and u q i represents the dq-axis stator voltage component. d and i q R is the dq-axis stator current component. s For the stator resistance, ω e L is the rotor electric angular velocity.d and L q For the dq axis stator inductance, It is a permanent magnet flux linkage.

[0016] Decoupling the mathematical model of the permanent magnet motor, its electromagnetic torque equation is expressed as:

[0017]

[0018] Where, p n This represents the number of pole pairs of the motor.

[0019] Based on the output characteristics of the current sensor feedback signal, sensor faults are classified into four typical modes: open circuit fault, jamming fault, gain fault, and offset fault.

[0020] Taking a phase a current sensor failure as an example, under normal operating conditions, the current sensor outputs current i. a Represented as:

[0021] i a =I m cos(ω e t+θ0)

[0022] The current sensor output current signal under four typical fault modes is represented as follows:

[0023] (1) Current sensor open circuit fault model:

[0024]

[0025] (2) Current sensor jamming fault model:

[0026]

[0027] (3) Current sensor gain fault model:

[0028]

[0029] (4) Current sensor offset fault model:

[0030]

[0031] Where C1, C2, and C3 are assumed to be constants, C1 is the output of the jamming fault, C2 is the gain factor, C3 is the offset, t0 is the time when the fault occurs, and I m ω represents the amplitude of the stator current. e θ is the rotor electric angular velocity, and θ0 is the initial angle of the A-phase current.

[0032] A model of a permanent magnet motor drive system powered by a three-level inverter based on vector control was constructed. Raw output current data for the three-phase current sensor in normal mode and four typical fault modes were obtained through MATLAB / Simulink simulation. a i b i c They are then labeled and used to form a dataset for subsequent network training and testing.

[0033] Furthermore, step 2 involves preprocessing the dataset and proportionally dividing it for network training and performance evaluation, specifically as follows:

[0034] When the load size or motor speed varies, the amplitude of the output current may not be in the same unit. To improve generalization ability, the original fault data is normalized to the range of [-1, 1] to eliminate the influence of the unit on network training. The mapminmax normalization function is defined as:

[0035]

[0036] Among them, y min and y max This represents the normalized interval of the expectation, i.e., y. min =-1, y max =1, x is the input data, x min and x max Let represent the minimum and maximum values ​​in the input data, respectively, and y be the normalized individual value.

[0037] Next, a random overlap sampling method is used for data augmentation, where L is the total length of the original data, N is the window size (i.e., the length of the new samples), and S is the step size of the sliding window. Each time the window slides through the original long-term data series, a new dataset of fault samples is generated. The total number of new samples is represented as:

[0038]

[0039] The fix function represents rounding down to zero.

[0040] The processed dataset is used as a training and validation set with 80% of the total number of samples, and a test set with the remaining 20% ​​for subsequent network training and performance evaluation.

[0041] Furthermore, the three-channel global pooling one-dimensional convolutional neural network constructed in step 3 is specifically as follows:

[0042] It includes an input layer, three convolutional modules, a global max pooling layer, and an output layer; the convolutional module consists of a convolutional layer, an activation function layer, a batch normalization layer, and a pooling layer.

[0043] The information transfer formula for convolutional layers is expressed as:

[0044]

[0045] Where * represents the convolution operation; k represents the index of the layer; and i represents the index of the kernel; X represents the output of the i-th feature in the k-th layer; k-1 This represents the output of the (k-1)th layer; and These are the weights and biases of the i-th convolutional kernel in the k-th layer, respectively.

[0046] The information transfer formula for the ReLU activation function layer is expressed as:

[0047]

[0048] The information transfer formula for the batch normalization layer is expressed as:

[0049]

[0050]

[0051]

[0052] Where, μ B and Here, represents the mean and standard deviation of the input batch samples, respectively; m is the training batch size; and x... i For the input sample, These are normalized samples, where ε is a very small constant to avoid the denominator being zero, and y i The reconstructed data for the output equations, γ and β are learnable parameters.

[0053] The pooling layer uses max pooling, and its information transfer formula is expressed as:

[0054]

[0055] in, This is the output value after the max pooling operation. For the i-th output feature map of the k-th convolutional layer, t x The value of a pixel.

