Multi-feature fusion power field effect transistor online state monitoring and evaluation method and system

Abstract

The invention belongs to but is not limited to the technical field of state monitoring, and particularly relates to a multi-feature fusion power field effect transistor online state monitoring and evaluation method and system, which are used for understanding the aging mechanism of a device by selecting a proper aging signal and calculating the features of the aging signal. Then, effective feature information strongly correlated with the life sequence is screened out by using a mutual information correlation coefficient algorithm; next, dimension reduction fusion is performed on the screened features by using a KPCA polynomial kernel function algorithm, so that a health index is constructed, and the degradation state of the device is reflected; and monitoring and evaluating the health state of the device by using CNN, XGBoost and RF classifiers according to the constructed health indexes so as to determine the health state of the device. In addition, the RF classification model with the best classification effect is deployed on the DSP development board, the performance of the prediction model is further optimized, and the prediction time of the model is shortened. According to the method, the health state of the power device is described by integrating the multi-feature information.

Description

Technical Field

[0001] The present invention belongs to but is not limited to the field of state monitoring and evaluation technology, and in particular relates to a multi-feature fusion power field effect transistor online state monitoring and evaluation method and system. Background Art

[0002] Electricity is now widely used in industrial production, human activities, and to improve quality of life. Efficient and reliable power electronics systems, which enable the conversion and processing of electrical energy, are crucial for smart grids and renewable energy systems. As a core technology in power electronics systems, MOSFETs play a key role in system efficiency, size, and cost, providing crucial support for optimal energy utilization and management.

[0003] Most current power device reliability research focuses on offline prediction, which is broadly divided into two areas: physical model-based and data-driven reliability prediction methods. Prediction methods based on physical failure models are typically implemented through finite element simulation, which provides in-depth analysis of aging failure mechanisms and damage evolution. This process allows for the characterization of performance parameter changes and stress distribution characteristics of bond wires and solder layers during power device failure simulations. Combined with lifetime prediction models, the service life of power devices under specific operating conditions can be further estimated. Data-driven reliability prediction methods utilize in-depth data mining during power device operation and combine it with prediction models to implement condition monitoring and assessment. Unlike traditional methods that rely on the internal workings of power devices, data-driven prediction methods leverage massive amounts of data from historical records, real-time operations, and simulation results. These data come from a wide range of sources, are diverse in type, and have complex structures.

[0004] Existing research shows that offline prediction relies on historical data, making it difficult to capture dynamic changes in devices or equipment in real time, and thus unable to identify potential gradual deterioration trends. Furthermore, offline prediction is limited by data coverage and cannot cover all operating conditions, making it difficult to fully monitor and assess equipment status. Online prediction, through continuous data collection, can identify gradual deterioration trends in equipment in real time, shifting maintenance from reactive repair to proactive prevention, effectively extending equipment life.

[0005] Given the above analysis, the urgent technical challenges facing existing technologies are: Finding an accurate and real-time state prediction method that fully utilizes existing technologies and data resources while taking into account the device's actual operating environment and application requirements. Furthermore, establishing a comprehensive data collection and management system is necessary to improve data quality and reliability, providing reliable support for the establishment and verification of state prediction models.

[0006] The difficulty of solving the above technical problems: The technology often lacks real-time performance and cannot timely assess the health status of the device, resulting in the accuracy and real-time performance of the prediction being affected. Therefore, further research and improvement of existing technologies are needed to improve the accuracy and real-time performance of MOSFET status monitoring.

[0007] The significance of solving the above technical problems: It can effectively address the problems existing in existing technologies, improve the accuracy and real-time nature of predictions, and provide important reference for device design, manufacturing, and application. By more accurately and in real time monitoring and evaluating the health status of devices, the probability of risks in power electronics systems can be reduced, maintenance costs can be lowered, and system stability can be improved. Summary of the Invention

[0008] In response to the problems existing in the prior art, the present invention provides a multi-feature fusion power field effect transistor online status monitoring and evaluation method and system.

