GaN device power cycle failure prediction method based on long short-term memory network

Abstract

The invention discloses a GaN device power cycle failure prediction method based on a long short-term memory network, and the method comprises the steps: firstly collecting a power cycle dynamic parameter of a GaN device, then carrying out the preprocessing of the power cycle dynamic parameter of the GaN device, obtaining the preprocessing data, constructing a multi-dimensional time sequence input feature tensor according to the preprocessing data, and carrying out the prediction of the power cycle failure of the GaN device. The method comprises the following steps of: inputting a multi-dimensional time sequence input feature tensor into a GaN device, optimizing a long-short-term memory network to obtain a power cycle failure prediction model, inputting the multi-dimensional time sequence input feature tensor into the power cycle failure prediction model, performing power cycle failure prediction on the GaN device, and adaptively updating the power cycle failure prediction model. According to the method, the power cycle test of the GaN device can be terminated in advance through failure prediction, the data requirement of the target working condition is reduced, and the overall test time is shortened.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of gallium nitride power devices, and particularly relates to a GaN device power cycle failure prediction method based on a long short-term memory network. BACKGROUND

[0002] Power cycle (PC) test is an important means to evaluate the long-term reliability of GaN devices, and its core is to simulate the thermal stress in the actual working environment by repeatedly applying temperature fluctuations (such as junction temperature fluctuations Δ T j ) until the device fails. However, the existing technology has the following shortcomings: (1) PC test cycle is too long: GaN power device life model relies on a large amount of experimental data, and the traditional PC test method needs to be tested continuously until the preset failure standard is reached, and the test time is as long as several weeks or months. In the case of low junction temperature fluctuation (such as Δ T j <50K> and the increase of the number of test samples, the increase of test time is more prominent, which needs to spend a lot of economic and time cost.

[0003] (2) Low universality of failure prediction method: existing prediction models are mostly based on a single failure precursor (such as on-state resistance R DS(on) ) or specific test conditions (such as fixed Δ T j 、 T on / T off parameters, T on representing the on-time of the device, T off representing the off-time of the device), resulting in poor model migration ability. Most of the current prediction modeling data are for Si-based and SiC devices, and the dynamic Ron degradation, current collapse and other characteristics of GaN devices are not fully considered. Different prediction algorithms mainly consider prediction accuracy, and the selection basis and explainability of the prediction method are not considered, and the migration ability of the algorithm is insufficient, which is difficult to capture the timing degradation rule. SUMMARY

[0004] The purpose of the present application is to solve the problems of long test cycle and weak migration ability of existing GaN device failure analysis technology, and to propose a GaN device power cycle failure prediction method based on a long short-term memory network.

[0005] The technical scheme of the present application is: a GaN device power cycle failure prediction method based on a long short-term memory network, comprising the following steps: S1, collect power cycle dynamic parameters of the GaN device.

[0006] S2, preprocess the power cycle dynamic parameters of the GaN device to obtain preprocessed data.

[0007] S3, construct a multi-dimensional time sequence input feature tensor according to the preprocessed data.

[0008] S4, optimize the long short-term memory network to obtain a power cycle failure prediction model.

[0009] S5, input the multi-dimensional time sequence input feature tensor into the power cycle failure prediction model, predict the power cycle failure of the GaN device, and update the power cycle failure prediction model adaptively.

[0010] Further, step S1 is specifically: under the condition of constant on / off time T on / T off Under the mode, the power cycle dynamic parameters of the GaN device are obtained through the power cycle test platform, simulation data and public data set, and the power cycle dynamic parameters include gate voltage, on voltage, load current, maximum junction temperature, minimum junction temperature, junction temperature fluctuation, threshold voltage and thermal resistance.

[0011] Further, step S2 includes the following steps: S21, superimpose Gaussian noise on the power cycle dynamic parameters of the GaN device, and perform time interpolation to simulate different heat dissipation conditions and load fluctuations of the GaN device.

[0012] S22, expand the low junction temperature fluctuation data amount in the power cycle dynamic parameters through simulation, and combine the power cycle dynamic parameters to form a mixed data set.

[0013] S23, temperature decoupling is performed on the data in the mixed data set to establish a mapping relationship between temperature and electrical parameters, and the influence of temperature rise on aging electrical parameters is excluded.