[0056] The information transfer formula for the global max pooling layer is expressed as:

[0057]

[0058] in, It is the i-th feature map output from the k-th layer. It is the overall feature map of the i-th channel in the (k-1)-th layer.

[0059] The three-phase current signal i a i b i c Three channels are input separately as the input layer of the convolutional neural network; the probability of the input data being classified into different fault types is the output layer of the network. Backpropagation training is performed based on the collected dataset and combined with the Adam algorithm. The cross-entropy cost function is used as the evaluation index. The convolutional structure and hyperparameters are adjusted according to the training results to select the optimal network.

[0060] Furthermore, step 4 verifies the accuracy and robustness under different operating conditions as follows:

[0061] Different rotor speeds N were obtained on the experimental platform. ref and load torque T e The data under the operating conditions is processed according to step 2, and then the divided dataset is put into the trained network model for testing to achieve high-precision fault diagnosis of current sensors under different operating conditions and verify the accuracy and robustness under different operating conditions.

[0062] The beneficial technical effects of this invention are as follows:

[0063] (1) The diagnostic method proposed in this invention uses MATLAB software to construct a three-channel global pooling one-dimensional convolutional neural network. The network is trained by forward propagation and Adam back propagation. Compared with the traditional CNN, this invention uses three-channel input and replaces the traditional fully connected layer with a global max pooling layer, which can achieve high-precision diagnosis of four typical faults of current sensors. At the same time, the number of network parameters is reduced by more than 60%, which greatly improves the diagnostic speed and has good online potential.

[0064] (2) The present invention performs normalization processing and random overlapping sampling on the collected data. Normalization processing eliminates the influence of the output current amplitude under different operating conditions, accelerates the convergence speed of the training network, and avoids the generation of singular samples that affect the training of the CNN model; random overlapping sampling is used to enhance the dataset, which can improve the robustness of the network and avoid overfitting.

[0065] (3) This invention is a data-driven method that relies on fault sample design and network training for fault location and identification. It does not require complex mathematical models of the drive system or prior knowledge, thus simplifying diagnostic complexity and exhibiting good robustness to data differences. This invention is universal; even if the historical database covers fault data from other control strategies or different motors, the method of this invention can still effectively diagnose faults.

[0066] (4) This invention obtains several sets of different rotor speeds N through experiments. ref and load torque T eThe data under the working conditions is processed, and then the divided dataset is put into the trained network model for testing. It is compared with four intelligent algorithms, BP, RBF, LSTM and SVM, to verify the superior accuracy and generalization ability of the method under complex operating conditions. Attached Figure Description

[0067] Figure 1 This is a flowchart of the current sensor fault diagnosis method based on a multi-channel convolutional neural network according to the present invention.

[0068] Figure 2 This is a block diagram of a permanent magnet motor drive system with a three-level inverter.

[0069] Figure 3 This is a data processing flowchart of the present invention.

[0070] Figure 4 This is a structural diagram of the convolutional neural network model of the present invention.

[0071] Figure 5 This is a training progress diagram of the three-channel global pooling one-dimensional convolutional neural network of the present invention.

[0072] Figure 6 This is a diagnostic result diagram of the three-channel global pooling one-dimensional convolutional neural network of the present invention.

[0073] Figure 7 This is a performance comparison chart of the three-channel global pooling one-dimensional convolutional neural network of the present invention. Detailed Implementation

[0074] The present invention will be further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0075] This invention focuses on a permanent magnet motor drive system powered by a three-level active neutral-point clamped (ANPC) inverter. It provides a current sensor fault diagnosis method based on a multi-channel convolutional neural network. Based on typical operating conditions observed in experiments, a database of the three-phase current output from the current sensor under normal and fault modes is established. The feature extraction capability of the convolutional neural network is used to classify the input data. The training of the convolutional neural network is implemented offline to avoid additional computational burden. This method enables the location and identification of four typical current sensor faults under different operating conditions. The process is as follows: Figure 1 As shown, specifically:

[0076] Step 1: Establish mathematical models for the permanent magnet motor and four typical fault modes: current sensor disconnection, jamming, gain, and zero offset. Build a model of a permanent magnet motor drive system powered by a three-level inverter based on vector control. Obtain raw three-phase current data for normal mode and the four typical fault modes through MATLAB / Simulink simulation.a i b i c Next, the data is labeled to form a dataset, which is then used for subsequent network training and testing.

