A method for predicting the lifetime of SiC MOSFETs based on gate oxide aging.

Abstract

The application 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. First, real threshold voltage sequences under different working conditions are obtained through SiC MOSFET accelerated aging test and pretreated; then, a life prediction model is constructed 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 taken as conditional labels, and the pretreated real threshold voltage sequences are sequentially subjected to the encoding network and the reverse recovery network to obtain real threshold voltage sequences embedded with the working conditions; the working conditions are taken as conditional labels, and a randomly generated Gaussian noise sequence is subjected to the generator to obtain simulated threshold voltage sequences embedded with the working conditions; the real threshold voltage sequences embedded with the working conditions and the simulated threshold voltage sequences are simultaneously input into the common discriminator and the conditional discriminator for discrimination; finally, the life prediction model is trained, and the trained generator is used to generate threshold voltage sequences under expected working conditions. The method can predict the life of SiC MOSFET under different working conditions offline, and is more applicable.

Description

Technical Field

[0001] This invention belongs to the field of power semiconductor device lifetime prediction technology, specifically a method for predicting the lifetime of SiC MOSFETs based on gate oxide aging. Background Technology

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

[0003] Compared to traditional silicon-based MOSFETs, SiC MOSFETs offer superior performance due to the physical properties of their materials, including low on-resistance, high frequency, and high voltage, making them highly popular in high-power and high-temperature applications. However, the lower barrier height at the gate oxide interface of SiC MOSFETs allows charge carriers in the channel to more easily cross the barrier and reach the oxide layer. The carbon elements remaining at the interface during SiC oxidation generate a high interface state density at the SiC / SiO2 interface, significantly reducing the quality of the gate oxide layer. This makes the gate of SiC MOSFETs more susceptible to Fowler-Nordheim (FN) tunneling, potentially leading to performance degradation or even functional failure under long-term high temperatures and gate bias stress. Therefore, lifetime prediction based on gate oxide aging is crucial for the safe and efficient application of SiC MOSFETs. Since the actual operating environment of SiC MOSFETs is complex and real-time data is difficult to obtain, predicting the lifetime of SiC MOSFETs under actual operating conditions through accelerated aging tests helps to understand their degradation. Traditional lifetime prediction methods include two types: those based on failure physics models and those based on data. Failure physics model-based prediction requires specific analysis of operating conditions under different failure environments, calculating the lifetime of SiC MOSFETs using formulas. However, some physical information is difficult to extract, making it difficult to construct an accurate and complete physical model, resulting in poor versatility. Data-driven prediction, based on neural networks or statistical analysis, predicts the remaining lifetime of SiC MOSFETs. This requires online monitoring and prediction using historical data. However, SiC MOSFETs typically operate under high temperature, high pressure, and high frequency conditions, relying on highly reliable sensors and data acquisition systems for monitoring, communication, and calculation. This results in strong dependence on networks and communication, high costs, and poor security. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is to provide a method for predicting the lifetime of SiC MOSFETs based on gate oxide aging.

[0005] The present invention solves the aforementioned technical problem by adopting the following technical solution:

[0006] A method for predicting the lifetime of SiC MOSFETs based on gate oxide layer aging, characterized by the following steps:

[0007] Step 1: Obtain the true threshold voltage sequence under different operating conditions through SiC MOSFET accelerated aging test, and preprocess the true threshold voltage sequence;

[0008] Step 2: Construct a lifetime prediction model based on conditional generative adversarial networks, including a generator, a general discriminator, a conditional discriminator, a coding network, and a reverse recovery network;

[0009] The operating conditions are used as condition labels and the preprocessed true threshold voltage sequence are input into the encoding network to generate multiple two-dimensional spatial tensors. These two-dimensional spatial tensors are then processed by the inverse recovery network to obtain the true threshold voltage sequence embedded with the operating conditions.

[0010] The generator includes an encoding network, an encoder, and a reverse recovery network. The operating conditions are used as condition labels and a randomly generated Gaussian noise sequence are input into the encoding network to generate multiple simulated two-dimensional space tensors. These simulated two-dimensional space tensors are input into the encoder for encoding to obtain multiple encoded features. These encoded features are processed by the reverse recovery network to obtain a simulated threshold voltage sequence embedded with the operating conditions.

[0011] The real threshold voltage sequence and the simulated threshold voltage sequence with embedded operating conditions are simultaneously input into the ordinary discriminator and the conditional discriminator for discrimination. The ordinary 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 operating conditions.