[0014] S24, filter the data after temperature decoupling to eliminate the influence of noise.

[0015] Further, step S3 is specifically: a time sequence matrix is constructed according to the four features of junction temperature, load current, on voltage and thermal resistance, and is used as a sample. Time sequence segments are extracted through a sliding window. The number of samples, time step and feature quantity are used as dimensions to construct a multi-dimensional time sequence input feature tensor.

[0016] Further, step S4 includes the following steps: S41, an adaptive forgetting gate and input gate weight adjustment strategy is introduced in the long short-term memory network, interference is suppressed and the memory ability of key features is strengthened by dynamically adjusting the gating parameters.

[0017] S42, a Bayesian optimization algorithm is used to search for the optimal network depth, number of units and learning rate of the long short-term memory network, and an early stopping strategy is used to prevent overfitting.

[0018] S43, a hybrid loss function is used to train the long short-term memory network to obtain a power cycle failure prediction model.

[0019] Further, the adaptive forgetting gate and input gate weight adjustment strategy in step S41 is specifically: A1, for each time step t , based on the input of the current long short-term memory network and the hidden state of the long short-term memory network at the previous time, the attention score is calculated: wherein represents an attention score function.

[0020] A2, the attention score is normalized to obtain the normalized attention score : wherein represents the attention score at time step k , and T represents the total time.

[0021] A3, the gating parameters of the forgetting gate and the input gate are dynamically adjusted according to the normalized attention score : wherein and represent the weight matrix of the forgetting gate and the input gate, respectively, and represent the bias term of the forgetting gate and the input gate, respectively, represents a Sigmoid function, represents a Hadamard product.

[0022] Further, the Bayesian optimization algorithm in step S42 is specifically: B1, the network depth, number of units and learning rate of the long short-term memory network are initialized and set as a combination of hyperparameters.

[0023] B2, constructing an objective function based on a Gaussian process : wherein represents a Gaussian process function, represents a mean function, represents a kernel function, represents a signal variance, represents a length scale, and respectively represent two different sets of hyperparameters.

[0024] B3, constructing an acquisition function : wherein represents a current optimal performance, represents a cumulative distribution function of a standard normal distribution, represents a probability density function of a standard normal distribution, represents an expected improvement variable, represents a standard deviation, represents a mean.

[0025] B4, selecting an optimal candidate hyperparameter combination through the acquisition function.

[0026] B5, training a long short-term memory network using an early stopping strategy, and recording the validation set performance through the objective function.

[0027] B6, determining whether a maximum number of iterations is reached or the validation set performance converges to a preset result, if so, obtaining the optimal network depth, number of units and learning rate of the long short-term memory network, otherwise updating the objective function with the optimal candidate hyperparameter combination as new observation data, and returning to step B4 for the next iteration.

[0028] Further, the early stopping strategy specifically is: in response to the validation set performance of the long short-term memory network not improving for consecutive patience training cycles, triggering early stopping to stop training, wherein patience represents a patience value.

[0029] Further, the specific formula of the mixed loss function in step S43 is: wherein , and are weight coefficients, represents a cross-entropy loss, denotes the KL divergence loss, denotes the mean absolute error.

[0030] Further, the step S5 comprises the following sub-steps: S51, inputting the multi-dimensional time sequence input feature tensor into the power cycle failure prediction model, outputting the predicted value of the power cycle data of each time step, and thus performing power cycle failure prediction on the GaN device.

[0031] S52, performing reliability evaluation on the power cycle failure prediction result in combination with a preset confidence interval.

[0032] S53, periodically comparing the newly collected power cycle dynamic parameters with the power cycle failure prediction result, updating the parameters of the power cycle failure prediction model through online learning, and avoiding model drift.

[0033] S54, cross-verifying the power cycle failure prediction result in combination with a physical failure model, and ensuring that the power cycle failure prediction result conforms to the physical law.

[0034] The beneficial effects of the present application are: (1) The power cycle test of the GaN device can be terminated in advance through failure prediction, the data requirement of the target working condition is reduced, and the overall test time is shortened by 40%-60%.

