MOSFET health state prediction method based on improved physical information neural network

Abstract

The application provides a MOSFET health state prediction method based on an improved physical information neural network, relates to the field of semiconductor microelectronic technology, and comprises the following steps: S1: multi-dimensional data acquisition and preprocessing; S2: customized IPINN network architecture; S3: construction of a multi-constraint adaptive total loss function; S4: training the model by adopting a phased strategy with the minimization of the total loss function as the only target; and S5: deploying the trained model to the equipment end to realize health state prediction. The application designs an exclusive non-redundant network architecture with an input layer-shared feature encoder-double health index independent prediction head, the double prediction heads correspond to the two core health factors of on-resistance and junction temperature respectively, adopts a layer-by-layer feature compression structure, and strictly limits the prediction output to a reasonable interval of MOSFET physical parameters, thereby solving the problems of mismatch between a general PINN network and power device degradation characteristics and missing of single-head prediction coupling correlation.

Description

Technical Field

[0001] This invention relates to the field of semiconductor microelectronics technology, and more specifically to a method for predicting the health status of MOSFETs based on an improved physical information neural network. Background Technology

[0002] In modern aircraft power supply systems, power converters play a crucial role. Their core function is to achieve efficient power distribution and precise quality control to ensure the stable and reliable operation of various onboard electronic devices. Among these complex power converters, the metal-oxide-semiconductor field-effect transistor (MOSFET) serves as the core power switching device, and its performance and reliability directly determine the overall lifespan and safety of the entire converter and even the aircraft power supply system.

[0003] Facing the challenge of MOSFET health prediction, the general approach is to construct a prediction model. This model takes various relevant parameters as inputs and calculates the final health status. Traditional prediction models are mainly divided into two categories: physics-based models and data-driven models, but both have significant limitations.

[0004] Pure physics models attempt to predict device lifetime by establishing a system of differential equations describing the electro-thermal-degradation process, starting from first principles. These models offer good interpretability and extrapolation capabilities, but their accuracy heavily relies on a precise mathematical description of complex physical mechanisms. However, microscopic degradation mechanisms, such as crack propagation paths in bond wire fatigue, are highly stochastic and nonlinear, making them difficult to express precisely with simple analytical formulas. This often results in insufficient accuracy of pure physics models in practical applications.

[0005] On the other hand, traditional pure data-driven methods, such as multilayer perceptrons (MLPs) and long short-term memory networks (LSTMs), while capable of learning complex nonlinear mappings directly from data, are essentially "black box" models lacking physical interpretability. These models typically require extensive training with fault data covering various operating conditions, but in engineering practice, obtaining sufficient and complete failure data is extremely costly and time-consuming. Furthermore, data-driven models exhibit poor generalization ability outside the training data distribution, making them ill-suited to adapt to changing real-world working environments.

[0006] Therefore, how to integrate the interpretability of physical models with the flexibility of data-driven models to overcome the prediction challenges under small sample sizes and multiple operating conditions is a core issue that urgently needs to be addressed in the field of MOSFET health prediction. Summary of the Invention

[0007] To address the shortcomings of the prior art, the present invention aims to provide a MOSFET health state prediction method based on an improved physical information neural network. This prediction method can integrate a semi-physical model with deterministic physical laws and data-driven empirical terms, inheriting the physical interpretability and extrapolation robustness of PINN. At the same time, by introducing learnable degradation characterization terms, it significantly enhances the fitting ability to complex degradation dynamics.

[0008] Specifically, the present invention provides a MOSFET health status prediction method based on an improved physical information neural network, comprising the following steps: S1: Multi-dimensional data acquisition and preprocessing: Eight-dimensional physical parameters of the MOSFET during operation are acquired through a synchronous measurement system, including: time step. Drain current Drain-source voltage Shell temperature Ambient temperature Switching losses Junction-Environmental Thermal Resistance and on-resistance The Z-score standardization method was used to unify the dimensions of each parameter; S2: Custom-built IPINN network architecture; S3: Construct a multi-constraint adaptive total loss function, including the following steps: S31: Define the basic loss terms of the total loss function, including: data fitting loss term. Electrical loss residual loss item Thermal balance residual loss term Degradation kinetic residual loss term Boundary / Initial Condition Loss Term and regularization loss term ; S32: Constructing an adaptive weighted total loss function The basic loss term in S31 is adjusted using adaptive weighting coefficients. Combined into a total loss function: ; in, All weight coefficients are initially set to 1 / 6, and are adaptively adjusted during training using gradient descent to minimize the total loss function. S4: With minimizing the total loss function as the sole objective, a phased strategy is employed to train the model, including the following steps: S41: Pre-training and physical parameter initialization: Fix the inherent physical parameters, and use only the data fitting loss term as the training target for pre-training to obtain the initial values ​​of the degradation parameters, where the data fitting loss = total loss function; S42: Initial training of physical constraints: Load the initial values ​​of all network learnable parameters and degradation parameters from S41, use the total loss function as the training target, and complete the initial learning of physical constraints; S43: Refined Co-training: Loads all model parameters optimized in S42, and dynamically adjusts them during training using gradient descent. This minimizes the total loss function until the set independent test set no longer decreases after iteration; S5: Deploy the trained IPINN model to the device to achieve real-time prediction of MOSFET health status.

