Three-level converter composite fault online diagnosis method and system

Abstract

The invention discloses a three-level converter composite fault online diagnosis method and system, relates to the technical field of three-level converter fault diagnosis, and solves the problems that the fault types processed by the existing three-level converter fault diagnosis technology are limited and the accuracy is not high. And real-time state information of the converter cannot be provided for a fault-tolerant control algorithm. The method comprises the steps of constructing a two-dimensional convolutional neural network, making a data set, training the two-dimensional convolutional neural network by using the data set in combination with transfer learning to obtain a composite fault diagnosis model, and deploying the composite fault diagnosis model to edge equipment. The faults of the three-level converter are diagnosed in real time through the collected data of the three-level converter, wherein the faults of the three-level converter comprise a power switch tube open-circuit fault and a sensor fault; according to the method, the generalization ability of the network model is improved by utilizing the transfer learning training strategy, and secondary development of the existing diagnosis model in other diagnosis tasks is facilitated.

Description

Technical Field

[0001] The present invention relates to the technical field of three-level converter fault diagnosis, and in particular to a three-level converter composite fault online diagnosis method and system. Background Art

[0002] The three-level converter has the advantages of low energy consumption, high working efficiency, low voltage stress of power devices, small voltage change rate and small waveform distortion. It is widely used in photovoltaic power generation, rail transportation, aerospace, DC microgrid, energy storage system, high-voltage inverter and other power transmission occasions, and has significant advantages in medium-high voltage and high-power application scenarios.

[0003] Compared with two-level converters, three-level converters have more switching devices and more complex circuit structures, making them more prone to system failures. Three-level converters mainly have current sensor failures and power device open circuit failures. The current sensor provides real-time current signals to the closed-loop control of the converter drive system, and its reliability and accuracy directly affect the stability of the drive system. As one of the sensitive components in the traction motor drive system, the current sensor is easily affected by factors such as electromagnetic interference and load mutation in a complex working environment, resulting in missing or deviation of the measurement signal and seriously affecting the control performance.

[0004] In addition, open-circuit failure of power devices will significantly increase the total harmonic distortion rate of current, torque pulsation and voltage stress on power devices, further aggravating problems such as voltage imbalance of converter bus capacitors, and inducing a chain reaction, leading to secondary failures, and ultimately having a negative impact on system safety.

[0005] In order to realize converter fault diagnosis, domestic and foreign scholars have conducted a lot of research, which can be divided into signal-based methods, model-based methods and data-driven methods. Model-based methods can realize the composite diagnosis of sensor faults and power device open-circuit faults, but they need to build accurate mathematical models, which is heavily dependent on the accuracy of system model parameters, and the diagnosis speed is slow under low-speed conditions. Signal-based methods do not require accurate mathematical models and have parameter robustness, but signal processing technology has a large amount of calculation and is more sensitive to changes in motor operating conditions.

[0006] Data-driven methods do not require an accurate mathematical model of the diagnostic object, nor do they require additional sensors. Only historical data is needed to realize the composite fault diagnosis of converter power devices and sensors. However, knowledge-based diagnostic strategies are currently mostly used for two-level topologies, and are rarely used for fault diagnosis of three-level converters. Compared with two-level topologies, three-level converters have more power devices and more similar open fault characteristics. In addition, most data-driven methods are offline diagnosis and cannot provide real-time converter status information to the fault-tolerant control algorithm. Summary of the invention

[0007] In order to solve the problems existing in the above-mentioned prior art, the present invention provides a three-level converter composite fault online diagnosis method and system to solve the technical problems that the existing three-level converter fault diagnosis technology has limited fault types to handle and low accuracy, and cannot provide real-time converter status information to the fault-tolerant control algorithm.

[0008] A three-level converter composite fault online diagnosis method includes constructing a two-dimensional convolutional neural network, preparing a data set, using the data set in combination with transfer learning to train the two-dimensional convolutional neural network to obtain a composite fault diagnosis model, deploying the composite fault diagnosis model to an edge device, and performing real-time diagnosis of the three-level converter faults through collected data of the three-level converter, wherein the three-level converter faults include power switch tube open circuit faults and sensor faults.