[0035] (2) The multi-dimensional time sequence input feature tensor constructed through the gating mechanism optimization and feature fusion avoids the problem of inaccurate prediction model of single feature, and significantly improves the capture ability of the nonlinear degradation trajectory.

[0036] (3) The power cycle failure prediction model is trained using a large amount of data under standard or common working conditions, and when the failure of the GaN device under different junction temperature fluctuations (Δ T j ) needs to be predicted, the trained power cycle failure prediction model is directly used for migration, without the need to redesign the experiment or train the model from scratch for each target working condition, so that the failure prediction method can adapt to the failure prediction of GaN devices with a wide range of junction temperature fluctuations Δ T j , and the cost of repeated experiments is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The figure shows the GaN device power cycle failure prediction method flowchart provided by the embodiment of the present application based on the long short-term memory network. DETAILED DESCRIPTION

[0038] Exemplary embodiments of the present application will now be described in detail with reference to the accompanying drawings. It should be understood that the embodiments illustrated and described herein are merely exemplary of the principles and spirit of the present application, and are not intended to limit the scope of the present application.

[0039] The embodiment of the present application provides a GaN device power cycle failure prediction method based on a long short-term memory network, as shown in the figure, comprising the following steps S1-S5: Figure 1 S1, collecting power cycle dynamic parameters of the GaN device.

[0040] In the embodiment of the present application, in the constant on / off time T on / T off mode, the power cycle dynamic parameters of the GaN device are obtained through a power cycle test platform, simulation data and a public data set, and the power cycle dynamic parameters include gate voltage, on voltage, load current, maximum junction temperature, minimum junction temperature, junction temperature fluctuation, threshold voltage and thermal resistance.

[0041] The constant on / off time T on / T off mode is a control strategy in the power cycle test, and the core is to maintain stable periodic thermal cycle in the test process by fixing the on time T on and the off time T off .

[0042] In the embodiment of the present application, the simulation data is obtained through an electro-thermal-force simulation model, which is a multi-physical field coupling model used to simulate the dynamic behavior of the GaN device in the power cycle test, and the core is to combine electrical characteristics (such as current, voltage, on resistance, etc.), thermal characteristics (such as junction temperature, shell temperature, thermal resistance, etc.) and mechanical characteristics (such as thermal stress distribution, material fatigue, etc.) to predict the failure risk of the device under thermal-mechanical stress.

[0043] S2, preprocessing the power cycle dynamic parameters of the GaN device to obtain preprocessed data.

[0044] Step S2 includes the following sub-steps S21-S24: S21, superimposing Gaussian noise on the power cycle dynamic parameters of the GaN device, and performing time interpolation to simulate different heat dissipation conditions and load fluctuations of the GaN device.

[0045] ​S22, simulate the low junction temperature fluctuation data in the extended power cycle dynamic parameter through simulation, and combine the power cycle dynamic parameter to form a hybrid data set.

[0046] S23, temperature decoupling is performed on the data in the hybrid data set, a mapping relationship between temperature and electrical parameters (on voltage and threshold voltage in the power cycle dynamic parameter) is established, and the influence of temperature rise on the aging electrical parameter is excluded.

[0047] S24, filtering processing is performed on the data after temperature decoupling, noise influence is eliminated, irrelevant features are avoided to be learned by the long short-term memory network, and the accuracy of the power cycle aging trend analysis is enhanced.

[0048] S3, a multi-dimensional time sequence input feature tensor is constructed according to the preprocessed data.

[0049] In the embodiment of the application, considering the strong correlation between the load current, the on voltage, the thermal resistance and the temperature in the power cycle test, the junction temperature T j , the load current I ds , the on voltage V ds and the thermal resistance are used to construct a time sequence matrix, and are used as samples, a time sequence segment is extracted through a sliding window (100-500 sampling points are set in the embodiment of the application, and the time step is 10), and a multi-dimensional time sequence input feature tensor is constructed with the sample number, the time step and the feature number as dimensions. The multi-dimensional time sequence input feature tensor constructed in the embodiment of the application can balance the capturing ability of the power cycle failure prediction model to short-term fluctuations and long-term trends, and ensure that the power cycle failure prediction model can learn the dynamic correlation between parameters.

