IGBT module aging failure diagnosis system based on optimized random forest model

Abstract

An IGBT module aging fault diagnosis system based on an optimized random forest model includes: a data acquisition module, a data processing module, a model building module, a model optimization module, and an aging fault diagnosis module. This invention uses the parameter data during the operation of the IGBT module as diagnostic signals to construct an aging fault diagnosis dataset to achieve the training, construction, and optimization of the random forest model. Finally, the IGBT state diagnosis results are obtained by optimizing the random forest model.

Description

Technical Field

[0001] This invention belongs to the field of IGBT module fault diagnosis, specifically relating to an IGBT module aging fault diagnosis system based on an optimized random forest model. Background Technology

[0002] IGBT modules operate in harsh environments year-round, bearing complex cyclic stresses, leading to the continuous accumulation of fatigue damage and eventual complete aging failure. If not addressed promptly and effectively, this can progress to severe failures such as open circuits and short circuits, causing catastrophic failures. These failures can range from damaging electrical equipment and causing significant economic losses to endangering personal safety and posing serious safety hazards. Therefore, accurate diagnosis of IGBT module aging failures is crucial for improving system reliability, safety, and availability.

[0003] Currently, technologies for diagnosing IGBT module aging faults are mainly based on direct measurement methods or historical data-driven methods. Direct measurement methods rely on observations of device aging using equipment such as X-rays or acoustic microscopes to estimate the module's aging failure process. However, this approach lacks universality and struggles to adapt to real-world scenarios involving large datasets and numerous samples. Historical data-driven methods, on the other hand, build machine learning models based on the changing patterns of historical data to effectively diagnose the degree of device aging failure. However, compared to this invention, existing historical data-driven methods for IGBT module aging fault diagnosis suffer from several unresolved technical problems, including overfitting, sensitivity to spatial data density differences, cumbersome modeling processes, slow convergence, sensitivity to kernel functions, and low diagnostic accuracy on high-dimensional datasets. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes an IGBT module aging fault diagnosis system based on an optimized random forest model. The system uses parameter data during the operation of the IGBT module as diagnostic signals to construct an aging fault diagnosis dataset to train, build, and optimize the random forest model. Finally, the system obtains IGBT state diagnosis results by optimizing the random forest model.

[0005] This invention is achieved through the following technical solution:

[0006] This invention relates to an IGBT module aging fault diagnosis system based on an optimized random forest model, comprising: a data acquisition module, a data processing module, a model building module, a model optimization module, and an aging fault diagnosis module. Specifically: the data acquisition module collects IGBT module temperature, collector current, collector voltage, gate current, and gate voltage data during IGBT module operation via sensors, generating IGBT module aging fault diagnosis data signals, and outputs them to the data processing module; the data processing module standardizes the IGBT module aging fault diagnosis data signals to obtain an IGBT module aging fault diagnosis dataset, and outputs the IGBT module aging fault diagnosis dataset to the model building module; the model building module trains and constructs a random forest model using a CART decision tree as the base evaluator; the model optimization module optimizes the model using pre-pruning, cross-validation, learning curves, and Bagging resampling methods to form an optimized random forest model; and the aging fault diagnosis module outputs the aging fault diagnosis results of the IGBT module based on the optimized random forest model.

[0007] The sensors include an infrared sensor, a current sensor, and a voltage sensor, wherein the infrared sensor acquires the temperature of the IGBT module, and the current and voltage sensors acquire the collector current, collector voltage, gate current, and gate voltage data signals during the operation of the IGBT module.

[0008] The standardization process refers to centering the collected raw data signal by its mean and then scaling it by its standard deviation, so that the processed signal data follows a standard normal distribution with a mean of 0 and a standard deviation of 1. Where: X is the original data signal. For the standardized data signal, μ is the mean of all sample data signals, and σ is the standard deviation of all sample data signals.

[0009] The aforementioned IGBT module aging fault diagnosis dataset is a collection of data signals obtained after standardization processing of diagnostic signals collected by sensors.

[0010] The base estimators are multiple weak classification models of the same type that together constitute the random forest model. CART decision tree is preferred as the base estimator of the constructed random forest model. The CART decision tree is a classification and regression tree that uses Gini coefficient as a feature evaluation method.