[0009] Furthermore, the transfer learning method includes: using multiple two-dimensional convolutional neural networks with different numbers of convolution kernels D as pre-trained networks, training the pre-trained networks for m epochs in sequence through a data set to generate multiple pre-trained models, and then loading the pre-trained models in sequence to perform transfer learning training using the data set and n epochs, automatically adjusting the hyperparameters to be adjusted until the optimal hyperparameters of each pre-trained model are obtained, and then comparing the performance of each pre-trained model with the optimal hyperparameters to obtain the optimal pre-trained model as a composite fault diagnosis model, m>n.

[0010] Furthermore, the hyperparameters to be adjusted are the maximum value, initial value and minimum value of the learning rate of the training algorithm that is dynamically adjusted, and the training algorithm is a back propagation algorithm.

[0011] Furthermore, it is determined whether it is necessary to perform lightweight design on the composite fault diagnosis model according to the situation of the edge device. If necessary, the composite fault diagnosis model is processed into a lightweight diagnosis model through model lightweight design and then deployed to the edge device.

[0012] Furthermore, the making of the data set includes: obtaining simulation data through a simulation experiment, and then performing data fusion on the simulation data to obtain a two-dimensional input feature map.

[0013] Furthermore, the two-dimensional convolutional neural network includes an input layer, a first convolutional module, a second convolutional module, a third convolutional module, an expansion layer, a fully connected layer and softmax, and an output layer. The first convolutional module includes a convolutional layer and a batch normalization layer, an activation function layer, and a maximum pooling layer.

[0014] Furthermore, the composite fault diagnosis model is provided with an anti-misdiagnosis threshold, and only when a certain type of fault is diagnosed consecutively for a number of times that reaches the anti-misdiagnosis threshold is this type of fault determined to have occurred.

[0015] A three-level converter composite fault online diagnosis system comprises a sensor monitoring network, a controller and an edge device, wherein a composite fault diagnosis model is deployed in the edge device, the sensor monitoring network is used to detect the current and voltage data of the three-level converter in real time, the controller extracts the status information of the three-level converter according to the current and voltage data, and the edge device inputs the status information of the three-level converter into the composite fault diagnosis model to obtain the diagnosis result and feeds it back to the controller.

[0016] The beneficial effects of the present invention include: (1) Construction of composite fault diagnosis model: A composite fault diagnosis model is constructed based on the pre-trained network model, which simplifies the design difficulty of composite fault modeling and diagnosis model. In addition, multi-source and multi-dimensional information fusion and noise enhancement strategies are used to improve the fault feature extraction efficiency and noise robustness of the composite fault diagnosis model. (2) Transfer learning training: The transfer learning strategy is used to improve the offline training efficiency of the composite fault diagnosis model and simplify the hyperparameter adjustment process. At the same time, the transfer learning training strategy improves the generalization ability of the network model and facilitates the secondary development of existing diagnostic models in other diagnostic tasks; (3) Anti-misdiagnosis design: To improve the reliability of the diagnosis results, an anti-misdiagnosis threshold is set for codes prone to misdiagnosis based on the generalization ability evaluation results. Although the setting of the diagnostic threshold prolongs the diagnosis time, this method avoids misdiagnosis and missed diagnosis as much as possible, and the diagnostic threshold is only set for fault codes with misdiagnosis, thereby weakening the impact of this method on the overall diagnosis speed of the deployed model. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of a method for online diagnosis of a three-level converter composite fault involved in an embodiment of the present application.

[0018] Figure 2 This is a topological diagram of a three-level ANPC converter involved in an embodiment of the present application.

[0019] Figure 3 A schematic diagram of fusing multi-source multi-dimensional data with an input feature map according to an embodiment of the present application.

[0020] Figure 4 It is an overall structural diagram of the two-dimensional convolutional network involved in the embodiment of the present application.

[0021] Figure 5 This is a flowchart of network training and hyperparameter adjustment based on transfer learning involved in an embodiment of the present application.

[0022] Figure 6 This is a cross entropy comparison between the transfer learning training method involved in the embodiment of the present application and other methods.

[0023] Figure 7 This is a comparison of the accuracy of the transfer learning training method involved in the embodiment of the present application and other methods.

[0024] Figure 8 This is a verification result of the generalization ability of the composite fault diagnosis network involved in the embodiment of the present application.

[0025] Fig. 9 This is a flowchart of the lightweight deep neural network edge deployment involved in an embodiment of the present application.

[0026] Fig.10 Schematic diagram of a three-level converter composite fault online diagnosis system involved in an embodiment of the present application.

[0027] Fig.11 Schematic diagram of a real-time diagnostic test platform of a three-level converter composite fault online diagnostic system involved in an embodiment of the present application.