[0012] The third step is to train the lifetime prediction model and use the trained generator to generate a threshold voltage sequence under the desired 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 lifetime prediction.

[0013] Compared with the prior art, the present invention has at least the following advantages:

[0014] 1. The lifetime prediction model based on conditional generative adversarial networks can be used to directly predict the lifetime of SiC MOSFETs. It can predict the threshold voltage sequence of SiC MOSFETs under different operating conditions and Gaussian noise sequences, realizing offline lifetime prediction of SiC MOSFETs instead of online prediction of the remaining lifetime of SiC MOSFETs. Therefore, in actual prediction, it does not need to rely on the historical data of SiC MOSFETs under operating conditions, and can also realize lifetime prediction under various operating conditions, with strong applicability.

[0015] 2. To enable the model to better learn the correlation between conditional labels and threshold voltage sequences, a general discriminator and a conditional discriminator were implemented. The general discriminator determines whether the input threshold voltage sequence is real or generated, while the conditional discriminator determines whether the input threshold voltage sequence is a real threshold voltage sequence under the input conditional labels. Inputting the conditional labels and threshold voltage sequences into the encoding network, and embedding the conditional labels into the threshold voltage sequences, achieves better feature extraction.

[0016] 3. This invention obtains the threshold voltage sequence of SiC MOSFET through accelerated aging tests, which is used to predict the lifetime under actual operating conditions, realizing offline SiC MOSFET lifetime prediction and avoiding the high hardware and computing power requirements of real-time data acquisition. Attached Figure Description

[0017] Figure 1 This is a structural diagram of the lifetime prediction model of the present invention. Detailed Implementation

[0018] Specific embodiments are given below with reference to the accompanying drawings. These specific embodiments are only used to describe the technical solution of the present invention in detail, and are not intended to limit the scope of protection of this application.

[0019] This invention provides a method for predicting the lifetime of SiC MOSFETs based on gate oxide layer aging (hereinafter referred to as the method, see [link]). Figure 1 ), including the following steps:

[0020] Step 1: Accelerated aging test of SiC MOSFET was carried out using a dynamic high-temperature gate bias test platform. Real threshold voltage sequences under different operating conditions were collected to form a dataset, which was then preprocessed.

[0021] The dynamic high-temperature gate bias experimental platform includes a host computer, SiC MOSFET, heating stage, test circuit, power supply, adapter board, and DSP (digital signal processor). The SiC MOSFET is connected to the test circuit and fixed on the heating stage, with the temperature of the SiC MOSFET controlled by the heating stage. The test circuit is connected to the power supply and DSP via the adapter board. The DSP is connected to the host computer, and sends a high-level signal as the gate voltage for the SiC MOSFET aging test, and a low-level signal as a -10V square wave to control the cyclic switching on and off of the SiC MOSFET. During the experiment, the threshold voltage is measured every 10 seconds. When the SiC MOSFET threshold voltage exceeds the rate of change relative to the initial value exceeding a specified value (20%), the SiC MOSFET is considered damaged, and the experiment is stopped. By changing the operating conditions (including gate voltage, switching frequency, temperature, etc.), the true threshold voltage sequence under different operating conditions is obtained.

[0022] Preprocessing includes data smoothing and axis flipping; firstly, due to the instability of device connections, one or more outliers with large differences may occasionally appear. Therefore, a data filtering method is used to identify outliers, and the values ​​of outliers are calculated according to the following formula to replace the original values ​​of the outliers.

[0023] (1)

[0024] In the formula, It is the value of the outlier. and It is the value of the two points before the outlier. and These are the values ​​of the two points following the outlier;

[0025] Then, the moving average method is used to calculate the moving average and standard deviation of the window. According to formula (2), 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 average value is used to replace the original value of the outlier.

[0026] (2)

[0027] In the formula, It is the value of any point within the window. , These are the moving average and standard deviation of the window. It is a constant coefficient, in this embodiment ;

[0028] Gaussian filtering is then used to remove noise and smooth the data.

[0029] Since the threshold voltage sequence varies greatly in the time dimension under different operating conditions, making it difficult to predict, the coordinate axis of the threshold voltage sequence is flipped, that is, the horizontal axis is converted to threshold voltage and the vertical axis is converted to time.

