A method for assessing the health status and predicting the remaining useful life of power MOSFETs

Abstract

This invention belongs to the field of power device degradation monitoring technology and discloses a method for assessing the health status and predicting the remaining lifespan of power MOSFETs. The method involves selecting appropriate aging signals based on aging mechanisms to calculate their characteristics; using the Pearson algorithm to filter out effective features strongly correlated with lifespan; employing the KPCA algorithm to perform dimensionality reduction and fusion on the filtered features, and constructing health indicators from the dimensionality-reduced results; using the k-means++ algorithm to assess the state of the constructed health indicators; and using a BILSTM neural network to predict the remaining lifespan of the device from the filtered effective feature parameters strongly correlated with lifespan. This invention achieves the characterization of the degradation state and prediction of lifespan of power devices through the fusion of multi-source information.

Description

Technical Field

[0001] This invention belongs to the field of power device degradation monitoring technology, and in particular relates to a method for assessing the health status and predicting the remaining service life of power MOSFETs based on multi-source data fusion. Background Technology

[0002] Power MOSFETs (Metal-Oxide-Semiconductor Field-Effect Transistors) play a crucial role in electronic systems, and their reliability directly impacts the stable operation of the entire system. Neglecting the aging and failure issues of power MOSFETs can lead to system malfunctions and even safety accidents, posing significant safety hazards to people's lives. Therefore, reliability analysis and remaining lifespan prediction of power MOSFET devices are of great importance. Typically, during the device aging process, signals such as voltage, current, and resistance measured by sensors contain MOSFET aging-related information and can therefore be used to monitor the device's aging status.

[0003] Currently, methods for predicting the remaining lifespan of power semiconductor devices, both domestically and internationally, can be broadly categorized into two types. The first type is data-driven, which trains the model on a selected normal modeling dataset to reveal the degradation patterns of power semiconductor devices and thus predict their lifespan. Data-driven prediction methods employ multiple sensors to monitor device degradation depending on the operating environment. However, current technology typically uses a single sensor signal with a specific characteristic parameter to characterize the device's degradation state. The accuracy of this monitoring is limited by the precision of a single sensor and cannot effectively detect device degradation. Data quality is crucial to the accuracy of the model. Poor data quality or a lack of representativeness can distort the model's predictions.

[0004] The second approach is based on finite element method (FEM) simulation analysis. It utilizes FEM software to construct a multiphysics coupled model of the power semiconductor device and employs the software's built-in fatigue damage criteria to predict the device's remaining lifespan. Building an accurate physical model requires a deep understanding of the device's structure, materials, and operating conditions, factors that may not always be fully clear or reliable. Therefore, the model's accuracy can be affected by assumptions and simplifications made during the modeling process. FEM simulations typically require significant computational resources and time, especially for complex devices and large-scale simulations. This can make the analysis process very time-consuming and place high demands on computational resources. Verifying the model's accuracy may require extensive experimental data for comparison, but the availability of experimental data may be limited. Furthermore, model verification may require testing under numerous different operating conditions and environments, which can increase costs and time.

[0005] Based on the above analysis, the existing technologies have the following problems and shortcomings: Existing data-driven methods have limitations in characterizing the degradation state of power devices. Health indicators used to characterize device degradation cannot accurately represent the device's health status due to the single data source and the single characteristic aging parameters. Finite element simulation analysis methods require significant computational resources and time, as well as large amounts of data and testing under various operating conditions and environments, which greatly increases experimental costs.

[0006] The difficulty in solving the above technical problems lies in effectively utilizing the signals collected by multiple sensors, extracting multiple feature parameters from the signals, and establishing a health index that can better characterize the degradation state of the device.

[0007] The significance of solving the above technical problems: The accuracy of existing technologies is limited by the precision of a single sensor, and cannot monitor the degradation process of devices from multiple perspectives. Effectively fusing information from multiple sensors can provide a better understanding of the MOSFET degradation process. Summary of the Invention

[0008] To address the problems existing in the prior art, this invention provides a method for assessing the health status and predicting the remaining service life of power MOSFETs.

[0009] This invention is implemented as follows: a method for assessing the health status and predicting the remaining service life of power MOSFETs, comprising the following steps:

[0010] The first step is to acquire training data and test data with obvious aging trends, and to segment and denoise the MOSFET drain-source on-state voltage, current and resistance acquired by the sensor to obtain data that can better characterize MOSFET degradation.

