A reliability prediction method for power field effect transistors

Abstract

A reliability prediction method for a power field effect transistor belongs to the technical field of power MOSFET reliability modeling. The present invention is to solve the problem of low work efficiency caused by manually adjusting model hyperparameters in the existing deep learning model for reliability prediction of power devices. It includes: S1, collecting sensor data as initial data and performing data processing on the initial data; S2, optimizing the hyperparameters of the Transformer model using a genetic algorithm to construct a GA-Transformer model; S3, transmitting the data obtained by S1 to the GA-Transformer model obtained by S2, evaluating the accuracy of the prediction based on performance indicators, predicting the remaining service life RUL, and outputting the reliability prediction result. The present invention is used for power MOSFET reliability prediction.

Description

Technical Field

[0001] The invention relates to a reliability prediction method for a power field effect transistor, and belongs to the technical field of power MOSFET reliability modeling. Background Art

[0002] Power MOSFET (Metal-Oxide-Semiconductor Field-Effect Transistor) is an indispensable semiconductor device in modern power electronic systems, playing a key role in current switching and amplification. With the increasing demand for high-efficiency and high-density power conversion in applications such as renewable energy technology, smart grids, and electric vehicles, the reliability research of power MOSFET is becoming increasingly important. However, during its service life, power MOSFET may suffer from a variety of stress factors (such as voltage shock, thermal cycle, and load fluctuation), causing its electrical performance to degrade or even fail, thus affecting the stability and safety of the entire power system.

[0003] In recent years, machine learning, especially deep learning algorithms, has gradually become a hot topic of research due to its advantages in feature learning, adaptability, large-scale data processing and complex nonlinear modeling. During the on-orbit operation and online monitoring of satellites, machine learning and artificial intelligence algorithms are used to identify abnormal patterns and predict the occurrence of faults, thereby achieving more intelligent and efficient fault diagnosis and predictive maintenance. Deep learning has shown great potential in the field of power device reliability prediction. By constructing multi-layer nonlinear processing units, deep learning models can learn complex representations from large amounts of data and capture deep features of fault data. It has been proven to have significant advantages in different engineering disciplines. Deep learning models such as RNN and LSTM are also gradually being applied to the reliability prediction of power devices. However, in order to meet the accuracy requirements of the model, trial and error methods are often used to manually adjust the model hyperparameters, which usually takes a lot of time and effort. Summary of the invention

[0004] The purpose of the present invention is to solve the problem of manual adjustment of model hyperparameters in the existing reliability prediction of power devices using deep learning models, resulting in low work efficiency, and provide a reliability prediction method for power field effect transistors.

[0005] A reliability prediction method for a power field effect transistor according to the present invention comprises:

[0006] S1, collecting sensor data as initial data, and performing data processing on the initial data;

[0007] S2, using genetic algorithm to optimize the hyperparameters of the Transformer model and constructing the GA-Transformer model;

[0008] S3, transmits the data obtained by S1 to the GA-Transformer model obtained by S2, evaluates the accuracy of the prediction based on the performance indicators, predicts the remaining useful life RUL, and outputs the reliability prediction results.

[0009] Preferably, the specific method for processing the initial data described in S1 includes:

[0010] S1-1, perform data cleaning, data resampling and data noise reduction on the initial data in turn;

[0011] S1-2, divide the data obtained in S1-1 into a training set and a test set;

[0012] S1-3. Normalize the divided data.

[0013] Preferably, the Transformer model in S2 specifically includes:

[0014] Select ΔR DS(on) As the failure characteristic parameter for model establishment, the failure threshold of the power field effect transistor is selected as:

[0015] ΔR DS(on) =0.18Ω;

[0016] The data is normalized using the minimum-maximum normalization calculation method:

[0017]

[0018] Among them, ΔR DS(on)k ′ represents the normalized on-resistance value, ΔR DS(on)k Indicates the on-resistance value before normalization, min(ΔR DS(on) ) represents the minimum value of the on-resistance before normalization, max(ΔR DS(on) ) represents the maximum value of the on-resistance before normalization, and k represents the number of on-resistances;

[0019] The encoder layer uses multi-head attention mechanism and layer normalization to process the input sequence;

[0020] The feed-forward network layer uses a 1D convolutional layer and a dropout layer to increase the nonlinearity of the model;

[0021] The output layer uses a flatten layer, a fully connected layer, and a dropout layer.