[0028] Fig.12 The power switch tube under steady-state conditions involved in the embodiment of the present application S a1 Diagnostic result of open circuit.

[0029] Fig.13 The power switch tube under steady-state conditions involved in the embodiment of the present application S b1 Diagnostic result of open circuit.

[0030] Fig.14 This is the diagnosis result of the A-phase sensor offline fault under steady-state conditions involved in the embodiment of the present application.

[0031] Fig.15 This is the diagnosis result of the noise fault under the steady-state condition involved in the embodiment of the present application.

[0032] Fig.16 The power switch tube under transient conditions involved in the embodiment of the present application S a3 Diagnostic result of open circuit.

[0033] Fig.17 The power switch tube under transient conditions involved in the embodiment of the present application S b3 Diagnostic result of open circuit.

[0034] Fig.18 This is the diagnosis result of the B-phase sensor gain fault under the transient condition involved in the embodiment of the present application.

[0035] Fig.19 This is the diagnosis result of the sensor noise fault under the transient condition involved in the embodiment of the present application.

[0036] Fig. 20 A schematic diagram of a multi-layer stack of CNN operators involved in an embodiment of the present application. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

[0038] A three-level converter composite fault online diagnosis method, such as Figure 1 As shown, it includes constructing a two-dimensional convolutional neural network, making a data set, using the data set in combination with transfer learning to train the two-dimensional convolutional neural network to obtain a composite fault diagnosis model, deploying the composite fault diagnosis model to the edge device, and performing real-time diagnosis of the fault of the three-level converter through the collected data of the three-level converter, wherein the fault of the three-level converter includes an open circuit fault of a power switch tube and a sensor fault.

[0039] Specifically, the topological structure of the three-level converter is shown in the figure below: Figure 2 The corresponding setting of the three-level converter composite fault code is shown in Table 1: Table 1 Converter composite fault codes

[0040] The making of the data set includes: obtaining simulation data through simulation experiments, and then fusing the simulation data to obtain a two-dimensional input feature map.

[0041] After determining the composite fault code, define the dimension and size of the diagnostic network input feature map.

[0042] Since the fault categories shown in Table 1 include open circuit faults and sensor faults, there are many types of compound faults and the diagnosis is difficult. Therefore, the two-dimensional convolutional network is more suitable for the compound fault diagnosis of the converter. Figure 3 The multi-source and multi-dimensional information fusion strategy of the designed network is demonstrated. Figure 3 As shown, the input feature map has 7 types of signals and W and H are the width and height of the feature map.

[0043] Then, the working conditions of simulation and experimental data are defined to determine the values ​​of W and H. Specifically, the composite fault waveform of the traction converter is periodic and the fault period is the fundamental period of the stator current, where the fundamental period of the stator current is shown in the following formula:

[0044] In the formula, T S is the fundamental wave cycle time; N and p are the PMSM speed and pole pair number respectively.

[0045] Generally speaking, the number of pole pairs of the motor controlled by the inverter is 2.

[0046] To facilitate the periodic sampling of data samples, the motor speed can be set to 300 r / min, so that T S The sampling frequency of the converter is usually 20kHz, so the number of sampling points of each sensor signal within 0.1s is 2000.

[0047] because Figure 3 There are 7 types of signals in the input feature map, and the designed network input feature map contains a total of 14,000 sampling points. The W and H of the feature map are defined as 140 and 100 respectively.

[0048] After determining the converter operating conditions and fault parameters, the samples of the composite fault diagnosis model need to be set in combination with the fault codes shown in Table 1. The number of experimental and simulation samples corresponding to the fault codes is shown in Table 2. According to Table 2, the network data set has a total of 31,500 groups of samples, and the number of simulation samples and experimental samples are 25,200 and 6,300 respectively.

[0049] Table 2 Number of compound fault experiments and simulation samples

[0050] The data set in Table 2 needs to be normalized to eliminate the signal amplitude differences caused by different working conditions. Subsequently, the normalized data set is split into a training set, a cross-validation set, and a test set at a ratio of 70%, 15%, and 15%, respectively. The training set is used for diagnostic network training, the cross-validation set is used to evaluate the network training effect under each Epoch (training round), and the test set is used to test the generalization ability of the network. After completing data collection, the training data can be used for network training.

