CNN-ISSA-SVM-based inverter open-circuit fault diagnosis method

Abstract

The invention discloses an inverter open-circuit fault diagnosis method based on CNN-ISSA-SVM, and aims to solve the problem that in a current inverter circuit fault diagnosis method, subjectivity of manually selecting fault features and limitation of fault small sample data have important influences on accuracy of a fault diagnosis result. The invention provides an inverter open-circuit fault diagnosis algorithm based on the combination of a convolutional neural network (CNN) and a support vector machine (SVM), and provides an inverter open-circuit fault diagnosis algorithm based on the combination of the CNN and the SVM. The CNN is used to carry out adaptive feature extraction on three-phase current data output by the load side of the inverter circuit, and then the SVM is used to classify CNN full connection layer output. In order to improve the classification performance of the SVM, an improved Sparrow Search Algorithm (ISSA) is used to realize optimization selection of a penalty parameter c and a kernel parameter g of the SVM. Experimental results show that the fault diagnosis accuracy of the SVM after parameter optimization is improved by 2.5% compared with that of a standard SVM and improved by 3.8% compared with that of a traditional CNN method, and the method shows better robustness.

Description

Technical Field

[0001] The invention relates to the technical field of fault diagnosis of power electronic equipment, and in particular to an inverter open circuit fault diagnosis method based on CNN-ISSA-SVM. Background Art

[0002] The three-phase inverter is a high-power inverter used for power supply and is widely used in transportation, aerospace, industry and other fields. However, due to the influence of high voltage, high temperature, high current and other factors in the working environment, the inverter is prone to aging or damage, which can cause inverter switch failure. If the inverter failure cannot be diagnosed and repaired in time, the secondary failure caused by it may cause the power system to paralyze, resulting in irreparable property losses and safety hazards. Therefore, it is extremely important to diagnose inverter failures quickly and accurately.

[0003] At present, data-driven methods that combine sensor detection data with machine learning algorithms have become a research hotspot in the field of fault diagnosis. Generally speaking, data-driven fault diagnosis methods first use signal processing technology to extract fault features from signals, and then use machine learning algorithms to classify the extracted fault features.

[0004] In recent years, the application of deep learning models in the field of fault diagnosis has gradually increased. In particular, convolutional neural networks (CNNs), as a key technology in the field of deep learning, have demonstrated excellent feature extraction capabilities. They can automatically extract deep features from input signals and reduce dependence on traditional signal processing methods. However, the classification layer of traditional CNNs usually consists of a fully connected layer and a Softmax classifier, which has certain limitations. The fully connected layer has many parameters and is prone to overfitting problems. On the other hand, the performance of the Softmax classifier is often inferior to that of the support vector machine (SVM) when dealing with complex multi-classification problems. Summary of the invention

[0005] The purpose of the present invention is to provide an inverter open circuit fault diagnosis method based on CNN-ISSA-SVM, which uses CNN to adaptively extract the deep features of the output current of the three-phase inverter circuit, and inputs the features into the SVM classifier optimized by ISSA for accurate classification. In addition, it also involves the diagnosis capability of the inverter open circuit fault under different noise environments, thereby demonstrating the generalization ability of the model.

[0006] The above technical objectives of the present invention are achieved through the following technical solutions: An inverter open circuit fault diagnosis method based on CNN-ISSA-SVM includes the following steps: (a) Obtain the three-phase current data output by the inverter load side; (b) extracting features from the three-phase current data by CNN, wherein the CNN includes at least two convolutional layers, a maximum pooling layer, a batch normalization layer, and a fully connected layer, and the convolutional layer uses a convolution kernel to process the input data to form an output feature map, and the output expression is: Where: and Represent the output and input feature maps of the k-th layer network respectively, represents a collection of feature maps, represents the convolution kernel weight matrix, represents the bias term, represents the activation function; (c) The penalty parameter c and kernel parameter g of SVM are optimized by using the improved ISSA algorithm. The ISSA algorithm initializes the population through the Sin chaos model to improve the diversity of the initial population, thereby enhancing the global search ability of the algorithm. The one-dimensional self-mapping of the Sin chaos model is: In the formula, in order to prevent the generation of fixed points and zero points in [-1,1], the initial value cannot be 0; (d) During the iteration process, the ISSA algorithm introduces a dynamic weight factor adaptive adjustment strategy. The weight coefficient gradually decreases with the increase of the number of iterations to enhance the efficiency and accuracy of the local search. The calculation formula of the weight coefficient is:

[0007] Where: Represents the current iteration number; represents the position information of the i-th sparrow in the j-th dimension at the t-th iteration; Indicates the warning limit; For safety margins; is a random value following a normal distribution; is the global optimal solution of the jth dimension of the previous generation; when , indicating that the current location is safe and the search range can be expanded in this area; when When , it indicates that the individual sparrow needs to move its position away from the current area; (e) The ISSA algorithm combines the Cauchy mutation strategy and the reverse learning strategy to dynamically update the target position with a certain probability, enhance the algorithm performance, and ensure that the algorithm pursues the global optimal solution through the greedy principle. Specifically, if the fitness value of the new position is lower than that of the old position, the current position is updated; (f) The optimized SVM is used to classify the features extracted by CNN to realize the inverter open circuit fault diagnosis. The SVM performs classification by maximizing the interval between the decision hyperplane and the sample. For linear inseparable problems, the slack variable and penalty factor c are introduced to convert it into a soft margin SVM. The decision hyperplane can be expressed as: Where: is a normal vector, indicating the direction of the hyperplane; is a sample feature vector; is the bias term.

[0008] Further preferably, the CNN also includes a pooling layer and a fully connected layer. The pooling layer is used to reduce redundant information in the feature map and reduce model complexity, and the fully connected layer further extracts features and inputs them into the Softmax function for classification processing.

[0009] Further preferably, when optimizing the SVM parameters, the ISSA algorithm dynamically adjusts the weight factors, integrates the Cauchy mutation and reverse learning strategies, and adopts the greedy principle to decide whether to update the current position, thereby ensuring the global and local search capabilities of the algorithm.

[0010] Further preferably, the inverter open circuit fault diagnosis method also includes a data enhancement step, which increases the diversity and quantity of the data set through signal cutting technology to enhance the generalization ability of the model.

[0011] Further preferably, the inverter open circuit fault diagnosis method also includes labeling processing of fault data to facilitate SVM training and classification.

[0012] Further preferably, the inverter open circuit fault diagnosis method constructs a three-phase inverter circuit simulation model in a Matlab / Simulink module, and sets simulation parameters, including a DC voltage source of 100V, an initial phase interval of phase-to-phase current, a carrier frequency of 6KHz, a sinusoidal modulation wave frequency of 30Hz, and a simulation time of 0.2s.

[0013] Compared with the prior art, the present invention has the following beneficial effects: First, it improves the fault diagnosis accuracy: Compared with the standard support vector machine (SVM) method, the use of SVM with optimized parameters improves the fault diagnosis accuracy by 2.5%; compared with the traditional convolutional neural network (CNN) method, the accuracy is improved by 3.8%.

[0014] Second, the robustness of the model is enhanced: this method can still maintain good diagnostic capabilities in different noise environments, indicating that it has strong robustness and can stably perform fault diagnosis in complex working environments.

[0015] Third, adaptive feature extraction is achieved: CNN is used to perform adaptive feature extraction on the three-phase current data output on the load side of the inverter circuit, which reduces the dependence on traditional signal processing methods and can automatically extract deep features from the input signal.

[0016] Fourth, the SVM parameter selection was optimized: the improved sparrow search algorithm (ISSA) was used to optimize the penalty parameter c and kernel parameter g of the SVM, which significantly improved the classification performance of the SVM and further improved the accuracy and reliability of fault diagnosis.

[0017] Fifth, it improves diagnostic efficiency: By combining the powerful feature extraction capability of CNN and the efficient classification capability of SVM, the entire fault diagnosis process is not only accurate but also efficient, which helps to quickly locate and repair open circuit faults in inverters. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a convolutional neural network structure; Figure 2 is the CNN-SVM model structure and parameters; Figure 3 It is the flow chart of CNN-ISSA-SVM; Figure 4 It is a three-phase inverter circuit simulation model; Figure 5 are the three-phase current waveforms under different faults; Figure 6 Model classification results; Figure 7 Normal operating data with noise; Figure 8 Model accuracy at different signal-to-noise ratios. DETAILED DESCRIPTION

[0019] The following is a further detailed description of an inverter open circuit fault diagnosis method based on CNN-ISSA-SVM proposed by the present invention in conjunction with the accompanying drawings and specific embodiments. According to the following description, the advantages and features of the present invention will become clearer. It should be noted that the accompanying drawings are all in a very simplified form and are not in precise proportions, which are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention. The same or similar reference numerals in the accompanying drawings represent the same or similar components.