[0022] Preferably, the specific method of optimizing the hyperparameters of the Transformer model using a genetic algorithm in S2 includes:

[0023] S2-1. Define key parameters, including population size, number of iterations, dimension, retention rate, random selection rate, and mutation probability;

[0024] S2-2. Generate a new population based on selection, crossover and mutation operations, and determine whether the termination condition is reached. If not, repeat this step. If yes, stop the iteration and output the optimal solution.

[0025] Preferably, the population size described in S2-1 is set to 50, that is, each generation has 50 different candidate solutions;

[0026] The number of iterations is set to 50, that is, the algorithm will evolve for 50 generations to find the optimal or approximately optimal solution;

[0027] The dimension is set to 8, that is, each candidate solution consists of 8 values;

[0028] The retention rate is set to 0.3, that is, 30% of the best individuals will be directly passed to the next generation without being generated through crossover or mutation;

[0029] The random selection rate is set to 0.5, that is, 50% of the inferior individuals will be randomly selected and retained to the next generation to maintain genetic diversity;

[0030] The mutation probability is set to 0.6, that is, a 60% mutation probability is set for each gene.

[0031] Preferably, the termination conditions include:

[0032] t≥T max ;

[0033] f(x best )≤F thresh ;

[0034] Among them, t represents the current iteration number, T max represents the maximum number of iterations, f(x best ) represents the validation loss function, F thresh represents the validation loss threshold, which is set to 0.001.

[0035] Preferably, the optimized hyperparameters in S2 include: sliding window w, attention head size h, number of attention heads m, feedforward dimension f, batch size b, learning rate l, drop rate d and activation function a.

[0036] Preferably, when the population size is 50, the initial population generated is P0, and each individual is a hyperparameter set:

[0037] P0={x1,x2,·...,x 50};

[0038] Among them, x i represents an individual, x i =[w i ,h i ,m i ,f i ,b i ,l i ,d i ,a i ] represents the sliding window w of the i-th individual i , attention head size h i , number of attention heads m i , feedforward dimension f i , batch size b i , learning rate l i , discard rate d i And the activation function a i Settings; i = 1 ~ 50;

[0039] Define the fitness function f(x i ), evaluate the quality of individuals:

[0040] f(x i )=Loss(x i );

[0041] Among them, Loss(x i ) represents the validation loss on the Transformer model validation set;

[0042] The tournament selection method is used to select based on fitness. A group of individuals is randomly selected and the individual with the highest fitness is selected from them. The size of the tournament is k = 2. For each tournament, k individuals are randomly selected from the population, which is expressed as randomly selecting a subset T from the population P, where |T| = k. For each tournament subset T, the individual with the highest fitness is selected to enter the next generation:

[0043]

[0044] Among them, x i ∈T,x best represents the optimal individual, For x i , so that the function f(x i ) maximizes x i The value of

[0045] Randomly select two individuals as parents to generate one or more offspring, where the parameters of the offspring are the combination of the parameters of the parents;

[0046] x p =[w p ,h p ,m p ,f p ,b p ,l p ,d p ,a p ] and x q =[w q ,h q ,m q ,f q ,b q ,l q ,d q ,a q ] as the parent individual, and obtain the offspring x through single-point crossover child =[w p ,h p ,m p ,f q ,b q ,l q ,d q ,a q ];

[0047] Randomly adjust the parameters of individuals to increase the diversity of the population:

[0048] x i ′=[w i ,h i ,m i ,f i ,b i ,l i ,d i ,a i ]+Δ

[0049] Among them, x i ′ represents the individual after mutation, and Δ represents random disturbance.

