An online diagnosis method and system for compound faults of a three-level converter

Abstract

The present invention discloses an online diagnosis method and system for compound faults of a three-level converter, which relates to the technical field of fault diagnosis of three-level converters, and solves the technical problems that the existing fault diagnosis technology of three-level converters has limited types of faults to be processed, low accuracy, and cannot provide real-time state information of the converter for the fault-tolerant control algorithm; the present invention includes constructing a two-dimensional convolutional neural network, making a data set, training the two-dimensional convolutional neural network by combining the data set with transfer learning to obtain a compound fault diagnosis model, deploying the compound fault diagnosis model to an edge device, and performing real-time diagnosis on the faults of the three-level converter through the collected data of the three-level converter, wherein the faults of the three-level converter include open-circuit faults of power switching tubes and sensor faults; the present invention uses a transfer learning training strategy to improve the generalization ability of the network model, which helps the existing diagnosis model to be redeveloped in other diagnosis tasks.

Description

Technical Field

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

[0002] Three-level converters have advantages such as low energy consumption, high working efficiency, low voltage stress of power devices, small voltage change rate, and small waveform distortion. They are widely used in power transmission applications such as photovoltaic power generation, rail transit, aerospace, DC microgrids, energy storage systems, and high-voltage inverters, and have significant advantages in medium-voltage and high-power application scenarios.

[0003] Compared with two-level converters, three-level converters have more switching devices and a more complex circuit structure, making them more prone to system failures. The three-level converter mainly has current sensor faults and power device open-circuit faults. The current sensor provides real-time current signals for 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 vulnerable to electromagnetic interference and load mutations in a complex working environment, resulting in missing or deviated measurement signals and seriously affecting the control performance.

[0004] In addition, the power device open-circuit fault will significantly increase the total harmonic distortion rate of the current, torque ripple, and the voltage stress borne by the power device, further deteriorating problems such as the imbalance of the converter bus capacitor voltage, and inducing a chain effect, leading to secondary faults, and ultimately having a negative impact on the system safety.

[0005] To achieve converter fault diagnosis, domestic and foreign scholars have conducted a large number of studies, which can be divided into signal-based methods, model-based methods, and data-driven methods. The model-based method can realize the composite diagnosis of sensor faults and power device open-circuit faults, but it needs to build an accurate mathematical model, highly depends on the accuracy of system model parameters, and has a slow diagnosis speed under low-speed working conditions. The signal-based method does not require an accurate mathematical model and has parameter robustness, but the signal processing technology has a large amount of calculation and is more sensitive to changes in motor working conditions.

[0006] The data-driven method neither requires an accurate mathematical model of the diagnosis object nor additional sensors, and can realize the composite fault diagnosis of converter power devices and sensors only with historical data. However, most of the current knowledge-based diagnosis strategies are used for two-level topologies and are less applied to the 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 of the data-driven methods are offline diagnoses and cannot provide real-time state information of the converter for the fault-tolerant control algorithm. Summary of the Invention

[0007] To solve the problems existing in the above-mentioned prior art, the present invention provides a method and system for online diagnosis of composite faults of a three-level converter, which solves the technical problems that the fault types processed by the existing fault diagnosis technology of three-level converters are limited, the accuracy is not high, and the real-time state information of the converter cannot be provided for the fault-tolerant control algorithm.

[0008] A method for online diagnosis of composite faults of a three-level converter includes constructing a two-dimensional convolutional neural network, making a data set, training the two-dimensional convolutional neural network by using the data set combined with transfer learning to obtain a composite fault diagnosis model, deploying the composite fault diagnosis model to an edge device, and performing real-time diagnosis on the faults of the three-level converter through the collected data of the three-level converter. The faults of the three-level converter include open-circuit faults of power switching tubes and sensor faults.

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

[0010] Further, the hyperparameters to be adjusted are the maximum value, initial value, and minimum value of the learning rate dynamically adjusted by the training algorithm, and the training algorithm is the backpropagation algorithm.

[0011] Further, it is judged whether the composite fault diagnosis model needs to be lightweight designed 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] Further, the making of the data set includes: obtaining simulation data through simulation experiments, and then performing data fusion on the simulation data to obtain a two-dimensional input feature map.