[0011] The second step involves extracting 14 time-domain features from the processed data, forming a 42-dimensional feature set. The correlation between each feature and lifetime is calculated using the Pearson algorithm, and features with strong correlation to lifetime are selected. The selected features are then subjected to dimensionality reduction using the KPCA (Kernel Principal Component Analysis) algorithm. The first-order principal components with a contribution rate of 88% are used as the health indicators of the MOSFET. The selected features and the constructed health indicators in the training set are then used to perform state division using a clustering algorithm. At the same time, the selected feature data is used to train a BILSTM (Bidirectional Long Short-Term Memory) neural network.

[0012] The third step involves inputting the features selected from the test set and the constructed health indicators into the trained BILSTM neural network and clustering algorithm model to further evaluate the device's status and predict its remaining lifespan.

[0013] Furthermore, the second step involves data fusion of the matrix Q, composed of features with obvious trends obtained through kernel principal component analysis, including the following steps:

[0014] (1) Combine features with obvious trends into matrix Q;

[0015] (2) The centered data matrix Q is obtained as matrix Q. * Then, for matrix Q * Transpose to obtain matrix (Q) * ) T ;

[0016] (3) Based on matrix Q * Sum matrix (Q) * ) T Calculate the kernel matrix K, and let the eigenvalues ​​of matrix K be arranged in descending order as λ1, λ2, ..., λ n The eigenvectors corresponding to its eigenvalues ​​are α1, α2, ..., α n ;

[0017] (4) Construct the projection matrix W based on the eigenvectors corresponding to the first three largest eigenvalues. Typically, the eigenvalues ​​and eigenvectors are arranged in descending order, and the eigenvectors corresponding to the first three largest eigenvalues ​​are selected.

[0018] (5) The centered data matrix Q * Projecting the data onto the selected first three feature vectors yields the dimensionality-reduced data matrix Y;

[0019] (6) The first-order principal component normalization value is selected as the health index to characterize the degradation state of the MOSFET.

[0020] Furthermore, the third step involves inputting the health indicators constructed from the training set data into the k-means++ (K-means improved) algorithm to obtain the cluster centers for each training data point, and then averaging the cluster centers as the initial cluster centers for the test set. This includes the following steps:

[0021] (1) Randomly select the first center point c1 from the dataset;

[0022] (2) For each data point x i Calculate the distance d(x) between it and the nearest center point in the already selected center point set C. i ,C);

[0023] (3) Select the next center point c from the dataset using a probability distribution. i This ensures that the probability of selecting each point is proportional to the square of its distance from the nearest selected center point;

[0024] (4) Repeat steps 2 and 3 until k center points are selected, where k is the number of clusters specified by the user;

[0025] (5) Perform the standard iterative process of the k-means clustering algorithm using the selected initial cluster centers.

[0026] The distance from each data point to the selected center point set C is calculated as shown in Equation 1:

[0027]

[0028] This invention performs k-means++ clustering on the HI (health indicators) constructed from the MOSFET data in the training set. After obtaining the final cluster centers of the training set data, the average value is taken, and then the averaged cluster centers are used as the initial cluster centers for k-means clustering of the test set data. This not only guides the state classification of the test devices using historical data from other devices, but also avoids the serious error in clustering results caused by improper initialization of cluster centers in the k-means algorithm.

[0029] Furthermore, the method for constructing the remaining useful life prediction model based on the BILSTM neural network in the third step includes:

[0030] (1) Establish a BILSTM neural network model, which includes a three-layer BILSTM neural network with an input layer, a hidden layer and an output layer. In this model, the input layer consists of feature parameters that are strongly related to lifetime selected by the Pearson algorithm, and the output layer has 1 unit. The input layer unit is a feature that characterizes the degradation state of the device, and the output layer unit is the lifetime value of the device.

[0031] (2) Train the BILSTM neural network. The transfer function of the hidden layer neurons in the BILSTM neural network is the sigmoid function, and the transfer function of the output layer neurons is the ReLU (Rectified Linear Unit) function, which is used to output the prediction results of the network. The training function adopts the Adam (Adaptive Moment Estimation) gradient descent algorithm, and the initial weights are selected as random numbers between [0,1]. The established BILSTM neural network is trained by continuously improving the weights and thresholds in the BILSTM neural network model until convergence.