[0050] Preferably, the performance indicators in S3 include mean absolute percentage error (MAPE), root mean square error (RMSE) and mean square error (MSE);

[0051] The calculation method of the mean absolute percentage error MAPE includes:

[0052]

[0053] Where N is the total number of samples of on-resistance, x i is the actual on-resistance value, x i ' is the predicted on-resistance value;

[0054] The calculation method of the root mean square error RMSE includes:

[0055]

[0056] The calculation method of the mean square error MSE includes:

[0057]

[0058] Preferably, the calculation method of the remaining useful life RUL in S3 includes:

[0059] RUL p =t f -t p ;

[0060] RUL a =t a -t p ;

[0061] Error=RUL a -RUL p ;

[0062]

[0063] Among them, RUL p Represents the predicted remaining life, RUL a represents the actual remaining life, Error represents the deviation between the predicted remaining life and the actual remaining life, RA represents the relative accuracy of the prediction result, and t p Indicates the start time of the prediction, t f It indicates the time when the prediction reaches the failure threshold, t a Indicates the actual failure time point.

[0064] The reliability prediction method of a power field effect transistor proposed in the present invention has the following advantages:

[0065] 1. The proposed Transformer model uses a one-dimensional convolutional layer as a feedforward network to flatten the data processed by the Transformer encoder block, and then further processes it through a fully connected layer. Finally, a linear activation function is used as the output layer. Compared with the traditional Transformer model, this structure is more suitable for processing time series data and regression tasks, and can effectively capture local patterns and dependencies in time series, reducing computational complexity and improving model flexibility.

[0066] 2. The GA-Transformer model is proposed. This model combines the global search advantages of genetic algorithms and the powerful time series data processing capabilities of the Transformer model. It surpasses existing technologies in various performance evaluation indicators and solves the problem of time-consuming and labor-intensive model adjustment. The Ga-Transformer model can handle large-scale data sets and complex nonlinear relationships, and adapt to the ever-changing data environment of satellites in orbit. It is conducive to real-time monitoring and fault diagnosis of satellite systems.

[0067] 3. Through the optimized calculation process, the demand for computing resources is reduced, and the running time is reduced by 43.54% compared with the LSTM model, thereby saving energy consumption, which is an important advantage for energy-constrained satellite systems;

[0068] 4. Achieve higher accuracy in reliability prediction of power MOSFET, with a prediction accuracy of over 95%, which is significantly improved compared with RNN and LSTM methods, effectively reducing the misjudgment rate of fault prediction;

[0069] 5. With high robustness, the model is applicable to different power MOSFET devices and can effectively plan the maintenance and replacement cycle of satellite systems. At the same time, it can also be extended to similar reliability prediction problems in other fields, such as automobiles and avionics systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a principle block diagram of a reliability prediction method for a power field effect transistor according to the present invention;

[0071] Figure 2 It is a structural diagram of the Transformer model described in the present invention. DETAILED DESCRIPTION

[0072] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0073] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0074] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0075] Embodiment 1:

[0076] Combine the following Figure 1and Figure 2 This embodiment describes a reliability prediction method for a power field effect transistor, which includes:

[0077] S1, collecting sensor data as initial data, and performing data processing on the initial data;

[0078] S2, using genetic algorithm to optimize the hyperparameters of the Transformer model and constructing the GA-Transformer model;

[0079] S3, transmits the data obtained by S1 to the GA-Transformer model obtained by S2, evaluates the accuracy of the prediction based on the performance indicators, predicts the remaining useful life RUL, and outputs the reliability prediction results.

[0080] In this embodiment, in order to quantitatively analyze the prediction performance of the model, a performance index is used as an evaluation index of the model. For the reliability of the device, the remaining useful life (RUL) is used as an evaluation index. The RUL represents the time that the device is expected to operate normally in the current state. The RUL is helpful for the reliability maintenance and resource allocation of the device.

[0081] Furthermore, the specific method of processing the initial data described in S1 includes:

[0082] S1-1, perform data cleaning, data resampling and data noise reduction on the initial data in turn;

[0083] S1-2, divide the data obtained in S1-1 into a training set and a test set;

[0084] S1-3. Normalize the divided data.

[0085] In this implementation, data cleaning of the initial data can remove outliers, data resampling can ensure consistency, and data denoising can improve data quality.