[0030] The second step involves constructing a lifetime prediction model based on a conditional generative adversarial network (GAN), which includes a generator, a general discriminator, a conditional discriminator, an encoding network, and a reverse recovery network. The generator is used to generate a simulated threshold voltage sequence under operating conditions. The general 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 operating conditions. The encoding network is used to embed conditional labels into the sequence and encode the one-dimensional sequence into a two-dimensional spatial tensor. The reverse recovery network is used to recover the input data into a one-dimensional sequence.

[0031] The operating conditions are used as condition labels and the preprocessed true threshold voltage sequence are input into the encoding network. First, the condition labels are concatenated to the true threshold voltage sequence. Then, the concatenated sequence is subjected to Fast Fourier Transform to achieve recursive decomposition of the sequence, thereby decomposing the true 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 the K threshold voltage sequences with the largest amplitudes are selected for reshaping operation, mapping the one-dimensional threshold voltage sequence into a two-dimensional space tensor, resulting in a total of K two-dimensional space tensors.

[0032] K two-dimensional spatial tensors are input into the inverse recovery network. Convolution is performed on each of the K two-dimensional spatial tensors to extract high-dimensional features, resulting in K high-dimensional features. Each high-dimensional feature is assigned a weight, and all high-dimensional features are concatenated to obtain concatenated features. The concatenated features are then reshaped to map from two-dimensional space to one-dimensional space, resulting in the true threshold voltage sequence embedded with the operating conditions.

[0033] The generator includes an encoding network, an encoder, and a reverse recovery network. The operating conditions are used as condition labels and a randomly generated Gaussian noise sequence are input into the encoding network to obtain K analog two-dimensional space tensors. The K analog two-dimensional space tensors are input into the encoder for encoding to obtain K encoded features. The K encoded features are processed by the reverse recovery network to obtain an analog threshold voltage sequence embedded with the operating conditions.

[0034] The real threshold voltage sequence and the simulated threshold voltage sequence, embedded with the operating conditions, are simultaneously input into a general discriminator and a conditional discriminator for discrimination. The general discriminator includes an encoder, a fully connected layer, and a sigmoid activation function. The input threshold voltage sequence is processed by the encoder to extract encoded features, which are then 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.

[0035] The conditional discriminator comprises a GRU, a fully connected layer, and a Sigmoid activation function connected in sequence, used to calculate the probability that the input threshold voltage sequence is the true threshold voltage sequence under the input operating conditions.

[0036] Step 3: Train the lifetime prediction model and use the trained generator to generate the threshold voltage sequence under the desired operating conditions to achieve SiC MOSFET lifetime prediction.

[0037] The generator and discriminator are evaluated using the WGAN-GP loss function and the binary cross-entropy loss function. The generator's loss function is as follows:

[0038] (3)

[0039] In the formula, Let be the generator loss function. , These represent the ordinary discriminator and the conditional discriminator for the analog threshold voltage sequence, respectively. The judgment result, The generator's adversarial loss with respect to the ordinary discriminator, It is the generator's adversarial loss with respect to the conditional discriminator;

[0040] The loss function of the ordinary discriminator is:

[0041] (4)

[0042] In the formula, The loss function of a standard discriminator. For a general discriminator to the true threshold voltage sequence The judgment result, For the adversarial loss of a normal discriminator, For hyperparameters, This is the threshold voltage interpolation sequence located between the true threshold voltage sequence and the simulated threshold voltage sequence; It is a gradient penalty term. For the threshold voltage interpolation sequence of a general discriminator The judgment result; To penalize the gradient magnitude and prevent overfitting by the ordinary discriminator, it is specifically expressed as follows:

[0043] (5)

[0044] In the formula, It is a threshold voltage interpolation sequence The One value;

[0045] Threshold voltage interpolation sequence Represented as:

[0046] (6)

[0047] In the formula, From uniform distribution Weights for random sampling in the middle;

[0048] The loss function of the conditional discriminator is:

[0049] (7)

[0050] In the formula, Let the loss function of the conditional discriminator be... For the conditional discriminator on the true threshold voltage sequence The judgment result, For hyperparameters, The generator regularization gradient penalty term is expressed as:

[0051] (8)

[0052] (9)

[0053] In the formula, It is a generator for conditional tags gradient, For the first The simulated threshold voltage sequence with respect to the first Conditional tags The partial derivatives, The current simulated threshold voltage sequence length. The number of conditional tags;

[0054] Finally, the total loss function is:

[0055] (10)

[0056] In the formula, These are the weighting coefficients.

