SiC MOSFET service life prediction method based on gate oxide layer aging

Abstract

The invention belongs to the technical field of power semiconductor device life prediction, and particularly relates to a SiC MOSFET life prediction method based on gate oxide layer aging. The method comprises the following steps: firstly, acquiring a real threshold voltage sequence under different working conditions through a SiC MOSFET accelerated aging test, and preprocessing the real threshold voltage sequence; then, a life prediction model is constructed based on the conditional generative adversarial network, and the life prediction model comprises a generator, a common discriminator, a conditional discriminator, a coding network and a reverse recovery network; enabling the working condition as a condition label and the preprocessed real threshold voltage sequence to sequentially pass through a coding network and a reverse recovery network to obtain a real threshold voltage sequence embedded with the working condition; using the working condition as a condition label and a randomly generated Gaussian noise sequence to pass through a generator to obtain an analog threshold voltage sequence embedded with the working condition; inputting the real threshold voltage sequence and the simulation threshold voltage sequence with the embedded working conditions into a common discriminator and a condition discriminator at the same time for discrimination; and finally, training the life prediction model, and generating a threshold voltage sequence under an expected working condition by using a trained generator. According to the method, the service life of the SiC MOSFET under different working conditions can be predicted offline, and the method is more applicable.

Description

Technical Field

[0001] The present invention belongs to the technical field of life prediction of power semiconductor devices, and in particular is a method for predicting the life of a SiC MOSFET based on gate oxide layer aging. Background Art

[0002] With the vigorous development of the chip industry, power semiconductor devices (such as MOSFET) as the core components of power electronic systems are widely used in various fields such as energy storage systems and new energy vehicles. Due to the complex working environment, the aging process of power semiconductor devices is affected by many factors, so its life prediction is of great significance to improving the reliability of power electronic systems.

[0003] Compared with traditional silicon-based MOSFET, SiC MOSFET has superior performance such as low on-resistance, high frequency, and high voltage due to the physical properties of the materials used, making it very popular in high-power and high-temperature applications. However, the barrier height at the interface of the gate oxide layer of SiC MOSFET is low, making it easier for carriers in the channel to cross the barrier to reach the oxide layer. The C element remaining at the interface during the oxidation process of SiC will produce a high interface state density at the interface of SiC / SiO2, resulting in a significant reduction in the quality of the gate oxide layer, making the gate of SiC MOSFET more susceptible to Fowler-Nordheim (FN) tunneling, causing performance degradation or even functional failure under long-term high temperature and gate bias stress. Therefore, life prediction based on gate oxide layer aging helps to achieve safe and efficient application of SiC MOSFET. Since the actual working environment of SiC MOSFET is relatively complex and it is difficult to obtain real-time data, predicting the life of SiC MOSFET under actual working conditions through accelerated aging tests helps to understand its degradation. Traditional life prediction methods include those based on failure physical models and those based on data-driven. The prediction based on failure physical models requires a specific analysis of the working conditions under different failure environments and the calculation of the life of SiCMOSFET through formulas. However, some physical information is difficult to extract, so it is difficult to build an accurate and complete physical model, and the versatility is poor. Data-driven prediction is based on neural networks or statistical analysis to predict the remaining life of SiC MOSFET. It needs to be monitored online and the remaining life is predicted through historical data. However, SiC MOSFET usually works under high temperature, high pressure and high frequency conditions, and relies on high-reliability sensors, data acquisition systems, etc. for monitoring, communication and calculation. It has a strong dependence on networks and communications, is costly, and has poor security. Summary of the invention

[0004] In view of the deficiencies in the prior art, the technical problem to be solved by the present invention is to provide a method for predicting the life of a SiC MOSFET based on gate oxide layer aging.