[0011] The model optimization module uses grid search and pre-pruning methods to define the optimal parameters of the base estimator in the random forest model, and uses learning curves to determine the optimal number of base estimators in the subsequently built random forest model.

[0012] The aging fault diagnosis results are output after comprehensive calculation based on the voting results of each base evaluator in the optimized random forest model, including: the IGBT module is in normal working condition, labeled as T0; the IGBT module is in the early stage of aging, labeled as T1; and the IGBT module is in an aging fault state, labeled as T2.

[0013] Technical effect

[0014] This invention collects IGBT module aging fault diagnosis data signals from various sensors in the data acquisition module and inputs these signals into the data processing module. Next, the data processing module standardizes the IGBT module aging fault diagnosis data signals to obtain an IGBT module aging fault diagnosis dataset, which is then output to the model building module. Next, the model building module uses the IGBT module aging fault diagnosis dataset to build a traditional random forest model. Based on the traditional random forest model built in the model building module, the model optimization module applies pre-pruning and grid search methods to improve modeling efficiency and accuracy. Then, the learning curve method is used to determine the number of base estimators in the random forest model, avoiding model fitting misalignment and increased resource consumption due to improper base estimator settings. Finally, the Bagging resampling method is used to lower the average correlation coefficient between base estimators, resulting in a higher diagnostic accuracy for the random forest model, thus completing the optimization of the random forest model. Finally, the optimized random forest model is applied to the aging fault diagnosis module, which outputs the diagnostic results of IGBT aging faults using the optimized random forest model.

[0015] Based on the error distribution from multiple tests, the optimized random forest model exhibits low and stable mean squared error and mean absolute error in IGBT module aging fault diagnosis, and has the highest correlation with the dataset. Its application in IGBT device aging fault diagnosis systems has the advantages of high prediction accuracy and good fitting, demonstrating good practical value and application prospects. Attached Figure Description

[0016] Figure 1 Flowchart for diagnosing aging faults in IGBT modules;

[0017] Figure 2 Create a flowchart for the traditional random forest model;

[0018] Figure 3 A schematic diagram illustrating the comprehensive evaluation of grid search results using the base evaluator;

[0019] Figure 4 The base evaluator comprehensively evaluates the 3D surface plot;

[0020] Figure 5Optimization curve based on the number of evaluators;

[0021] Figure 6 This is a basic structural diagram of the Bagging resampling method;

[0022] Figure 7 This is a diagram of a K-fold cross-validation structure;

[0023] Figure 8 To optimize the comparison chart of evaluation results between Random Forest and other models;

[0024] Figure 9 To optimize the learning curve of random forests compared to other models. Detailed Implementation

[0025] like Figure 1 As shown in this embodiment, an IGBT module aging fault diagnosis method based on an optimized random forest model of the system is used. The method involves acquiring IGBT module aging fault diagnosis data signals through sensors in the data acquisition module; inputting the acquired data signals into the data processing module for standardization to obtain an IGBT module aging fault diagnosis dataset; inputting the obtained IGBT aging fault diagnosis dataset into the model building module to train and establish a traditional random forest model; then, in the model optimization module, optimizing the traditional random forest model obtained in the previous step using pre-pruning, grid search, learning curves, and Bagging resampling methods; thus, the establishment of the optimized random forest model is completed. Finally, the optimized random forest model is applied to the aging fault diagnosis module to output the IGBT module aging fault diagnosis results, thereby realizing the IGBT module aging fault diagnosis.

[0026] The specific steps in this embodiment include:

[0027] S1. IGBT Module Aging Fault Diagnosis Data Acquisition: In the data acquisition module, voltage, current, and infrared sensors are used to collect collector current, gate current, collector voltage, gate voltage, and module temperature of the IGBT module under normal operation, early aging, and complete aging fault states, respectively. These data are used as samples of IGBT module aging fault diagnosis data signals. Each diagnostic signal is collected at 301,680 sampling points. The IGBT module under normal operation is labeled as T0, the early aging state is labeled as T1, and the complete aging fault state is labeled as T2. The collected IGBT module aging fault diagnosis data signals are then input to the data processing module.