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

[0014] Further, the composite fault diagnosis model is provided with an anti-misdiagnosis threshold, and a certain type of fault is determined to occur only when the continuous diagnosis times of a certain fault reach the anti-misdiagnosis threshold.

[0015] A three-level converter composite fault online diagnosis system includes a sensor monitoring network, a controller, and edge devices. A composite fault diagnosis model is deployed in the edge devices. 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 state information of the three-level converter based on the current and voltage data. The edge devices input the state information of the three-level converter into the composite fault diagnosis model to obtain a diagnosis result and feedback it to the controller.

[0016] The beneficial effects of the present invention include:

[0017] (1) Composite fault diagnosis model construction: A composite fault diagnosis model is constructed based on a pre-trained network model, which simplifies the design difficulty of composite fault modeling and diagnosis models. By using a multi-source multi-dimensional information fusion and noise enhancement strategy, the fault feature extraction efficiency and noise robustness of the composite fault diagnosis model are improved.

[0018] (2) Transfer learning training: By using a transfer learning strategy, the offline training efficiency of the composite fault diagnosis model is improved, and the adjustment process of hyperparameters is simplified. At the same time, the transfer learning training strategy enhances the generalization ability of the network model, which is helpful for the secondary development of existing diagnosis models in other diagnosis tasks.

[0019] (3) Anti-misdiagnosis design: To improve the reliability of the diagnosis result, an anti-misdiagnosis threshold is set for the codes prone to misdiagnosis according to the generalization ability evaluation result. Although the setting of the diagnosis threshold prolongs the diagnosis time, this method avoids misdiagnosis and missed diagnosis as much as possible, and the diagnosis threshold setting is only for the fault codes with misdiagnosis phenomena, thus weakening the impact of this method on the overall diagnosis speed of the deployed model. Brief Description of the Drawings

[0020] Figure 1 It is a flowchart of a three-level converter composite fault online diagnosis method according to an embodiment of the present application.

[0021] Figure 2 It is a topological structure diagram of a three-level ANPC type converter according to an embodiment of the present application.

[0022] Figure 3 It is a schematic diagram of input feature map fusion of multi-source multi-dimensional data according to an embodiment of the present application.

[0023] Figure 4 It is an overall structure diagram of a two-dimensional convolutional network according to an embodiment of the present application.

[0024] Figure 5 It is a flowchart of network training and hyperparameter adjustment based on transfer learning according to an embodiment of the present application.

[0025] Figure 6 Cross - entropy comparison between the transfer learning training method involved in the embodiments of this application and other methods.

[0026] Figure 7 Accuracy comparison between the transfer learning training method involved in the embodiments of this application and other methods.

[0027] Figure 8 Verification result of the generalization ability of the composite fault diagnosis network involved in the embodiments of this application.

[0028] Figure 9 Flowchart of the edge deployment of the lightweight deep neural network involved in the embodiments of this application.

[0029] Figure 10 Schematic diagram of an online diagnosis system for composite faults of a three - level converter involved in the embodiments of this application.

[0030] Figure 11 Schematic diagram of the real - time diagnosis test platform for an online diagnosis system for composite faults of a three - level converter involved in the embodiments of this application.

[0031] Figure 12 Diagnosis result of the open - circuit of the power switch tube under steady - state conditions involved in the embodiments of this application S a1

[0032] Figure 13 Diagnosis result of the open - circuit of the power switch tube under steady - state conditions involved in the embodiments of this application S b1

[0033] Figure 14 Diagnosis result of the disconnection fault of the A - phase sensor under steady - state conditions involved in the embodiments of this application.

[0034] Figure 15 Diagnosis result of the noise fault under steady - state conditions involved in the embodiments of this application.

[0035] Figure 16 Diagnosis result of the open - circuit of the power switch tube under transient conditions involved in the embodiments of this application S a3

[0036] Figure 17 Diagnosis result of the open - circuit of the power switch tube under transient conditions involved in the embodiments of this application S b3

[0037] Figure 18 Diagnosis result of the gain fault of the B - phase sensor under transient conditions involved in the embodiments of this application.