[0009] Furthermore: The method for customizing the IPINN network architecture in S2 includes the following steps: S21: Build the input layer: Receive the eight-dimensional physical parameters preprocessed by S1 as the original input; S22: Building a shared encoder: Extracting common characteristics of MOSFET electro-thermal-degradation coupling from 8-dimensional physical parameters; S23: Set up dual prediction heads: Map the shared feature vector output by the shared encoder to the on-resistance prediction values ​​respectively. and predicted junction temperature This enables decoupled prediction of two health indicators.

[0010] Furthermore: The methods for setting up a shared encoder in S22 include: Layer 1: Input 8-dimensional, output 384-dimensional, configured with linear transformation, layer normalization, Tanh activation function and Dropout regularization; Layer 2: Input 384-dimensional, output 384-dimensional, configuration is the same as Layer 1, used to enhance nonlinear representation capabilities; Layer 3: Input 384-dimensional, output 384-dimensional, configuration is the same as Layer 1, used to fully capture coupling features; Layer 4: Input 384-dimensional, output 192-dimensional, configured with linear transformation and layer normalization, compressing feature dimensions while retaining coupling information, including core information on MOSFET electrical-thermal-degradation coupling.

[0011] Furthermore: Linear transformation method: In the formula, This is the weight matrix. These are bias vectors, all of which are learnable parameters; tanh activation function: Output range ; Layer normalization method: , , These are learnable parameters; Dropout regularization method: Randomly disable 20% of neurons during training and restore all neurons during testing.

[0012] Furthermore, the method for building a dual prediction head in S23 includes the following sub-steps: S231: Prediction Head 1, Prediction: The 192-dimensional shared feature vector output by the shared encoder is compressed layer by layer: 192-dimensional → 96-dimensional → 32-dimensional → 1-dimensional. Each layer is configured with linear transformation, layer normalization, tanh activation function, and 8% Dropout regularization. The output of the last layer is constrained by the Sigmoid function to map the output value to the physically reasonable range of MOSFET on-resistance. The constraint formula is: ; in, This is the predicted value of the on-resistance. Output range ; 0 is the lower physical limit, 2 is the upper physical limit; S232: Predicting Head 2, Prediction: Ensure the prediction head structure and parameters are consistent with prediction head 1. The output of the last layer is constrained by the Sigmoid function and mapped to the physically reasonable range of MOSFET junction temperature. The constraint formula is: ; in, The value represents the predicted junction temperature, with 25 being the lower physical limit and 150 being the upper physical limit.

[0013] Furthermore, the methods for defining the basic loss term in S31 include: Data fitting loss term: ; In the formula, Indicates the first One data sample, The physical quantity used to indicate junction temperature. This represents the total number of data samples; Electrical loss residual loss term: ; In the formula, This is the predicted value for total power loss. , and These represent the on-resistance at the reference temperature, temperature coefficient, and reference temperature, respectively, which are obtained from the device datasheet. Thermal equilibrium residual loss: ; In the formula, These represent the resistances at thermal fusion and junction to the environment, respectively, and are obtained from the device datasheet. Indicates switching losses. The junction temperature sequence predicted by the model is calculated using first-order difference. Degradation kinetic residual loss term: ; In the formula, For the degradation parameters of the IPINN network's adaptive learning, Boltzmann's constant, The activation energy is n, and the degradation index is n. , , These are inherent parameters of the device, obtained from the datasheet / physical manual; Boundary / Initial Condition Loss Term: ; Regularization loss term : ; In the formula, W is the set of all learnable weight matrices in the IPINN network, with no bias term.

[0014] Furthermore, the pre-training methods in S41 include: employing the Adam optimizer and setting the initial learning rate. The IPINN network backbone was pre-trained for 2000 iterations until the rate of change of the fitted loss term was <1e-6. Using the output of the pre-trained model, the degradation parameters of adaptive learning are adjusted using a nonlinear least squares method. Conduct preliminary identification, and use the identification results as... The initial value.

[0015] Furthermore, the initial training method for physical constraints in S42 includes: using the Adam optimizer and setting the learning rate decay to... The iteration count is 3000 times until the rate of change of the fitting loss term is <1e-5; Every 100 iterations, the configuration points are resampled uniformly in the time-condition data space to ensure that the physical constraints are uniformly satisfied throughout the entire input domain, thus avoiding overfitting to fixed configuration points.

[0016] Furthermore, the refined collaborative training method in S43 includes: using the L-BFGS optimizer to find local optima using its second derivative information; Set up an independent test set and monitor the fitting loss term of the test set in real time. During the training process, if the test set loss does not decrease for 50 consecutive iterations, immediately terminate the training to prevent the model from overfitting.