[0005] The present invention solves the technical problem by adopting the following technical solution: A method for predicting the life of a SiC MOSFET based on gate oxide layer aging, characterized in that the method comprises the following steps: Step 1: Obtain the real threshold voltage sequence under different working conditions through SiC MOSFET accelerated aging test, and pre-process the real threshold voltage sequence; Step 2: Construct a lifespan prediction model based on a conditional generative adversarial network, including a generator, a common discriminator, a conditional discriminator, an encoding network, and a reverse recovery network; The working conditions are input as conditional labels and the preprocessed real threshold voltage sequence into the encoding network to generate multiple two-dimensional space tensors; these two-dimensional space tensors are passed through the reverse recovery network to obtain the real threshold voltage sequence embedded with the working conditions; The generator includes a coding network, an encoder and a reverse recovery network; the working condition is input into the coding network as a condition label and a randomly generated Gaussian noise sequence to generate a plurality of simulated two-dimensional space tensors; these simulated two-dimensional space tensors are input into the encoder for encoding to obtain a plurality of coding features; these coding features are passed through the reverse recovery network to obtain a simulated threshold voltage sequence embedded with the working condition; The real threshold voltage sequence and the simulated threshold voltage sequence embedded in the working condition are simultaneously input into the common discriminator and the conditional discriminator for discrimination, the common discriminator is used to calculate the probability that the input threshold voltage sequence is the real threshold voltage sequence, and the conditional discriminator is used to calculate the probability that the input threshold voltage sequence is the real threshold voltage sequence under the input working condition; Step 3: Train the life prediction model and use the trained generator to generate a threshold voltage sequence under the expected operating conditions. When the rate of change of the SiC MOSFET threshold voltage relative to the initial value exceeds the specified value, the SiC MOSFET is considered to be damaged, thus achieving life prediction.

[0006] Compared with the prior art, the present invention has at least the following advantages: 1. The life prediction model based on conditional generative adversarial network is used to directly predict the life of SiC MOSFET. The threshold voltage sequence of SiC MOSFET under different working conditions can be predicted according to different working conditions and Gaussian noise sequences, so as to realize offline life prediction of SiC MOSFET instead of online prediction of the remaining life of SiC MOSFET. Therefore, in actual prediction, there is no need to rely on the historical data of SiC MOSFET under working conditions, and life prediction under various working conditions can be realized, which has strong applicability.

[0007] 2. In order to make the model better learn the correlation between the conditional label and the threshold voltage sequence, a normal discriminator and a conditional discriminator are set. The normal discriminator is used to judge whether the input threshold voltage sequence is real or generated, and the conditional discriminator is used to judge whether the input threshold voltage sequence is a real threshold voltage sequence under the input conditional label. The conditional label and the threshold voltage sequence are input into the encoding network, and the conditional label is embedded in the threshold voltage sequence to achieve better feature extraction.

[0008] 3. The present invention obtains the SiC MOSFET threshold voltage sequence through accelerated aging test, which is used to predict the life under actual working conditions, realizes offline SiC MOSFET life prediction, and avoids the high hardware requirements and high computing power requirements for real-time acquisition. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 It is a structural diagram of the life prediction model of the present invention. DETAILED DESCRIPTION

[0010] Specific embodiments are given below in conjunction with the accompanying drawings. The specific embodiments are only used to introduce the technical solutions of the present invention in detail and are not intended to limit the protection scope of the present application.

[0011] The present invention provides a method for predicting the life of SiC MOSFET based on gate oxide aging (hereinafter referred to as method, see Figure 1 ), including the following steps: Step 1: Use the dynamic high-temperature gate bias test platform to conduct SiC MOSFET accelerated aging tests, collect real threshold voltage sequences under different working conditions, form a data set, and perform preprocessing; The dynamic high-temperature gate bias experimental platform includes a host computer, SiC MOSFET, a heating platform, a test circuit, a power supply, an adapter board and a DSP (digital signal processor); the SiC MOSFET is connected to the test circuit and fixed on the heating platform at the same time, and the temperature of the SiC MOSFET is controlled by the heating platform; the test circuit is connected to the power supply and DSP through the adapter board at the same time, and the DSP is connected to the host computer. The high level is sent by the DSP as the gate voltage of the SiC MOSFET aging test, and the low level is a -10V square wave to control the SiC MOSFET to turn on and off cyclically. During the experiment, the threshold voltage is measured every 10s. When the rate of change of the SiC MOSFET threshold voltage exceeds the specified value (20%) relative to the initial value, it is considered that the SiC MOSFET has been damaged, and the experiment is stopped; by changing the operating conditions (including gate voltage, switching frequency, temperature, etc.), the real threshold voltage sequence under different operating conditions is obtained.