[0077] Taking a permanent magnet motor fed by a three-level active midpoint clamped inverter as the controlled object, a mathematical model of the permanent magnet motor in a synchronous rotating reference coordinate system (dq) is established, which is expressed as:

[0078]

[0079]

[0080] Among them, u d and u q i represents the dq-axis stator voltage component. d and i q R is the dq-axis stator current component. s For the stator resistance, ω e L is the rotor electric angular velocity. d and L q For the dq axis stator inductance, It is a permanent magnet flux linkage.

[0081] Decoupling the mathematical model of the permanent magnet motor, its electromagnetic torque equation is expressed as:

[0082]

[0083] Where, p n This represents the number of pole pairs of the motor.

[0084] Based on the output characteristics of the current sensor feedback signal, sensor faults are classified into four typical modes: open circuit fault, jamming fault, gain fault, and offset fault.

[0085] Taking a phase a current sensor failure as an example, under normal operating conditions, the current sensor outputs current i. a Represented as:

[0086] i a =I m cos(ω e t+θ0)

[0087] The current sensor output current signal under four typical fault modes is represented as follows:

[0088] (1) Current sensor open circuit fault model:

[0089]

[0090] (2) Current sensor jamming fault model:

[0091]

[0092] (3) Current sensor gain fault model:

[0093]

[0094] (4) Current sensor offset fault model:

[0095]

[0096] Where C1, C2, and C3 are assumed to be constants, C1 is the output of the jamming fault, C2 is the gain factor, C3 is the offset, t0 is the time when the fault occurs, and I m ω represents the amplitude of the stator current. e θ is the rotor electric angular velocity, and θ0 is the initial angle of the A-phase current.

[0097] Figure 2 The diagram shows a block diagram of a permanent magnet motor drive system using a vector-controlled three-level neutral-point clamping inverter. Raw output current data (i) for the three-phase current sensor in normal mode and four typical fault modes were obtained through MATLAB / Simulink simulation. a i b i c They are then labeled and used to form a dataset for subsequent network training and testing.

[0098] Step 2: Preprocess the dataset. First, normalize the data within the range of [-1, 1]. Then, use random overlapping sampling to augment the data and generate a new data sample. Divide the new data sample into training set, cross-validation set, and test set according to the proportion, which are used for network training, network parameter design, and network performance evaluation, respectively.

[0099] Because the amplitude of the output current may not be in the same unit of measurement when the load size or motor speed varies, in order to eliminate the interference of different units and amplitudes, avoid generating singular samples that affect network training, and thus improve the training speed of the convolutional network, the original fault data is normalized to the range of [-1,1] to eliminate the influence of the unit of measurement on network training. Figure 3 (a) shows the flowchart of data normalization processing, where the mapminmax normalization function is defined as:

[0100]

[0101] Among them, y min and y max This represents the normalized interval of the expectation, i.e., y. min =-1, y max =1, x is the input data, xmin and x max Let represent the minimum and maximum values ​​in the input data, respectively, and y be the normalized individual value.

[0102] Next, a random overlap sampling method is used to augment the data, improving the network's robustness and preventing overfitting. For example... Figure 3 (b) shows the flowchart of random overlap sampling, where L is the total length of the original data, N is the window size (i.e., the length of the new sample), and S is the step size of the sliding window. Each time the window slides through the original long-term data series, a new dataset of fault samples is generated. The total number of new samples is represented as:

[0103]

[0104] The fix function represents rounding down to zero.

[0105] The processed dataset is used as a training and validation set with 80% of the total number of samples, and a test set with the remaining 20% ​​for subsequent network training and performance evaluation.

[0106] After normalizing, randomly overlapping, and dividing the data, this invention will construct a three-channel global pooling convolutional neural network based on the dataset using MATLAB simulation software.