[0032] Furthermore, the method for predicting the trained BILSTM neural network model in the third step includes: given the corresponding prediction sensor data, obtaining a feature dataset with obvious trends, and then calculating the output layer value through the BILSTM neural network model, which is the prediction result.

[0033] Another object of the present invention is to provide a system for assessing the condition and predicting the remaining useful life of a numerically controlled power device using the aforementioned power MOSFET health status assessment and remaining useful life prediction method, the system comprising:

[0034] The data acquisition and processing module is used to acquire training data and test data with obvious aging trends, and to segment and reduce noise in the MOSFET drain-source on-state voltage, current and resistance acquired by the sensor to obtain data that can better characterize MOSFET degradation.

[0035] The feature extraction and processing module is used to extract 14 time-domain features from the processed data, form a 42-dimensional feature set, calculate the correlation between each feature and lifetime using the Pearson algorithm, select features with strong correlation to lifetime, reduce the dimensionality of the selected features using the KPCA algorithm, and use the first-order principal component as the health index of the MOSFET.

[0036] The neural network training module is used to divide the state of the features selected in the training set and the constructed health indicators through a clustering algorithm, and to train the BILSTM neural network using the selected feature data.

[0037] The device evaluation and prediction module is used to input the features selected from the test set and the constructed health indicators into the trained BILSTM neural network and clustering algorithm model to further evaluate the device's status and predict its remaining lifespan.

[0038] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0039] First, addressing the technical problems existing in the prior art and the difficulty of solving them, this paper closely analyzes, in conjunction with the technical solution to be protected by this invention and the results and data obtained during the research and development process, how the technical solution of this invention solves the technical problems, and the inventive technical effects brought about by solving these problems. The specific description is as follows:

[0040] The main technical problem addressed by this invention is that a single sensor signal cannot accurately characterize the degradation state of a MOSFET or predict its remaining lifetime. Furthermore, the thermal cycling generated during MOSFET operation under power cycling causes continuous degradation. In this process, multiple sensors measure various signals. The problem this invention aims to solve is how to utilize these multiple signals to construct a health index that accurately characterizes the device's degradation state and uses these signals to predict the remaining lifetime of the MOSFET.

[0041] This invention acquires training and test data with obvious aging trends, and segments and denoises the MOSFET drain-source on-state voltage, current, and resistance collected by sensors to obtain data that can better characterize MOSFET degradation. Time-domain features are extracted from the above data, and the Pearson algorithm is used for feature selection. The KPCA algorithm is then used to fuse the selected features and select the first-order principal components as the construction of health indicators. The constructed health indicators are then used for state assessment, and the selected features are input into a BILSTM neural network to predict lifetime. This not only effectively improves the model's predictive ability but also broadens its application scope. After adjustments, it can be widely applied to the state assessment and remaining lifetime prediction of various power devices.

[0042] Second, considering the technical solution as a whole or from the perspective of the product, the technical effects and advantages of the technical solution to be protected by this invention are specifically described as follows:

[0043] This invention analyzes the data monitored during the MOSFET degradation process and extracts and filters characteristic parameters suitable for the MOSFET aging process.

[0044] The proposed method for assessing the health status and predicting the remaining lifespan of power MOSFETs effectively utilizes multi-source monitoring data to construct health indicators characterizing the degradation state of MOSFETs and uses multiple characteristic parameters to predict the remaining lifespan of the devices.

[0045] This invention fully utilizes the characteristics of sensor signals, is simple to operate, has wide applicability, and can be widely applied to the condition assessment and remaining service life prediction of various power devices.

[0046] Third, as supplementary evidence of the inventive step of the claims of this invention, it is also reflected in the following important aspects:

[0047] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0048] Improving Device Reliability and Safety: This invention utilizes multi-source data fusion technology to effectively monitor the degradation state of power devices and predict their remaining lifespan, thereby improving device reliability and safety. In practical applications, this will reduce the risk of system failures and safety incidents, providing users with more reliable products and services. Reducing Maintenance Costs: By predicting the remaining lifespan of power devices in advance, maintenance plans can be rationally scheduled, reducing unnecessary maintenance and replacement costs. This will result in significant cost savings and improve the economic efficiency of the equipment.

[0049] (2) Does the technical solution of the present invention overcome technical bias?