[0086] Furthermore, the Transformer model described in S2 specifically includes:

[0087] Select ΔR DS(on) As the failure characteristic parameter for model establishment, the failure threshold of the power field effect transistor is selected as:

[0088] ΔR DS(on) =0.18Ω;

[0089] The data is normalized using the minimum-maximum normalization calculation method:

[0090]

[0091] Among them, ΔR DS(on)k ′ represents the normalized on-resistance value, ΔR DS(on)k Indicates the on-resistance value before normalization, min(ΔR DS(on) ) represents the minimum value of the on-resistance before normalization, max(ΔR DS(on) ) represents the maximum value of the on-resistance before normalization, and k represents the number of on-resistances;

[0092] The encoder layer uses multi-head attention mechanism and layer normalization to process the input sequence;

[0093] The feed-forward network layer uses a 1D convolutional layer and a dropout layer to increase the nonlinearity of the model;

[0094] The output layer uses a flatten layer, a fully connected layer, and a dropout layer.

[0095] Furthermore, the specific method of optimizing the hyperparameters of the Transformer model using the genetic algorithm described in S2 includes:

[0096] S2-1. Define key parameters, including population size, number of iterations, dimension, retention rate, random selection rate, and mutation probability;

[0097] S2-2. Generate a new population based on selection, crossover and mutation operations, and determine whether the termination condition is reached. If not, repeat this step. If yes, stop the iteration and output the optimal solution.

[0098] Furthermore, the population size described in S2-1 is set to 50, that is, there are 50 different candidate solutions in each generation;

[0099] The number of iterations is set to 50, that is, the algorithm will evolve for 50 generations to find the optimal or approximately optimal solution;

[0100] The dimension is set to 8, that is, each candidate solution consists of 8 values;

[0101] The retention rate is set to 0.3, that is, 30% of the best individuals will be directly passed to the next generation without being generated through crossover or mutation;

[0102] The random selection rate is set to 0.5, that is, 50% of the inferior individuals will be randomly selected and retained to the next generation to maintain genetic diversity;

[0103] The mutation probability is set to 0.6, that is, a 60% mutation probability is set for each gene.

[0104] Furthermore, the termination conditions include:

[0105] t≥T max ;

[0106] f(x best )≤F thresh ;

[0107] Among them, t represents the current iteration number, T max represents the maximum number of iterations, f(x best ) represents the validation loss function, F thresh represents the validation loss threshold, which is set to 0.001.

[0108] Furthermore, the optimized hyperparameters described in S2 include: sliding window w, attention head size h, number of attention heads m, feedforward dimension f, batch size b, learning rate l, drop rate d and activation function a.

[0109] Furthermore, when the population size is 50, the initial population generated is P0, and each individual is a set of hyperparameters:

[0110] P0={x1,x2,·...,x 50};

[0111] Among them, x i represents an individual, x i =[w i ,h i ,m i ,f i ,b i ,l i ,d i ,a i ] represents the sliding window w of the i-th individual i , attention head size h i , number of attention heads m i , feedforward dimension f i , batch size b i , learning rate l i , discard rate d i And the activation function a i Settings; i = 1 ~ 50;

[0112] Define the fitness function f(x i ), evaluate the quality of individuals:

[0113] f(x i )=Loss(x i );

[0114] Among them, Loss(x i ) represents the validation loss on the Transformer model validation set;

[0115] The tournament selection method is used to select based on fitness. A group of individuals is randomly selected and the individual with the highest fitness is selected from them. The size of the tournament is k = 2. For each tournament, k individuals are randomly selected from the population, which is expressed as randomly selecting a subset T from the population P, where |T| = k. For each tournament subset T, the individual with the highest fitness is selected to enter the next generation:

[0116]

[0117] Among them, x i ∈T,x best represents the optimal individual, For x i , so that the function f(x i ) maximizes x i The value of

[0118] Randomly select two individuals as parents to generate one or more offspring, where the parameters of the offspring are the combination of the parameters of the parents;

[0119] x p =[w p ,h p ,m p ,f p ,b p ,l p ,d p ,a p ] and x q =[w q ,h q ,m q ,f q ,b q ,l q ,d q ,a q ] as the parent individual, and obtain the offspring x through single-point crossover child =[w p ,h p ,m p ,f q ,b q ,l q ,d q ,a q ];

[0120] Randomly adjust the parameters of individuals to increase the diversity of the population:

[0121] x i ′=[w i ,h i ,m i ,f i ,b i ,li ,d i ,a i ]+Δ

[0122] Among them, x i ′ represents the individual after mutation, and Δ represents random disturbance.