[0028] S2. Establishment of IGBT Module Aging Fault Diagnosis Dataset: The IGBT module aging fault diagnosis data signal input to the data processing module is centered by mean and then scaled by standard deviation, so that the processed data follows a standard normal distribution with a mean of 0 and a standard deviation of 1. That is, the IGBT module aging fault diagnosis signal data sample is standardized to obtain the IGBT aging fault diagnosis dataset, which helps to improve the accuracy and convergence speed of the final random forest model. Where: X is the original data signal. For the standardized data signal, μ is the mean of all sample data signals, and σ is the standard deviation of all data signal samples.

[0029] The standardized IGBT module aging fault diagnosis signal data set is the IGBT module aging fault diagnosis dataset, as shown in Table 1.

[0030] Table 1

[0031]

[0032] Input the obtained IGBT module aging fault diagnosis dataset into the model building module.

[0033] S3, such as Figure 2 The diagram shows the process of training and building a traditional random forest model with CART decision trees as the base evaluator using the IGBT module aging fault diagnosis dataset obtained in the previous step in the model building module. Specifically, the traditional random forest model with CART decision trees as the base evaluator is essentially a collection of multiple CART decision trees. When the number of decision trees in the random forest model is not limited and the growth process of any CART decision tree in the random forest model is not interfered with, the traditional random forest model is completed when all CART decision trees in the random forest model have completed their growth.

[0034] The aforementioned base evaluator, the CART decision tree, is a decision tree constructed based on a classification and regression tree algorithm. Specifically, the classification and regression tree algorithm is a rule that uses the Gini coefficient as the criterion for the optimal split node to construct the CART decision tree model. Taking the IGBT module aging fault diagnosis dataset as an example, the details are as follows: Where: p(x i |t) represents a sample randomly selected from the dataset at node t, and the sample belongs to category x. i The probability of tini(t) is given by tini(t), which is the tini coefficient at node t, representing the probability that two randomly drawn samples from the dataset belong to different classes. x is a calculated constant. i It contains three categories: T0, T1, and T2. Node t can be any branch node in the decision tree.

[0035] Step 3 specifically includes:

[0036] 3.1) Randomly sample the IGBT module aging fault diagnosis dataset to obtain the sample set for each CART decision tree. The number of samplings is random, and the generated sample set and the number of final generated decision trees are random. Divide the sample set of each decision tree into training set and test set in an 8:2 ratio.

[0037] 3.2) Calculate the Gini coefficients of different categories in the sample set of each CART decision tree. Taking sample set D as an example: take feature A with the smallest Gini coefficient value as the root node of the decision tree.

[0038] 3.3) Starting from the root node, divide the dataset D into two subsets according to feature A. Then, under the condition of feature A, the Gini coefficient of the sample set D is:

[0039] 3.4) The node with the minimum Gini(D, A) value is taken as the optimal splitting node of the root node of the decision tree. This node continues to split downwards until the conditions for splitting at a subsequent node cannot be met or its Gini coefficient has reached the minimum value and cannot continue to split. At this point, the decision tree stops growing, that is, the decision tree stops growing on its own rather than being manually intervened. The classification result of the node where the decision tree stops growing is the final classification result output by the decision tree.

[0040] 3.5) Repeat steps 3.3) and 3.4). After all decision trees in the model have finished growing, gather all the decision trees together to form the final traditional random forest model.

[0041] S4. In the model optimization module, optimize the traditional random forest model established in step S3, specifically including:

[0042] 4.1) First, the grid search method is used to query the parameters of the decision tree pre-pruning algorithm: the possible values ​​of each growth parameter are arranged and combined, all possible combination results are listed to build a grid, and decision trees are built one by one, and the grid search results of the prediction accuracy and running time of a single decision tree are output.

[0043] The growth parameters of the decision tree include: maximum depth, minimum number of branch nodes, and minimum number of branch samples. In this embodiment, taking an IGBT module aging fault diagnosis dataset as an example, the decision tree depth is set to a range of [1, 50] with a step size of 1; the minimum number of branch nodes is set to a range of [2, 25] with a step size of 1; the minimum number of branch samples is set to a range of [2, 25] with a step size of 1; and the values ​​of other parameters are determined based on the sample conditions. The grid search results for the prediction accuracy and running time of a single decision tree are output as follows: Figure 3 As shown.

