Traction converter multi-fault intelligent diagnosis method based on edge AI

Abstract

The invention discloses a traction converter multi-fault intelligent diagnosis method based on edge AI, and the method comprises the following steps: building a permanent magnet motor mathematical model and a three-level active neutral point clamped converter mathematical model, and carrying out the simulation and experiment of a permanent magnet motor control system fed by a three-level ANPC converter, acquiring a three-phase current signal, a midpoint voltage signal and a rotating speed signal to form an original data set; performing multi-source and multi-dimensional information fusion and normalization on simulation and experimental data, and dividing a training set, a cross validation set and a test set; constructing a two-dimensional convolutional neural network model, and performing offline training; and carrying out lightweight design and edge deployment on the trained neural network model to realize real-time diagnosis of multiple fault modes of the converter under different working conditions. According to the method, the edge calculation efficiency is improved through the lightweight design of the network model, intelligent diagnosis of multiple faults can be realized, and the method has relatively high diagnosis precision and diagnosis speed.

Description

Technical Field

[0001] The present invention belongs to the field of converter fault diagnosis, and in particular relates to an edge AI-based traction converter multi-fault intelligent diagnosis method. Background Art

[0002] Multilevel converters have the advantages of low device voltage stress, low current harmonic content, and high efficiency. They are widely used in medium-voltage and high-power power conversion fields, such as high-voltage direct current transmission and AC motor drive. Among the multilevel converter topologies, the three-level active neutral point clamped (ANPC) topology has broad application prospects in high-efficiency motor drive applications due to its uniform loss distribution and strong fault tolerance. However, with the increase in the number of power switch tubes, the risk of switch tube failure in the converter has also greatly increased. Switch tube failures are usually divided into short-circuit faults and open-circuit faults. Since short-circuit faults usually cause abnormal overcurrent in the system, they can be detected and isolated through hardware protection. In addition, the current and voltage distortion caused by open-circuit faults are not as obvious as short-circuit faults, which makes fault diagnosis more difficult. If the open-circuit fault cannot be detected and isolated in time, it will cause secondary damage to the drive system and higher maintenance costs. Therefore, it is urgent to design a fast and accurate fault diagnosis strategy for the open-circuit fault of the three-level ANPC converter to improve the stability and safety of the motor drive system.

[0003] At present, the diagnostic strategies for converter open circuit faults can be roughly divided into model-based, signal-based and data-driven methods. Model-based methods have high versatility, but their implementation depends on accurate system models, which may be limited by parameter uncertainty and poor scalability. Signal-based methods do not require accurate mathematical models, but are more sensitive to load changes and have a large computational burden. To address the shortcomings of the above methods, data-driven methods use a large amount of historical data to extract fault features, which can achieve fault diagnosis simply and efficiently.

[0004] With the development of machine learning, data-driven methods have been successfully applied to converter fault diagnosis, but most of the current diagnostic strategies are implemented offline. In order to achieve online fault diagnosis, the offline trained neural network can be deployed to edge devices for fast and accurate real-time reasoning and online diagnosis. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention provides a traction inverter multi-fault intelligent diagnosis method based on edge AI.

[0006] The present invention provides an edge AI-based traction converter multi-fault intelligent diagnosis method, comprising the following steps:

[0007] Step 1: Construction of dataset.

[0008] A mathematical model of the permanent magnet motor drive system fed by a three-level active neutral point clamped ANPC converter is established. In the MATLAB / Simulink environment, a simulation model of the permanent magnet motor control system fed by a three-level active neutral point clamped ANPC converter is built. Multiple fault modes such as single switch open circuit and multiple switch open circuit of the three-level ANPC converter are simulated. The original data sets consisting of three-phase current signals, midpoint voltage signals and speed signals in normal mode and fault mode are obtained and labeled for subsequent network training and testing.

[0009] Step 2: Preprocess the data.