[0051] The two-dimensional convolutional neural network is as follows Figure 4 As shown in the figure, it consists of three convolution modules and a maximum pooling layer, and ReLU is the activation function layer in the figure.

[0052] The two-dimensional convolutional network includes an input layer, a first convolutional module, a second convolutional module, a third convolutional module, an expansion layer, a fully connected layer and a softmax, and an output layer. The first convolutional module includes a convolutional layer and a batch normalization layer, an activation function layer, and a maximum pooling layer.

[0053] For two-dimensional convolutional layers and pooling layers, the output size calculation of the above operators can be expressed as follows:

[0054]

[0055] Where ( W in , H in )and( W out , H out ) are the width and height of the two-bit input feature map and output feature map respectively; P , S.K. They are padding size, convolution stride, and convolution kernel size. Generally speaking, the stride of the convolution layer and the stride of the pooling layer are set to 1 and 2 respectively, while the padding of the convolution layer and the padding of the pooling layer are set to 1 and 0 respectively, so as to simplify the calculation of the output size of the convolution layer and the pooling layer.

[0056] 2D Convolutional Neural Network Utilization Fig. 20 The operator hierarchy stacking method shown is similar to Figure 3 The multi-source and multi-dimensional information fusion technology shown in the figure perceives the local feature map and shares parameters, realizing the automatic extraction and learning of input features from shallow to deep. At the same time, the two-dimensional convolutional neural network combines operators such as batch normalization to further improve the training stability and generalization ability of the two-dimensional convolutional network, which helps the two-dimensional convolutional network to accurately detect the complex faults of the three-level converter.

[0057] Specifically: Convolutional neural network is mainly composed of input feature map, convolution layer, activation function, pooling layer, fully connected layer, Softmax function and output layer, and CNN realizes feature extraction of different depths through multi-layer stacking of the above operators, such as Fig. 20As shown. Specifically, the input feature map receives the original data and passes it to the next layer, and then the convolution layer extracts local features through convolution kernel operations. The extracted local features introduce nonlinear transformations with the help of activation functions to achieve nonlinear feature extraction, which helps the convolutional network learn more complex feature relationships. At the same time, the pooling layer is used to reduce the spatial dimension of the feature map and suppress its noise, aiming to reduce the amount of calculation and the risk of overfitting. After multiple convolutional layers, activation functions, pooling layers, nonlinear transformations, and spatial dimensionality reduction and other network operations, the extracted feature map will be flattened into a one-dimensional vector and combined with the fully connected layer and Softmax function, where the output layer will map the output value of Softmax to a specific category to cope with multi-target classification tasks.

[0058] Based on the previous analysis of convolutional network operators, CNN uses convolutional layers and input data for local perception, reducing the number of parameters and computations. At the same time, when CNN performs convolution operations on local windows, neurons in the same layer share parameter weights and offsets, further reducing the amount of parameter training and improving computational efficiency. According to the different dimensions of the pooling layer and the convolution layer, the convolutional network can be divided into one-dimensional, two-dimensional, and three-dimensional convolutional neural networks, which extract effective information from one-dimensional, two-dimensional, and three-dimensional input feature maps, respectively. In addition, the size calculation of the operator's output shows that the convolution kernel and pooling layer can also use padding (Padding) to increase the data boundary or adjust the stride (Stride) to control the size of the output feature map, allowing CNN to capture features of different scales and categories, and improve the flexibility of CNN in classification tasks. It is worth mentioning that convolutional neural networks can be used Figure 3 The input feature diagram shown integrates the electrical signals (voltage, current, etc.) and mechanical signals (speed, load torque, etc.) of the converter into the feature diagram, thereby realizing multi-source and multi-dimensional information fusion, which is helpful for extracting the features of weak signals in the early stage of converter faults and further improving the fault feature signal-to-noise ratio in a strong interference environment.

[0059] In addition, convolutional neural networks often use batch normalization layers to reduce the sensitivity of the network to the initial network parameters during training, thereby improving the stability of model training and accelerating the convergence of training. Since the activation function of CNN is usually ReLU and the output of this function is non-negative, the batch normalization layer cannot transform the output data characteristics into a standard normal distribution. Therefore, the batch normalization layer is usually placed after the convolution layer and before the ReLU function to ensure that the normalization layer effectively transforms the data distribution.

[0060] The specific two-dimensional convolutional network parameters are shown in Table 3, where D is the number of convolution kernels, and D It will be used in hyperparameter adjustment. After the network structure is determined, the back propagation algorithm will be selected and the data set will be used for network training and hyperparameter adjustment.