[0050] This invention makes full use of information collected by multiple sensors and adopts data processing techniques such as Pearson algorithm and KPCA algorithm, overcoming the limitations of existing technologies that mostly rely on data from a single sensor, and improving the accuracy and comprehensiveness of the degradation state of power devices. Attached Figure Description

[0051] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart of the method for assessing the health status and predicting the remaining service life of a power MOSFET provided in an embodiment of the present invention.

[0053] Figure 2 This is the aging feature map selected from the embodiments of the present invention.

[0054] Figure 3 This is the first three principal component diagrams after data fusion provided in the embodiments of the present invention.

[0055] Figure 4 This is a state evaluation diagram provided in an embodiment of the present invention.

[0056] Figure 5 This is a comparison chart of the prediction of remaining useful life using three different neural networks, provided in an embodiment of the present invention.

[0057] Figure 6 This is a structural diagram of a power MOSFET health status assessment and remaining service life prediction system based on multi-source data fusion provided in an embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0059] like Figure 1 As shown, the method for assessing the health status and predicting the remaining useful life of a power MOSFET provided in this embodiment of the invention includes the following steps:

[0060] S1: The MOSFET drain-source on-state voltage, current and resistance acquired by the sensor are segmented and noise-reduced to obtain data that can better characterize MOSFET degradation;

[0061] S2: Extract 14 temporal features from the data processed above;

[0062] S3: The extracted features are formed into a 42-dimensional feature set. The correlation between each feature and lifespan is calculated using the Pearson algorithm, and features with strong correlation to lifespan are selected.

[0063] S4: The selected features are dimensionality reduced using the KPCA algorithm. First-order principal components are used as the health indicators of the MOSFET. The health indicators constructed from the training data are divided into states using a clustering algorithm. At the same time, the selected feature data is used to train the BILSTM neural network.

[0064] S5: Input the features selected from the test set and the constructed health indicators into the trained BILSTM neural network and clustering algorithm model to further evaluate the device's status and predict its remaining lifespan.

[0065] In a preferred embodiment of the present invention, step S4 uses KPCA to fuse the obtained features that are strongly correlated with lifetime, and inputs the fused data into a clustering algorithm for training to obtain health indicators that can characterize the degradation state of the device, as follows:

[0066] Step 1: Combine features with obvious trends into matrix Q;

[0067] Step 2: Centralize the data matrix Q to obtain matrix Q. * Then, for matrix Q * Transpose to obtain matrix (Q) * ) T ;

[0068] Step 3, based on matrix Q * Sum matrix (Q) * ) T Calculate the kernel matrix K, and let the eigenvalues ​​of matrix K be arranged in descending order as λ1, λ2, ..., λ nThe eigenvectors corresponding to its eigenvalues ​​are α1, α2, ..., α n ;

[0069] Step four: Construct the projection matrix W based on the eigenvectors corresponding to the first three largest eigenvalues. Typically, the eigenvalues ​​and eigenvectors are arranged in descending order, and the eigenvectors corresponding to the first three largest eigenvalues ​​are selected.

[0070] Step 5, center the data matrix Q * Projecting the data onto the selected first three feature vectors yields the dimensionality-reduced data matrix Y;

[0071] Step 6: Select the first-order principal component normalized value as a health indicator to characterize the degradation state of the MOSFET.

[0072] In a preferred embodiment of the present invention, the method for constructing the BILSTM neural network model in step S5 is as follows: a BILSTM neural network model is established, including a three-layer BILSTM neural network with an input layer, a hidden layer and an output layer. In this model, the input layer consists of feature parameters that are strongly correlated with lifetime selected by the Pearson algorithm, and the output layer has 1 unit. The input layer unit is a feature that characterizes the degradation state of the device, and the output layer unit is the lifetime value of the device.

[0073] The BILSTM neural network is trained using the sigmoid function for the hidden layer neurons and the ReLU function for the output layer neurons, which are used to output the network's prediction results. The training function uses the Adam gradient descent algorithm, and the initial weights are selected as random numbers between [0,1]. The established BILSTM neural network is trained by continuously improving the weights and thresholds in the BILSTM neural network model until it converges.

[0074] like Figure 6 As shown in the embodiment of the present invention, a system for assessing the health status and predicting the remaining useful life of a numerically controlled power device, as described in the power MOSFET health status assessment and remaining useful life prediction method, is provided. The system includes:

[0075] The data acquisition and processing module is used to acquire training data and test data with obvious aging trends, and to segment and reduce noise in the MOSFET drain-source on-state voltage, current and resistance acquired by the sensor to obtain data that can better characterize MOSFET degradation.