[0009] The present invention is implemented as follows: a multi-feature fusion power field effect tube online status monitoring and evaluation method, comprising:

[0010] S1: Extract features from the dataset, select training data and test data with obvious aging trends, segment and denoise the MOSFET drain-source on-state voltage and current in the dataset, and obtain data that can better characterize MOSFET degradation;

[0011] S2: Extract 10 time-domain features from the processed data, form a 20-dimensional feature set, calculate the correlation coefficient between each feature and lifespan, screen out features with strong correlation with lifespan, perform dimensionality reduction on these features using the KPCA algorithm, and use the first-order principal component with a contribution of more than 95% as the health indicator of the MOSFET;

[0012] S3: Build CNN, XGBoost and RF classifier models;

[0013] S4: The constructed health indicators are used to classify the extracted data into normal data (0) and abnormal data (1) by setting thresholds. The filtered training set features are input into the three classifier models mentioned above for status monitoring. The classification results of the three models are compared;

[0014] S5: According to the normal data (0) in the health index, three thresholds are set to divide the state into NH1, NH2, NH3 and NH4. According to the results of the state division, the filtered training set features are input into the above three classifier models for state evaluation. The classification results of the three models are compared;

[0015] S6: Convert the RF model with the best classification performance into C code. Compile and debug the converted C code in the CCS6 compiler environment and program it onto a Texas Instruments 32-bit fixed-point 150MHz DSP TMS320F2812. The digital signal processor's digital display alternately displays the number of correct predictions, the prediction accuracy, and the run time.

[0016] Furthermore, in S2, the polynomial kernel principal component analysis method is used to perform data fusion on the matrix X composed of features with obvious trends obtained by screening, and the MOSFET health index is constructed, which includes the following steps:

[0017] (1) Combine the features with obvious trends obtained through screening into a feature matrix X;

[0018] (2) Perform a centralization operation on the feature matrix X to obtain the centralization matrix X * , eliminate the offset effect between features and obtain its transposed matrix (X * ) T , prepare to perform kernel matrix calculation;

[0019] (3) Based on the centralized matrix X * and the transposed matrix (X * ) T , and use the polynomial kernel function to calculate the kernel matrix K, centralize the kernel matrix, and generate a symmetric and centralized kernel matrix. Solve the eigenvalues ​​and corresponding eigenvectors of the centralized kernel matrix, and arrange them from large to small according to the eigenvalues ​​as λ1,λ2,...,λ n , and their corresponding eigenvectors are v1,v2,...,v n ;

[0020] (4) Arrange the eigenvalues ​​and eigenvectors in descending order, select the eigenvectors corresponding to the first three principal components, and construct the projection matrix Q.

[0021] (5) Centralize the matrix X * Project it onto the projection matrix Q formed by the selected eigenvectors to obtain the reduced-dimensional data matrix Y

[0022] (6) The first-order principal component is extracted from the data matrix Y after dimensionality reduction and normalized as a health indicator to characterize the degradation state of MOSFET.

[0023] Furthermore, S3's prediction method for the trained CNN convolutional neural network model includes:

[0024] (1) Define a CNN convolutional neural network model, including an input layer, a convolution layer, a pooling layer, and an output layer. The number of units in the input layer should be based on the characteristic parameters that are strongly correlated with lifespan, which are screened by the mutual information correlation coefficient algorithm. The convolution layer is used to extract local features in the data, and the pooling layer is used to reduce the dimensionality of features and retain important information. The number of units in the output layer is 1, which is used to output the health status category of the device.

[0025] (2) Selecting the activation function: The activation function of the convolutional layer neurons is set to the ReLU function, which is used to transform the neural network from linear transformation to nonlinear mapping, enabling the network to learn complex features and patterns. The transfer function of the output layer neurons can be selected as the softmax function to output the network's prediction results.

[0026] Furthermore, S3's prediction methods for the trained XGBoost model include:

[0027] (1) An XGBoost model is constructed based on the feature parameters that are strongly correlated with lifespan, screened using the mutual information correlation coefficient algorithm. The input features serve as the model's training data, and the weak classifier (base learner) uses a decision tree. The core of the model makes predictions through the weighted accumulation of multiple trees, with each tree focusing on correcting the error of the previous round, ultimately building a powerful classifier for predicting the device's health status category.

[0028] (2) Set the objective function of the XGBoost model to a binary:logistic function. The output is the probability of the health status category.

[0029] (3) The learning rate, maximum tree depth, number of trees, and subsample ratio are selected as model hyperparameters. Based on the training set data, the model is trained using an iterative optimization method, gradually reducing the model error through multiple rounds of improvement. The validation set is used to evaluate the model performance in real time, monitor whether overfitting occurs, and optimize the training process through early stopping.

[0030] Furthermore, S3's prediction method for the trained RF classification model includes:

[0031] (1) Obtain the data set required for the classification task and divide the data into a training set and a test set;

[0032] (2) Set model parameters, including the number of decision trees, maximum depth, splitting criteria, maximum feature tree, and random seed hyperparameters;

[0033] (3) Input the training set into the model, fit the data, and use the validation set to evaluate the model performance.