[0012] Preprocessing includes data smoothing and coordinate axis flipping. First, due to the instability of device connection, one or multiple consecutive outliers with large differences occasionally appear. Therefore, a data screening method is used to identify outliers. The value of the outlier is calculated according to the following formula to replace the original value of the outlier. (1) In the formula, is the value of the outlier, and is the value of the two points before the outlier, and is the value of the two points after the outlier; Then, the moving average method is used to calculate the moving average and standard deviation of the window. According to formula (2), the outliers that are significantly different from the moving average of the window are identified. The values ​​of the two adjacent points before and after the outlier are averaged, and the original value of the outlier is replaced by the average value. (2) In the formula, is the value of any point in the window, , are the moving mean and standard deviation of the window, is a constant coefficient, in this embodiment ; Gaussian filtering is further used to remove noise and complete data smoothing.

[0013] Since the threshold voltage sequences under different working conditions vary greatly in the time dimension and are difficult to predict, the coordinate axes of the threshold voltage sequence are flipped, that is, the horizontal axis is converted to the threshold voltage, and the vertical axis is converted to time.

[0014] Step 2: Construct a life prediction model based on a conditional generative adversarial network, including a generator, a common discriminator, a conditional discriminator, an encoding network, and a reverse recovery network; the generator is used to generate a simulated threshold voltage sequence under working conditions, the common discriminator is used to determine whether the input threshold voltage sequence is real or generated, the conditional discriminator is used to determine whether the input threshold voltage sequence is a real threshold voltage sequence under the working conditions, the encoding network is used to embed the conditional label into the sequence and encode the one-dimensional sequence into a two-dimensional space tensor, and the reverse recovery network is used to restore the input data to a one-dimensional sequence; The working conditions are input into the encoding network as conditional labels and the preprocessed real threshold voltage sequence. The conditional labels are first concatenated to the real threshold voltage sequence, and then the concatenated sequence is subjected to fast Fourier transform to realize recursive decomposition of the sequence, thereby decomposing the real threshold voltage sequence into multiple threshold voltage sequences in the frequency domain. Then, the amplitude of the threshold voltage sequence in each frequency domain is calculated, and K threshold voltage sequences with large amplitudes are selected for reshaping operations, and the one-dimensional threshold voltage sequence is mapped to a two-dimensional space tensor, and a total of K two-dimensional space tensors are obtained. K two-dimensional space tensors are input into the inverse recovery network, and the K two-dimensional space tensors are convolved respectively to extract high-dimensional features to obtain K high-dimensional features. A weight is assigned to each high-dimensional feature and all high-dimensional features are spliced ​​to obtain spliced ​​features. The spliced ​​features are reshaped to map the spliced ​​features from the two-dimensional space to the one-dimensional space to obtain the real threshold voltage sequence embedded in the working conditions.

[0015] The generator includes a coding network, an encoder and a reverse recovery network; the working condition is input into the coding network as a condition label and a randomly generated Gaussian noise sequence to obtain K simulated two-dimensional space tensors; the K simulated two-dimensional space tensors are input into the encoder for encoding to obtain K coding features; the K coding features are passed through the reverse recovery network to obtain a simulated threshold voltage sequence embedded with the working condition; The real threshold voltage sequence and the simulated threshold voltage sequence embedded in the working condition are simultaneously input into the common discriminator and the conditional discriminator for discrimination. The common discriminator includes an encoder, a fully connected layer and a Sigmoid activation function. The input threshold voltage sequence is passed through the encoder to extract the encoding features, and the encoding features are sequentially passed through the fully connected layer and the Sigmoid activation function to obtain the probability that the input threshold voltage sequence is the real threshold voltage sequence.