[0020] Example 1 Basic principles of CNN-SVM algorithm 1.1 Convolutional Neural Networks As a complex multi-layer neural network model, convolutional neural network mainly includes: convolution layer, pooling layer and fully connected layer. Among them, the convolution layer uses filters to process the signal in a local range to form an output feature map. The output expression is:

[0021] Where: and Represent the output and input feature maps of the k-th layer network respectively; Represents a collection of feature maps; Represents the convolution kernel weight matrix; represents the bias term; Represents the activation function.

[0022] The pooling layer, also known as the downsampling layer, has the main function of reducing redundant information in the feature map, reducing the complexity of the model, and effectively preventing overfitting.

[0023] The fully connected layer further extracts the output features of the pooling layer and inputs these features into the Softmax function to achieve classification processing. The specific expression is as follows:

[0024] Where: is the value input to each unit of Softmax.

[0025] Figure 1 This is a common convolutional neural network structure: 1.2 Support Vector Machine SVM is a widely used linear classifier that aims to find the best decision hyperplane that maximizes the distance between samples of two categories. The decision hyperplane can be expressed as:

[0026] Where: is a normal vector, indicating the direction of the hyperplane; is a sample feature vector; is the bias term.

[0027] For linearly inseparable problems, by introducing slack variables and penalty factor c to transition to soft margin support vector machine:

[0028]

[0029] Where: For sample points The slack variables, is the regularization parameter used to balance the weight of margin and misclassification.

[0030] 2 Algorithm Improvement 2.1 Sin Chaos Initialization Population In the conventional SSA algorithm, the initial value is generated by a random function, but this method may cause an imbalance in the distribution of the population, which in turn affects the efficiency of the algorithm. This imbalanced distribution may slow down the convergence of the algorithm and even cause the algorithm to fall into a local optimal solution. Yang Haidong et al. have shown that the Sin chaos model is a chaotic model with multiple iterative mappings, which is superior to the Logistic model in terms of chaotic characteristics. Therefore, in order to improve the diversity of the initial population, thereby enhancing the global search ability of the algorithm and improving the efficiency of finding the optimal solution, the Sin chaos mapping strategy is used to initialize the population. Its one-dimensional self-mapping is as follows:

[0031] In formula (6), in order to prevent the generation of fixed points and zero points in [-1,1], the initial value cannot be 0.

[0032] 2.2 Dynamic Adaptive Weights In order to avoid the SSA algorithm falling into the local optimum or failing to converge effectively for a long time in the process of finding the optimal solution, this paper introduces a dynamic weight factor when updating the discoverer position. Adaptive adjustment strategy. The specific implementation is as follows: In the early stage of iteration, the weight factor It is assigned a higher value in order to enhance the algorithm's global search capability and improve the search speed; in the later stages of iteration, The value of is gradually reduced, the purpose is to enhance the efficiency of local search and improve the accuracy of search. The calculation formula and the improved finder position update formula are as follows:

[0033]

[0034] Where: Represents the current iteration number; represents the position information of the i-th sparrow in the j-th dimension at the t-th iteration; Indicates the warning limit; For safety margins; is a random value following a normal distribution; is the global optimal solution of the jth dimension of the previous generation; when , indicating that the current location is safe and the search range can be expanded in this area; when , it indicates that the individual sparrow needs to move its position away from the current area.

[0035] 2.3 Fusion of Cauchy mutation and reverse fusion strategies In order to prevent the sparrow search algorithm from falling into the local optimal solution during the iteration process, the algorithm is improved through the reverse learning strategy:

[0036]

[0037]

[0038] Where: represents the optimal solution of the tth generation; It represents the reverse solution of the optimal solution; and are the lower and upper bounds of the variable respectively; is a random matrix that obeys the (0, 1) uniform distribution; Control parameters for information exchange.

[0039] By integrating the Cauchy mutation strategy, the sparrow search algorithm can introduce disturbances during the iteration process, increase the diversity of the search, and help the algorithm escape the local optimal solution. Integrate this strategy into the sparrow algorithm:

[0040] Where: is the standard Cauchy distribution.