[0076] The feature extraction and processing module is used to extract 14 time-domain features from the processed data, form a 42-dimensional feature set, calculate the correlation between each feature and lifetime using the Pearson algorithm, select features with strong correlation to lifetime, reduce the dimensionality of the selected features using the KPCA algorithm, and use the first-order principal component as the health index of the MOSFET.

[0077] The neural network training module is used to divide the state of the features selected in the training set and the constructed health indicators through a clustering algorithm, and to train the BILSTM neural network using the selected feature data.

[0078] The device evaluation and prediction module is used to input the features selected from the test set and the constructed health indicators into the trained BILSTM neural network and clustering algorithm model to further evaluate the device's status and predict its remaining lifespan.

[0079] The technical effects of the present invention will be described in detail below with reference to the test results.

[0080] The data used in this embodiment of the invention comes from accelerated aging test data of MOSFET devices from NASAAmes Center of Prediction of Excellence. This dataset represents aging tests of power MOSFETs under power cycling. To evaluate the degradation state of the MOSFETs, three different types of data are used: voltage sensor data, current sensor data, and resistance data calculated from current and voltage. The dataset records different sensor data from the start of MOSFET operation until it reaches its aging limit.

[0081] The data under the same working condition were divided into training and test sets. The results obtained according to the steps described in this invention are as follows. Figure 4 The results obtained from the state assessment, Figure 5 This is a comparison chart showing the results of predicting remaining lifespan using three different neural networks.

[0082] The results show that the method proposed in this invention can better classify the state of MOSFETs and predict the remaining lifetime more accurately than the results obtained using the other two neural networks.

[0083] II. Application Examples. To demonstrate the inventiveness and technical value of the technical solution of this invention, this section provides application examples of the technical solution of the claims on specific products or related technologies.

[0084] The multi-source data fusion power MOSFET health status assessment and remaining lifetime prediction method provided in the application embodiment of the present invention is applied to a computer device. The computer device includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor performs the steps of the multi-source data fusion power MOSFET health status assessment and remaining lifetime prediction method.

[0085] The multi-source data fusion power MOSFET health status assessment and remaining service life prediction method provided in the application embodiment of the present invention is applied to an information data processing terminal, which is used to implement the multi-source data fusion power MOSFET health status assessment and remaining service life prediction system.

[0086] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as 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 using hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or using software executed by various types of processors, or using a combination of the above-described hardware circuitry and software, such as firmware.

[0087] III. Evidence of the Relevant Effects of the Embodiments. The embodiments of the present invention have achieved some positive effects during research and development or use, and indeed possess significant advantages compared to existing technologies. The following description, in conjunction with data, charts, and other materials from the experimental process, illustrates these advantages.

[0088] The data used in this embodiment of the invention comes from accelerated aging test data of MOSFET devices from NASAAmes Center of Prediction of Excellence. This dataset represents aging tests of power MOSFETs under power cycling. To evaluate the degradation state of the MOSFETs, three different types of data are used: voltage sensor data, current sensor data, and resistance data calculated from current and voltage. The dataset records different sensor data from the start of MOSFET operation until it reaches its aging limit.

[0089] The acquired data from the three types of sensors were analyzed, and 14 time-domain features suitable for these three types of sensor data were extracted. Then, the Pearson algorithm was used to select aging parameters with a correlation greater than 0.8. Figure 2 As shown; the selected features are fused using the KPCA algorithm, and the first three principal components of the fused data are retained as follows. Figure 3 As shown; finally, the first-order principal components are used as health indicators for state division. After obtaining the cluster centers of MOSFETs 8 and 11, the average value is taken as the initial cluster center of MOSFET 9 for clustering, as shown below. Figure 4 As shown; finally, the selected features are input into three neural networks to predict the remaining lifespan, as shown. Figure 5 As shown.