[0107] Step 3: Construct a three-channel global pooling one-dimensional convolutional neural network. Take the labeled three-phase current data as input and the probability of correctly predicting the classification of each fault category as output. Train the network based on the dataset and the Adam algorithm. Adjust the network structure and hyperparameters using the cross-entropy cost function as the evaluation index, and select the network with the optimal index.

[0108] The constructed three-channel global pooling one-dimensional convolutional neural network is as follows: Figure 4 As shown, it includes an input layer, three convolutional modules, a global max pooling layer, and an output layer; the convolutional module consists of a convolutional layer, an activation function layer, a batch normalization layer, and a pooling layer.

[0109] The information transfer formula for convolutional layers is expressed as:

[0110]

[0111] Where * represents the convolution operation; k represents the index of the layer; and i represents the index of the kernel; X represents the output of the i-th feature in the k-th layer; k-1 This represents the output of the (k-1)th layer; and These are the weights and biases of the i-th convolutional kernel in the k-th layer, respectively.

[0112] The information transfer formula for the ReLU activation function layer is expressed as:

[0113]

[0114] The information transfer formula for the batch normalization layer is expressed as:

[0115]

[0116]

[0117]

[0118] Where, μ B and Here, represents the mean and standard deviation of the input batch samples, respectively; m is the training batch size; and x... i For the input sample, These are normalized samples, where ε is a very small constant to avoid the denominator being zero, and y i The reconstructed data for the output equations, γ and β are learnable parameters.

[0119] The pooling layer uses max pooling, and its information transfer formula is expressed as:

[0120]

[0121] in, This is the output value after the max pooling operation. For the i-th output feature map of the k-th convolutional layer, t x The value of a pixel.

[0122] The information transfer formula for the global max pooling layer is expressed as:

[0123]

[0124] in, It is the i-th feature map output from the k-th layer. It is the overall feature map of the i-th channel in the (k-1)-th layer.

[0125] The three-phase current output by the current sensor is used as the input layer of the convolutional neural network. The input data is processed by three convolutional modules to extract features, and then fed into the softmax layer through a global max pooling layer. Finally, the probability of the input data being classified into different fault types is used as the output layer of the network.

[0126] Compared to traditional convolutional networks, two-dimensional three-phase current data will be processed according to i a i b i cBy dividing the input into three channels and converting it into a one-dimensional convolution, the computational load of the network can be reduced. Furthermore, by replacing the traditional fully connected layer with a global max pooling layer, it can not only achieve high-precision diagnosis of four typical current sensor faults, but also reduce the number of network parameters by more than 60%, greatly improving the network's diagnostic speed and demonstrating good online potential.

[0127] Convolutional neural networks are trained using backpropagation based on a collected dataset and the Adam algorithm. This process automatically updates the network parameters, and the cross-entropy cost function is used as an evaluation metric to adjust the network's hyperparameters. The cross-entropy cost function is expressed as:

[0128]

[0129] Where J(w,b) is the cost function, L m Let M be the loss function, K be the number of samples, and I be the number of fault categories. m,k Let I be a logical indicator function; if the predicted class of the m-th sample is true, then... m,k =1, if false, then I m,k =0. P m,k It is the probability that sample m is assigned to class k.

[0130] To avoid the network getting stuck in local optima, the convolutional network is trained multiple times with random initial parameters. The convolutional structure and hyperparameters are adjusted based on the training results to select the optimal network. Figure 5 The training progress diagram of the three-channel one-dimensional convolutional network shows that as the number of network iterations increases, the network's evaluation metrics improve, namely, the cross-entropy loss is smaller and the prediction accuracy is higher.

[0131] Using MATLAB / Simulink simulation software, we collected three-phase current signals of 13 types, including normal conditions and four typical faults from three current sensors, to build a database. This database served as the input to a convolutional network, and the prediction and classification were used as the network's output. Figure 6 The image shows the diagnostic results of a three-channel global pooling one-dimensional convolutional neural network. The confusion matrix shows that only two samples were misclassified, with a diagnostic accuracy of 99.87%. This indicates that the trained network can accurately classify faults.