[0123] Furthermore, the performance indicators in S3 include mean absolute percentage error MAPE, root mean square error RMSE and mean square error MSE;

[0124] The calculation method of the mean absolute percentage error MAPE includes:

[0125]

[0126] Where N is the total number of samples of on-resistance, x i is the actual on-resistance value, x i ′ is the predicted on-resistance value;

[0127] The calculation method of the root mean square error RMSE includes:

[0128]

[0129] The calculation method of the mean square error MSE includes:

[0130]

[0131] In this implementation, the mean absolute percentage error (MAPE) can intuitively represent the accuracy of the prediction, the root mean square error (RMSE) can reduce the impact of large error values, and the mean square error (MSE) can evaluate the degree of change in the data.

[0132] Furthermore, the calculation method of the remaining useful life RUL in S3 includes:

[0133] RUL p =t f -t p ;

[0134] RUL a =t a -t p ;

[0135] Error=RUL a -RUL p ;

[0136]

[0137] Among them, RUL p Represents the predicted remaining life, RUL arepresents the actual remaining life, Error represents the deviation between the predicted remaining life and the actual remaining life, RA represents the relative accuracy of the prediction result, and t p Indicates the start time of the prediction, t f It indicates the time when the prediction reaches the failure threshold, t a Indicates the actual failure time point.

[0138] In this implementation, the remaining useful life RUL is used as an evaluation index for device reliability. RUL represents the time the device is expected to operate normally under the current state. RUL helps with device reliability maintenance and resource allocation.

[0139] In the present invention, a reliability prediction method for a power field effect transistor is proposed, such as Figure 1 As shown in the figure, in the first stage, we collect raw sensor readings or data points and use data processing methods, including: data cleaning to remove outliers, resampling to ensure consistency, noise reduction to improve data quality, etc. At the same time, we divide the data into training and test sets and perform normalization to facilitate a more efficient deep learning training process; in the second stage, we use "genetic algorithm to optimize the hyperparameters of the Transformer model. The algorithm iteratively refines the hyperparameter space based on evolutionary principles such as selection, crossover and mutation, and fitness evaluation based on mean square error (MSE). The optimization cycle continues until the predefined stopping condition is reached; in the third stage, the accuracy of the prediction is evaluated by various performance indicators such as MAPE, RMSE and MAE, and RUL prediction is performed to output accurate reliability prediction results.

[0140] The Tensorflow library was used to develop the model, and GPU was used for accelerated computing. The Adam algorithm was used to train the deep learning model. The main software configurations used for the experiment were Python 3.9.18, keras2.6.0, Tensorflow2.6.0, Scikit-learn 1.3.0, and numpy1.19.5. Considering the randomness of the model parameter initialization, we ran each model 10 times and took the average value.

[0141] To eliminate the effect of the initial value difference of the devices, select ΔR DS(on) As the failure characteristic parameter for model establishment, the failure threshold of the MOSFET device is selected as ΔR DS(on) =0.18Ω, in order to accelerate the convergence speed of the model and improve the stability of the model, the data is normalized and the minimum-maximum normalization calculation method is adopted:

[0142]

[0143] In the formula, ΔR DS(on)k ′ is the normalized on-resistance value, ΔR DS(on)k is the normalized on-resistance value, min(ΔR DS(on) ) is the minimum value of the on-resistance before normalization, max(ΔR DS(on) ) is the maximum value of the on-resistance before normalization, and k is the number of on-resistances.

[0144] The sliding window method is uniformly adopted, with a window size of 3 and a sliding step of 1. The model training cycle is uniformly set to 200. The early stopping technology is used to improve the training efficiency. The training cycle of the model early stopping technology is set to 20 to avoid overfitting of the model and save time and resources. For the Transformer network model, the encoder layer uses layer normalization and multi-head attention layer to process the sequence input. In the multi-head attention layer, the model can learn the dependencies between all positions within the input sequence. The feedforward network consists of a 1-dimensional convolutional layer and a dropout layer, which increases the nonlinear ability of the model. The output layer is composed of a flat layer followed by a fully connected layer and a dropout layer. The early stopping technology is used to prevent overfitting during the model training process, that is, the verification loss is monitored through the early stopping callback. When the verification loss has not improved in 20 consecutive epochs, the training will be stopped.