[0016] The conditional discriminator includes a GRU, a fully connected layer and a Sigmoid activation function connected in sequence, and is used to calculate the probability that the input threshold voltage sequence is a true threshold voltage sequence under the input working condition.

[0017] Step 3: Train the life prediction model and use the trained generator to generate the threshold voltage sequence under the expected working conditions to achieve SiC MOSFET life prediction. The WGAN-GP loss function and the binary cross entropy loss function are used to evaluate the generator and discriminator. The loss function of the generator is: (3) In the formula, is the generator loss function, , They represent the normal discriminator and the conditional discriminator for the simulated threshold voltage sequence The judgment result of is the adversarial loss of the generator with respect to the ordinary discriminator, is the adversarial loss of the generator with respect to the conditional discriminator; The loss function of the ordinary discriminator is: (4) In the formula, is the loss function of the ordinary discriminator, is the real threshold voltage sequence of the ordinary discriminator The judgment result of is the adversarial loss of the ordinary discriminator, is a hyperparameter, is a threshold voltage interpolation sequence between the real threshold voltage sequence and the simulated threshold voltage sequence; is the gradient penalty term, Interpolation sequence of threshold voltage for ordinary discriminator The judgment result of To penalize the gradient modulus and avoid overfitting of the ordinary discriminator, it is specifically expressed as: (5) In the formula, is the threshold voltage interpolation sequence No. Values; Threshold voltage interpolation sequence It is expressed as: (6) In the formula, From a uniform distribution The weight of random sampling in ; The loss function of the conditional discriminator is: (7) In the formula, is the loss function of the conditional discriminator, is the conditional discriminator for the true threshold voltage sequence The judgment result of is a hyperparameter, is the regularized gradient penalty term for the generator, expressed as: (8) (9) In the formula, is the generator for the conditional label The gradient of For the The simulated threshold voltage sequence is about Conditional tags The partial derivative of is the current analog threshold voltage sequence length, is the number of conditional tags; Finally, the total loss function is: (10) In the formula, is the weight coefficient.

[0018] During the training process, the Adam optimizer is used to alternately optimize the generator, the ordinary discriminator, and the conditional discriminator. When the number of iterations is the weight coefficient When the value is a multiple of , the parameters of the generator and the conditional discriminator are updated, otherwise the parameters of the ordinary discriminator are updated until the ordinary discriminator cannot distinguish whether the input threshold voltage sequence is a real threshold voltage sequence or a simulated threshold voltage sequence, and the conditional discriminator cannot distinguish between the real threshold voltage sequence and the simulated threshold voltage sequence under the current conditional label, that is, the loss functions of the generator, the ordinary discriminator and the conditional discriminator all converge to stable values, and the model training is completed.

[0019] In the actual prediction process, the expected operating conditions are input into the trained generator as conditional labels and randomly generated Gaussian noise sequences, and the threshold voltage sequence under the expected operating conditions is output. When the change rate of the SiC MOSFET threshold voltage relative to the initial value exceeds 20%, the SiC MOSFET is considered to be damaged, and life prediction is achieved.