[0132] Step 4: Obtain data under different motor speeds and load conditions on the experimental platform, process it according to Step 2, put the divided dataset into the trained model for testing, realize high-precision fault diagnosis under different operating conditions, and verify the accuracy and robustness under different operating conditions.

[0133] The experiment was conducted at a load torque (T) e The values ​​are 5 N·m and 10 N·m, with a reference speed (N). refThe tests were conducted at 200 rpm and 300 rpm, collecting multiple sets of fault test data. The data were then processed and divided according to steps 2 and 3. The experimental variable T was controlled as follows: e and N ref The network can be trained using a combination of different operating conditions. For example, data with a load torque of 5 N·m and a reference speed of 200 rpm can be used as a training set to train the network, and data with a load torque of 5 N·m and a reference speed of 300 rpm can be used as a validation set to test the trained network model. This can verify the effect of speed variation on the diagnostic effect of the present invention. Similarly, the effect of load torque variation on the diagnostic effect can also be verified. Figure 7 The figure shows a comparison of the diagnostic performance of the method of this invention with four intelligent algorithms: BP, RBF, LSTM, and SVM. The figure shows that the classification accuracy of the method of this invention is the highest under any working condition. This indicates that the trained convolutional neural network can accurately diagnose the fault of the current sensor and identify the fault type under the conditions of variable load torque and variable speed. This further illustrates the accuracy and robustness of the method, and it is more suitable for the complex application conditions of high-speed train traction systems.

Claims

1. A fault diagnosis method for current sensors based on multi-channel convolutional neural networks, characterized in that, Includes the following steps: Step 1: Establish mathematical models for the permanent magnet motor and four typical fault modes for the current sensor: open circuit, jamming, gain, and zero offset; build a model of the permanent magnet motor drive system powered by a three-level inverter based on vector control; and obtain the original three-phase current data i under normal mode and the four typical fault modes through MATLAB / Simulink simulation. a i b i c Next, the data is labeled to form a dataset for subsequent network training and testing; Step 2: Preprocess the dataset. First, normalize the data within the range of [-1, 1]. Then, use random overlapping sampling to augment the data and generate a new data sample. Divide the new data sample into training set, cross-validation set, and test set according to the proportions, which are used for network training, network parameter design, and network performance evaluation, respectively. Step 3: Construct a three-channel global pooling one-dimensional convolutional neural network. Take the labeled three-phase current data as input and the probability of correctly predicting the classification of each fault category as output. Train the network based on the dataset and the Adam algorithm. Use the cross-entropy cost function as the evaluation index to adjust the network structure and hyperparameters, and select the network with the optimal index. Step 4: Obtain data under different motor speeds and load conditions on the experimental platform, process it according to Step 2, put the divided dataset into the trained model for testing, realize high-precision fault diagnosis under different operating conditions, and verify the accuracy and robustness under different operating conditions.

2. The current sensor fault diagnosis method based on a multi-channel convolutional neural network according to claim 1, characterized in that, The establishment of the mathematical model of the permanent magnet motor and the mathematical models of the four typical faults in step 1 is specifically as follows: Taking a permanent magnet motor fed by a three-level active midpoint clamped inverter as the controlled object, a mathematical model of the permanent magnet motor in a synchronous rotating reference coordinate system is established, which is expressed as: Among them, u d and u q i represents the dq-axis stator voltage component. d and i q R is the dq-axis stator current component. s For the stator resistance, ω e L is the rotor electric angular velocity. d and L q For the dq axis stator inductance, For permanent magnet flux linkage; Decoupling the mathematical model of the permanent magnet motor, its electromagnetic torque equation is expressed as: Where, p n This represents the number of pole pairs of the motor. Based on the output characteristics of the current sensor feedback signal, sensor faults are classified into four typical modes: open circuit fault, jamming fault, gain fault, and offset fault. For phase a current sensor, under normal operating conditions, the current sensor outputs current i. a Represented as: I a =I m cos(ω e t+θ0) The current sensor output current signal under four typical fault modes is represented as follows: (1) Current sensor open circuit fault model: (2) Current sensor jamming fault model: (3) Current sensor gain fault model: (4) Current sensor offset fault model: Where C1, C2, and C3 are assumed to be constants, C1 is the output of the jamming fault, C2 is the gain factor, C3 is the offset, t0 is the time when the fault occurs, and I m ω represents the amplitude of the stator current. e Let θ be the rotor electric angular velocity, and θ0 be the initial angle of the A-phase current; The same principle applies to the b-phase current sensor and the c-phase current sensor; A model of a permanent magnet motor drive system powered by a three-level inverter based on vector control was constructed. Raw output current data for the three-phase current sensor in normal mode and four typical fault modes were obtained through MATLAB / Simulink simulation. a i b i c They are then labeled and used to form a dataset for subsequent network training and testing.