[0050] In the embodiment of the application, the junction temperature T j , the load current I ds , the on voltage V ds and the thermal resistance are all variables changing with the power cycle time, so each parameter itself is a time sequence, but in order to construct a multivariate time sequence, the measurement values of the four parameters need to be aligned by time to form a multivariate time sequence matrix, wherein each row corresponds to a time point and each column corresponds to a feature.

[0051] S4, the long short-term memory network is optimized to obtain a power cycle failure prediction model.

[0052] Step S4 includes the following sub-steps S41-S43: S41. Introducing adaptive forget gate and input gate weight adjustment strategies into the long short-term memory network, suppressing interference by dynamically adjusting the gate parameters, and strengthening the memory capacity of key features.

[0053] In the embodiment of the present invention, the adaptive forget gate and input gate weight adjustment strategy can dynamically adjust the gating parameters by calculating the importance score of each time step, specifically: A1. For each time step t , based on the input of the current long short-term memory network and the hidden state of the LSTM network at the previous moment , calculate the attention score : in represents the attention score function.

[0054] A2. Attention score Perform normalization operation to obtain normalized attention score : in Represents the time step k The attention score, T Indicates the total time.

[0055] A3. Based on the normalized attention score Dynamically adjust the forget gate and input gate Gating parameters: in and Represent the weight matrices of the forget gate and input gate respectively, and Represent the bias terms of the forget gate and input gate respectively, represents the Sigmoid function, represents the Hadamard product.

[0056] S42. A Bayesian optimization algorithm is used to search for the optimal network depth (2-4 layers), number of units (64-256), and learning rate (1e-4 to 1e-3) of the long short-term memory network, and an early stopping strategy is used to prevent overfitting.

[0057] In the embodiment of the present invention, the Bayesian optimization algorithm is specifically: B1. Initialize and set the network depth, number of units, and learning rate of the long short-term memory network as a hyperparameter combination.

[0058] B2, constructing the objective function based on the Gaussian process : wherein represents a Gaussian process function, represents a mean function, usually set to 0, represents a kernel function, represents a signal variance, represents a length scale, and respectively represent two different sets of hyperparameters.

[0059] B3, constructing the acquisition function : wherein represents the current optimal performance, represents the cumulative distribution function of the standard normal distribution, represents the probability density function of the standard normal distribution, represents the expected improvement variable, represents the standard deviation, represents the mean.

[0060] B4, selecting the optimal candidate hyperparameter combination through the acquisition function.

[0061] B5, training the long short-term memory network with an early stopping strategy, and recording the validation set performance through the objective function.

[0062] In the embodiment of the application, the early stopping strategy is specifically: in response to the fact that the validation set performance of the long short-term memory network has not improved for consecutive patience training cycles, early stopping is triggered, and training is stopped, wherein patience represents a patience value, which is used to control the number of consecutive non-improvement cycles allowed by the model when the validation set performance does not improve, and its core role is to determine the length of time that the model tolerates performance stagnation on the validation set, thereby avoiding missing the best model due to premature termination of training, and in the embodiment of the application, the value is patience=20.

[0063] B6, determining whether the maximum number of iterations is reached or the validation set performance converges to a preset result, if yes, obtaining the optimal network depth, the number of units and the learning rate of the long short-term memory network, otherwise updating the objective function with the optimal candidate hyperparameter combination as new observation data, and returning to step B4 for the next iteration.

[0064] S43, training the long short-term memory network with a hybrid loss function to obtain a power cycle failure prediction model.

[0065] In the embodiment of the present application, the local timing error and the global degradation trend can be balanced by the mixed loss function, and the mixed loss function The specific formula is: Wherein , and are weight coefficients, represents the cross-entropy loss (used for classification tasks, such as judging whether to fail), represents the KL divergence loss (constraining the Dirichlet distribution to prevent overfitting), represents the mean absolute error (used for regression tasks, such as predicting the remaining life).

[0066] S5, inputting the multi-dimensional time sequence input feature tensor into the power cycle failure prediction model, performing power cycle failure prediction on the GaN device, and adaptively updating the power cycle failure prediction model.