[0017] Furthermore: S4 also includes S44: inputting the test set into the trained IPINN model and outputting the predicted values ​​of on-resistance and junction temperature; Using root mean square error (RMSE) and mean absolute error (MAE) as evaluation indicators, the error between predicted and measured values ​​is calculated. Multilayer perceptron, long short-term memory network (LSTM), and standard PINN are used as comparison models. When the RMSE and MAE of the model in predicting on-resistance and junction temperature are reduced by more than 40% compared with the comparison model, the model is deemed to have passed the verification.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention abandons the existing general-purpose PINN network structure and designs a dedicated, non-redundant network architecture for MOSFET devices, consisting of an "input layer - shared feature encoder - dual independent health indicator prediction heads". The shared encoder adopts a multi-layer fully connected + layer normalization structure to deeply extract common features of electrical-thermal-degradation coupling and eliminate redundant data noise. The dual prediction heads correspond to the two core health factors of on-resistance and junction temperature, respectively. They adopt a layer-by-layer feature compression structure and strictly limit the prediction output to the reasonable range of MOSFET physical parameters through the Sigmoid activation function, achieving dual constraints of decoupled prediction of dual indicators and physical consistency.

[0019] This invention addresses the mismatch between general-purpose PINN networks and the degradation characteristics of power devices, as well as the lack of single-head prediction coupling correlation. It eliminates non-physical invalid prediction results, accurately captures the multi-physics coupling characteristics during MOSFET degradation, and improves the pertinence and accuracy of core health indicator predictions, providing a new technical path for predicting MOSFET health status and remaining lifetime.

[0020] 2. This invention breaks through the limitations of traditional PINN's single physical constraint and manually fixed weights. It embeds three core physical mechanisms—the law of electrical loss, the differential equation of thermal balance, and the Arrhenius degradation kinetic equation—into the model as rigid constraints in the form of residual loss. This achieves a dynamic and synergistic balance between physical mechanism constraints and data fitting, freeing it from dependence on large-scale datasets and possessing complete physical interpretability. It solves the technical problem of mutual interference between physical model simplification errors and measured data noise. It ensures that the prediction results fully conform to the physical operating laws of power devices and has data-driven nonlinear fitting capabilities, balancing model interpretability and prediction accuracy. Compared with traditional PINN, the adaptability of physical constraints is greatly improved.

[0021] 3. This invention constructs a multi-constraint adaptive total loss function and fuses the basic loss terms through adaptive weight coefficients, building a six-fold composite loss system that includes data fitting loss, electrical loss, thermal loss, degradation kinetics loss, boundary initial condition loss, and regularization loss. It introduces adaptively learnable weight coefficients to replace manually set weights, with the weights dynamically optimized during model training. Furthermore, it adds device-specific adaptively learnable degradation parameters to the degradation kinetics loss term. The adaptive fitting of the micro-stochastic degradation mechanism determines the final total loss function, thereby improving the accuracy of the prediction model.

[0022] 4. This invention addresses the industry pain points of high cost of MOSFET fault data acquisition and difficulty in training with small samples. It designs a three-stage gradient training process combining data pre-training, physical constraint initial training, and refined collaborative training. In the pre-training stage, the Adam optimizer is used to quickly learn data degradation trends and initialize degradation parameters for adaptive learning. In the physical constraint initial training stage, a full loss function is introduced to achieve basic convergence. In the refined training stage, the L-BFGS second-order optimizer is switched, combined with dynamic point sampling and early stopping regularization mechanisms to achieve synchronous optimization of network parameters and physical degradation parameters. This overcomes the technical challenges of unstable model training, slow convergence speed, and overfitting in dynamic scenarios with small samples and multiple operating conditions. It eliminates the dependence on large-scale fault datasets, significantly improving model generalization ability and training stability. Compared to traditional data-driven models, the amount of training data required is greatly reduced, and prediction accuracy is improved by more than 40%.

[0023] 5. This invention establishes a high-precision, multi-dimensional, synchronous data acquisition and standardized preprocessing workflow, adaptable to the characteristics of industrial field measured data. Based on model prediction results, it establishes a quantitative health status judgment standard for MOSFETs, and combines nonlinear fitting and Monte Carlo uncertainty quantification methods to derive the remaining service life and provide a 95% confidence interval. This forms a closed-loop technical solution encompassing data acquisition, model prediction, health judgment, service life estimation, and operation and maintenance decision-making, supporting real-time deployment at the device end. It achieves seamless integration from theoretical models to engineering applications, with prediction results directly guiding equipment operation and maintenance decisions. This solves the problem that existing technologies remain only in theoretical research and are difficult to implement in actual engineering, broadening the application scenarios of the technology and making it suitable for power device health management scenarios in multiple fields such as aerospace, new energy, and industrial power supply. Attached Figure Description

[0024] Figure 1 This is a model structure diagram of the customized IPINN network architecture in the MOSFET health state prediction method disclosed in this invention. Figure 2 This is a comparison curve of the prediction of on-resistance and junction temperature and the prediction error of different models in the MOSFET health state prediction method disclosed in this invention. Detailed Implementation

[0025] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0026] The physical quantities, model symbols, and their meanings used in this manual are uniformly defined as follows, and all symbols are universally applicable throughout: Example 1: This invention discloses a MOSFET health status prediction method based on an improved physical information neural network, comprising the following steps: S1: Multi-dimensional data acquisition and preprocessing, including the following sub-steps: S11: Data Acquisition; Eight-dimensional physical parameters of the MOSFET during operation are acquired using a synchronous measurement system with a sampling frequency of 10Hz and a measurement accuracy of ±0.1%, specifically including: time step. Drain current Drain-source voltage Shell temperature Ambient temperature Switching losses Junction-Environmental Thermal Resistance On-resistance ; S12: Use the Z-score standardization method to unify the dimensions of each parameter; ; in, The mean of the parameters ( ), The standard deviation of the parameter; , The total number of measured data points for the parameter.