[0145] In the process of running the genetic algorithm for optimization, the key parameters of the algorithm are defined. The population size is set to 50, indicating that 50 different candidate solutions will be considered in each generation. The number of iterations is set to 50, and the algorithm will evolve for 50 generations to find the optimal or approximately optimal solution. The dimension is set to 8, indicating that each candidate solution will consist of 8 values. The retention rate is set to 0.3, so 30% of the best individuals will be directly passed to the next generation without being produced by crossover or mutation. The random selection rate is 0.5, which specifies that 50% of the poor individuals will be randomly selected and retained to the next generation to maintain genetic diversity and prevent the algorithm from converging to the local optimal solution too early. The mutation probability is set to 0.6, and a 60% mutation probability is set for each gene. It is worth mentioning that multiple termination criteria are used to determine whether the algorithm should stop iterating based on the maximum number of iterations and verification loss. The maximum number of iterations is 50, and the verification loss threshold is 0.001.

[0146] In the genetic algorithm, eight hyperparameters of the Transformer model are selected for optimization, namely: sliding window (w), attention head size (h), number of attention heads (n), feedforward dimension (f), batch size (b), learning rate (l), dropout rate (d), and activation function (a).

[0147] For a population size of 50, generate an initial population P0 where each individual is a possible set of hyperparameters:

[0148] P0={x1,x2,·...,x 50}

[0149] Among them, x i =[w i ,h i ,m i ,f i ,b i ,d i ,a i ], x i Represents the settings of sliding window, attention head size, number of attention heads, feedforward layer dimension, batch size, learning rate, dropout rate and activation function for individual i, i = 1 to 50.

[0150] Define the fitness function f(x) to evaluate the quality of individuals:

[0151] f(x)=Loss(x)

[0152] Among them, Loss(x) represents the validation loss on the Transformer model validation set.

[0153] The selection is based on fitness. The tournament selection method is used to randomly select a group of individuals and select the individual with the highest fitness from them. The size of the tournament is k = 2. For each tournament, k individuals are randomly selected from the population, which is expressed as randomly selecting a subset T from the population P, where |T| = k. For each tournament subset T, the individual with the highest fitness is selected to enter the next generation:

[0154]

[0155] Among them, x best represents the optimal individual, f(x i ) is the individual x i The fitness of x i ∈T, For x i , find the function f(x i ) maximizes the x i The value of .

[0156] Randomly select two individuals (parents) and generate one or more offspring, where the parameters of the offspring are some combination of the parameters of the parents, x p =[w p ,h p ,m p ,f p ,b p ,d p ,a p ] and x q =[wq ,h q ,m q ,f q ,b q ,d q ,a q ] as the parent individual, and obtain the offspring x through single-point crossover child =[w p ,h p ,m p ,f q ,b q ,d q ,a q ].

[0157] Randomly adjust the parameters of individuals to increase the diversity of the population:

[0158] x i ′=[w i ,h i ,m i ,f i ,b i ,d i ,a i ]+Δ

[0159] Among them, x i ′ represents the individual after mutation, and Δ represents a small random disturbance.

[0160] A new population is generated based on selection, crossover and mutation operations. This process is repeated until the stopping condition is reached:

[0161] t≥T max

[0162] f(x best )≤F thresh

[0163] Among them, t represents the current iteration number, T max represents the maximum number of iterations, F thresh represents the validation loss threshold, which is 0.001.

[0164] In order to quantitatively analyze the prediction performance of the model, the mean absolute percentage error (MAPE), root mean square error (RMSE) and mean square error (MSE) are used as evaluation indicators of the model. The mean absolute percentage error (MAPE) can intuitively represent the accuracy of the prediction, the root mean square error (RMSE) can reduce the impact of large error values, and the mean square error (MSE) can evaluate the degree of change of the data. The calculation formula is:

[0165]

[0166]

[0167]

[0168] N represents the total number of samples of on-resistance, x i is the actual on-resistance value, x i ' is the predicted on-resistance value.