[0090] The above description is merely 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 those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for assessing the health status and predicting the remaining useful life of a power MOSFET, characterized in that, Includes the following steps: The first step is to acquire training data and test data with obvious aging trends. The MOSFET drain-source conduction voltage, current and resistance calculated from voltage and current are segmented and noise-reduced from the sensor to obtain data that can better characterize MOSFET degradation. The second step involves extracting 14 time-domain features from the processed data, forming a 42-dimensional feature set. The correlation between each feature and lifetime is calculated using the Pearson algorithm, and features with strong correlation to lifetime are selected. The selected features are then dimensionality-reduced using the KPCA algorithm. The first-order principal components with a contribution rate of 88% are used as the health indicators of the MOSFET. The selected features and the constructed health indicators in the training set are used to perform state division using a clustering algorithm. At the same time, the selected feature data is used to train a BILSTM neural network. The initial cluster centers of the clustering algorithm are the average cluster centers of the same type of MOSFET training set. The step of using clustering algorithms to divide the selected features and constructed health indicators from the training set into states, and simultaneously training a BILSTM neural network with the selected feature data, includes: (1) Establish a BILSTM neural network model, which includes a three-layer BILSTM neural network with an input layer, a hidden layer and an output layer. In this model, the input layer consists of feature parameters that are strongly correlated with lifetime selected by the Pearson algorithm, and the output layer has 1 unit. The input layer unit is a feature that characterizes the degradation state of the device, and the output layer unit is the lifetime value of the device. (2) Train the BILSTM neural network. The transfer function of the hidden layer neurons in the BILSTM neural network is the sigmoid function, and the transfer function of the output layer neurons is the ReLU function, which is used to output the prediction results of the network. The training function adopts the Adam gradient descent algorithm, and the initial weights are selected as random numbers between [0, 1]. The established BILSTM neural network is continuously improved by adjusting the weights and thresholds in the BILSTM neural network model until it converges and completes the training. The third step is to input the features selected from the test set and the constructed health indicators into the trained BILSTM neural network and clustering algorithm model to further evaluate the device's status and predict its remaining lifespan. In the second step, kernel principal component analysis is used to analyze the matrix formed by the selected features with obvious trends. Q Data fusion includes the following steps: (1) Combine features with obvious trends into a matrix Q ; (2) Centralized data matrix Q Obtain the matrix Q *, then on the matrix Q * Transpose to obtain the matrix ( Q *) T ; (3) According to the matrix Q * and matrix ( Q *) T Calculate the kernel matrix K Let the matrix K The eigenvalues ​​are arranged from largest to smallest as λ1, λ2, ..., λ3. n The eigenvectors corresponding to its eigenvalues ​​are ; (4) Construct the projection matrix W based on the eigenvectors corresponding to the first 3 largest eigenvalues; usually, the eigenvalues ​​and eigenvectors are arranged in descending order, and the eigenvectors corresponding to the first 3 largest eigenvalues ​​are selected. (5) The centered data matrix Q * Projecting onto the selected first 3 feature vectors yields the dimensionality-reduced data matrix Y; (6) The first-order principal component normalization value is selected as the health index to characterize the degradation state of the MOSFET. The third step involves inputting the health indicators constructed from the training set data into the k-means++ algorithm to obtain the cluster center of each training data point, and then taking the average value of each cluster center as the initial cluster center of the test set for k-means clustering.

2. The method for assessing the health status and predicting the remaining service life of a power MOSFET as described in claim 1, characterized in that, The third step of predicting the trained BILSTM neural network model includes: given the corresponding prediction sensor data, a feature dataset with obvious trends is obtained, and then the value of the output layer is the prediction result after calculation by the BILSTM neural network model.

3. A system for assessing the condition and predicting the remaining useful life of a numerically controlled power device, based on the power MOSFET health status assessment and remaining useful life prediction method as described in any one of claims 1 to 2, characterized in that, The system includes: The data acquisition and processing module is used to acquire training data and test data with obvious aging trends, and to segment and reduce noise in the MOSFET drain-source on-state voltage, current and resistance acquired by the sensor to obtain data that can better characterize MOSFET degradation. The feature extraction and processing module is used to extract 14 time-domain features from the processed data, form a 42-dimensional feature set, calculate the correlation between each feature and lifetime using the Pearson algorithm, select features with strong correlation to lifetime, reduce the dimensionality of the selected features using the KPCA algorithm, and use the first-order principal component as the health index of the MOSFET. The neural network training module is used to divide the state of the features selected in the training set and the constructed health indicators through a clustering algorithm, and to train the BILSTM neural network using the selected feature data. The device evaluation and prediction module is used to input the features selected from the test set and the constructed health indicators into the trained BILSTM neural network and clustering algorithm model to further evaluate the device's status and predict its remaining lifespan.

Images