[0027] Step 2: Customized construction of the IPINN network architecture; Based on the standard PINN, this design addresses the electro-thermal-degradation coupling characteristics of MOSFETs and the dual health indicator on-resistance. Predicting demand and designing a customized architecture with no redundant modules, all layers' dimensions, activation functions, and regularization methods are customized for MOSFETs. Specific construction steps are as follows: Figure 1 As shown, the specific steps include the following: S21: Input layer setup; The input layer is a fully connected layer with 8 neurons. It directly receives the 8-dimensional feature vector after preprocessing in step 1, providing the original input to the network. There is no activation function or regularization; it only performs feature propagation.

[0028] S22: Setting up a shared encoder; Common features of MOSFET electro-thermal-degradation coupling are extracted from the eight-dimensional original features, and redundant noise is removed to provide a unified feature basis for dual prediction heads, thus solving the problem of mismatch between standard PINN feature extraction and device characteristics.

[0029] The setup steps include: Layer 1: Input 8-dimensional → Output 384-dimensional, configured with linear transformation, layer normalization, Tanh activation function and Dropout regularization.

[0030] Layer 2: Input 384 dimensions → Output 384 dimensions, with the same configuration as Layer 1, used to enhance nonlinear representation capabilities.

[0031] Layer 3: Input 384 dimensions → Output 384 dimensions, with the same configuration as Layer 1, used to fully capture coupling features.

[0032] Layer 4: Input 384 dimensions → Output 192 dimensions, only configured with linear transformation and layer normalization, no activation function, no Dropout, compresses feature dimensions and retains coupling information, including the core information of MOSFET electrical-thermal-degradation coupling.

[0033] Core function description: Linear transformation: ; This is the weight matrix. These are bias vectors, all of which are learnable parameters; tanh activation function: Output range The formula enhances nonlinearity; Layer Normalization, that is, LN: , , These are learnable parameters; This represents the mean of a single sample across all features. It represents the variance of a single sample across all feature dimensions.

[0034] Dropout: Randomly disables 20% of neurons during training and restores all neurons during testing.

[0035] S23: Dual Prediction Head Setup: The 192-dimensional shared feature vector output by the shared encoder is mapped to... Predicted value and Predicted value This achieves decoupled prediction of two health indicators; at the same time, it ensures the physical coupling consistency between the two by sharing features at the bottom layer. The two prediction heads have completely identical structures, both being 3-layer fully connected networks. The specific construction steps are as follows: S231: Prediction Header 1, On-Resistance predict: Layer-by-layer feature compression: 192-dimensional → 96-dimensional → 32-dimensional → 1-dimensional, each layer configured with linear transformation + LN + tanh activation function + 8% Dropout regularization; the output of the last layer is constrained by the Sigmoid function, mapping the output value to the physically reasonable range of MOSFET on-resistance. The constraint formula is: ; in, Output range ; 0 represents the lower physical limit, and 2 represents the upper physical limit.

[0036] S232: Prediction head 2, junction temperature predict: The structure and parameters are completely consistent with the prediction head 1, i.e., 192-dimensional → 96-dimensional → 32-dimensional → 1-dimensional, linear transformation + LN + tanh + 8% Dropout; the output of the last layer is constrained by the Sigmoid function and mapped to the physically reasonable range of MOSFET junction temperature. The constraint formula is: ; Where 25 is the physical lower limit and 150 is the physical upper limit, the Sigmoid function expression is the same as in S231.

[0037] Output result: ,Right now Predicted value ,Right now The predicted values ​​are output independently, and both are constrained to a physically reasonable range with no irrational outputs.

[0038] S3: Construction of Multi-Constraint Adaptive Total Loss Function: An adaptive weighted total loss function is constructed by embedding mature physical formulas in the field of power electronics as rigid constraints into the loss function, and data fitting requirements, initial / boundary conditions, and regularization requirements as flexible constraints. The loss function is the sole objective function for model training, without any additional constraints. First, six basic loss terms are defined, encompassing three types of physical formula constraints. These are then fused into the total loss function using adaptive weight coefficients. All basic loss terms are based on mature formulas. Physical constraints are not added additionally to the network structure, dataset, or input layer; instead, they are embedded into the entire training process through the loss function. This is the core logic of the physical information neural network. Physical laws themselves are equations / differential equations, which cannot be directly embedded into the network for training. They need to be transformed into physical residual loss: Residual = Network Predicted Value - Theoretical Value of Physical Formula. The smaller the residual, the more the network output conforms to physical laws. The squared residual is used as the loss term, penalizing illogical outputs during training. This step improves the combination method of loss terms and the embedding form of physical formulas, i.e., the adaptive weight mechanism. The specific steps are as follows: S31: Define 6 types of basic loss terms: All loss terms are scalars; the smaller the value, the better the model output meets the constraints. The physical formulas are all mature formulas in the field of power electronics / thermodynamics, and their specific definitions and functions are as follows: 1) Data fitting loss term : ; Constrained model predictions , The deviation from the measured values ​​ensures the model's basic data fitting ability and is the foundation of all loss terms; It calculates directly based on model output and measured data, without embedding physical formulas, and only performs data constraints.