[0044] To ensure both prediction accuracy and efficiency, the value 1 is taken as... Figure 3 The difference in prediction accuracy between each point is added to the decision tree running time, and the grid search results are output again as follows: Figure 4 As shown; take Figure 3 The 3D surface plot is generated by normalizing and connecting the data at the lowest points of all cross-sections, as shown in the figure. Figure 4 As shown, the prediction accuracy and time taken by the base estimator on this dataset are affected by the maximum depth, the number of minimum branch samples, and the number of minimum branch nodes. The 3D surface plot generally shows a trend of first decreasing and then increasing. The minimum value is obtained at the lowest point of the surface concavity. As the base estimator of the random forest algorithm, the output is the data at the lowest point of the 3D surface concavity. That is, when the maximum depth of the decision tree is 7, the number of minimum branch nodes is 15, and the number of minimum branch samples is 3, the surface reaches the minimum value. That is, the random forest model achieves the best in terms of prediction accuracy and time taken.

[0045] The normalization process involves performing a linear transformation on the data at the lowest points of all cross sections, mapping them to the range [0,1]. Where: x ′ denoted as the data at the lowest point of the cross section after normalization, x represents the data at the lowest point of the original cross section, max(x) represents the maximum value of the data at the lowest point of the original cross section, and min(x) represents the minimum value of the data at the lowest point of the original cross section.

[0046] 4.2) Apply grid search results to pre-prune CART decision trees: By setting the generation parameters of the decision tree during the tree construction process, the full growth of the tree is limited. When the growth of the base evaluator is limited, the complexity of the random forest model will also be effectively controlled, and a balance can be achieved between the training error and complexity of the decision tree. This is beneficial for the detection and correction optimization of the dataset categories: that is, extract the main rules of the dataset and discard abnormal rules, and correct hidden errors, noise and outliers in the manually set dataset labels.

[0047] Specifically, the decision tree is pre-pruned using the maximum depth of 7, the minimum number of branch nodes of 15, and the minimum number of branch samples of 3 as limiting growth parameters. That is, the decision tree stops growing when any one or more of these conditions are simultaneously met.

[0048] 4.3) After the decision trees within the traditional random forest have undergone pre-pruning, the learning curve method is used to output the framework parameter n_estimators in the random forest modeling process. This parameter represents the number of base estimators in the random forest model and is one of the important factors determining the complexity of the random forest model. An excessively large value for n_estimators will cause model misfitting, and with the increase in the number of base estimators, the time and resources consumed by the model to run normally will also increase significantly.

[0049] like Figure 5 As shown, when the value range of n_estimators is set to [1,200] and the step size is 1, the learning curve of the prediction accuracy of the output model is displayed as the number of base estimators increases. The horizontal axis represents the number of decision trees in the random forest model after applying pre-pruning, and the vertical axis represents the prediction accuracy of the random forest model. Figure 5 It can be seen that the random forest model performs best when the base estimator parameters are already optimal and the frame parameter n_estimators is 24, without needing to increase the number of base estimators. Therefore, after completing the pre-pruning of the decision trees within the traditional random forest, the number of base estimators in the random forest model is set to 24.

[0050] 4.4) After setting the number of base estimators in the random forest model, perform Bagging resampling on the random forest model: (e.g.) Figure 6 As shown, the training set is resampled with replacement during the modeling process to form multiple new datasets that are similar in size to the original training set but different from each other. Due to the randomness and independence of the resampling of the training sample set, the differences between the base estimators formed on this basis increase, and the correlation between any two base estimators is significantly reduced, thereby improving the prediction accuracy of the random forest model and forming the final optimized random forest model.

[0051] S5. In the aging fault diagnosis module, the optimized random forest model obtained in step 4 is used to output the aging fault diagnosis results of the IGBT module: the classification results of all decision trees in the optimized random forest model are voted on, and the category with the most votes in the decision tree classification results is taken as the final output result of the optimized random forest model: Where: Y(x) is the output of the optimized random forest model; yn(x) is the output of the nth decision tree in the optimized random forest, and the expression in parentheses indicates that the final classification result of the decision tree is i; λ(*) is the number of decision trees that satisfy the expression in parentheses; z is the number of categories in the optimized random forest model.

[0052] Experimental data collected through specific experiments were used for validation in a Python 3.8 and Tensorflow 2.3 environment. Cross-validation was employed to compare and analyze the prediction accuracy and fitting performance of the random forest regression model, the XGboost model, the traditional random forest model, and the optimized random forest model used in this invention.