[0041] In order to further improve the efficiency and accuracy of the algorithm, this paper adopts a dynamic update strategy, which combines the reverse learning strategy and the Cauchy mutation method, and dynamically updates the target position with a certain probability to enhance the performance of the algorithm:

[0042] Where: is parameter adjustment, and its value is 0.05; Indicates the maximum number of iterations.

[0043] When performing perturbations, we first use Matlab to generate random numbers between 0 and 1, denoted as .according to The value determines which strategy to adopt. , implement the Cauchy mutation strategy; on the contrary, if , a reverse learning strategy is adopted.

[0044] The main purpose of implementing perturbations is to prevent the algorithm from wandering only in the local optimal solution, thereby improving its global search ability. However, perturbations may cause the quality of the solution to decline, affecting the overall performance of the algorithm. To solve this problem, the greedy principle is introduced to determine whether the current position needs to be updated by comparing the fitness values ​​of the new and old positions, thereby ensuring that the algorithm will not reduce efficiency due to frequent invalid perturbations in the process of pursuing the global optimal solution. The greedy principle is as follows:

[0045] Where: Represents the fitness value of x.

[0046] 2.4 ISSA algorithm optimizes SVM parameters In the SVM classification task, the kernel function g and the penalty parameter c have an important influence on its classification effect. Therefore, the improved sparrow search algorithm (ISSA) is used to optimize the SVM parameters in space.

[0047] The specific optimization process can be broken down into the following steps: a) Use the Sin chaos method to initialize the sparrow population and calculate the optimal and worst fitness values ​​and their corresponding positions; b) Update the positions of discoverers, followers and sentinels; c) According to the size of rand and Ps, select the optimal solution perturbation strategy, and decide whether to update the optimal solution position according to the greedy principle; d) Determine whether the iteration termination condition is met. If so, the algorithm terminates and outputs the optimal result; otherwise, return to step a.

[0048] 2.5 CNN-SVM Model This paper proposes an inverter fault diagnosis model based on CNN-SVM adaptive feature extraction. This model combines the advantages of CNN and SVM, aiming to use the powerful feature extraction ability of CNN and the efficient classification ability of SVM to realize inverter open circuit fault diagnosis.

[0049] The model workflow is as follows: First, the original data is subjected to convolution and pooling operations in CNN to learn and extract features that best reflect the characteristics of the original data; then, the features extracted by CNN are passed as input data to SVM for classification. The model structure is as follows: Figure 2 shown.

[0050] 2.6 CNN-ISSA-SVM fault diagnosis process Based on the algorithm improvement steps described above, this paper designs a new CNN-ISSA-SVM fault diagnosis algorithm process, such as Figure 3As shown in the figure. This algorithm combines convolutional neural network (CNN), improved sparrow search algorithm (ISSA) and support vector machine (SVM) to achieve more efficient and accurate fault diagnosis. Specific implementation process Figure 3 shown.

[0051] Step 1: The output fault sample data is divided into training set and test set according to a certain ratio; Step 2: The training set data is first processed by the convolution layer and normalization layer of CNN. The processed data forms a set containing rich fault feature information. Step 3: The extracted feature information passes through the fully connected layer, and the features are further integrated and processed to be input into the SVM model; Step 4: Use the improved sparrow search algorithm to optimize the SVM parameters to determine the best combination of c and g; Step 5: Input the test set into the trained CNN-ISSA-SVM fault diagnosis model and obtain the classification results.

[0052] 3 Fault diagnosis examples 3.1 Inverter circuit simulation and data acquisition Construct a three-phase inverter circuit simulation model in Matlab / Simulink module as follows Figure 4 As shown. Set the simulation parameters, including the DC voltage source 100V, the initial phase interval of the phase current , carrier frequency is 6KHz, sinusoidal modulation wave frequency is 30Hz, and simulation time is 0.2s.

[0053] In actual situations, the inverter open circuit fault is usually caused by a single power tube open circuit, so this article considers at most two power tubes failing at the same time. For the six power tubes in the inverter, there will be There are 22 open circuit fault states and one normal working state. In order to simulate the open circuit state of the inverter, the fault generation module is set by making the corresponding trigger pulse of the inverter always zero. By adjusting the stator inductance and stator current of the motor in the inverter circuit, data samples in different states are obtained.