[0010] Firstly, the collected data are fused with multi-source and multi-dimensional information. Specifically, seven types of signals, including three-phase current, dq-axis current, midpoint error and speed, are fused into a two-dimensional feature map to improve the efficiency of feature extraction of the model. Subsequently, the data are normalized in the range of [-1, 1] to eliminate the influence of working condition changes on the diagnostic performance and generate data samples. Then, the data samples are divided into training set, cross-validation set and test set according to a certain ratio, which are used for training, network parameter design and network performance evaluation respectively.

[0011] Step 3: Construct a two-dimensional convolutional neural network model and use the established data set for offline training; input the divided training set and cross-validation set into the network, train and evaluate the network according to the AdamW (Adam with decoupled weight decay, AdamW) training algorithm, use the cross entropy loss function as the evaluation indicator, select the optimal network parameters, and finally use the test set to verify the generalization ability of the trained network.

[0012] Step 4: Lightweight design and edge deployment.

[0013] The trained neural network model is converted into an ONNX model (Open Neural Network Exchange, ONNX). ONNX is an open file format for storing and exchanging trained machine learning models. It enables different artificial intelligence frameworks (such as PyTorch and TensorFlow) to share models, facilitating the migration and reuse of models between different platforms. TensorRT is used to perform a lightweight design on it; the lightweight network is then deployed to the edge device (GPU) to achieve real-time diagnosis of converter open-circuit faults under different working conditions, and diagnostic thresholds are designed to prevent misdiagnosis and improve diagnostic accuracy.

[0014] Furthermore, the multiple fault modes in step 1 include both single module open circuit faults and double module open circuit faults, so there are 9 single module open circuit faults and 36 (i.e. )Open circuit fault of two power modules.

[0015] Furthermore, in step 2, the training set, cross-validation set, and test set are 70%, 15%, and 15%, respectively.

[0016] Furthermore, the two-dimensional convolutional neural network in step 3 is composed of four convolutional modules, a flat layer and a fully connected layer; wherein, the framework of the convolutional module is composed of a convolutional layer, a batch normalization layer, an activation layer and a maximum pooling layer, wherein the convolutional layer extracts the features of the input layer, the batch normalization layer standardizes the output of the convolutional layer to improve the stability of the two-dimensional convolutional neural network, and the activation layer enables the convolutional layer to perform nonlinear feature extraction; the maximum pooling layer reduces the size of the feature map to eliminate redundant information; the flat layer and the fully connected layer convert the extracted features into classification results, and then pass these results to the output layer.

[0017] The beneficial technical effects of the present invention compared with the prior art are:

[0018] (1) The diagnostic method proposed in the present invention uses the Pytorch deep learning development tool to construct a two-dimensional convolutional neural network, and trains the network through forward propagation and AdamW back-propagation algorithms. Before edge deployment, the lightweight design of the two-dimensional convolutional neural network is completed using TensorRT, which significantly reduces the size of network parameters, and the computational requirements of the lightweight model are only 5% of the original training network, which improves computing efficiency, reduces edge deployment costs, and speeds up diagnosis.

[0019] (2) The present invention integrates the collected electrical and mechanical signals into a feature map dedicated to a two-dimensional convolutional neural network to achieve multi-source and multi-dimensional information fusion, thereby improving the effectiveness and comprehensiveness of fault feature extraction and further improving diagnostic accuracy.

[0020] (3) The present invention is a data-driven method, which is not limited by the mathematical models and prior knowledge in the traditional method. It uses the trained network and real-time operation data to locate and identify the fault. The method is simple and robust to changes in operating conditions. In addition, the present invention is universal. When the historical database covers other different converter fault data, the method of the present invention is still effective.

[0021] (4) The present invention deploys a lightweight two-dimensional convolutional network to edge devices to achieve real-time diagnosis of single and multiple switch open circuit faults in three-level ANPC converters. The feasibility of the proposed online diagnosis and misdiagnosis prevention strategy is fully verified through experiments. At the same time, the experimental results prove the effectiveness and superiority of this method in fault diagnosis under different working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flow chart of the multi-fault intelligent diagnosis method for traction converter of the present invention.

[0023] Figure 2 This is a topological structure diagram of a three-level active neutral point clamped converter according to the present invention.