[0053] like Figure 7 The diagram illustrates the principle of the cross-validation method. Essentially, cross-validation divides each sample set into K equal parts, sequentially using each part as the test set, and the remaining K-1 parts as the training set. After K training iterations, the mean of the test set results is used as the final model prediction output. Cross-validation allows the entire dataset to be used simultaneously as both a training and test set, thus enabling effective evaluation of the model's prediction accuracy and generalization ability.

[0054] The mean squared error (MSE), mean absolute error (MAE), and model determination coefficient (R²) of the model are used to measure the mean squared error (MSE), mean absolute error (MAE), and model determination coefficient (R²). 2 The mean squared error of the model's prediction accuracy on the training and test sets over multiple training iterations is used as the evaluation metric to compare the performance of the model established in this paper with other models in the fault diagnosis system. Mean Absolute Error Model decision coefficients Where: y i For the true value, These are the model's predicted values. This is the mean of the original dataset.

[0055] Specifically, the smaller the values ​​of MSE and MAE, the lower the R... 2 The larger the value, the higher the correlation between the model's prediction accuracy and the test data.

[0056] The output model evaluation data is shown in Table 2.

[0057] Table 2

[0058]

[0059] like Figure 8The figure shows a comparison between the model used in this invention and other models in IGBT aging fault diagnosis. As can be seen from the figure, the optimized random forest model used in this invention achieves the highest prediction accuracy on both the training and test sets. The prediction accuracy on the test set is 17.35%, 17.27%, and 1.09% higher than that of the XGboost model, random forest regression model, and traditional random forest classification model under the same conditions, respectively. From the error distribution of multiple tests, the optimized random forest model has lower and more stable mean squared error and mean absolute error, and has the highest correlation with the dataset.

[0060] The optimized random forest model and other fault diagnosis models used in this invention, after multiple cross-validation training, should have learning curves for IGBT aging fault diagnosis, as shown in the figure. Figure 9 As shown: the red and blue lines represent the changing trends of the model's prediction accuracy on the training and test sets, respectively. The vertical axis represents the prediction accuracy value, and the horizontal axis represents the number of samples in the training set.

[0061] Depend on Figure 9 It can be seen that: the random forest regression model has high prediction accuracy on the training set but low prediction accuracy on the test set, which is an overfit; the XGBoost model has poor prediction accuracy on both the training and test sets, which is an underfit; the traditional random forest model has a moderate degree of fit, but its prediction accuracy on both the training and test sets is lower than that of the optimized random forest model; the optimized random forest model has high prediction accuracy on both the training and test sets and can achieve a perfect fit on the training set, with the final difference between the two prediction curves being only 1.19%.

[0062] Compared with existing technologies, the optimized random forest model used in this invention does not suffer from redundant base estimators. The model complexity is reasonable and appropriate, enabling rapid and reliable diagnosis of IGBT module aging faults and accurate and reliable acquisition of IGBT module state information. This invention's model boasts high prediction accuracy and good fit: by selecting a traditional random forest model from numerous machine learning models and employing methods such as grid search, pre-pruning, learning curves, and Bagging resampling, this invention reduces the average correlation coefficient between decision trees in the random forest model. This paper optimizes the traditional random forest model from three aspects: improving the prediction accuracy and modeling efficiency of individual base estimators; evaluating and setting the number and parameters of base estimators in the model according to actual conditions to obtain the optimal fitting effect of the model; and achieving a balance between training difficulty and final output accuracy. An optimized random forest is then established for IGBT aging fault diagnosis. Ultimately, this system achieves 100% accuracy on the training set and 98.81% prediction accuracy on the test set. A comparison of the output learning curves of the optimized random forest model used in this invention with other types of models under the same conditions shows that the optimized random forest model used in this invention does not exhibit overfitting, underfitting, or other abnormalities, and can achieve complete fitting on the training set. The difference between the two prediction curves on the training set and the test set is only 1.19%. Thirdly, in power systems under extreme environments where operation and maintenance are inaccessible or difficult to implement, this invention can combine IGBT operating condition data collected in real-time monitoring of the power system to diagnose IGBT module aging faults, demonstrating good practical value and application prospects.