[0061] Given that there are many fault codes in Table 1, the amount of network data in Table 3 is large, and Figure 4 The convolutional network structure shown is relatively complex, so AdamW is used as the back propagation algorithm for transfer learning training. Since the training data shown in Table 2 is large in scale, the data set needs to be split into small batches to train the network, and the batch size is set to 128.

[0062] When the optimal hyperparameters are selected, any small batch size can achieve the same training performance. Therefore, the Batch Size for convolutional network training is fixed to 128, and there is no need to adjust BatchSize in hyperparameter optimization. In addition, the hyperparameters of the two-dimensional convolutional network in training can be concentrated on the learning rate of AdamW, because the learning rate directly affects the training performance and diagnostic accuracy of the convolutional network.

[0063] Set the maximum value of the dynamic adjustment of the learning rate ( η max )、initial value( η de ) and minimum value ( η min ), and use these parameters as hyperparameters to be adjusted for network optimization.

[0064] The two-dimensional convolutional network includes an input layer, a first convolutional module, a second convolutional module, a third convolutional module, an expansion layer, a fully connected layer and softmax, and an output layer. The first convolutional module includes a convolutional layer and a batch normalization layer, an activation function layer, and a maximum pooling layer. The second convolutional module and the third convolutional module have the same structure as the first convolutional module, but different structural parameters.

[0065] Table 3 Structural parameters of 2D convolutional network

[0066] The hyperparameters to be adjusted are the maximum value, initial value and minimum value of the learning rate of the back propagation algorithm dynamically adjusted.

[0067] The transfer learning method includes: using multiple two-dimensional convolutional networks with different numbers of convolution kernels D as pre-trained networks, training the pre-trained networks in sequence through data sets for m Epochs (training rounds) to generate multiple pre-trained models, and then loading the pre-trained models in sequence and using data sets and n Epochs (training rounds) for transfer learning training, using Optuna to automatically adjust the hyperparameters to be adjusted until the optimal hyperparameters of each pre-trained model are obtained, and then the performance of each pre-trained model with the optimal hyperparameters set is compared to obtain the optimal pre-trained model as a composite fault diagnosis model. It should be noted that m is greater than n, thereby reducing the number of Epochs required for diagnostic model training.

[0068] like Figure 5 As shown in the figure, the designed transfer learning strategy uses a two-dimensional convolutional network as a pre-trained network, and generates pre-trained models with different numbers of convolution kernels (D) through training sets, validation sets and 30 Epochs, where the parameters such as the dynamic learning rate of the pre-trained model use the default values ​​of PyTorch. Subsequently, PyTorch is used to read these pre-trained models, and the training set, validation set and 10 Epochs are used again for transfer learning training, and Optuna is used to automatically adjust hyperparameters such as the dynamic learning rate. Since the number of data samples in Table 2 is large, the data set can be split into small batches (Batch Size) for training, thereby reducing the hardware resources required for network training and improving training efficiency. It is worth mentioning that when the optimal hyperparameters are determined, small batches of data of any size can achieve the same performance, and the size of small batches of data for transfer learning training will be constant at 256. In summary, the designed transfer learning simplifies the hyperparameter adjustment process, and Optuna's automatic adjustment parameters do not include the number of convolution kernels D. In addition, transfer learning performs secondary training based on the pre-trained model, so that the Epoch of network training is reduced to 10, thereby improving the model training efficiency and convergence speed.

[0069] Table 4 Hyperparameter value range and optimal value of pre-trained model

[0070] according to Figure 5 In the network training and hyperparameter adjustment method based on transfer learning, Table 4 shows the hyperparameter adjustment range of the pre-trained model and the optimal value of Optuna's automatic optimization. The number of convolution kernels is set before transfer learning, while other parameters such as learning rate are set in Optuna's optimization. Therefore, there are 5 types of pre-trained networks for transfer learning training and Optuna's automatic parameter adjustment. Then the network model with the best performance is selected, and the corresponding learning rate and number of convolution kernels are recorded.

[0071] In order to verify the advantages of the transfer learning strategy, after completing the transfer learning training and hyperparameter adjustment, the training performance of the designed composite fault diagnosis model can be compared with the typical convolutional network, thereby showing the advantages of the composite fault diagnosis model and the transfer learning model in diagnostic accuracy and training efficiency. The corresponding comparison results are shown in Figure 2. Figure 6-7 shown.