[0067] Step S5 includes the following steps S51-S54: S51, inputting the multi-dimensional time sequence input feature tensor into the power cycle failure prediction model, outputting the predicted value of each time step power cycle data (turn-on voltage, thermal resistance, etc.), and performing power cycle failure prediction on the GaN device.

[0068] S52, combining the preset confidence interval (95% in the embodiment of the present application) to perform reliability evaluation on the power cycle failure prediction result.

[0069] In the embodiment of the present application, the confidence interval is an estimated interval of the population parameter based on sample data, which means that under a given confidence level (such as 95%), the true parameter value has a 95% probability of falling within the interval. The goal of reliability evaluation is to judge the credibility of the power cycle failure prediction result through the confidence interval, and then evaluate the reliability of the power cycle failure prediction model.

[0070] When the prediction result is within the confidence interval, it means that the consistency between the prediction result and the actual data is high, and the power cycle failure prediction model has good reliability. The width of the confidence interval needs to be further analyzed. If the confidence interval is narrow, it means that the prediction accuracy is high, and the power cycle failure prediction model is reliable (such as large sample size and small data fluctuation); if the confidence interval is wide, it means that the prediction accuracy is low, and the sample size needs to be increased or the power cycle failure prediction model needs to be optimized (such as reducing noise and improving feature selection).

[0071] When the prediction result is outside the confidence interval, it means that the prediction result is biased, the power cycle failure prediction model is overfitting / underfitting or the confidence interval is set unreasonably. If the prediction result exceeds the confidence interval for many times, the power cycle failure prediction model needs to be retrained or the parameters (such as the optimization patience value patience) need to be adjusted.

[0072] S53, periodically (set to every 100 test cycles in the embodiment of the present application) compare the newly collected power cycle dynamic parameters with the power cycle failure prediction results, update the parameters of the power cycle failure prediction model through online learning, and avoid model drift.

[0073] S54, cross-verify the power cycle failure prediction results in combination with the physical failure model (the Coffin-Manson equation is used in the embodiment of the present application), and ensure that the power cycle failure prediction results conform to the physical law.

[0074] Those skilled in the art will realize that the embodiments described herein are for the purpose of helping the reader understand the principles of the present application and should be understood as not limiting the scope of protection of the present application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations according to the technical inspiration disclosed in the present application without departing from the essence of the present application, and these modifications and combinations are still within the scope of protection of the present application.

Claims

1. A GaN device power cycle failure prediction method based on long short-term memory network, characterized by: The following steps are involved: S1. Collect the power cycle dynamic parameters of GaN devices; S2. preprocessing the power cycle dynamic parameters of the GaN device to obtain preprocessed data; S3. Construct a multi-dimensional time series input feature tensor based on the preprocessed data; S4. Optimizing the long short-term memory network to obtain a power cycle failure prediction model; S5. Input the multi-dimensional time series input feature tensor into the power cycle failure prediction model, perform power cycle failure prediction on the GaN device, and adaptively update the power cycle failure prediction model.

2. The method for predicting power cycle failure of GaN devices based on long short-term memory networks according to claim 1, characterized in that: The step S1 is specifically as follows: At constant on / off time T on / T off In this mode, the power cycling dynamic parameters of the GaN device are obtained through the power cycling test platform, simulation data and public data sets. The power cycling dynamic parameters include gate voltage, on-state voltage, load current, maximum junction temperature, minimum junction temperature, junction temperature fluctuation, threshold voltage and thermal resistance.

3. The method for predicting power cycle failure of GaN devices based on long short-term memory networks according to claim 1, characterized in that: The step S2 comprises the following sub-steps: S21. Superimposing Gaussian noise on the power cycle dynamic parameters of the GaN device and performing time interpolation to simulate different heat dissipation conditions and load fluctuations of the GaN device; S22. Expand the amount of low junction temperature fluctuation data in the power cycle dynamic parameters through simulation, and form a hybrid data set by combining the power cycle dynamic parameters; S23, performing temperature decoupling on the data in the mixed data set, establishing a mapping relationship between temperature and electrical parameters, and eliminating the influence of temperature rise on aging electrical parameters; S24. Filter the temperature decoupled data to eliminate noise effects.