[0063] The specific implementations described herein may be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of this invention. The scope of protection of this invention is defined by the claims and is not limited by the specific implementations described herein. All implementation schemes within the scope thereof are subject to the constraints of this invention.

Claims

1. An aging fault diagnosis system for IGBT modules based on an optimized random forest model, characterized in that, include: The system comprises a data acquisition module, a data processing module, a model building module, a model optimization module, and an aging fault diagnosis module. Specifically: the data acquisition module collects IGBT module temperature, collector current, collector voltage, gate current, and gate voltage data during IGBT module operation using sensors, generating aging fault diagnosis data signals for the IGBT module and outputting them to the data processing module; the data processing module standardizes the IGBT module aging fault diagnosis data signals to obtain an IGBT module aging fault diagnosis dataset, and outputs this dataset to the model building module; the model building module trains and constructs a random forest model using CART decision trees as the base evaluator; the model optimization module optimizes the model using pre-pruning, cross-validation, learning curves, and Bagging resampling methods to form an optimized random forest model; and the aging fault diagnosis module outputs the aging fault diagnosis results of the IGBT module based on the optimized random forest model. The base evaluators are multiple weak classification models of the same type that together constitute the random forest model, with CART decision trees serving as the base evaluators for the constructed random forest model. The model optimization module uses grid search and pre-pruning methods to define the optimal parameters of the base estimator in the random forest model, and uses learning curves to determine the optimal number of base estimators in the subsequently built random forest model. The aging fault diagnosis results are output after comprehensive calculation based on the voting results of each base evaluator in the optimized random forest model, including: the IGBT module is in normal working condition, labeled as T0; the IGBT module is in the early stage of aging, labeled as T1; and the IGBT module is in an aging fault state, labeled as T2.

2. The IGBT module aging fault diagnosis system based on an optimized random forest model according to claim 1, characterized in that, The sensors include an infrared sensor, a current sensor, and a voltage sensor, wherein the infrared sensor acquires the temperature of the IGBT module, and the current and voltage sensors acquire the collector current, collector voltage, gate current, and gate voltage data signals during the operation of the IGBT module.

3. The IGBT module aging fault diagnosis system based on an optimized random forest model according to claim 1, characterized in that, The standardization process refers to centering the collected raw data signal by its mean and then scaling it by its standard deviation, so that the processed signal data follows a standard normal distribution with a mean of 0 and a standard deviation of 1. ,in: The original data signal, For the standardized data signal, The mean of all sample data signals. The standard deviation of the signals for all sample data; The aforementioned IGBT module aging fault diagnosis dataset is a collection of data signals obtained after standardization processing of diagnostic signals collected by sensors.

4. A method for diagnosing aging faults of IGBT modules based on the system described in any one of claims 1-3, characterized in that, The data acquisition module collects IGBT module aging fault diagnosis data signals from sensors. The collected data signals are then input to the data processing module for standardization to obtain the IGBT module aging fault diagnosis dataset. This dataset is then input to the model building module to train and establish a traditional random forest model. In the model optimization module, the traditional random forest model obtained in the previous step is optimized using pre-pruning, grid search, learning curves, and Bagging resampling. This completes the establishment of the optimized random forest model. Finally, the optimized random forest model is applied to the aging fault diagnosis module to output the IGBT module aging fault diagnosis results, thus achieving IGBT module aging fault diagnosis.