[0072] The 1D-CNN and 2D-CNN in the figure are trained from scratch. 1D-CNN is a one-dimensional convolutional neural network, 2D-CNN is a two-dimensional convolutional neural network, and 2D-CNN-T is a two-dimensional convolutional neural network trained using transfer learning. In addition, the input feature maps of the three types of networks in the figure are all fused with Figure 3 The current, voltage and speed signals shown in the figure are all the optimal values ​​of the hyperparameters of these networks under Optuna automatic parameter adjustment, thus ensuring the fairness of network comparison. Figure 6-7 As shown in the figure, the accuracy and cross entropy of 2D-CNN-T are better than those of 1D-CNN and 2-DNN, and 2D-CNN-T reaches the optimal value at the 8th Epoch, while other networks need to reach the optimal value at around 23 Epochs. Therefore, the transfer learning strategy reduces the network model training time to 34.7% of the original time. This speeds up the convergence speed of the network model in the new task and improves the training performance and generalization ability of the network model on the new task.

[0073] Similarly, after completing network training and hyperparameter adjustment, the test set will be used to evaluate the generalization ability of transfer learning after training. Figure 8 The corresponding generalization ability verification results are shown, among which fault codes 16 to 20 and 27 to 32 have misdiagnosis. other Represents other fault codes in Table 1.

[0074] according to Figure 8 It can be seen that the accuracy of the designed transfer learning network in the test set is 97.56%, which is consistent with Figure 7 There is no significant difference in the optimal accuracy of the two models, which verifies that the network model trained by transfer learning has good generalization ability. In addition, in order to improve the real-time diagnosis accuracy of the deployed model and the credibility of the output results, Figure 8 Anti-misdiagnosis design is carried out for fault categories that may have misdiagnosis phenomena.

[0075] In another embodiment, whether it is necessary to perform lightweight design on the composite fault diagnosis model is determined according to the edge device situation. If necessary, the composite fault diagnosis model is processed into a lightweight diagnosis model through model lightweight design and then deployed to the edge device.

[0076] The composite fault diagnosis model is deployed to the edge device using Fig. 9The lightweight design and edge deployment strategy of the diagnostic model shown in the figure, where ONNX (Open Neural Network Exchange) is an open file format designed for deep learning, which is used to store trained models and helps deep learning models to be effectively converted and deployed between different frameworks and hardware platforms. Since the GPU and operating system used for deep learning model training are different from the edge deployment platform, the trained neural network needs to be converted into the ONNX format to improve the cross-platform compatibility of the deep learning model and facilitate TensorRT optimization and model edge deployment. After completing the network format conversion, use C++ or Python language to set the optimized numerical precision of TensorRT, and generate an inference engine to be deployed in the embedded GPU. The data set is then used to evaluate the performance of inference engines with different numerical precisions (such as prediction accuracy and real-time diagnosis speed), and the inference engine with the best performance is selected for edge deployment.

[0077] Using the 31,500 sample data sets in Table 2, the optimal numerical precision is selected with diagnostic accuracy and parameter scale as indicators. Table 5 shows the parameter scale and accuracy of the deployment model under different numerical precisions.

[0078] Table 5 Parameter scale and accuracy of lightweight composite fault diagnosis model

[0079] According to the comparison results in the table, the lightweight network model with numerical precision of FP16 has the smallest parameter size and FLOPs, and there is no significant difference in diagnostic accuracy between FP32 and the original network model. Therefore, the lightweight network with numerical precision of FP16 will be used as the diagnostic model for converter compound faults. Since the precision and recall rates of misdiagnosis codes (16-20 and 27-32) are both greater than 80%, and the sum of misdiagnosis categories and missed diagnosis categories is 2, Figure 7 The confusion matrix in satisfies the two requirements for setting the diagnostic threshold. Subsequently, the diagnostic threshold can be set for the misdiagnosis codes in the confusion matrix, thereby improving the reliability of the output results of the diagnostic model. Therefore, Figure 7 The fault codes and corresponding diagnostic thresholds and times are shown in Table 6, and the accuracy in the table is calculated from 500 sample data sets corresponding to the misdiagnosis codes in Table 2.

[0080] Table 6 Misdiagnosis prevention threshold and accuracy of the composite fault diagnosis model

[0081] Since the accuracy rates corresponding to the fault codes shown in Table 6 are all 100% when the diagnostic threshold is 2, the diagnostic thresholds of the above composite fault codes can be set to 2, thereby minimizing the probability of misdiagnosis and maximizing the diagnostic speed of the deployed model.