[0039] 2) Electrical loss residual loss item : Based on Ohm's law and the formula for conduction loss The formula for physical constraint loss is given: ; In the formula, This is the predicted value for total power loss. , and These represent the on-resistance, temperature coefficient, and reference temperature at the reference temperature, respectively, obtained from the device datasheet; Ohm's law and the conduction loss formula are used as rigid constraints to ensure the predicted total power loss. Consistent with theoretical conduction losses and preventing model outputs from violating the laws of electrical physics is one of the core terms of physical constraints.

[0040] 3) Thermal balance residual loss term : Embedded mature physical formulas: ; The formula for the loss term is the sum of squared residuals: ; In the formula, These represent the resistances at thermal fusion and junction to the environment, respectively, and are obtained from the device datasheet. Indicates switching losses. The junction temperature sequence predicted by the model is calculated using first-order difference; the law of conservation of energy and the thermal balance equation are used as rigid constraints to force the junction temperature changes predicted by the model to conform to the laws of heat conduction, and to prevent the model output from violating the laws of thermodynamics and physics. This is the second core term of the physical constraint.

[0041] 4) Degradation kinetic residual loss term : Embedded mature physical formulas: ; Construct the formula for the loss term: ; In the formula, For the degradation parameters of the IPINN network's adaptive learning, Boltzmann's constant, The activation energy is n, and the degradation index is n. , , These are intrinsic parameters of the device, obtained from the datasheet / physical manual; the Arrhenius degradation relation is used as a rigid constraint framework, and the degradation parameters are adaptively learned. Fitting complex degradation terms that are difficult to model accurately , The first-order difference of the predicted on-resistance sequence, ; This approach preserves physical interpretability while enhancing the model's ability to fit complex degradation mechanisms, preventing model outputs from contradicting the actual degradation patterns of MOSFETs. The loss term introduces device-specific learnable parameters into the mature Arrhenius formula. This solves the problem that pure physical formulas cannot fit the complex degradation of microscopic features.

[0042] 5) Boundary / Initial Condition Loss Term : formula: ; Utilizing the known initial health state of the MOSFET, where , The result is obtained from the factory static test and serves as a rigid constraint to force the initial output of the model to be consistent with the actual result.

[0043] 6) Regularization loss term : formula: ; In the formula, W is the set of all learnable weight matrices in the IPINN network, with no bias term.

[0044] To prevent overfitting, ensure the smoothness of the physical parameters learned by the network, and improve the model's generalization ability.

[0045] S32: Constructing an adaptive weighted total loss function : The above 6 types of basic loss terms are weighted by adaptive coefficients. The weights are combined into a total loss function, and the weight coefficients are automatically learned during the model training process. The formula for the total loss function is: ; in, All weight coefficients are initially set to 1 / 6 and adaptively adjusted during training using gradient descent to minimize the total loss function; the sole objective of model training is to minimize the total loss function. During training, various loss terms are dynamically balanced through weight coefficients, forcing the model output to simultaneously meet the "data fitting requirements" and "all physical formula constraints". The loss function plays a core constraining role throughout the entire model training process, and there is no stage where the model deviates from the constraints of the loss function.

[0046] Step S3 has the following beneficial effects: 1. For MOSFET characteristics, three types of mature physical formulas—electrical, thermodynamic, and degradation kinetic—are embedded in the total loss function in the form of residual loss terms to achieve rigid constraints of physical laws, which is different from the single physical formula embedding of standard PINN.

[0047] 2. Introduce device-specific adaptive learning degradation parameters By fitting complex microscopic degradation mechanisms within the framework of mature physical formulas, we can solve the problems of low accuracy of pure physical formulas and lack of interpretability of pure data-driven approaches.

[0048] 3. Design adaptive weight coefficients to achieve a dynamic balance between physical constraints and data fitting, and solve the constraint imbalance problem caused by the manual fixed weights in standard PINN.

[0049] 4. Add initial MOSFET constraints to better fit engineering practice and improve the model's feasibility.

[0050] S4: Staged model training; To minimize the total loss function With the sole objective of addressing the technical challenges of standard PINN training instability, slow convergence, and mismatch between physical and network parameters under limited sample conditions, a phased training strategy is employed. Throughout the training process, the loss function acts as a constraint, and all optimizers, learning rates, and iteration counts are customized for use with limited MOSFET data. The specific training steps are as follows: S41: Pre-training and physical parameter initialization; fixed , , Physically inherent parameters, only the data fitting loss term As the training target, among which The Adam optimizer is used, with an initial learning rate of... The IPINN network backbone was pre-trained, with 2000 iterations, until... The rate of change is <1e-6; using the output of the pre-trained model, the degradation parameters of adaptive learning are adjusted using nonlinear least squares. Conduct preliminary identification, and use the identification results as... The initial value; only through The macroscopic degradation trend of the constraint model learning data lays the foundation for subsequent physical constraint training.