3. The current sensor fault diagnosis method based on a multi-channel convolutional neural network according to claim 1, characterized in that, Step 2 involves preprocessing the dataset and proportionally dividing it for network training and performance evaluation. The original fault data is normalized to the range [-1, 1] to eliminate the influence of units on network training. The mapminmax normalization function is defined as follows: Among them, y min and y max This represents the normalized interval of the expectation, i.e., y. min =-1, y max =1, x is the input data, x min and x max Let represent the minimum and maximum values ​​in the input data, respectively, and y be the normalized individual value; Next, a random overlap sampling method is used for data augmentation, where L is the total length of the original data, N is the window size (i.e., the length of the new samples), and S is the step size of the sliding window. Each time the window slides through the original long-term data series, a new dataset of fault samples is generated. The total number of new samples is represented as: The fix function represents rounding down to zero; The processed dataset is used as a training and validation set with 80% of the total number of samples, and a test set with the remaining 20% ​​for subsequent network training and performance evaluation.

4. The current sensor fault diagnosis method based on a multi-channel convolutional neural network according to claim 1, characterized in that, The three-channel global pooling one-dimensional convolutional neural network constructed in step 3 is specifically as follows: It includes an input layer, three convolutional modules, a global max pooling layer, and an output layer; the convolutional module consists of convolutional layers, activation function layers, batch normalization layers, and pooling layers; The information transfer formula for convolutional layers is expressed as: Where * represents the convolution operation; k represents the index of the layer; and i represents the index of the kernel; X represents the output of the i-th feature in the k-th layer; k-1 This represents the output of the (k-1)th layer; and These are the weights and biases of the i-th convolutional kernel in the k-th layer, respectively. The information transfer formula for the ReLU activation function layer is expressed as: The information transfer formula for the batch normalization layer is expressed as: Where, μ B and Here, represents the mean and standard deviation of the input batch samples, respectively; m is the training batch size; and x... i For the input sample, These are normalized samples, where ε is a very small constant to avoid the denominator being zero, and y i The reconstructed data for the output equation, γ and β are learnable parameters; The pooling layer uses max pooling, and its information transfer formula is expressed as: in, This is the output value after the max pooling operation. For the i-th output feature map of the k-th convolutional layer, t x The value of a pixel; The information transfer formula for the global max pooling layer is expressed as: in, It is the i-th feature map output from the k-th layer. It is the overall feature map of the i-th channel in the (k-1)-th layer; The three-phase current signal i a i b i c Three channels are input separately as the input layer of the convolutional neural network; the probability of the input data being classified into different fault types is the output layer of the network. Backpropagation training is performed based on the collected dataset and combined with the Adam algorithm. The cross-entropy cost function is used as the evaluation index. The convolutional structure and hyperparameters are adjusted according to the training results to select the optimal network.

5. The current sensor fault diagnosis method based on a multi-channel convolutional neural network according to claim 1, characterized in that, Step 4, verifying the accuracy and robustness under different operating conditions, specifically involves: Different rotor speeds N were obtained on the experimental platform. ref and load torque T e The data under the operating conditions is processed according to step 2, and then the divided dataset is put into the trained network model for testing to achieve high-precision fault diagnosis of current sensors under different operating conditions and verify the accuracy and robustness under different operating conditions.

Images