[0082] In another embodiment, a three-level converter composite fault online diagnosis system is provided. Fig.10 As shown, it includes a sensor monitoring network, a controller and an edge device, a composite fault diagnosis model is deployed in the edge device, the sensor monitoring network is used to detect the current and voltage data of the three-level converter in real time, the controller extracts the status information of the three-level converter according to the current and voltage data, and the edge device inputs the status information of the three-level converter into the composite fault diagnosis model to obtain the diagnosis result and feeds it back to the controller.

[0083] In order to further verify the designed composite fault diagnosis model, it is necessary to use Fig.11 The test platform shown experimentally verifies the feasibility of the method, in which the transfer learning model is deployed in the embedded GPU, and the converter control strategy is burned into the converter controller. In addition, the diagnostic results of the deployment model interact with the real-time status information of the converter through serial communication, in which the converter status information is used as the online diagnosis basis of the deployment model, and the real-time diagnostic results of the deployment model can be used as the execution basis of the subsequent adaptive fault-tolerant strategy. It is worth mentioning that in order to show the feasibility of the diagnostic algorithm under various working conditions, the experimental working conditions should include steady-state working conditions and transient working conditions with different speeds and torques. To achieve the above working conditions, this embodiment uses a load asynchronous motor as the load under different working conditions of the tested permanent magnet synchronous motor, and uses the tested permanent magnet synchronous motor as the driving object of the three-level converter. The load asynchronous motor is controlled by a commercial frequency converter and the torque loading instruction of the load motor is issued by the host computer. At the same time, the speed instruction of the tested permanent magnet motor is adjusted by the converter controller to achieve transient changes in different speeds of the motor.

[0084] (1) Steady-state performance verification Figure 12-15 The experimental results of the designed composite fault diagnosis strategy under steady-state conditions are presented. T i For diagnosis time, Count For the designed recording function, when Count When the diagnostic threshold set in Table 6 is reached, D e Only then will the corresponding deployment model diagnosis results be output, thereby minimizing the probability of misdiagnosis.

[0085] According to the experimental results, the designed transfer learning deployment model can achieve accurate identification of power device and sensor compound faults under various steady-state conditions by normalizing input data and fusion of multi-source and multi-dimensional data. Fig.13 and Fig.15 The experimental results in also verify the feasibility of the designed anti-misdiagnosis strategy under various steady-state conditions, which helps to improve the credibility of the model's diagnosis results.

[0086] Specifically, the curve in the figure is: three-phase current waveform ( i a , i b , i c ), torque waveform ( T m )、Speed ​​waveform( N )、Converter midpoint voltage deviation waveform( V er )、fault-triggered signal (De), diagnosis result (flag), anti-misdiagnosis threshold count (count), fault diagnosis waveform ( T i Indicates the diagnosis time, i.e., the time difference between De and flag, reflecting the lag time between fault occurrence and fault diagnosis. The smaller it is, the better the real-time nature of the diagnosis. Specifically, Fig.12 The experimental conditions are N=400RPM, T m =10N·m, SNR=20dB, S a1 open circuit; Fig.13 The experimental conditions are N=500RPM, T m =5N·m, SNR=22.5dB, S b1 open circuit; Fig.14 The experimental conditions are N=500RPM, T m =10N·m, SNR=20dB, phase A is disconnected; Fig.15 The experimental conditions are N=500RPM, T m =10N·m, SNR=20dB, noise fault; (2) Transient performance verification In addition to the steady-state experiments, the experimental results of the designed composite fault diagnosis strategy under transient conditions are as follows: Figure 16-19 As shown in the figure, the sensor and power device faults are triggered at a speed of 400 r / min. According to the experimental results, under various types of transient conditions, the designed diagnostic algorithm can quickly detect and locate the combined faults of sensors and power devices, and Fig.17 and Fig.19 The experimental results verify the feasibility of the anti-misdiagnosis strategy under transient conditions.