[0169] For device reliability, we use the remaining useful life (RUL) as an evaluation indicator. RUL represents the time that the device is expected to operate normally under the current state. RUL helps with device reliability maintenance and resource allocation. The calculation formula is:

[0170] RUL p =t f -t p

[0171] RUL a =t a -t p

[0172] Error=RUL a -RUL p

[0173]

[0174] Where RUL p Represents the predicted remaining life, RUL a represents the actual remaining life, Error represents the deviation between the predicted remaining life and the actual remaining life, RA represents the relative accuracy of the prediction result, and t p Indicates the start time of the prediction, t f It indicates the time when the prediction reaches the failure threshold, t a Indicates the actual failure time point.

[0175] In the present invention, a Transformer model is proposed. The model uses a one-dimensional convolutional layer as a feedforward network to flatten the data processed by the Transformer encoder block, and then further processes it through a fully connected layer. Finally, a linear activation function is used as the output layer to predict the reliability of power MOSFET, thereby expanding the application scope of the Transformer model. At the same time, a GA-Transformer model is proposed, which combines the global search advantage of the genetic algorithm and the powerful time series data processing capability of the Transformer model, surpassing the existing technology in various performance evaluation indicators and solving the problem of time-consuming and labor-intensive model adjustment.

[0176] In the present invention, an application framework based on deep learning reliability prediction is established. Through the end-to-end prediction framework, from data collection, data processing to model training, the remaining service life of MOSFET is calculated, and finally the reliability prediction of MOSFET is realized.

[0177] In this invention, a comprehensive analysis and evaluation framework for deep learning models is established. It not only focuses on the accuracy of the model, but also conducts a comprehensive analysis of the model performance from multiple aspects, such as computational efficiency, generalization ability, stability, etc., and conducts a comparative analysis of the most widely used deep learning models to better understand the performance of the model in practical applications.

[0178] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in a manner different from that described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be used in other described embodiments.

Claims

1. A reliability prediction method for a power field effect transistor, characterized in that: It includes: S1, collecting sensor data as initial data, and performing data processing on the initial data; S2, using genetic algorithm to optimize the hyperparameters of the Transformer model and constructing the GA-Transformer model; S3, transmits the data obtained by S1 to the GA-Transformer model obtained by S2, evaluates the accuracy of the prediction based on the performance indicators, predicts the remaining useful life RUL, and outputs the reliability prediction results; The Transformer model described in S2 specifically includes: Select ΔR DS(on) As the failure characteristic parameter for model establishment, the failure threshold of the power field effect transistor is selected as: ΔR DS(on) =0.18Ω; The data is normalized using the minimum-maximum normalization calculation method: Among them, ΔR DS(on)k ′ represents the normalized on-resistance value, ΔR DS(on)k Indicates the on-resistance value before normalization, min(ΔR DS(on) ) represents the minimum value of the on-resistance before normalization, max(ΔR DS(on) ) represents the maximum value of the on-resistance before normalization, and k represents the number of on-resistances; The encoder layer uses multi-head attention mechanism and layer normalization to process the input sequence; The feed-forward network layer uses a 1D convolutional layer and a dropout layer to increase the nonlinearity of the model; The output layer uses a flatten layer, a fully connected layer, and a dropout layer.

2. A reliability prediction method for a power field effect transistor according to claim 1, characterized in that: The specific method for processing the initial data described in S1 includes: S1-1, perform data cleaning, data resampling and data noise reduction on the initial data in turn; S1-2, divide the data obtained in S1-1 into a training set and a test set; S1-3. Normalize the divided data.

3. The reliability prediction method of a power field effect transistor according to claim 1, characterized in that: The specific method of optimizing the hyperparameters of the Transformer model using the genetic algorithm described in S2 includes: S2-1. Define key parameters, including population size, number of iterations, dimension, retention rate, random selection rate, and mutation probability; S2-2. Generate a new population based on selection, crossover and mutation operations, and determine whether the termination condition is reached. If not, repeat this step. If yes, stop the iteration and output the optimal solution.