[0024] Figure 3 It is a schematic diagram of the multi-source information fusion feature of the present invention.

[0025] Figure 4 It is a structural diagram of the two-dimensional convolutional neural network model of the present invention.

[0026] Figure 5 The figure is a comparison chart of the training performance of different convolutional neural network models of the present invention ((a) accuracy, (b) cross entropy loss).

[0027] Figure 6 The confusion matrix for verifying the generalization ability of the two-dimensional convolutional neural network of the present invention.

[0028] Figure 7 This is a TensorRT-based lightweight optimization strategy diagram of the present invention.

[0029] Figure 8 Figure 2 is the experimental results of multi-fault intelligent diagnosis under different working conditions of the present invention (a) a2 and M a3 Fault (torque 10 N·m, speed 400 rpm), (b) M a3 and M b2 Fault (speed 10N·m, torque 300rpm), (c) M a1 and M a2 Fault (speed 5N·m, torque 200rpm), (d) M b1 and M b2 Fault (speed 5N·m, torque 300rpm)). DETAILED DESCRIPTION

[0030] The present invention is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0031] The present invention takes the permanent magnet motor drive system fed by a three-level active neutral point clamped converter as the object, and provides a multi-fault intelligent diagnosis method for traction converters based on edge AI. According to simulations and experiments under different working conditions, a database of electrical and mechanical signals under normal and fault conditions is established, and the input data is classified using the powerful feature extraction capability of the two-dimensional convolutional neural network. The network is lightweight designed through TensorRT so that the network can be successfully deployed to the edge device (GPU) to realize the open circuit fault diagnosis of single switch tube and multiple switch tubes of the converter under different working conditions. The process is as follows Figure 1 As shown, specifically:

[0032] Step 1: Establish mathematical models of permanent magnet motor and three-level ANPC converter. Use MATLAB / Simulink to build a simulation model of permanent magnet motor drive system fed by three-level ANPC converter based on finite set model predictive control, analyze 45 fault modes of single switch tube fault and multi-switch tube fault, and conduct experiments to obtain the original data set consisting of current signals, midpoint voltage signals and speed signals in normal mode and fault mode, and label them for subsequent network training and testing.

[0033] The topology of the three-level ANPC converter is as follows: Figure 2 As shown, each phase bridge arm (x = a, b, c) of the converter consists of three half-bridge modules (M x1 -M x3 ). Among them, (S x1 , S x2 )、(S x3 , S x4 ) and (S x5 , S x6 ) are complementary switch pairs. The commonly used switch states of three-level ANPC converters are shown in Table 1, where U x represents the output voltage of phase x. P and N represent the positive and negative switching states respectively. The present invention considers two commonly used zero switching states of the ANPC inverter, which are represented by [OU] and [OL] respectively.

[0034] Table 1 Switching states and output voltages of three-level ANPC converters

[0035] Switch Status <![CDATA[S x1 ]]> <![CDATA[S x2 ]]> <![CDATA[S x3 ]]> <![CDATA[S x4 ]]> <![CDATA[S x5 ]]> <![CDATA[S x6 ]]> <![CDATA[U x ]]> [P] 1 1 0 0 0 1 <![CDATA[V dc / 2]]> [OU] 1 0 1 0 1 0 0 [OL] 0 1 0 1 1 0 0 [N] 0 0 1 1 1 0 <![CDATA[-V dc / 2]]>

[0036] The mathematical model of the permanent magnet synchronous motor in the dq synchronous rotating reference coordinate system is established and expressed as:

[0037]

[0038] Among them, (u d 、u q ) and (i d、i q ) are the stator voltage and stator current on the dq axis, R s is the stator resistance, ψ f represents the flux linkage of the permanent magnet, L d and L q is the stator inductance on the dq axis, ω e Represents electrical angular velocity.

[0039] In addition, the electromagnetic torque T of the permanent magnet synchronous motor m It can be expressed as:

[0040] T m =1.5p n i q [(L d -L q )i d +ψ f ]

[0041] Among them, p n Represents the number of permanent magnet pole pairs.