[0054] The signal cutting technique is used to enhance the diversity and quantity of the data set, and the length after cutting is set to 8000, and the overlap ratio of the overlapping window is 0.1. After the above processing, the size of The 1320 sample sets are divided into The data is divided into 924 training sets and 396 test sets. Each sample is given a classification label of 0-21, where 0 represents the normal state and 1-21 represents 21 different failure modes.

[0055] Figure 5 As shown in the figure, during the normal operation of the inverter, the three-phase current on the load side presents an obvious sinusoidal waveform, and there is a phase difference of 120 degrees between the currents of each phase. When the power tube T1 fails, the corresponding A-phase current will lose the waveform characteristics of the positive half cycle, causing the currents of the other two phases to be distorted. When the power tubes T1 and T2 fail at the same time, the positive half cycle of the A-phase and B-phase currents will disappear, and the C-phase current will also be distorted accordingly. When the power tubes T1 and T4 fail at the same time, the A-phase current will become zero, and the other two-phase currents will also be distorted accordingly. When the power tubes T4 and T5 fail at the same time, the negative half cycle of the A-phase and B-phase current waveforms will be lost, and the C-phase current will be distorted accordingly with the positive half cycle of the A-phase and B-phase currents.

[0056] 3.2 Establishment of CNN-SVM fault diagnosis model The CNN feature extraction module and SVM are used to build the CNN-SVM fault diagnosis model. The detailed parameters are shown in Table 1.

[0057] During the network training process, the Adam optimizer algorithm was used and executed in a GPU environment. The maximum training batch was 200 times, and the initial learning rate was 0.01. After every 150 training cycles, the learning rate was reduced by a factor of 0.1. In order to increase the diversity of the data, the data was shuffled in each cycle. In addition, in order to control the complexity of the model, L2 regularization was also used with a regularization parameter of 0.001. At the same time, to avoid overfitting of the model, Dropout was introduced in CNN and set to 0.5.

[0058] Table 1 Parameters of CNN-SVM fault diagnosis model Network Layer parameter Input Layer Feature size: 8000×1×1 Convolutional layer 1 Convolution kernel size: 1×1 Number of convolution kernels: 32 Max Pooling Layer 1 Pooling window: 2×1 Step length: 2 Batch Normalization Layer 1 Batch size: 128 Convolutional Layer 2 Convolution kernel size: 1×1 Number of convolution kernels: 32 Max Pooling Layer 2 Pooling window: 2×1 Step length: 2 Batch Normalization Layer 2 Batch size: 128 Fully connected layer Output size: 22 SVM Classification Feature size: 22 3.3 Fault diagnosis results After the input data is processed by the fully connected layer, it is input into the SVM classifier for fault classification. In order to verify the advantages of the model proposed in this paper in classification performance, five methods will be used for classification comparison: CNN-Softmax: Use the traditional CNN-Softmax classifier for classification. CNN is used to extract the features of the input data, which are then fed into the Softmax classifier for final classification.

[0059] CNN-SVM: Unoptimized CNN-SVM is used for classification, where CNN is responsible for feature extraction and SVM is used for classification. The penalty parameter c is set to 1, the kernel function is the radial basis kernel function, and its parameter g is set to 0.1.

[0060] CNN-SSA-SVM: Based on CNN-SVM, this method introduces the sparrow search algorithm (SSA) to optimize the parameters of SVM. The population size of SSA is set to 30, the dimension is 2, and the maximum number of iterations is 30. The penalty parameter c and the kernel parameter g of SVM are optimized and searched. The specific search range is: .

[0061] CNN-ISSA-SVM: The improved sparrow search algorithm (ISSA) is used to optimize the SVM model for fault classification. The parameter settings are the same as those of method 3.

[0062] PCA-ISSA-SVM: PCA method is used for feature extraction, the variance contribution rate is set to 0.95, and ISSA-SVM is combined for classification.