4. The method for predicting power cycle failure of GaN devices based on long short-term memory networks according to claim 2, characterized in that: The step S3 is specifically as follows: A time series matrix is ​​constructed based on the four features of junction temperature, load current, on-state voltage, and thermal resistance. These features are used as samples, and time series fragments are extracted through a sliding window. A multidimensional time series input feature tensor is constructed with the number of samples, time step, and number of features as dimensions.

5. The method for predicting power cycle failure of GaN devices based on long short-term memory networks according to claim 1, characterized in that: The step S4 comprises the following sub-steps: S41. Introducing adaptive forget gate and input gate weight adjustment strategies into the long short-term memory network, suppressing interference by dynamically adjusting the gate parameters, and strengthening the memory capacity of key features; S42. Use Bayesian optimization algorithm to search for the optimal network depth, number of units and learning rate of long short-term memory network, and use early stopping strategy to prevent overfitting; S43. A hybrid loss function is used to train the long short-term memory network to obtain a power cycle failure prediction model.

6. The method for predicting power cycle failure of GaN devices based on long short-term memory networks according to claim 5, characterized in that: The adaptive forget gate and input gate weight adjustment strategy in step S41 is specifically as follows: A1. For each time step t , based on the input of the current long short-term memory network and the hidden state of the LSTM network at the previous moment , calculate the attention score : in represents the attention score function; A2. Attention score Perform normalization operation to obtain normalized attention score : in Represents the time step k The attention score, T Indicates the total time; A3. Based on the normalized attention score Dynamically adjust the forget gate and input gate Gating parameters: in and Represent the weight matrices of the forget gate and input gate respectively, and Represent the bias terms of the forget gate and input gate respectively, represents the Sigmoid function, represents the Hadamard product.

7. The method for predicting power cycle failure of GaN devices based on long short-term memory networks according to claim 5, characterized in that: The Bayesian optimization algorithm in step S42 is specifically: B1. Initialize and set the network depth, number of units, and learning rate of the long short-term memory network as a hyperparameter combination; B2. Constructing the objective function based on Gaussian process : in represents the Gaussian process function, represents the mean function, represents the kernel function, represents the signal variance, represents the length scale, and Represent two different sets of hyperparameters; B3. Constructing the acquisition function : in Indicates the current optimal performance, represents the cumulative distribution function of the standard normal distribution, represents the probability density function of the standard normal distribution, Indicates the expected improvement variable, represents the standard deviation, represents the mean; B4. Select the optimal candidate hyperparameter combination through the acquisition function; B5. Use the early stopping strategy to train the long short-term memory network and record the performance of the validation set through the objective function; B6. Determine whether the maximum number of iterations has been reached or the performance of the validation set has converged to the preset result. If so, the optimal network depth, number of units, and learning rate of the long short-term memory network are obtained. Otherwise, the optimal candidate hyperparameter combination is used as the new observation data to update the objective function, and return to step B4 for the next iteration.

8. The method for predicting power cycle failure of GaN devices based on long short-term memory networks according to claim 7, characterized in that: Specifically, the early stopping strategy is as follows: in response to the long short-term memory network's validation set performance not improving for consecutive patience training cycles, early stopping is triggered to stop training, where patience represents the patience value.

9. The method for predicting power cycle failure of GaN devices based on long short-term memory networks according to claim 5, characterized in that: The mixed loss function in step S43 The specific formula is: in 、 and are weight coefficients, represents the cross entropy loss, represents the KL divergence loss, represents the mean absolute error.

10. The method for predicting power cycle failure of GaN devices based on long short-term memory networks according to claim 1, characterized in that: The step S5 comprises the following sub-steps: S51, inputting the multi-dimensional time series input feature tensor into the power cycle failure prediction model, outputting the predicted value of the power cycle data at each time step, thereby predicting the power cycle failure of the GaN device; S52. Perform reliability assessment on the power cycle failure prediction result in combination with a preset confidence interval; S53, regularly comparing newly collected power cycle dynamic parameters with power cycle failure prediction results, and updating parameters of the power cycle failure prediction model through online learning to avoid model drift; S54. Cross-validate the power cycle failure prediction results in combination with the physical failure model to ensure that the power cycle failure prediction results conform to physical laws.

Images