[0057] During training, the Adam optimizer is used to alternately optimize the generator, ordinary discriminator, and conditional discriminator. The iteration count is equal to the weight coefficients. When the threshold voltage is a multiple of the given value, update the parameters of the generator and the conditional discriminator; otherwise, update the parameters of the ordinary discriminator. Continue this process until the ordinary discriminator can no longer distinguish between the input threshold voltage sequence and the simulated threshold voltage sequence, and the conditional discriminator can no longer distinguish between the true threshold voltage sequence and the simulated threshold voltage sequence under the current condition label. In other words, the loss functions of the generator, the ordinary discriminator, and the conditional discriminator all converge to a stable value, and the model training is complete.

[0058] In the actual prediction process, the desired operating conditions are used as condition labels and a randomly generated Gaussian noise sequence are input into the trained generator to output the threshold voltage sequence under the desired operating conditions. When the SiC MOSFET threshold voltage changes by more than 20% relative to the initial value, the SiC MOSFET is considered to be damaged, thus achieving lifetime prediction.

[0059] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A method for predicting the lifetime of SiC MOSFETs based on gate oxide aging, characterized in that, The method includes the following steps: Step 1: Obtain the true threshold voltage sequence under different operating conditions through SiC MOSFET accelerated aging test, and preprocess the true threshold voltage sequence; Step 2: Construct a lifetime prediction model based on conditional generative adversarial networks, including a generator, a general discriminator, a conditional discriminator, a coding network, and a reverse recovery network; The operating conditions are used as condition labels and the preprocessed true threshold voltage sequence are input into the encoding network to generate multiple two-dimensional spatial tensors. These two-dimensional spatial tensors are then processed by the inverse recovery network to obtain the true threshold voltage sequence embedded with the operating conditions. The generator includes an encoding network, an encoder, and a reverse recovery network. The operating conditions are used as condition labels and a randomly generated Gaussian noise sequence are input into the encoding network to generate multiple simulated two-dimensional space tensors. These simulated two-dimensional space tensors are input into the encoder for encoding to obtain multiple encoded features. These encoded features are processed by the reverse recovery network to obtain a simulated threshold voltage sequence embedded with the operating conditions. The real threshold voltage sequence and the simulated threshold voltage sequence with embedded operating conditions are simultaneously input into the ordinary discriminator and the conditional discriminator for discrimination. The ordinary 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 operating conditions. The third step is to train the lifetime prediction model and use the trained generator to generate a threshold voltage sequence under the desired 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 realizing lifetime prediction.

2. The SiC MOSFET lifetime prediction method based on gate oxide aging according to claim 1, characterized in that, The coding network concatenates the conditional labels to the real threshold voltage sequence, and then performs a fast Fourier transform on the concatenated 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 spatial tensors.

3. The SiC MOSFET lifetime prediction method based on gate oxide aging according to claim 1 or 2, characterized in that, The inverse recovery network convolves each two-dimensional spatial tensor to obtain high-dimensional features; weights are assigned to each high-dimensional feature and then they are concatenated; the concatenated features are then reshaped to obtain the true threshold voltage sequence embedded with the operating conditions.

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

5. The SiC MOSFET lifetime prediction method based on gate oxide aging according to claim 1, characterized in that, During model training, the training loss is calculated using the following formula: （10） In the formula, Let be the generator loss function. The loss function of a standard discriminator. Let the loss function of the conditional discriminator be... These are the weighting coefficients; The loss function of the generator is: （3） In the formula, , These represent the ordinary discriminator and the conditional discriminator for the analog threshold voltage sequence, respectively. The judgment result, The generator's adversarial loss with respect to the ordinary discriminator, It is the generator's adversarial loss with respect to the conditional discriminator; The loss function of the ordinary discriminator is: （4） In the formula, For a general discriminator to the true threshold voltage sequence The judgment result, For the adversarial loss of a standard discriminator, For hyperparameters, This is the threshold voltage interpolation sequence located between the true threshold voltage sequence and the simulated threshold voltage sequence; It is a gradient penalty term. For the threshold voltage interpolation sequence of a general discriminator The judgment result; The loss function of the conditional discriminator is: （7） In the formula, For the conditional discriminator on the true threshold voltage sequence The judgment result, For hyperparameters, This is a gradient penalty term for generator regularization.

6. The SiC MOSFET lifetime prediction method based on gate oxide aging according to claim 1 or 5, characterized in that, The preprocessing includes data smoothing and coordinate axis flipping; data smoothing includes outlier handling, anomaly handling, and Gaussian filtering.

Images