[0063] Figure 6 The classification results of the above five classification methods in the fault diagnosis model are shown. As can be seen from the figure, the CNN-Softmax fault diagnosis model has an overall accuracy of 95.7071%. However, it performs poorly in the identification of simultaneous faults of two power tubes. Since the optimization goal of the Softmax classifier is to minimize the empirical risk, its generalization ability is slightly worse than that of the SVM structural risk minimization, and its ability to identify complex fault conditions is limited. In contrast, the accuracy of the CNN-SVM fault diagnosis model has improved to 96.9697%. Compared with CNN-Softmax, CNN-SVM performs better in dealing with complex situations. Although the CNN-SVM model has a high accuracy, it should be noted that selecting appropriate SVM parameters is key to maintaining this performance. If the parameters are not set properly, the model classification ability may be affected. The CNN-SSA-SVM fault diagnosis model has an accuracy of 98.2323%. This shows that optimizing SVM parameters using the SSA algorithm can effectively improve performance. The CNN-ISSA-SVM fault diagnosis model shows a higher accuracy of 99.4949%. Compared with the CNN-SSA-SVM model, the accuracy rate increased by about 1.3%. This result confirms the efficiency of the ISSA method in finding the best parameter combination. However, among the five methods, the PCA-ISSA-SVM fault diagnosis model has the lowest accuracy rate, which is only 90.9091%. For data sets with complex features, using PCA may not be able to obtain effective representations. In contrast, CNN can effectively extract and utilize multi-scale information in the data through repeated application of convolution kernels and mapping pooling, so as to better adapt to various features. At the same time, through the layer-by-layer connection of the network, CNN can also learn higher-level abstract features, thereby improving its feature extraction capabilities, making CNN superior to PCA in terms of performance and adaptability.

[0064] From the data in Table 2, the accuracy of the four models, CNN-Softmax, CNN-SVM, CNN-SSA-SVM and CNN-ISSA-SVM, on the training set is not much different. However, the performance of the CNN-ISSA-SVM model is better on the test set. This shows that the CNN-ISSA-SVM model has stronger generalization ability and higher accuracy when processing samples outside the training set. It is also worth noting that the CNN-SSA-SVN and CNN-ISSA-SVM models require a large number of parameter attempts in the process of finding the optimal parameter combination, which undoubtedly increases the training time. Despite this, there is no significant difference in test time compared with other models. This ensures the real-time use of the CNN-ISSA-SVM algorithm in actual inverter circuit fault diagnosis.

[0065] Table 2 Performance comparison of different fault diagnosis methods

[0066] 3.4 Robustness Verification Based on CNN-ISSA-SVM Noise resistance is one of the important indicators for evaluating the circuit fault diagnosis capability. In view of the fact that the output noise in the actual engineering environment includes high-frequency noise, electromagnetic interference, environmental noise, sensor noise and other forms. In order to verify the ability of the CNN-ISSA-SVM model to process Gaussian white noise with different signal-to-noise ratios, Gaussian white noise with signal-to-noise ratios of 5, 10, 15, 20, 25 and 30 dB is added to the data set. The waveforms with signal-to-noise ratios of 5, 10, 20 and 30 dB are shown in Figure 1. Figure 7 shown.

[0067] The CNN-ISSA-SVM and CNN-SSA-SVM methods are used to diagnose faults on signals with different signal-to-noise ratios. The experimental results are shown in Figure 8 As shown in the figure, it can be seen that when the signal-to-noise ratio is greater than 20, both methods can effectively diagnose the open circuit fault of the three-phase inverter circuit. However, when the signal-to-noise ratio is less than 10, the diagnostic accuracy of CNN-SSA-SVM for the open circuit fault of the three-phase inverter circuit is lower than 90%, while the accuracy of CNN-ISSA-SVM remains above 95%. This shows that the CNN-ISSA-SVM model proposed in this paper can still perform good diagnosis when the noise interference intensity is high.

[0068] 4 Conclusion This paper constructs a CNN-ISSA-SVM inverter fault diagnosis model and applies it to the diagnosis of open circuit faults in three-phase inverter circuits. The following conclusions are obtained: (1) Compared with the traditional method, this model uses the three-phase current output at the load end as the diagnostic variable, which not only reduces the hardware cost but also avoids the introduction of potential fault points and achieves better diagnostic effect.

[0069] (2) When using SVM as a classifier, its performance is better than the traditional Softmax classifier. However, the classification performance of SVM is greatly affected by parameter selection.

[0070] (3) In the process of parameter optimization, there is a risk of falling into the local optimum when using the SSA algorithm. However, by introducing dynamic inertia weights, reverse learning strategies, or Cauchy mutation perturbation strategies, the ability of the SSA algorithm can be enhanced, enabling it to jump out of the local optimal solution and find the global optimal solution.

[0071] The fault diagnosis method proposed in this paper can accurately diagnose the fault mode according to the operating data when an inverter circuit fault occurs, and has strong anti-noise ability.