[0038] ​​​​Figure 19 This is the diagnostic result of the sensor noise fault under transient conditions involved in the embodiments of the present application.

[0039] Figure 20 This is the schematic diagram of the multi-layer stacking of the CNN operators involved in the embodiments of the present application. Detailed implementation manners

[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part rather than all of the embodiments of the present application. Therefore, the detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the claimed present application, but merely represents the selected embodiments of the present application. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0041] A method for online diagnosis of compound faults of a three-level converter, as Figure 1 shown, includes 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 compound fault diagnosis model, deploying the compound fault diagnosis model to an edge device, and performing real-time diagnosis of the faults of the three-level converter through the collected data of the three-level converter. The faults of the three-level converter include open-circuit faults of power switching tubes and sensor faults.

[0042] Specifically, the topological structure diagram of the three-level converter is as Figure 2 shown, and the compound fault coding of the three-level converter is set as shown in Table 1:

[0043] Table 1 Converter compound fault coding

[0044]

[0045] The making of the data set includes: obtaining simulation data through simulation experiments, and then performing data fusion on the simulation data to obtain a two-dimensional input feature map.

[0046] After determining the compound fault coding, define the dimension and size of the input feature map of the diagnostic network.

[0047] Since the fault categories shown in Table 1 include open-circuit faults and sensor faults, the number of compound fault categories is large and the diagnosis is difficult. Therefore, the two-dimensional convolutional network is more suitable for the compound fault diagnosis of the converter, where Figure 3 shows the multi-source and multi-dimensional information fusion strategy of the designed network. As Figure 3 shown, the input feature map has 7 types of signals, and W and H are the width and height of the feature map.

[0048] Then define the operating conditions of the simulation and experimental data 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 as follows:

[0049]

[0050] In the formula, T S is the fundamental period time; N and p are the rotational speed and the number of pole pairs of the PMSM respectively.

[0051] Generally speaking, the number of pole pairs of the motor controlled by the converter is 2.

[0052] For the convenience of periodic sampling of data samples, the motor speed can be set to 300 r / min, so that T S is 0.1 s. Given that the sampling frequency of the converter is usually 20 kHz, the number of sampling points of each sensor signal within 0.1 s is 2000.

[0053] Since Figure 3 the input feature map contains a total of 7 types of signals, the designed network input feature map contains a total of 14000 sampling points, and W and H of the feature map are defined as 140 and 100 respectively.

[0054] After determining the converter operating conditions and fault parameters, it is necessary to set the samples of the composite fault diagnosis model in combination with the fault codes shown in Table 1, where the number of experimental and simulation samples corresponding to the fault codes is shown in Table 2. According to Table 2, the network dataset has a total of 31500 groups of samples, and the number of simulation samples and experimental samples are 25200 and 6300 respectively.

[0055] Table 2 Number of composite fault experimental and simulation samples

[0056]

[0057] The dataset in Table 2 needs to be normalized to eliminate the signal amplitude differences caused by different operating conditions. Subsequently, the normalized dataset is split into a training set, a cross-validation set, and a test set according to the ratios of 70%, 15%, and 15% respectively. The training set is used for the training of the diagnosis network, 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 the data acquisition, the training data can be used for network training.

[0058] The two-dimensional convolutional neural network is as Figure 4As shown, it consists of three convolutional modules and a max pooling layer, and ReLU is the activation function layer in the figure.

[0059] The two-dimensional convolutional network includes an input layer, a first convolutional module, a second convolutional module, a third convolutional module, an unfolding layer, a fully connected layer, softmax, and an output layer. The first convolutional module includes a convolutional layer, a batch normalization layer, an activation function layer, and a max pooling layer.

[0060] For the two-dimensional convolutional layer and pooling layer, the output size calculation of the above operators can be expressed by the following formula:

[0061]

[0062]

[0063] In the formula, ( W in , H in ) and ( W out , H out ) are the widths and heights of the two-dimensional input feature map and output feature map respectively; P 、 S, K are the padding size, convolution stride, and convolution kernel size respectively. Generally speaking, the stride of the convolutional layer and the stride of the pooling layer are set to 1 and 2 respectively, while the padding of the convolutional 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 convolutional layer and the pooling layer.