[0051] S42: Initial training with physical constraints; Load all learnable parameters of the IPINN neural network learned during the S4 pre-training phase, including network parameters and... The initial values ​​include the learnable network parameters, which consist of the weight matrix W and bias vector b of all linear transformations of the shared encoder, and the layer-normalized learnable parameters γ and β. Also included are the weight matrix W and bias vector b of all linear transformations of each of the two predictors, and the layer-normalized learnable parameters γ and β.

[0052] Total loss function As the training objective, the weight coefficients Initial value 1 / 6, continue using Adam optimizer, learning rate decays to The iteration count is 3000 times, until... The rate of change is <1e-5; every 100 iterations, the configuration points are resampled uniformly in the time-condition data space to ensure that physical constraints are uniformly satisfied throughout the entire input domain, avoiding overfitting to fixed configuration points; through The model employs six loss terms to force its output to conform to both data patterns and physical laws, thus completing the initial learning of physical constraints.

[0053] S43: Refined Collaborative Training; Load the model parameters of S42. The model parameters refer to the optimized network parameters of S42, the degradation parameter η of adaptive learning, and the loss weights ω1 to ω6.

[0054] The Adam optimizer is switched to the L-BFGS optimizer, which utilizes the second derivative information to find a more accurate local optimum; during training, the optimizer is dynamically adjusted using gradient descent. ,make By continuously minimizing, a dynamic balance is achieved between physical constraints and data fitting.

[0055] Simultaneously, a separate test set is set up, and the test set is monitored in real time. Training should be terminated immediately if the test set loss does not decrease for 50 consecutive iterations to prevent the model from overfitting. As the core objective function, the network parameters and physical parameters are optimized together throughout the process, and the adaptive weight coefficients are dynamically adjusted according to the changes in the loss term to ensure constraint balance.

[0056] S44: Model Validation; After the multi-constraint adaptive total loss function is constructed and trained, the prediction accuracy of the prediction model needs to be evaluated and verified.

[0057] Input the test set into the trained IPINN model, and output... and Predicted values; using root mean square error (RMSE) and mean absolute error (MAE) as evaluation metrics, the error between predicted and measured values ​​is calculated, and the model is compared with multilayer perceptron (MLP), long short-term memory (LSTM), and standard PINN models; the model performs well in... and The validation is considered successful if the root mean square error (RMSE) and mean absolute error (MAE) of the predictions are reduced by more than 40% compared to the comparison model; based on the test set... and As the core evaluation criterion for model accuracy, it ensures the objectivity of the verification results.

[0058] To address the challenges of scarce failure data and difficulty in directly measuring junction temperature, and to verify the superiority of the improved physical information neural network method under small sample conditions, this invention uses a publicly available power MOSFET accelerated aging dataset as a verification dataset. This dataset, with its high authority and empirical validity, provides complete data support for the model from early degradation to final failure.

[0059] To comprehensively evaluate the performance of the proposed improved physical information neural network method, this invention systematically compares it with several baseline models that represent mainstream approaches in the current field of health prediction, including: 1) Multilayer Perceptron (MLP): A classic pure data-driven feedforward neural network used to learn the nonlinear mapping relationship between input and output.

[0060] 2) Long Short-Term Memory Network (LSTM): A recurrent neural network specifically designed for processing time-series data, capable of effectively capturing long-term dependencies in the degradation process.

[0061] 3) Standard Physical Information Neural Network: As the basis for the improved physical information neural network, the standard physical information neural network also incorporates basic physical constraints, but does not include adaptive weights or specific architectural improvements based on the improved physical information neural network.

[0062] All models were trained and tested on the same publicly available power MOSFET accelerated aging dataset to ensure a fair comparison.

[0063] exist Figure 2 In Figures a and b, we show the comparison between the predicted trajectories of different models for on-resistance and junction temperature and the actual degradation curves. It is clear that the predicted curve based on the improved physical information neural network (PIN) closely matches the actual values, maintaining excellent tracking ability. In contrast, while pure data-driven models such as Multilayer Perceptron (MLP) and Long Short-Term Memory (LSTM) networks perform reasonably well within the training data range, their prediction errors accumulate rapidly during long-term extrapolation, causing the predicted trajectory to deviate significantly from the actual degradation trend. The standard PIN, due to the introduction of physical constraints, outperforms pure data-driven models, but its prediction accuracy still falls short of the improved PIN, especially in capturing the details of electro-thermal coupling effects.

[0064] Figure 2 In the diagram, c and d further quantify the prediction errors of each model. The improved physical information neural network (PEN) shows significantly lower RMSE and MAE scores than other comparative models in predicting on-resistance and junction temperature, with errors reduced by approximately 40-60%. This advantage is mainly attributed to the semi-physical architecture design of the improved PSN: the shared feature extraction layer effectively captures the common patterns of the electro-thermal-degradation process, while the multi-task prediction head achieves this through a physical residual loss term. and The tight coupling ensures the physical consistency of the prediction results. The adaptive weighting mechanism dynamically balances the importance of data fitting and various physical constraints, avoiding training bias caused by the dominance of any one loss factor.

[0065] In summary, the experimental results fully verify the effectiveness of the improved physical information neural network in making high-precision and interpretable predictions of MOSFET health status under small sample and multi-condition conditions. Its performance comprehensively surpasses that of traditional data-driven methods and standard physical information neural network baseline models.