[0020] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A method for predicting SiC MOSFET life based on gate oxide aging, characterized in that: The method comprises the following steps: Step 1: Obtain the real threshold voltage sequence under different working conditions through SiC MOSFET accelerated aging test, and pre-process the real threshold voltage sequence; Step 2: Construct a lifespan prediction model based on a conditional generative adversarial network, including a generator, a common discriminator, a conditional discriminator, an encoding network, and a reverse recovery network; The working conditions are input as conditional labels and the preprocessed real threshold voltage sequence into the encoding network to generate multiple two-dimensional space tensors; these two-dimensional space tensors are passed through the reverse recovery network to obtain the real threshold voltage sequence embedded with the working conditions; The generator includes a coding network, an encoder and a reverse recovery network; the working condition is input into the coding network as a condition label and a randomly generated Gaussian noise sequence to generate a plurality of simulated two-dimensional space tensors; these simulated two-dimensional space tensors are input into the encoder for encoding to obtain a plurality of coding features; these coding features are passed through the reverse recovery network to obtain a simulated threshold voltage sequence embedded with the working condition; The real threshold voltage sequence and the simulated threshold voltage sequence embedded in the working condition are simultaneously input into the common discriminator and the conditional discriminator for discrimination, the common discriminator is used to calculate the probability that the input threshold voltage sequence is the real threshold voltage sequence, and the conditional discriminator is used to calculate the probability that the input threshold voltage sequence is the real threshold voltage sequence under the input working condition; Step 3: Train the life prediction model and use the trained generator to generate a threshold voltage sequence under the expected operating conditions. When the rate of change of the SiC MOSFET threshold voltage relative to the initial value exceeds the specified value, the SiC MOSFET is considered to be damaged, thus achieving life prediction.

2. The SiC MOSFET life prediction method based on gate oxide layer aging according to claim 1, characterized in that: After the encoding network splices the conditional labels to the real threshold voltage sequence, it performs fast Fourier transform on the spliced ​​sequence to decompose the real threshold voltage sequence into multiple threshold voltage sequences in the frequency domain; The amplitude of the threshold voltage sequence in each frequency domain is calculated, and multiple threshold voltage sequences with large amplitudes are selected for reshaping operations to obtain multiple two-dimensional space tensors.

3. The SiC MOSFET life prediction method based on gate oxide layer aging according to claim 1 or 2, characterized in that: The reverse recovery network convolves each two-dimensional spatial tensor to obtain high-dimensional features; each high-dimensional feature is weighted and then spliced. The spliced ​​features are reshaped to obtain the real threshold voltage sequence embedded in the working conditions.

4. The SiC MOSFET life prediction method based on gate oxide layer aging according to claim 3, characterized in that: The common discriminator includes an encoder, a fully connected layer and a Sigmoid activation function, and the conditional discriminator includes a GRU, a fully connected layer and a Sigmoid activation function.

5. The SiC MOSFET life prediction method based on gate oxide layer aging according to claim 1, characterized in that: During model training, the training loss is calculated according to the following formula: （10） In the formula, is the generator loss function, is the loss function of the ordinary discriminator, is the loss function of the conditional discriminator, is the weight coefficient; The loss function of the generator is: （3） In the formula, , They represent the normal discriminator and the conditional discriminator for the simulated threshold voltage sequence The judgment result of is the adversarial loss of the generator with respect to the ordinary discriminator, is the adversarial loss of the generator with respect to the conditional discriminator; The loss function of the ordinary discriminator is: （4） In the formula, is the real threshold voltage sequence of the ordinary discriminator The judgment result of is the adversarial loss of the ordinary discriminator, is a hyperparameter, is a threshold voltage interpolation sequence between the real threshold voltage sequence and the simulated threshold voltage sequence; is the gradient penalty term, Interpolation sequence of threshold voltage for ordinary discriminator The judgment result of The loss function of the conditional discriminator is: （7） In the formula, is the conditional discriminator for the true threshold voltage sequence The judgment result of is a hyperparameter, Regularize the gradient penalty term for the generator.

6. The SiC MOSFET life prediction method based on gate oxide layer aging according to claim 1 or 5, characterized in that: The preprocessing includes data smoothing and coordinate axis flipping; data smoothing includes outlier processing, abnormal point processing and Gaussian filtering.

Images