[0087] Specifically, the curve in the figure is: three-phase current waveform ( i a , i b , i c ), torque waveform ( T m )、Speed ​​waveform( N )、Converter midpoint voltage deviation waveform( V er )、fault-triggered signal (De), diagnosis result (flag), anti-misdiagnosis threshold count (count), fault diagnosis waveform ( T i Indicates the diagnosis time, i.e., the time difference between De and flag, reflecting the lag time between fault occurrence and fault diagnosis. The smaller it is, the better the real-time nature of the diagnosis. Specifically,

[0088] Fig.16 The experimental conditions are N=200-600RPM, T m =5N·m, SNR=20dB, S a3 open circuit; Fig.17 The experimental conditions are N=200-600RPM, T m =10N·m, SNR=22.5dB, S b3 open circuit; Fig.18 The experimental conditions are N=200-600RPM, T m =15N·m, SNR=22.5dB, G (amplification factor of gain failure)=2, B phase gain; Fig.19 The experimental conditions are N=200-600RPM, T m =10N·m, SNR=17.5dB, noise fault; according to Figures 12 to 19 The experimental results show that the designed lightweight deployment model can realize the composite fault diagnosis of converter and sensor under different steady-state and transient conditions.

[0089] In another embodiment, in order to improve the control performance of the three-level converter under complex fault conditions, a corresponding adaptive fault-tolerant strategy may be executed according to the real-time diagnosis results of the deployment model to improve the control performance of the predictive control under complex fault conditions of the converter.

[0090] The above-mentioned embodiments only express the specific implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the protection scope of the present application. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the technical solution concept of the present application, and these all belong to the protection scope of the present application.

Claims

1. A three-level converter composite fault online diagnosis method, characterized in that: The method includes constructing a two-dimensional convolutional neural network, making a data set, using the data set in combination with transfer learning to train the two-dimensional convolutional neural network to obtain a composite fault diagnosis model, deploying the composite fault diagnosis model to an edge device, and performing real-time diagnosis of the fault of the three-level converter through the collected data of the three-level converter, wherein the fault of the three-level converter includes an open circuit fault of a power switch tube and a sensor fault.

2. A three-level converter composite fault online diagnosis method according to claim 1, characterized in that: The transfer learning method includes: using multiple two-dimensional convolutional neural networks with different numbers of convolution kernels D as pre-trained networks, training the pre-trained networks for m epochs in sequence through a data set to generate multiple pre-trained models, then loading the pre-trained models in sequence and training n epochs through transfer learning using the data set, automatically adjusting the hyperparameters to be adjusted until the optimal hyperparameters of each pre-trained model are obtained, and then comparing the performance of each pre-trained model with the optimal hyperparameters set to obtain the optimal pre-trained model as a composite fault diagnosis model, m>n.

3. A three-level converter composite fault online diagnosis method according to claim 2, characterized in that: The hyperparameters to be adjusted are the maximum value, initial value and minimum value of the learning rate of the training algorithm for dynamic adjustment.

4. A three-level converter composite fault online diagnosis method according to claim 1, characterized in that: Determine whether it is necessary to perform lightweight design on the composite fault diagnosis model based on the situation of the edge device. If necessary, process the composite fault diagnosis model into a lightweight diagnosis model through model lightweight design and then deploy it to the edge device.

5. The method for online diagnosis of composite faults of a three-level converter according to claim 1, characterized in that: The making of the data set includes: obtaining simulation data through simulation experiments, and then fusing the simulation data to obtain a two-dimensional input feature map.

6. A three-level converter composite fault online diagnosis method according to claim 1, characterized in that: The two-dimensional convolutional neural network includes an input layer, a first convolutional module, a second convolutional module, a third convolutional module, an expansion layer, a fully connected layer and a softmax, and an output layer. The first convolutional module includes a convolutional layer and a batch normalization layer, an activation function layer, and a maximum pooling layer.

7. A three-level converter composite fault online diagnosis method according to claim 1, characterized in that: The composite fault diagnosis model is provided with an anti-misdiagnosis threshold, and only when a certain type of fault is diagnosed consecutively times reaching the anti-misdiagnosis threshold is this type of fault determined to have occurred.

8. A three-level converter composite fault online diagnosis system, characterized in that: It includes a sensor monitoring network, a controller and an edge device, wherein a composite fault diagnosis model obtained by a composite fault online diagnosis method for a three-level converter according to any one of claims 1 to 7 is deployed in the edge device, the sensor monitoring network is used to detect the current and voltage data of the three-level converter in real time, the controller extracts the status information of the three-level converter according to the current and voltage data, and the edge device inputs the status information of the three-level converter into the composite fault diagnosis model to obtain a diagnosis result and feeds it back to the controller.

Images