[0066] Step S4 has the following beneficial effects: 1) To address the small sample size characteristics of MOSFETs, a three-stage training strategy is designed, combining pre-training with physical constraint initial training and refined training, to solve the problems of unstable training and slow convergence of standard PINN.

[0067] 2) The Adam+L-BFGS serial optimizer is used, which combines the fast convergence of first-order optimization and the fine tuning of second-order optimization to match the characteristics of MOSFET data.

[0068] 3) Design dynamic sampling points combined with an early stopping mechanism to fit the engineering reality of power device timing data and improve the model's generalization ability.

[0069] Step 5: Health Status and Remaining Life (RUL) Prediction; The trained and validated IPINN model is deployed to the device to achieve real-time prediction of MOSFET health status. The loss function only acts as a constraint during the training phase and is not involved in the calculation during the prediction phase. This step provides quantitative and actionable decision-making basis for engineering maintenance through real-time data input, health indicator prediction, RUL derivation, and health status determination, including the following steps: S51: Real-time data input; The sensor collects the real-time 8-dimensional operating condition characteristics of the MOSFET to be predicted, i.e. , , , , , , , The Z-score standardization formula in step S12 is used for standardization, and the result is then input into the deployed IPINN model.

[0070] S52: Health indicator prediction, including the following sub-steps: S521: IPINN model outputs real-time predicted on-resistance value and predicted junction temperature Both indicators are quantitative values ​​within a physically reasonable range.

[0071] S522: Quantifies forecast uncertainty using the Monte Carlo Dropout method, providing 95% confidence intervals for two indicators, namely... This enhances the engineering reference value of the prediction results.

[0072] S53: Remaining lifetime; Predicted on-resistance value based on IPINN model output and predicted junction temperature The long-term degradation trend prediction sequence is used to derive the remaining lifetime (RUL) of the MOSFET and provide a probabilistic estimate by determining the failure threshold and regression fitting.

[0073] The specific steps are as follows: 1) Failure threshold definition: When Increase to its initial value twice as much, or When the temperature remains above 150°C, the MOSFET is considered to have reached a failure state.

[0074] 2) Degradation trend fitting: A nonlinear regression method is used to fit the degradation trend. and By fitting the long-term prediction sequence, a degradation trend equation is obtained. , .

[0075] 3) RUL Calculation: Determining the time to reach the failure threshold by solving the degradation trend equation. ,but .

[0076] 4) Probabilistic estimation: Combining the 95% confidence interval, the probability distribution range of RUL is given, providing fault tolerance space for maintenance decisions.

[0077] S54: Health Status Assessment; according to and The ratio of the predicted value to the failure threshold is used to classify the MOSFET health status into 5 quantifiable levels. These levels are clearly defined and quantifiable, directly aligning with engineering maintenance strategies. The specific classification criteria are shown in the table below: The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting the health status of MOSFETs based on an improved physical information neural network, characterized in that: Includes the following steps: S1: Multi-dimensional data acquisition and preprocessing: Eight-dimensional physical parameters of MOSFETs during operation are acquired through a synchronous measurement system, and the Z-score standardization method is used to unify the dimensions of each parameter. S2: Custom-built IPINN network architecture; S3: Construct a multi-constraint adaptive total loss function, including the following steps: S31: Define the basic loss terms of the total loss function, including: data fitting loss term. Electrical loss residual loss item Thermal balance residual loss term Degradation kinetic residual loss term Boundary / Initial Condition Loss Term and regularization loss term ; S32: Constructing an adaptive weighted total loss function The basic loss term in S31 is adjusted using adaptive weighting coefficients. Combined into a total loss function: ； in, All weight coefficients are initially set to 1 / 6, and are adaptively adjusted during training using gradient descent to minimize the total loss function. S4: With minimizing the total loss function as the sole objective, a phased strategy is employed to train the model, including the following steps: S41: Pre-training and physical parameter initialization: Fix the inherent physical parameters, and use only the data fitting loss term as the training target for pre-training to obtain the initial values ​​of the degradation parameters, where the data fitting loss = total loss function; S42: Initial training of physical constraints: Load the initial values ​​of all network learnable parameters and degradation parameters from S41, use the total loss function as the training target, and complete the initial learning of physical constraints; S43: Refined Co-training: Loads all model parameters optimized in S42, and dynamically adjusts them during training using gradient descent. This minimizes the total loss function until the set independent test set no longer decreases after iteration; S5: Deploy the trained IPINN model to the device to achieve real-time prediction of MOSFET health status.

2. The MOSFET health status prediction method based on an improved physical information neural network as described in claim 1, characterized in that: The method for customizing the IPINN network architecture in S2 includes the following steps: S21: Build the input layer: Receive the eight-dimensional physical parameters preprocessed by S1 as the original input; S22: Building a shared encoder: Extracting common characteristics of MOSFET electro-thermal-degradation coupling from 8-dimensional physical parameters; S23: Set up dual prediction heads: Map the shared feature vector output by the shared encoder to the on-resistance prediction values ​​respectively. and predicted junction temperature This enables decoupled prediction of two health indicators.