[0064] The two-dimensional convolutional neural network uses Figure 20 the operator hierarchical stacking method shown and Figure 3 the multi-source multi-dimensional information fusion technology shown to perceive local feature maps and share 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 composite faults of the three-level converter.

[0065] Specifically: The convolutional neural network mainly consists of an input feature map, a convolutional layer, an activation function, a pooling layer, a fully connected layer, a Softmax function, and an output layer. And CNN realizes feature extraction of different depths through the multi-layer stacking of the above operators, such as Figure 20As shown in the figure. Specifically, the input feature map receives the original data and passes it to the next layer. Subsequently, the convolutional layer performs operations through the convolutional kernel to extract local features. The extracted local features introduce non-linear transformations through the activation function to achieve non-linear 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 computational amount and the risk of overfitting. After network operations such as multiple convolutional layers, activation functions, pooling layers, non-linear transformations, and spatial dimensionality reduction, the extracted feature map will be flattened into a one-dimensional vector and combined with the fully connected layer and the Softmax function. Among them, the output layer will map out specific categories according to the output values of Softmax to handle multi-object classification tasks.

[0066] Based on the previous analysis of the convolutional network operators, CNN uses the convolutional layer to perform local perception on the input data, reducing the number of parameters and the computational amount. At the same time, when CNN performs convolutional operations in the local window, the neurons in the same layer share the parameter weights and offsets, further reducing the amount of parameter training and improving the computational efficiency. According to the different dimensions of the pooling layer and the convolutional layer, the convolutional network can be divided into one-dimensional, two-dimensional, and three-dimensional convolutional neural networks, which respectively extract the effective information of one-dimensional, two-dimensional, and three-dimensional input feature maps. In addition, the calculation of the output size of the operator shows that the convolutional kernel and the pooling layer can also use padding to increase the data boundary or adjust the stride to control the size of the output feature map, enabling CNN to capture features of different scales and categories and improving the flexibility of CNN in classification tasks. It is worth mentioning that the convolutional neural network can use Figure 3 the input feature map shown in the figure to integrate the electrical signals (such as voltage and current) and mechanical signals (such as rotational speed and load torque) of the converter into the feature map, thereby realizing multi-source and multi-dimensional information fusion, which helps to extract the features of weak signals in the early stage of converter faults and further improve the signal-to-noise ratio of fault features in a strong interference environment.

[0067] 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 training stability of the model and accelerating the training convergence speed. 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 located after the convolutional layer and before the ReLU function to ensure the effective transformation of the data distribution by the normalization layer.

[0068] The specific two-dimensional convolutional network parameters are shown in Table 3, where D is the number of convolutional kernels, and D will be used for hyperparameter adjustment. After the network structure is determined, the backpropagation algorithm will be selected and the dataset will be used for network training and hyperparameter adjustment.

[0069] Given that there are many fault codes in Table 1, the network data volume in Table 3 is large, and Figure 4 the shown convolutional network structure is relatively complex, so AdamW is used as the backpropagation algorithm for transfer learning training. Since the scale of the training data shown in Table 2 is large, the dataset needs to be split into small batches of data to train the network, and the Batch Size is set to 128.

[0070] Under the condition of selecting the optimal hyperparameters, any small batch data scale (Batch Size) can achieve the same training performance. Therefore, the Batch Size for convolutional network training is fixed at 128, and BatchSize does not need to be adjusted during hyperparameter optimization. In addition, the hyperparameters of the two-dimensional convolutional network during 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.

[0071] Set the maximum value ( η max ), initial value ( η de ), and minimum value ( η min ) of the learning rate dynamic adjustment, and use these parameters as the hyperparameters to be adjusted for network optimization.

[0072] The two-dimensional convolutional network includes an input layer, a first convolutional module, a second convolutional module, a third convolutional module, an unfolding layer, a fully connected layer, softmax, and an output layer. The first convolutional module includes a convolutional layer, a batch normalization layer, an activation function layer, and a max pooling layer. The second convolutional module and the third convolutional module have the same structure as the first convolutional module, but different structural parameters.