[0042] In order to set the output of the intelligent diagnosis model, it is first necessary to determine the converter fault type. By analyzing the converter commutation path and simulation results, the present invention decides to consider both single switch open circuit and multiple switch open circuit, a total of 45 fault modes, and defines fault labels, as shown in Table 2

[0043] Table 2 Different open circuit fault types and their labels

[0044]

[0045]

[0046] A simulation model of a permanent magnet motor drive system fed by a three-level ANPC converter based on finite set model predictive control was built, and experiments were carried out to collect current signals, midpoint voltage signals, and speed signals in normal mode and fault mode to form an original data set, which was then labeled for subsequent network training and testing.

[0047] Step 2: Multi-source and multi-dimensional information fusion is performed on the collected raw data to form a two-dimensional feature map. The data is then normalized within the range of [-1, 1] to eliminate the impact of working condition changes on diagnostic performance and generate periodic data samples. Next, the data samples are divided into training sets, cross-validation sets, and test sets according to a certain ratio, which are used for training, network parameter design, and generalization ability evaluation, respectively.

[0048] In order to extract more data features, the feature map of the input layer is designed as a two-dimensional structure. At the same time, in order to achieve the integration of multi-source information and effective feature extraction, a multi-source information fusion processing method is adopted to fuse the collected original electrical signals and mechanical signals. The specific process is as follows: Figure 3 shown.

[0049] Since the converter switch open circuit fault waveform is periodic, the frequency of the fault waveform is f oc Corresponding to the fundamental frequency of the stator current, the frequency of the fault waveform is expressed as:

[0050]

[0051] Among them, p n is the number of magnetic pole pairs of the permanent magnet motor considered in the experiment, N p is the permanent magnet motor speed. Since the sampling frequency of data generation is defined as 20kHz, each fault sample requires 2000 sampling points. Figure 3 It can be seen that each feature map consists of 7 different signals, so 14,000 sampling points are required. Since the kernel shape in a two-dimensional convolutional neural network is usually square, the width (W) and height (H) of the generated feature map are defined as 140 and 100, respectively, to enhance the feature extraction capability of the convolution kernel. In addition, the generated data set is normalized to eliminate the signal amplitude differences caused by different working conditions.

[0052] The normalized data set is divided into a training set (70%), a cross-validation set (15%), and a test set (15%). The training set is used to train the two-dimensional convolutional neural network, the cross-validation set is used to evaluate the trained network, and finally the test set is used to verify the generalization ability of the network.

[0053] Step 3: Build a two-dimensional convolutional neural network model and perform offline training. Input the divided training set and cross-validation set into the network, train and evaluate the network according to the AdamW training algorithm, use the cross entropy loss function as the evaluation index, select the optimal network parameters, and finally use the test set to verify the generalization ability of the trained network.

[0054] The designed two-dimensional convolutional neural network structure is as follows Figure 4 As shown in Figure 1, it mainly consists of four convolutional modules, a flat layer and a fully connected (FC) layer. Among them, the framework of the convolutional module consists of a convolutional layer, a batch normalization layer, an activation layer and a maximum pooling layer.

[0055] The convolution layer extracts the features of the input layer, the batch normalization layer normalizes the output of the convolution layer to improve the stability of the two-dimensional convolutional neural network, and the activation layer enables the convolution layer to perform nonlinear feature extraction. In addition, in order to prevent gradient loss and reduce the amount of calculation, the present invention selects the rectified linear unit (ReLU) as the activation function, and the expression of ReLU is:

[0056] ReLU(x)=max(0,x)

[0057] The max pooling layer reduces the size of the feature map to eliminate redundant information. The flattening layer and the fully connected layer convert the extracted features into classification results, which are then passed to the output layer. According to the PyTorch user guide, the output size of the convolutional layer is:

[0058]

[0059] Among them, O h and O w represents the size of the convolutional layer; H in and W in is the height and width of the input feature map; S c and k represent the step size and the size of the convolution kernel respectively; P c1 and P c2 is the size of the padding.

[0060] The output size of the max pooling layer is expressed as:

[0061]

[0062] Where r represents the kernel size of the maximum pooling layer; H p and W p The height and width of the input feature map for the pooling layer; S m Indicates the step size; P m1 and P m2 is the size of the padding.