[0034] Furthermore, the method in which S6 deploys the trained RF status monitoring and evaluation model on the DSP includes:

[0035] (1) In the Python compilation environment, save the trained model as a joblib model format;

[0036] (2) Use the sklearn-porter library in Python to export the model C code;

[0037] (3) Write the exported C code into .c file and .h file;

[0038] (4) Use the exported model C code to compile and debug using the Code Composer Studio 6.1.0 (CCS6.1) platform. Burn the prediction model to the DSP development board and perform online model prediction.

[0039] Another object of the present invention is to provide a multi-feature fusion power field effect transistor online state monitoring system that implements the multi-feature fusion power field effect transistor online state monitoring and evaluation method, comprising:

[0040] Data acquisition and processing module: used to collect training data and test data with obvious aging trends, and segment and reduce noise on the MOSFET drain-source on-state voltage and current to obtain data that can more accurately reflect the degradation of the MOSFET.

[0041] Feature extraction and processing module: Extracts 10 time-domain features from the processed data to form a 20-dimensional feature set. A correlation coefficient algorithm is used to select features with strong correlations with lifespan, and the KPCA algorithm is used for dimensionality reduction to generate MOSFET health indicators.

[0042] Classification model building module: The features of the training set and the constructed health indicators are classified into different states by setting the failure threshold, and CNN, XGBoost and RF classifiers are used for training.

[0043] Device evaluation and prediction module: Input the test set features and constructed health indicators into three classification models to monitor and evaluate the health status of the device.

[0044] Digital Signal Processor Module: DSP is deployed on the RF model with the best classification performance, and the number of correctly predicted samples, prediction time, and classification accuracy are alternately displayed on the digital tube to further optimize the prediction time of the model.

[0045] Another object of the present invention is to provide a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of the multi-feature fusion power field effect transistor online status monitoring and evaluation method.

[0046] Another object of the present invention is to provide a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the multi-feature fusion power field effect transistor online status monitoring and evaluation method.

[0047] Another object of the present invention is to provide an information data processing terminal, which includes the multi-feature fusion power field effect tube online status monitoring system.

[0048] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0049] First, the present invention is committed to solving the challenge of continuous degradation of power devices under power cycling, in which a single sensor signal cannot accurately reflect its degradation state and the real-time performance of offline prediction is low. By collecting training data and test data with obvious aging trends, and segmenting and denoising the various sensor signals of MOSFET, more accurate degradation data can be obtained. Subsequently, the mutual information correlation coefficient algorithm is used to screen features, the KPCA polynomial kernel function algorithm is used to fuse features, and the first-order principal component is selected to construct a health index. Thresholds are set according to health indicators, health status divisions are performed, and prediction effects are compared using CNN, XGBoost and RF classification models. The RF classification model with the best classification effect is deployed on the DSP. This method not only improves the prediction accuracy of the model, but also improves the real-time performance of model prediction, and has broad application prospects.

[0050] The present invention conducts in-depth analysis on the data monitored during the MOSFET degradation process, and selects key characteristic parameters suitable for describing the MOSFET aging law for extraction and screening.

[0051] The power field effect transistor health status monitoring and evaluation prediction method proposed in the present invention can use multiple characteristic data to construct a health indicator that can accurately characterize the degradation state of the MOSFET, and accurately evaluate the health status of the device through multiple characteristic parameter predictors.

[0052] The present invention also proposes an online platform deployment method, which uses Python's model library to convert the trained prediction model from Python language to C language, which can be effectively run on DSP

[0053] Second, the present invention uses multi-feature fusion technology to provide accurate monitoring and assessment of the degradation state of power devices, significantly improving the reliability and safety of the devices. In practical applications, this method can significantly reduce the occurrence of system failures and safety incidents, providing users with more stable and reliable products and services. In addition, by efficiently converting the offline trained prediction model into C code and deploying it in DSP hardware, the present invention achieves a significant improvement in real-time performance, adapts to embedded system environments, reduces device energy consumption, and thus extends the service life of the device. These technical advantages bring significant commercial value to the application of power devices.

[0054] This invention pioneers the integration of multi-feature data-driven analysis with a DSP hardware development board, providing a precise and effective solution for real-time monitoring and diagnosis of MOSFET health status. By incorporating a multi-layered classification architecture, this invention enables efficient assessment of multiple health states, significantly improving diagnostic accuracy and efficiency. This innovative technology fills a technological gap in power device health monitoring both domestically and internationally, providing a new direction for industry development.

[0055] This paper addresses the accelerated aging process of MOSFETs by developing an RF online condition monitoring and assessment model based on multi-feature fusion. This comprehensive feature extraction and data analysis encompasses key data points throughout the device's lifecycle, effectively improving the system's real-time performance and accuracy. This integrated solution enables users to fully and accurately understand the operational status of power devices, providing technical support for the health management and reliability maintenance of power electronics systems and resolving a long-standing industry challenge in real-time health monitoring.