[0073] Table 3 Structural Parameters of the Two-Dimensional Convolutional Network

[0074]

[0075] The hyperparameters to be adjusted are the maximum value, initial value, and minimum value of the learning rate dynamic adjustment of the backpropagation algorithm.

[0076] The method of transfer learning includes: using multiple two-dimensional convolutional networks with different numbers of convolutional kernels D as pre-trained networks, training the pre-trained networks with a data set for m Epochs (training rounds) in sequence to generate multiple pre-trained models, then loading the pre-trained models in sequence and performing transfer learning training with the data set and n Epochs (training rounds), using Optuna to automatically adjust the hyperparameters to be adjusted until the optimal hyperparameters of each pre-trained model are obtained, and then comparing the performances of each pre-trained model with the optimal hyperparameters set to obtain the optimal pre-trained model as the composite fault diagnosis model. It should be noted that m is greater than n, so as to reduce the number of Epochs required for training the diagnosis model.

[0077] As Figure 5 shown, the designed transfer learning strategy uses a two-dimensional convolutional network as the pre-trained network, and generates pre-trained models with different numbers of convolutional kernels (D) through the training set, validation set and 30 Epochs. The parameters such as the dynamic learning rate of the pre-trained model adopt the default values of PyTorch. Subsequently, use PyTorch to read these pre-trained models, and use the training set, validation set and 10 Epochs for transfer learning training again, and use Optuna 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, so as to reduce the hardware resources required for network training and improve the training efficiency. It is worth mentioning that, under the condition of determining the optimal hyperparameters, small batches of any size can achieve the same performance, and the size of the small batch for transfer learning training will be constant at 256. In summary, the designed transfer learning simplifies the hyperparameter adjustment process, and the automatic adjustment parameters of Optuna do not include the number of convolutional kernels D. In addition, transfer learning performs secondary training on the basis of the pre-trained model, reducing the number of Epochs for network training to 10, thereby improving the model training efficiency and convergence speed.

[0078] Table 4 Hyperparameter value ranges and optimal values of pre-trained models

[0079]

[0080] According to Figure 5 the network training and hyperparameter adjustment method based on transfer learning in, Table 4 presents the hyperparameter adjustment ranges of the pre-trained models and the optimal values obtained by automatic optimization of Optuna. The number of convolutional kernels is set before transfer learning, while other parameters such as the learning rate are set in the optimization of Optuna. Therefore, a total of 5 types of pre-trained networks are used for transfer learning training and automatic parameter adjustment of Optuna. Subsequently, the network model with the best performance is selected, and the corresponding learning rate and number of convolutional kernels are recorded.

[0081] To verify the advantages of the transfer learning strategy, after the transfer learning training and hyperparameter tuning are completed, the training performance of the designed composite fault diagnosis model can be compared with that of a typical convolutional network, so as to demonstrate the advantages of the composite fault diagnosis model and the transfer learning model in terms of diagnostic accuracy and training efficiency. And the corresponding comparison results are as Figures 6 - 7 shown.

[0082] In the figure, 1D-CNN and 2D-CNN adopt the training method of learning from scratch. 1D-CNN is a one-dimensional convolutional neural network, and 2D-CNN is a two-dimensional convolutional neural network. And 2D-CNN-T is a two-dimensional convolutional neural network using the transfer learning training method. In addition, the input feature maps of the three types of networks in the figure all fuse the Figure 3 current, voltage and rotational speed signals shown, and the hyperparameters of these networks are all the optimal values under the automatic hyperparameter tuning of Optuna, so as to ensure the fairness of network comparison. As Figures 6 - 7 shown, the accuracy and cross-entropy of 2D-CNN-T are both 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 about the 23rd Epoch. Therefore, the transfer learning strategy reduces the training time of the network model to 34.7% of the original. Thus, it speeds up the convergence rate of the network model in new tasks and improves the training performance and generalization ability of the network model in new tasks.

[0083] Similarly, after the network training and hyperparameter tuning are completed, the test set will be used to evaluate the generalization ability after transfer learning training. Figure 8 shows the corresponding verification results of the generalization ability, in which there are misdiagnosis phenomena for fault codes 16-20 and 27-32. In addition, the other in the figure represents other fault codes in Table 1.