[0063] Subsequently, the divided data set is input into the convolutional neural network, and back propagation is performed in combination with the AdamW algorithm to automatically update the parameters of the network. Compared with the ordinary Adam algorithm, the algorithm selected by the present invention can avoid the influence of the gradient on the weight decay, and has the characteristics of fast convergence speed and strong generalization ability. The AdamW algorithm can be expressed in pseudo code:

[0064]

[0065]

[0066] In Algorithm 1, Times represents the number of training iterations, using γ CLRTo optimize the training performance. According to the results of offline training, the value of Times is defined as 30 to avoid underfitting and reduce training time. It should be noted that during offline training, γ CLR The initial value (l ini )、Upper limit(l max ) and the lower bound (l min ) are included in hyperparameter tuning.

[0067] In order to verify the superiority of the two-dimensional convolutional neural network based on the multi-source information fusion strategy designed by the present invention, the training performance of different convolutional neural networks is compared, where 2D-CNN-P is a two-dimensional CNN whose input feature map only fuses the electrical signal and ignores the rotor speed, and 1D-CNN is a one-dimensional CNN whose input feature map is the same as that of the present invention. Figure 5 As shown. Figure 5 In (a), we can see that as the number of network iterations increases, the training and verification accuracy of the 2D convolutional neural network is better than that of other networks, with the best training accuracy and verification accuracy being 98.69% and 98.23% respectively. Figure 5 In (b), we can see that the network has better evaluation indicators, that is, the cross entropy loss is smaller and the diagnostic performance is higher.

[0068] After training and evaluating the 2D convolutional neural network, the test set is used to verify the generalization ability of the trained network. Figure 6 The confusion matrix of the test set is shown in Figure 1. The prediction accuracy of the test set is 98.31%, which proves that the designed network has a certain generalization ability and can accurately locate and identify faults. Figure 6 There are still some misclassifications in the results. In the future, we will use the threshold setting method to improve the fault diagnosis and reduce the misdiagnosis rate.

[0069] Step 4: Lightweight design and edge deployment. Convert the trained neural network model to an ONNX model and use TensorRT to perform lightweight design on it. Then deploy the lightweight network to the edge device (GPU) for experiments to achieve real-time diagnosis of converter open circuit faults under different working conditions, and design diagnostic thresholds to prevent misdiagnosis and improve diagnostic accuracy.

[0070] Figure 7The optimization strategy of TensorRT for lightweight design of training networks is described. The parallel computing architecture of GPU is effectively utilized through multi-stream execution, which improves the computing efficiency. At the same time, other optimizations are mainly focused on reducing the size of the trained two-dimensional convolutional network, thereby reducing the computational cost. The lightweight network is compared and analyzed with the original network. Floating point operations (FLOPs) represent the computational requirements of the network, and the prediction accuracy represents the accuracy of the diagnosis, as shown in Table 3. It can be seen that the lightweight optimization of TensoRT significantly reduces the parameter size and computational requirements of the network. At the same time, the accuracy difference between the lightweight network and the original network is not much, indicating that the optimization does not affect the diagnostic accuracy of the network.

[0071] Table 3 Comparison results of lightweight networks

[0072]

[0073] After being optimized by TensorRT, the lightweight 2D convolutional network is deployed to the edge device Jetson. At the same time, the diagnostic threshold is defined to avoid Figure 6 The misdiagnosis occurs in the circuit breaker, and the high-precision real-time diagnosis of the open circuit fault of the converter switch tube is realized. In addition, the diagnosis time will increase with the increase of the threshold value. Therefore, the setting of the threshold value should consider the balance between the diagnosis time and the prediction accuracy, and ensure the diagnosis speed while reducing the misdiagnosis. The threshold value of the present invention is selected as 3.

[0074] The proposed diagnostic method is experimentally verified on a permanent magnet motor test platform driven by a three-level ANPC converter. The finite set model predictive control algorithm is implemented by the MicroLabBox real-time controller, and a lightweight two-dimensional convolutional neural network is deployed on the edge device Jetson. The real-time transmission of data and diagnostic results is achieved between Jetson and MicroLabBox through serial communication.