[0056] This invention effectively overcomes the limitations of traditional technologies that rely on single-sensor data through multi-sensor data fusion and advanced algorithms (such as the mutual information correlation coefficient and the KPCA polynomial kernel function), significantly improving the accuracy and comprehensiveness of degradation state monitoring assessments. Furthermore, by combining multidimensional data collected by multiple sensors, this invention optimizes the overall performance of power device health state monitoring and provides a more reliable technical solution for online monitoring and fault assessment. This innovation breaks through technological bias and sets a new benchmark for the development of industry health monitoring technology.

[0057] Third, existing technologies for monitoring MOSFET degradation status typically rely on single features or traditional methods, making it difficult to fully and accurately reflect the health status of the device. This invention extracts multidimensional features such as the MOSFET drain-source on-state voltage and current, and uses noise reduction processing and feature correlation analysis to screen out features that are highly correlated with lifespan. KPCA dimensionality reduction is then used to form health indicators. This method extracts key features from multidimensional data, significantly improving the accuracy of identifying the MOSFET aging status and overcoming the deficiency of traditional methods that can only capture local information.

[0058] This paper combines CNN, XGBoost, and RF classifier models, utilizing health indicators for condition monitoring and assessment. Comparison of multiple classifier results verifies the superiority of the RF model in accuracy and performance. Finally, the RF model is converted into C code and deployed on DSP hardware, enabling efficient, real-time online condition monitoring and assessment. Directly displaying prediction results and runtime on the digital signal processor significantly improves the real-time performance of predictions while reducing hardware resource consumption, providing strong technical support for industrial deployment.

[0059] This invention combines multi-feature fusion, machine learning classification models, and embedded hardware to achieve precise monitoring and multi-stage classification of MOSFET health status through a data-driven approach. Compared to traditional monitoring methods based on empirical rules, this invention provides a more scientific and efficient monitoring method, providing a reliable basis for power device life management and fault prediction. Furthermore, by implementing real-time deployment of the algorithm model in DSP hardware, it further promotes the development of power device health monitoring technology towards intelligence and industrialization, creating significant economic and social benefits for industrial applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a flow chart of the multi-feature fusion power field effect transistor online status monitoring and evaluation method provided by an embodiment of the present invention.

[0061] Figure 2 4 is a diagram showing the calculation results of the mutual information value correlation coefficient provided by an embodiment of the present invention.

[0062] Figure 3 This is a HI trend diagram constructed by four MOSFETs provided in an embodiment of the present invention.

[0063] Figure 4 This is a schematic diagram of the confusion matrix results obtained by the CNN, XGBoost and RF state monitoring classification prediction model power field effect tube provided by the embodiment of the present invention.

[0064] Figure 5 This is a schematic diagram of the confusion matrix results obtained by the CNN, XGBoost and RF state assessment classification prediction model power field effect tube provided by an embodiment of the present invention.

[0065] Figure 6 This is a digital tube display result diagram of the RF status monitoring classification prediction model provided by an embodiment of the present invention deployed on a DSP.

[0066] Figure 7This is a digital tube display result diagram of the RF state assessment classification prediction model provided by an embodiment of the present invention deployed on a DSP.

[0067] Figure 8 This is a structural diagram of a multi-feature fusion power field effect transistor online status monitoring system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0069] like Figure 1 As shown, the multi-feature fusion MOSFET online state monitoring and evaluation prediction method provided by the embodiment of the present invention includes the following steps:

[0070] S1: Extract features from the dataset, select training data and test data with obvious aging trends, segment and denoise the MOSFET drain-source on-state voltage and current in the dataset, and obtain data that can better characterize MOSFET degradation;

[0071] S2: Extract 10 time-domain features from the processed data, form a 20-dimensional feature set, calculate the correlation coefficient between each feature and lifespan, screen out features with strong correlation with lifespan, perform dimensionality reduction on these features using the KPCA algorithm, and use the first-order principal component with a contribution of more than 95% as the health indicator of the MOSFET;

[0072] S3: Build CNN, XGBoost and RF classifier models;

[0073] S4: The constructed health indicators are used to classify the extracted data into normal data (0) and abnormal data (1) by setting thresholds. The filtered training set features are input into the three classifier models mentioned above for status monitoring. The classification results of the three models are compared;

[0074] S5: According to the normal data (0) in the health index, three thresholds are set to divide the state into NH1, NH2, NH3 and NH4. According to the results of the state division, the filtered training set features are input into the above three classifier models for state evaluation. The classification results of the three models are compared;

[0075] S6: Convert the RF model with the best classification performance into C code. Compile and debug the converted C code in the CCS6 compiler environment and program it onto a Texas Instruments TMS320F2812 32-bit fixed-point 150MHz DSP. The digital signal processor's LEDs alternately display the number of correct predictions, the prediction accuracy, and the run time.