[0072] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0073] The above description is only a description of the preferred embodiments of the present invention, and is not intended to limit the scope of the present invention. Any changes or modifications made by a person skilled in the art in the field of the present invention based on the above disclosure shall fall within the scope of protection of the claims.

Claims

1. A method for inverter open circuit fault diagnosis based on CNN-ISSA-SVM, It is characterized in that The following steps are involved: (a) Obtain the three-phase current data output by the inverter load side; (b) extracting features from the three-phase current data by CNN, wherein the CNN includes at least two convolutional layers, a maximum pooling layer, a batch normalization layer, and a fully connected layer, and the convolutional layer uses a convolution kernel to process the input data to form an output feature map, and the output expression is: Where: and Represent the output and input feature maps of the k-th layer network respectively, represents a collection of feature maps, represents the convolution kernel weight matrix, represents the bias term, represents the activation function; (c) The penalty parameter c and kernel parameter g of SVM are optimized by using the improved ISSA algorithm. The ISSA algorithm initializes the population through the Sin chaos model to improve the diversity of the initial population, thereby enhancing the global search ability of the algorithm. The one-dimensional self-mapping of the Sin chaos model is: In the formula, in order to prevent the generation of fixed points and zero points in [-1,1], the initial value cannot be 0; (d) During the iteration process, the ISSA algorithm introduces a dynamic weight factor adaptive adjustment strategy. The weight coefficient gradually decreases with the increase of the number of iterations to enhance the efficiency and accuracy of the local search. The calculation formula of the weight coefficient is: Where: Represents the current iteration number; represents the position information of the i-th sparrow in the j-th dimension at the t-th iteration; Indicates the warning limit; For safety margins; is a random value following a normal distribution; is the global optimal solution of the jth dimension of the previous generation; when , indicating that the current location is safe and the search range can be expanded in this area; when When , it indicates that the individual sparrow needs to move its position away from the current area; (e) The ISSA algorithm combines the Cauchy mutation strategy and the reverse learning strategy to dynamically update the target position with a certain probability, enhance the algorithm performance, and ensure that the algorithm pursues the global optimal solution through the greedy principle. Specifically, if the fitness value of the new position is lower than that of the old position, the current position is updated; (f) The optimized SVM is used to classify the features extracted by CNN to realize the inverter open circuit fault diagnosis. The SVM performs classification by maximizing the interval between the decision hyperplane and the sample. For linear inseparable problems, the slack variable and penalty factor c are introduced to convert it into a soft margin SVM. The decision hyperplane can be expressed as: Where: is a normal vector, indicating the direction of the hyperplane; is a sample feature vector; is the bias term.

2. According to claim 1, a method for diagnosing inverter open circuit faults based on CNN-ISSA-SVM, Features: The CNN also includes a pooling layer and a fully connected layer. The pooling layer is used to reduce redundant information in the feature map and reduce the complexity of the model. The fully connected layer further extracts features and inputs them into the Softmax function for classification processing.

3. According to the inverter open circuit fault diagnosis method based on CNN-ISSA-SVM in claim 1, Features: When optimizing SVM parameters, the ISSA algorithm dynamically adjusts the weight factor, integrates the Cauchy mutation and reverse learning strategies, and adopts the greedy principle to decide whether to update the current position, thereby ensuring the global and local search capabilities of the algorithm.

4. The inverter open circuit fault diagnosis method based on CNN-ISSA-SVM according to claim 1, Features: The inverter open circuit fault diagnosis method also includes a data enhancement step, which improves the diversity and quantity of the data set through signal cutting technology to enhance the generalization ability of the model.

5. The inverter open circuit fault diagnosis method based on CNN-ISSA-SVM according to any one of claims 1 to 4, Features: The inverter open circuit fault diagnosis method also includes labeling processing of fault data to facilitate SVM training and classification.

6. The inverter open circuit fault diagnosis method based on CNN-ISSA-SVM according to claim 1, Features: The inverter open circuit fault diagnosis method constructs a three-phase inverter circuit simulation model in a Matlab / Simulink module and sets simulation parameters, including a DC voltage source of 100V, an initial phase interval of phase-to-phase current, a carrier frequency of 6KHz, a sinusoidal modulation wave frequency of 30Hz, and a simulation time of 0.2s.

Images