[0084] According to Figure 8 it can be obtained that the accuracy of the designed transfer learning network in the test set is 97.56%, which has no obvious difference from the Figure 7 optimal accuracy in, verifying that the network model under transfer learning training has good generalization ability. In addition, to improve the real-time diagnosis accuracy of the deployed model and the credibility of the output results, it is necessary to design anti-misdiagnosis for the Figure 8 fault categories with misdiagnosis phenomena in.

[0085] In another embodiment, it is judged whether the composite fault diagnosis model needs to be lightweight designed 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.

[0086] Deploying the composite fault diagnosis model to the edge device adopts Figure 9The lightweight design and edge deployment strategy of the diagnostic model shown. Here, ONNX (Open Neural Network Exchange) is an open file format designed for deep learning, used to store trained models, which helps the effective conversion and deployment of deep learning models between different frameworks and hardware platforms. Since the GPUs and operating systems 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, facilitating 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. Subsequently, use the dataset to evaluate the performance of the inference engines with different numerical precisions (such as prediction accuracy and real-time diagnosis speed), and select the inference engine with the best performance for edge deployment.

[0087] Use the 31,500 groups of sample datasets in Table 2, with diagnostic accuracy and parameter scale as indicators, to select the optimal numerical precision. Table 5 shows the parameter scale and accuracy of the deployed models under different numerical precisions.

[0088] Table 5 Parameter Scale and Accuracy of the Lightweight Composite Fault Diagnosis Model

[0089]

[0090] According to the comparison results in the table, the lightweight network model with a numerical precision of FP16 has the smallest parameter scale and FLOPs, and there is no obvious difference in diagnostic accuracy between FP16 and FP32 and the original network model. Therefore, the lightweight network with a numerical precision of FP16 will be used as the diagnostic model for the converter composite fault. Since the precision and recall rates of the misdiagnosis codes (16 - 20 and 27 - 32) are both greater than 80%, and the sum of the misdiagnosis categories and missed diagnosis categories is 2, therefore 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 to improve the reliability of the output results of the diagnostic model. Therefore, Figure 7 The fault codes shown and the corresponding diagnostic thresholds and times are shown in Table 6, and the accuracy rates in the table are calculated from the 500 groups of sample datasets corresponding to the misdiagnosis codes in Table 2.

[0091] Table 6 Anti-Misdiagnosis Thresholds and Their Accuracy Rates of the Composite Fault Diagnosis Model

[0092]

[0093] 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 misdiagnosis probability while maximizing the diagnostic speed of the deployed model.

[0094] In another embodiment, a three-level converter composite fault online diagnosis system is involved. As Figure 10 shown, it includes a sensor monitoring network, a controller, and an edge device. The 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 state information of the three-level converter based on the current and voltage data. The edge device inputs the state information of the three-level converter into the composite fault diagnosis model to obtain a diagnosis result and feedback it to the controller.

[0095] To further verify the designed composite fault diagnosis model, it is necessary to use Figure 11 the test platform shown to experimentally verify the feasibility of this method. The transfer learning model is deployed in an embedded GPU, and the converter control strategy is burned into the converter controller. In addition, the diagnosis result of the deployed model and the real-time state information of the converter are interacted through serial communication. The converter state information is used as the online diagnosis basis for the deployed model, and the real-time diagnosis result of the deployed model can be used as the execution basis for subsequent adaptive fault tolerance strategies. It is worth mentioning that to demonstrate the feasibility of the diagnosis algorithm under various working conditions, the experimental working conditions should include steady-state and transient conditions with different speeds and torques. To achieve the above working conditions, in this embodiment, a load asynchronous motor is used as the load of the permanent magnet synchronous motor under different working conditions, and the permanent magnet synchronous motor to be measured is used 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 upper computer. At the same time, the speed instruction of the permanent magnet motor to be measured is adjusted by the converter controller to achieve the transient change of the motor at different speeds.

[0096] (1) Steady-state performance verification

[0097] Figures 12 - 15 Shows the experimental results of the designed composite fault diagnosis strategy under steady-state conditions, where T i is the diagnosis time, Count is the designed recording function. When Count reaches the diagnostic threshold set in Table 6, D e will output the corresponding diagnosis result of the deployed model, thereby minimizing the misdiagnosis probability.