[0076] The present invention first extracts the MOSFET drain-source on-state voltage and current signals from the data set, removes the noise through segmentation and noise reduction processing, and extracts data that can effectively characterize the degradation state of the MOSFET. Subsequently, 10 time domain features are extracted from the processed data, and the features with strong correlation with life are screened out to form a 20-dimensional feature set. The KPCA (kernel principal component analysis) algorithm is used to reduce the dimensionality of the features, and the health index of the MOSFET is constructed by the first-order principal component with a contribution of more than 95%, providing efficient feature expression for subsequent status monitoring and evaluation.

[0077] By constructing three classifier models—CNN, XGBoost, and RF—health indicators were used for classification. First, a threshold for the health indicator was set to classify the data into two categories: normal (0) and abnormal (1). The classifier model was trained and validated using the training set. The classification results were evaluated by comparing the accuracy, operational efficiency, and other performance indicators of the three models. This process enabled online monitoring of the MOSFET status and rapid identification of abnormal conditions.

[0078] Based on the normal data (0) in the health indicator, three thresholds are set to classify the MOSFET's state into four health states: NH1, NH2, NH3, and NH4. The classified data is then fed back into three classifier models to evaluate and predict different health states. By comparing the performance of the CNN, XGBoost, and RF models in terms of classification accuracy and runtime, the best performing classifier is selected for further deployment.

[0079] The best-performing RF model was converted into C code, compiled and debugged in the CCS6 compiler environment, and a compatible embedded program was generated. This program was then burned into a Texas Instruments TMS320F2812 32-bit fixed-point 150MHz DSP processor, which enables online status monitoring and evaluation. The digital signal processor's nixie tubes alternately display the number of predicted accuracy, prediction accuracy, and run time, providing users with real-time MOSFET status information and significantly improving the efficiency and accuracy of fault warnings.

[0080] In a preferred embodiment of the present invention, step S2 uses KPCA to perform data fusion on the features strongly correlated with the lifespan to obtain a health indicator that can characterize the degradation state of the device. Specifically, the health indicator is as follows:

[0081] Step 1: Combine the features with obvious trends obtained through screening into a feature matrix X;

[0082] Step 2: Perform a centralization operation on the feature matrix X to obtain the centralization matrix X * , eliminate the offset effect between features and obtain its transposed matrix (X * ) T , prepare to perform kernel matrix calculation;

[0083] Step 3: Based on the centralized matrix X * and the transposed matrix (X * ) T , and use the polynomial kernel function to calculate the kernel matrix K, centralize the kernel matrix, and generate a symmetric and centralized kernel matrix. Solve the eigenvalues ​​and corresponding eigenvectors of the centralized kernel matrix, and arrange them from large to small according to the eigenvalues ​​as λ1,λ2,...,λ n , and their corresponding eigenvectors are v1,v2,...,v n ;

[0084] Step 4: Arrange the eigenvalues ​​and eigenvectors in descending order, select the eigenvectors corresponding to the first three principal components, and construct the projection matrix Q;

[0085] Step 5: Centralize the matrix X * Project it onto the projection matrix Q composed of the selected eigenvectors to obtain the reduced-dimensional data matrix Y;

[0086] Step 6: Extract the first-order principal component from the reduced-dimensional data matrix Y and normalize it to serve as a health indicator to characterize the degradation state of the MOSFET.

[0087] In a preferred embodiment of the present invention, the CNN neural network model construction method in step S3 is as follows: a CNN convolutional neural network model is defined, comprising an input layer, a convolutional layer, a pooling layer, and an output layer. The number of units in the input layer should be based on characteristic parameters that are strongly correlated with lifespan, as screened using a mutual information correlation coefficient algorithm. The convolutional layer is used to extract local features from the data, while the pooling layer is used to reduce the dimensionality of features and retain important information. The output layer has 1 unit, which is used to output the device's health status category.

[0088] When training a CNN, the activation function of the convolutional layer neurons is set to the ReLU function. This transforms the neural network from a linear transformation to a nonlinear mapping, enabling the network to learn complex features and patterns. The transfer function of the output layer neurons can be selected as the softmax function, which is used to output the network's prediction results.

[0089] In a preferred embodiment of the present invention, the XGBoost classification model construction method in step S3 is:

[0090] In step 1, an XGBoost model is constructed based on the feature parameters strongly correlated with lifespan, identified through the mutual information correlation coefficient algorithm. The input features serve as training data for the model, and the weak classifier (base learner) uses a decision tree. The core of the model performs predictions through the weighted accumulation of multiple trees, with each tree focusing on correcting the previous round's errors. Ultimately, a robust classifier is constructed to predict the device's health status.