[0075] The experimental results of intelligent diagnosis of multiple switch tube faults under different working conditions are as follows: Figure 8 As shown in the figure, it can be seen that when multiple switch tubes fail at the same time, the three-phase current (i abc )、midpoint voltage deviation (U er ) and rotor speed (N p ) has a higher degree of distortion, which leads to serious degradation of converter performance. Therefore, timely diagnosis and location of multiple switch tube faults in the converter are crucial to ensure safe and stable operation and health management of the motor drive system. From the experimental results, it can be seen that at different speeds (N p ) and torque (T m) conditions, different switch tube faults can be successfully located and diagnosed, and the diagnosis time is less than 60ms, which proves that the edge AI-based intelligent diagnosis method proposed in the present invention can realize high-precision and high-efficiency diagnosis of multiple faults of converter switch tubes under different working conditions.

Claims

1. A multi-fault intelligent diagnosis method for traction converter based on edge AI, characterized in that: The following steps are involved: Step 1: Construction of dataset; A mathematical model of a permanent magnet motor drive system fed by a three-level active neutral point clamped ANPC converter is established. In the MATLAB / Simulink environment, a simulation model of a permanent magnet motor control system fed by a three-level active neutral point clamped ANPC converter is built. The single switch open circuit and multiple switch open circuit multiple fault modes of the three-level ANPC converter are simulated. The original data sets consisting of three-phase current signals, midpoint voltage signals, and speed signals in normal mode and fault mode are obtained and labeled for subsequent network training and testing. Step 2: Preprocess the data; First, the collected data is fused with multi-source and multi-dimensional information. Specifically, seven types of signals, including three-phase current, dq axis current, midpoint error and speed, are fused into a two-dimensional feature map to improve the efficiency of model feature extraction. Subsequently, the data is normalized within the range of [-1, 1] to eliminate the impact of operating condition changes on diagnostic performance and generate data samples. Next, the data samples are divided into a training set, a cross-validation set and a test set according to a certain ratio, which are used for training, network parameter design and network performance evaluation respectively. Step 3: Build a two-dimensional convolutional neural network model and use the established data set for offline training; input the divided training set and cross-validation set into the network, train and evaluate the network according to the AdamW training algorithm, use the cross entropy loss function as the evaluation index, select the optimal network parameters, and finally use the test set to verify the generalization ability of the trained network; Step 4: Lightweight design and edge deployment; The trained neural network model is converted into an ONNX model and lightweighted using TensorRT. The lightweight network is then deployed to edge devices to achieve real-time diagnosis of converter open-circuit faults under different working conditions. The diagnostic threshold is designed to prevent misdiagnosis and improve diagnostic accuracy.

2. The three-level ANPC converter single switch open circuit and multiple switch open circuit fault modes according to claim 2, characterized in that: The multiple fault modes in step 1 include both single module open circuit faults and two module open circuit faults, so there are 9 single module open circuit faults and 36 two power module open circuit faults.

3. The method for multi-fault intelligent diagnosis of traction converter based on edge AI according to claim 1 is characterized in that: In step 2, the training set, cross-validation set and test set are 70%, 15% and 15% respectively.

4. According to claim 1, a traction converter multi-fault intelligent diagnosis method based on edge AI is characterized in that: The two-dimensional convolutional neural network in step 3 is composed of four convolutional modules, a flat layer and a fully connected layer; wherein the framework of the convolutional module is composed of a convolutional layer, a batch normalization layer, an activation layer and a maximum pooling layer, the convolutional layer extracts the features of the input layer, the batch normalization layer standardizes the output of the convolutional layer to improve the stability of the two-dimensional convolutional neural network, and the activation layer enables the convolutional layer to perform nonlinear feature extraction; the maximum pooling layer reduces the size of the feature map to eliminate redundant information; the flat layer and the fully connected layer convert the extracted features into classification results, and then transmit these results to the output layer.

Images