4. The reliability prediction method of a power field effect transistor according to claim 3, characterized in that: The population size described in S2-1 is set to 50, that is, there are 50 different candidate solutions in each generation; The number of iterations is set to 50, that is, the algorithm will evolve for 50 generations to find the optimal or approximately optimal solution; The dimension is set to 8, that is, each candidate solution consists of 8 values; The retention rate is set to 0.3, that is, 30% of the best individuals will be directly passed to the next generation without being generated through crossover or mutation; The random selection rate is set to 0.5, that is, 50% of the inferior individuals will be randomly selected and retained to the next generation to maintain genetic diversity; The mutation probability is set to 0.6, that is, a 60% mutation probability is set for each gene.

5. A reliability prediction method for a power field effect transistor according to claim 4, characterized in that: The termination conditions include: t≥T max ; f(x best )≤F thresh ; Among them, t represents the current iteration number, T max represents the maximum number of iterations, f(x best ) represents the validation loss function, F thresh represents the validation loss threshold, which is set to 0.

001.

6. The reliability prediction method of a power field effect transistor according to claim 5, characterized in that: The optimized hyperparameters described in S2 include: sliding window w, attention head size h, number of attention heads m, feedforward dimension f, batch size b, learning rate l, dropout rate d and activation function a.

7. A reliability prediction method for a power field effect transistor according to claim 6, characterized in that: When the population size is 50, the initial population generated is P0, and each individual is a hyperparameter set: P0={x1,x2,·...,x 50 }; Among them, x i represents an individual, x i =[w i ,h i ,m i ,f i ,b i ,l i ,d i ,a i ] represents the sliding window w of the i-th individual i , attention head size h i , number of attention heads m i , feedforward dimension f i , batch size b i , learning rate l i , discard rate d i And the activation function a i Settings; i = 1 ~ 50; Define the fitness function f(x i ), evaluate the quality of individuals: f(x i )=Loss(x i ); Among them, Loss(x i ) represents the validation loss on the Transformer model validation set; The tournament selection method is used to select based on fitness. A group of individuals is randomly selected and the individual with the highest fitness is selected from them. The size of the tournament is k = 2. For each tournament, k individuals are randomly selected from the population, which is expressed as randomly selecting a subset T from the population P, where |T| = k. For each tournament subset T, the individual with the highest fitness is selected to enter the next generation: Among them, x i ∈T,x best represents the optimal individual, For x i , so that the function f(x i ) maximizes x i The value of Randomly select two individuals as parents to generate one or more offspring, where the parameters of the offspring are the combination of the parameters of the parents; x p =[w p ,h p ,m p ,f p ,b p ,l p ,d p ,a p ] and x q =[w q ,h q ,m q ,f q ,b q ,l q ,d q ,a q ] as the parent individual, and obtain the offspring x through single-point crossover child =[w p ,h p ,m p ,f q ,b q ,l q ,d q ,a q ]; Randomly adjust the parameters of individuals to increase the diversity of the population: x i ′=[w i ,h i ,m i ,f i ,b i ,l i ,d i ,a i ]+Δ Among them, x i ′ represents the individual after mutation, and Δ represents random disturbance.

8. The reliability prediction method of a power field effect transistor according to claim 7, characterized in that: The performance indicators in S3 include mean absolute percentage error (MAPE), root mean square error (RMSE) and mean square error (MSE); The calculation method of the mean absolute percentage error MAPE includes: Where N is the total number of samples of on-resistance, x i is the actual on-resistance value, x′ i is the predicted on-resistance value; The calculation method of the root mean square error RMSE includes: The calculation method of the mean square error MSE includes:

9. The reliability prediction method of a power field effect transistor according to claim 8, characterized in that: The calculation method of the remaining useful life RUL described in S3 includes: RUL p =t f -t p ; RUL a =t a -t p ; Error=RUL a -RUL p ; Among them, RUL p Indicates the predicted remaining life, RUL a represents the actual remaining life, Error represents the deviation between the predicted remaining life and the actual remaining life, RA represents the relative accuracy of the prediction result, and t p Indicates the start time of the prediction, t f It indicates the time when the prediction reaches the failure threshold, t a Indicates the actual failure time point.

Images