[0091] Step 2: Set the objective function of the XGBoost model to binary: the logistic function output is the probability of the health status category.

[0092] In step three, select the learning rate, maximum tree depth, number of trees, and subsample ratio as model hyperparameters. Based on the training set data, use an iterative optimization approach to train the model, gradually reducing the model error through multiple rounds of boosting. The validation set is used to evaluate model performance in real time, monitor for overfitting, and optimize the training process through early stopping.

[0093] In a preferred embodiment of the present invention, the RF classification model construction method in step S3 is:

[0094] Step 1: Obtain the data set required for the classification task and divide the data into a training set and a test set;

[0095] Step 2: Set the model parameters, including the number of decision trees, maximum depth, splitting criteria, maximum feature tree, and random seed hyperparameters.

[0096] Step 3: Input the training set into the model, fit the data, and use the validation set to evaluate the model performance.

[0097] In a preferred embodiment of the present invention, the RF classification model DSP deployment method in step S6 is:

[0098] Step 1: In the Python compilation environment, save the trained model as a joblib model format;

[0099] Step 2: Use the sklearn-porter library in Python to export the model C code;

[0100] Step 3: Write the exported C code into .c file and .h file;

[0101] Step 4: Use the exported model C code on the CCS6.1 platform to compile and debug. Burn the prediction model to the DSP development board and perform online model prediction.

[0102] The technical effects of the present invention are described in detail below in conjunction with tests.

[0103] The data used in this example comes from accelerated aging experiments on MOSFET devices conducted by the NASA Ames Center of Excellence in Prediction. This dataset represents aging experiments of power MOSFETs under power cycling. To assess the degradation of the MOSFET, two different types of data were obtained: voltage sensors and current sensors. The dataset records the different sensor data from the start of the MOSFET operation until it reaches the aging limit.

[0104] The data under the same working condition are divided into training set and test set, and the results obtained by the steps of the present invention are as follows: Figure 4 The confusion matrix results obtained by using three prediction models for health status monitoring are shown in the figure. Figure 5 The confusion matrix results obtained by using three prediction models to evaluate health status are shown in the figure. Figure 6 This is the digital tube display result of the RF classification model deployed on the DSP development board.

[0105] Application Examples: In order to demonstrate the creativity and technical value of the technical solution of the present invention, this section provides application examples of the claimed technical solution on specific products or related technologies.

[0106] The multi-feature fusion power field effect transistor health status monitoring and evaluation prediction method provided in the application embodiment of the present invention is applied to a computer device, and the computer device includes a memory and a processor, and the memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of the multi-feature fusion power field effect transistor health status monitoring and evaluation prediction method.

[0107] The multi-feature fusion power field effect transistor health status monitoring and evaluation prediction method provided by the application embodiment of the present invention is applied to an information data processing terminal, and the information data processing terminal is used to implement the multi-feature fusion power field effect transistor health status monitoring and evaluation prediction system.

[0108] The data used in this example comes from accelerated aging experiments on MOSFET devices conducted by the NASA Ames Center of Excellence in Prediction. This dataset represents aging experiments of power MOSFETs under power cycling. To assess the degradation of the MOSFET, two different types of data, voltage sensors and current sensors, were used. The dataset records the different sensor data from the start of the MOSFET operation until it reaches the aging limit.

[0109] The acquired two sensor data are analyzed, and 20 time domain features suitable for the two sensor data are selected for extraction. Then, the features with correlation greater than 4 are screened out through the mutual information correlation coefficient algorithm. The calculation results are as follows: Figure 2 As shown; and through the KPCA polynomial kernel function algorithm, the HI of the four MOSFET tubes constructed is as follows Figure 3 As shown in the figure; Condition monitoring and evaluation are carried out in three classification prediction models as shown in the figure. Figure 4 and Figure 5 shown.

[0110] like Figure 8 As shown, the multi-feature fusion power field effect transistor online status monitoring system provided by the embodiment of the present invention includes:

[0111] Data acquisition and processing module: used to collect training data and test data with obvious aging trends, and segment and reduce noise on the MOSFET drain-source on-state voltage and current to obtain data that can more accurately reflect the degradation of the MOSFET.

[0112] Feature extraction and processing module: Extracts 10 time-domain features from the processed data to form a 20-dimensional feature set. A correlation coefficient algorithm is used to select features with strong correlations with lifespan, and the KPCA algorithm is used for dimensionality reduction to generate MOSFET health indicators.