3. The MOSFET health status prediction method based on an improved physical information neural network as described in claim 2, characterized in that: The methods for setting up a shared encoder in S22 include: Layer 1: Input 8-dimensional, output 384-dimensional, configured with linear transformation, layer normalization, Tanh activation function and Dropout regularization; Layer 2: Input 384-dimensional, output 384-dimensional, configuration is the same as Layer 1, used to enhance nonlinear representation capabilities; Layer 3: Input 384-dimensional, output 384-dimensional, configuration is the same as Layer 1, used to fully capture coupling features; Layer 4: Input 384-dimensional, output 192-dimensional, configured with linear transformation and layer normalization, compressing feature dimensions while retaining coupling information, including core information on MOSFET electrical-thermal-degradation coupling.

4. The MOSFET health status prediction method based on an improved physical information neural network as described in claim 3, characterized in that: Linear transformation methods: In the formula, This is the weight matrix. These are bias vectors, all of which are learnable parameters; tanh activation function: Output range ; Layer normalization method: , , These are learnable parameters; Dropout regularization method: Randomly disable 20% of neurons during training and restore all neurons during testing.

5. The MOSFET health status prediction method based on an improved physical information neural network as described in claim 2, characterized in that: The method for building a dual prediction head in S23 includes the following sub-steps: S231: Prediction Head 1, Prediction: The 192-dimensional shared feature vector output by the shared encoder is compressed layer by layer: 192-dimensional → 96-dimensional → 32-dimensional → 1-dimensional. Each layer is configured with linear transformation, layer normalization, tanh activation function, and 8% Dropout regularization. The output of the last layer is constrained by the Sigmoid function to map the output value to the physically reasonable range of MOSFET on-resistance. The constraint formula is: ； in, This is the predicted value of the on-resistance. Output range ; 0 is the lower physical limit, 2 is the upper physical limit; S232: Predicting Head 2, Prediction: Ensure the prediction head structure and parameters are consistent with prediction head 1. The output of the last layer is constrained by the Sigmoid function and mapped to the physically reasonable range of MOSFET junction temperature. The constraint formula is: ； in, The value represents the predicted junction temperature, with 25 being the lower physical limit and 150 being the upper physical limit.

6. The MOSFET health status prediction method based on an improved physical information neural network as described in claim 1, characterized in that: The methods for defining the basic loss term in S31 include: Data fitting loss term: ; In the formula, Indicates the first One data sample, The physical quantity used to indicate junction temperature. This represents the total number of data samples; Electrical loss residual loss term: ; In the formula, This is the predicted value for total power loss. , and These represent the on-resistance at the reference temperature, temperature coefficient, and reference temperature, respectively, which are obtained from the device datasheet. Thermal equilibrium residual loss: ; In the formula, These represent the resistances at thermal fusion and junction to the environment, respectively, and are obtained from the device datasheet. Indicates switching losses. The junction temperature sequence predicted by the model is calculated using first-order difference. Degradation kinetic residual loss term: ; In the formula, For the degradation parameters of the IPINN network's adaptive learning, Boltzmann's constant, The activation energy is n, and the degradation index is n. , , These are inherent parameters of the device, obtained from the datasheet / physical manual; Boundary / Initial Condition Loss Term: ; Regularization loss term : In the formula, W is the set of all learnable weight matrices in the IPINN network, with no bias term.

7. The MOSFET health status prediction method based on an improved physical information neural network as described in claim 1, characterized in that: The pre-training methods in S41 include: using the Adam optimizer and setting the initial learning rate. The IPINN network backbone was pre-trained for 2000 iterations until the rate of change of the fitted loss term was <1e-6. Using the output of the pre-trained model, the degradation parameters of adaptive learning are adjusted using a nonlinear least squares method. Conduct preliminary identification, and use the identification results as... The initial value.

8. The MOSFET health status prediction method based on an improved physical information neural network as described in claim 1, characterized in that: The initial training method for physical constraints in S42 includes: using the Adam optimizer and setting the learning rate decay to... The iteration count is 3000 times until the rate of change of the fitting loss term is <1e-5; Every 100 iterations, the configuration points are resampled uniformly in the time-condition data space to ensure that the physical constraints are uniformly satisfied throughout the entire input domain, thus avoiding overfitting to fixed configuration points.

9. The MOSFET health status prediction method based on an improved physical information neural network as described in claim 1, characterized in that: The refined collaborative training methods in S43 include: using the L-BFGS optimizer and utilizing its second derivative information to find local optimal solutions; Set up an independent test set and monitor the fitting loss term of the test set in real time. During the training process, if the test set loss does not decrease for 50 consecutive iterations, immediately terminate the training to prevent the model from overfitting.

10. The MOSFET health status prediction method based on an improved physical information neural network as described in any one of claims 1-9, characterized in that: S4 also includes S44: inputting the test set into the trained IPINN model and outputting predicted values ​​for on-resistance and junction temperature; Using root mean square error (RMSE) and mean absolute error (MAE) as evaluation indicators, the error between predicted and measured values ​​is calculated. Multilayer perceptron, long short-term memory network (LSTM), and standard PINN are used as comparison models. When the RMSE and MAE of the model in predicting on-resistance and junction temperature are reduced by more than 40% compared with the comparison model, the model is deemed to have passed the verification.

Images