[0098] According to the results of the experimental diagrams, the designed transfer learning deployment model achieved accurate identification of compound faults of power devices and sensors under various steady-state operating conditions by means of input data normalization and multi-source multi-dimensional data fusion. In addition, Figure 13 and Figure 15 The experimental results in also verified the feasibility of the designed anti-misdiagnosis strategy under various steady-state operating conditions, which helps to improve the credibility of the model diagnosis results.

[0099] Specifically, the curves in the figure are: three-phase current waveforms ( i a , i b , i c ), torque waveforms ( T m ), rotational speed waveforms ( N ), converter midpoint voltage deviation waveforms ( V er ), fault trigger signals (De), diagnosis results (flag), anti-misdiagnosis threshold count (count), and fault diagnosis waveforms ( T i represents the diagnosis time, i.e., the time difference between De and flag, which reflects the lag time between the occurrence of the fault and the fault diagnosis. The smaller it is, the better the real-time performance of the diagnosis). Specifically,

[0100] Figure 12 The experimental conditions for are N = 400 RPM, T m = 10 N·m, SNR = 20 dB, S a1 open circuit;

[0101] Figure 13 The experimental conditions for are N = 500 RPM, T m = 5 N·m, SNR = 22.5 dB, S b1 open circuit;

[0102] Figure 14 The experimental conditions for are N = 500 RPM, T m = 10 N·m, SNR = 20 dB, phase A dropout;

[0103] Figure 15 The experimental conditions for are N = 500 RPM, T m = 10 N·m, SNR = 20 dB, noise fault;

[0104] (2) Transient performance verification

[0105] In addition to the steady-state experiments, the experimental results of the designed compound fault diagnosis strategy under transient operating conditions are as shown in Figures 16 - 19As shown, the sensor and power device failures are triggered at a rotational speed of 400 r / min. According to the experimental results, in various transient operating conditions, the designed diagnostic algorithm can quickly detect and locate the composite faults of the sensor and power device, and Figure 17 and Figure 19 The experimental results verify the feasibility of the anti-misdiagnosis strategy under transient operating conditions.

[0106] Specifically, the curves in the figure are: three-phase current waveforms ( i a , i b , i c ), torque waveforms ( T m ), rotational speed waveforms ( N ), converter midpoint voltage deviation waveforms ( V er ), fault-triggering signals (De), diagnostic results (flag), anti-misdiagnosis threshold count (count), and fault diagnosis waveforms ( T i represents the diagnostic time, i.e., the time difference between De and flag, reflecting the lag time between the occurrence of the fault and the fault diagnosis. The smaller it is, the better the real-time performance of the diagnosis). Specifically,

[0107] Figure 16 The experimental conditions for m are N = 200 - 600 RPM, T a3 is open;

[0108] Figure 17 The experimental conditions for m are N = 200 - 600 RPM, T b3 is open;

[0109] Figure 18 The experimental conditions for m are N = 200 - 600 RPM, T

[0110] Figure 19 are N = 200 - 600 RPM, T m = 10 N·m, SNR = 17.5 dB, noise fault;

[0111] According to Figures 12 - 19 The experimental results show that the designed lightweight deployment model has achieved the composite fault diagnosis of the converter and sensor under different steady-state and transient operating conditions.

[0112] In another embodiment, to improve the control performance of the three-level converter under compound fault conditions, an adaptive fault-tolerant strategy can be executed according to the real-time diagnosis results of the deployment model to improve the control performance of the predictive control under the compound fault conditions of the converter.

[0113] The above-described embodiments only represent the specific implementation manners of the present application, and the description thereof is relatively specific and detailed. However, it should not be construed as a limitation on the protection scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the technical solution of the present application, several modifications and improvements can still be made, 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, 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 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; The transfer learning method comprises: using a plurality of 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 a plurality of 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; 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.

2. A three-level converter composite fault online diagnosis method according to claim 1, 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.

3. The method for online diagnosis of composite faults of a three-level converter 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.

4. A three-level converter composite fault online diagnosis method 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.

5. The method for online diagnosis of composite faults of a three-level converter 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.

6. 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 5 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