[0113] Classification model building module: The features of the training set and the constructed health indicators are classified into different states by setting the failure threshold, and CNN, XGBoost and RF classifiers are used for training.

[0114] Device evaluation and prediction module: Input the test set features and constructed health indicators into three classification models to monitor and evaluate the health status of the device.

[0115] Digital Signal Processor Module: DSP is deployed on the RF model with the best classification performance, and the number of correctly predicted samples, prediction time, and classification accuracy are alternately displayed on the digital tube to further optimize the prediction time of the model.

[0116] An application embodiment of the present invention provides a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of a multi-feature fusion power field effect transistor online status monitoring and evaluation method.

[0117] An application embodiment of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of a method for online status monitoring and evaluation of a power field effect transistor with multi-feature fusion.

[0118] An application embodiment of the present invention provides an information data processing terminal, which includes a multi-feature fusion power field effect transistor online status monitoring system.

[0119] The application embodiments of the present invention can be applied to power electronic equipment, new energy vehicles, aerospace, industrial automation, renewable energy and other related fields, and have broad application prospects in improving system reliability and reducing maintenance costs.

[0120] The embodiment of the present invention first performs data preprocessing on the acquired raw data, including noise reduction and normalization, and then filters the processed data, and then extracts 10 time domain features from the filtered signals. This embodiment uses the mutual information method to filter signals to obtain features with a correlation with the life sequence greater than 4. The filtered features are fused with KPCA features and HI is constructed. The health status is divided according to HI and labeled. The filtered features are input into the classification model as a feature set for classification. The embodiment of the present invention uses three models, namely: CNN, XGBoost and RF. The classification results obtained are compared. The RF model is better than the other three models in terms of accuracy and time. Figure 4 Display RF status monitoring model R 2 The prediction result is 0.965. Figure 5 Display RF status evaluation model R 2 The prediction result was 0.9236. Therefore, the RF model was ultimately deployed on a digital signal processor for online health status monitoring and assessment. The classification results and time were displayed on a digital tube. A certain number of LEDs were turned on in each state to distinguish different health states, allowing staff to take appropriate maintenance measures in a timely manner.

[0121] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art will appreciate that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or contained in processor control code, for example, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0122] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A multi-feature fusion power field effect tube online status monitoring and evaluation method, characterized in that: include: S1: Extract features from the dataset, select training data and test data with obvious aging trends, segment and denoise the MOSFET drain-source on-state voltage and current in the dataset, and obtain data that can better characterize MOSFET degradation; S2: Extract 10 time-domain features from the processed data, form a 20-dimensional feature set, calculate the correlation coefficient between each feature and lifespan, screen out features with strong correlation with lifespan, perform dimensionality reduction on these features using the KPCA algorithm, and use the first-order principal component with a contribution of more than 95% as the health indicator of the MOSFET; S3: Build CNN, XGBoost and RF classifier models; S4: The constructed health indicators are used to classify the extracted data into normal data (0) and abnormal data (1) by setting thresholds, and the filtered training set features are input into the above three classifier models for status monitoring; the classification results of the three models are compared; S5: According to the normal data (0) in the health index, three thresholds are set to divide the state into NH1, NH2, NH3 and NH4; according to the results of the state division, the filtered training set features are input into the above three classifier models to perform state evaluation; and the classification results of the three models are compared; S6: Convert the RF model with the best classification performance into C code; compile and debug the converted C code in the CCS6 compilation environment, and burn the program into the 32-bit fixed-point 150MHz DSP TMS320F2812 developed by Texas Instruments; alternately display the number of accurate predictions, prediction accuracy, and running time on the digital signal processor's digital tube.

2. The multi-feature fusion power field effect tube online status monitoring and evaluation method according to claim 1, characterized in that: In S2, the polynomial kernel principal component analysis method is used to perform data fusion on the matrix X composed of features with obvious trends obtained by screening, and the MOSFET health index is constructed, which includes the following steps: (1) Combine the features with obvious trends obtained through screening into a feature matrix X; (2) Perform a centralization operation on the feature matrix X to obtain the centralization matrix X * , eliminate the offset effect between features and obtain its transposed matrix (X * ) T , prepare to perform kernel matrix calculation; (3) Based on the centralized matrix X * and the transposed matrix (X * ) T , and use the polynomial kernel function to calculate the kernel matrix K, centralize the kernel matrix, and generate a symmetric and centralized kernel matrix; solve the eigenvalues ​​and corresponding eigenvectors of the centralized kernel matrix, and arrange them from large to small according to the eigenvalues ​​as λ1,λ2,...,λ n , and their corresponding eigenvectors are v1,v2,...,v n ; (4) Arrange the eigenvalues ​​and eigenvectors in descending order, select the eigenvectors corresponding to the first three principal components, and construct the projection matrix Q; (5) Centralize the matrix X * Project it onto the projection matrix Q formed by the selected eigenvectors to obtain the reduced-dimensional data matrix Y (6) The first-order principal component is extracted from the data matrix Y after dimensionality reduction and normalized as a health indicator to characterize the degradation state of MOSFET.