5. The method according to claim 4, characterized in that, specifically include: S1. IGBT Module Aging Fault Diagnosis Data Acquisition: In the data acquisition module, voltage, current, and infrared sensors are used to collect collector current, gate current, collector voltage, gate voltage, and module temperature of the IGBT module in normal operation, early aging, and complete aging fault states, respectively. These data are used as samples of IGBT module aging fault diagnosis data signals. Each diagnostic signal is collected at 301,680 sampling points. The normal operation state of the IGBT module is labeled as T0, the early aging state as T1, and the complete aging fault state as T2. The collected IGBT module aging fault diagnosis data signals are then input to the data processing module. S2. Establishment of IGBT Module Aging Fault Diagnosis Dataset: The IGBT module aging fault diagnosis data signal input to the data processing module is centered by mean and then scaled by standard deviation, so that the processed data follows a standard normal distribution with a mean of 0 and a standard deviation of 1. That is, the IGBT module aging fault diagnosis signal data sample is standardized to obtain the IGBT aging fault diagnosis dataset, which helps to improve the accuracy and convergence speed of the final random forest model. ,in: The original data signal, For the standardized data signal, The mean of all sample data signals. The standard deviation of all data signal samples; S3. The model building module uses the IGBT module aging fault diagnosis dataset to train and build a traditional random forest model with CART decision trees as the base evaluator. Specifically, the traditional random forest model with CART decision trees as the base evaluator is essentially a collection of multiple CART decision trees. When the number of decision trees in the random forest model is not limited and the growth process of any CART decision tree in the random forest model is not interfered with, the traditional random forest model is completed when all CART decision trees in the random forest model have completed growth. S4. In the model optimization module, optimize the traditional random forest model established in step S3, specifically including: S5. In the aging fault diagnosis module, the optimized random forest model obtained in step 4 is used to output the aging fault diagnosis results of the IGBT module: the classification results of all decision trees in the optimized random forest model are voted on, and the category with the most votes in the decision tree classification results is taken as the final output result of the optimized random forest model: ,in: To optimize the output of the random forest model; To optimize the first in the random forest The output of this decision tree, the expression in parentheses represents the final classification result of this decision tree. ; The number of decision trees required to satisfy the expression in parentheses; To optimize the number of categories in the random forest model.

6. The method according to claim 5, characterized in that, Step 3 specifically includes: 3.1) Randomly sample the IGBT module aging fault diagnosis dataset to obtain the sample set for each CART decision tree. The number of samplings is random, and the generated sample set and the number of final generated decision trees are random. Divide the sample set of each decision tree into training set and test set in an 8:2 ratio. 3.2) Calculate the scores of different categories in each CART decision tree sample set. Coefficients, taking sample set D as an example: take the coefficient with the smallest value. The feature A of the coefficient values ​​serves as the root node of the decision tree; 3.3) Starting from the root node, divide the dataset D into two subsets according to feature A. Then, under the condition of feature A, the Gini coefficient of the sample set D is: ; 3.4) Take The node with the minimum value is the optimal splitting node of the root node of the decision tree. This node continues to split downwards until the conditions for splitting at a subsequent node can no longer be met or its Gini coefficient has reached the minimum value and cannot continue to split. At this point, the decision tree stops growing on its own rather than through human intervention. The classification result of the node where the decision tree stops growing is the final classification result output by the decision tree. 3.5) Repeat steps 3.3) and 3.4). After all decision trees in the model have finished growing, gather all the decision trees together to form the final traditional random forest model.

7. The method according to claim 5, characterized in that, Step 4 specifically includes: 4.1) First, the grid search method is used to query the parameters of the decision tree pre-pruning algorithm: the possible values ​​of each growth parameter are arranged and combined, all possible combination results are listed to build a grid, and decision trees are built one by one and the grid search results of prediction accuracy and running time of individual decision trees are output. 4.2) Apply grid search results to pre-prune CART decision trees: By setting the generation parameters of the decision tree during the tree construction process, the full growth of the tree is limited. When the growth of the base evaluator is limited, the complexity of the random forest model will also be effectively controlled, and a balance can be achieved between the training error and complexity of the decision tree. This is beneficial for the detection and correction optimization of the dataset categories: that is, extract the main rules of the dataset and discard abnormal rules, and correct hidden errors, noise and outliers in the manually set dataset labels. 4.3) After the decision trees inside the traditional random forest have completed the pre-pruning process, the learning curve method is used to output the framework parameter n_estimators in the random forest modeling process, which is the number of base estimators in the random forest model. This parameter is one of the important factors that determine the complexity of the random forest model. If the value of parameter n_estimators is too large, it will cause the model to lose its fit. Moreover, with the increase of the number of base estimators, the time and resources consumed by the normal operation of the model will also increase significantly. 4.4) After setting the number of base estimators in the random forest model, perform Bagging resampling on the random forest model: resample the training set with replacement to form multiple new datasets that are similar in size to the original training set but different from each other. Due to the randomness and independence of the training sample set resampling, the differences between the base estimators formed on this basis increase, and the correlation between any two base estimators is significantly reduced, thereby improving the prediction accuracy of the random forest model and forming the final optimized random forest model.

Images