3. The multi-feature fusion power field effect tube online status monitoring and evaluation method according to claim 1, characterized in that: S3's prediction methods for the trained CNN convolutional neural network model include: (1) Define a CNN convolutional neural network model, including input layer, convolution layer, pooling layer and output layer; the number of input layer units should be based on the characteristic parameters that are strongly correlated with lifespan screened by the mutual information correlation coefficient algorithm; the convolution layer is used to extract local features in the data, and the pooling layer is used to reduce the dimension of features and retain important information; the number of output layer units is 1, which is used to output the health status category of the device; (2) Select activation function: The activation function of the convolutional layer neurons is set to the relu function, which is used to transform the neural network from linear transformation to nonlinear mapping, so that the network can learn complex features and patterns; the transfer function of the output layer neurons can be selected as the softmax function to output the network's prediction results.

4. The multi-feature fusion power field effect tube online status monitoring and evaluation method according to claim 1, characterized in that: S3's prediction methods for the trained XGBoost model include: (1) An XGBoost model is constructed based on the characteristic parameters that are strongly correlated with lifespan, which are screened by the mutual information correlation coefficient algorithm. The input features are used as training data for the model, and the weak classifier uses a decision tree. The core of the model makes predictions through the weighted accumulation of multiple trees, with each tree focusing on correcting the error of the previous round, ultimately building a powerful classifier for predicting the health status category of the device. (2) The objective function of the XGBoost model is set to a binary:logistic function; the output is the probability of the health status category; (3) The learning rate, maximum tree depth, number of trees, and subsample ratio are selected as hyperparameters of the model; based on the training set data, the model is trained using an iterative optimization method, and the model error is gradually reduced through multiple rounds of improvement; the validation set is used to evaluate the model performance in real time, monitor whether overfitting occurs, and optimize the training process by early stopping.

5. The multi-feature fusion power field effect tube online status monitoring and evaluation method according to claim 1, characterized in that: S3's prediction methods for the trained RF classification model include: (1) Obtain the data set required for the classification task and divide the data into a training set and a test set; (2) Set model parameters, including the number of decision trees, maximum depth, splitting criteria, maximum feature tree, and random seed hyperparameters; (3) Input the training set into the model, fit the data, and use the validation set to evaluate the model performance.

6. The multi-feature fusion power field effect tube online status monitoring and evaluation method according to claim 1, characterized in that: S6 deploys the trained RF status monitoring and evaluation model on the DSP in the following ways: (1) In the Python compilation environment, save the trained model as a joblib model format; (2) Use the sklearn-porter library in Python to export the model C code; (3) Write the exported C code into .c file and .h file; (4) Using the Code Composer Studio 6.1.0 platform, the exported model C code is compiled and debugged; the prediction model is burned onto the DSP development board for online model prediction.

7. A system for implementing the multi-feature fusion power field effect transistor online status monitoring and evaluation method according to any one of claims 1 to 6, characterized in that: include: Data acquisition and processing module: used to collect training data and test data with obvious aging trends, and segment and reduce noise on the MOSFET drain-source on-state voltage and current to obtain data that more accurately reflects the MOSFET degradation; Feature extraction and processing module: Extracts 10 time-domain features from the processed data to form a 20-dimensional feature set. It uses a correlation coefficient algorithm to select features that are highly correlated with lifespan, and uses the KPCA algorithm for dimensionality reduction to generate MOSFET health indicators. Classification model building module: This module divides the features of the training set and the constructed health indicators into different states by setting failure thresholds, and uses CNN, XGBoost, and RF classifiers for training. Device evaluation and prediction module: Inputs the test set features and constructed health indicators into three classification models to monitor and evaluate the health status of the device; Digital Signal Processor Module: DSP is deployed on the RF model with the best classification performance, and the number of correctly predicted samples, prediction time, and classification accuracy are alternately displayed on the digital tube to further optimize the prediction time of the model.

8. A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the multi-feature fusion power field effect transistor online status monitoring and evaluation method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to execute the steps of the multi-feature fusion power field effect transistor online status monitoring and evaluation method as described in any one of claims 1 to 6.

10. An information data processing terminal, comprising the multi-feature fusion power field effect transistor online status monitoring and evaluation system according to claim 7.

Images