Life evaluation model establishing method and device, equipment and storage medium

Abstract

The application relates to a life evaluation model establishment method and device, equipment and a storage medium, which are applied in the field of integrated circuits, comprising determining an electrical stress condition of the power device, measuring a failure time of the power device under the electrical stress condition, calculating a Weibull cumulative distribution function of the power device under the failure time, fitting and calculating a characteristic life value of the power device under the electrical stress condition according to the Weibull cumulative distribution function and a preset global shape factor, constructing a deep neural network model, training the deep neural network model by using the characteristic life value obtained by fitting and calculation, and generating an accelerated life evaluation model. The application has the technical effect of performing failure analysis and life evaluation on a device with complex process and structure, and improving the evaluation accuracy.

Description

Technical Field

[0001] This application relates to the field of integrated circuit technology, and in particular to a method, apparatus, device and storage medium for establishing a lifetime assessment model. Background Technology

[0002] With the continuous advancement of semiconductor technology, solid-state power electronics technology has experienced rapid and sustained development. The development of solid-state devices is not only driving the technological revolution in the information industry but also prompting people to think and live in new ways. Semiconductor materials, as a fundamental discipline of semiconductor technology, play a crucial role in the development of energy and information technology.

[0003] In the development of semiconductor technology, GaN (gallium nitride) devices have become the best choice for high-frequency, high-efficiency power modules due to their higher electron saturation drift velocity and electron mobility. Natural GaN is a depletion-mode device, and the implementation of enhancement-mode GaN typically involves metal-insulator-semiconductor (MIS) structures and P-type gate structures, which may introduce additional reliability issues.

[0004] In realizing metal-insulator-semiconductor (MIS) structures, the prolonged plasma irradiation time causes severe damage to the semiconductor surface due to ultraviolet radiation. This surface damage leads to increased leakage current, decreased threshold voltage (Vth) stability, and current collapse (increased dynamic on-resistance). Furthermore, p-type gate gallium nitride (p-GaN) exhibits forward gate overvoltage failure and time-dependent breakdown. When a high electrical stress voltage is applied to the gate, a significant voltage drop and electric field appear in the p-GaN depletion region near the metal interface, promoting the formation of penetration paths. This degradation mechanism is consistent with time-dependent dielectric breakdown.

[0005] Time-dependent dielectric breakdown (TDDB) has been extensively studied in power devices manufactured using silicon and other processes, and it has become an essential evaluation method for transistor reliability. However, during dielectric breakdown, dielectric layer breakdown can lead to device destruction and loss of switching function, resulting in power module failure. Therefore, accurately assessing the lifetime of power devices and analyzing the failure mechanisms of new structures are crucial for improving device reliability and module quality. Currently, commonly used accelerated lifetime assessment models for power devices include exponential function models and power function models. By fitting the characteristic lifetime to exponential and power function models, the lifetime at the rated operating voltage or the typical operating voltage at the rated lifetime can be obtained.

[0006] However, traditional exponential and power function models can only fit the characteristic lifetime under a fixed distribution pattern. As system performance improves and breakdown voltage increases, the failure mechanisms and lifetime characteristics of newly developed device processes and structures become more complex, making it difficult to accurately assess failure mechanisms and lifetimes using traditional models. Summary of the Invention

[0007] To help solve the problem that traditional lifetime assessment models are difficult to accurately assess the failure mechanism and lifetime of power devices, this application provides a method, apparatus, device and storage medium for establishing a lifetime assessment model.

[0008] Firstly, this application provides a method for establishing a lifetime assessment model, employing the following technical solution: the method is applied to a lifetime assessment model establishment system, the lifetime assessment model establishment system including power devices, and the method includes:

[0009] The electrical stress condition of the power device is determined based on the gate leakage current of the power device, and the failure time of the power device under the electrical stress condition is measured.

[0010] Calculate the Wechsler cumulative distribution function of the power device at the failure time;

[0011] The characteristic lifetime value of the power device under the electrical stress condition is calculated by fitting the Vickers cumulative distribution function and the preset global shape factor.

[0012] A deep neural network model is constructed, and the feature lifetime values ​​obtained by fitting calculation are used to train the deep neural network model to generate an accelerated lifetime assessment model, wherein the accelerated lifetime assessment model is a partitioned deep neural network model.

[0013] In one specific implementation, determining the electrical stress condition of the power device based on its gate leakage current and measuring the failure time of the power device under the electrical stress condition includes:

[0014] Measure the steep increase point of the gate leakage current of the power device, and determine the steep increase stress range of the gate leakage current based on the steep increase point;

[0015] Several different electrical stress conditions are determined based on the steeply increasing stress range;

[0016] The failure time of the power device under several electrical stress conditions was measured using a dielectric breakdown full-life accelerated test.

[0017] In one specific implementation scheme, the calculation method of the Wechsler cumulative distribution function includes:

[0018]

[0019] Where t represents the failure time, β represents the shape factor, η represents the scaling factor, γ represents the time delay, and F(t) represents the Wechsler cumulative distribution function.

[0020] In one specific implementation, the step of calculating the characteristic lifetime value of the power device under the electrical stress condition by fitting the Vickers cumulative distribution function and a preset global shape factor includes:

[0021] The failure time is fitted with a Wechsler distribution based on the Wechsler cumulative distribution function;

[0022] The intercept value of the failure time under the electrical stress condition is calculated based on the Wechsler distribution of the failure time.

[0023] The characteristic lifetime value of the power device under the electrical stress condition is calculated based on the preset global shape factor and the intercept value.

[0024] The calculation method for the characteristic lifetime value includes:

[0025]

[0026] Where, intercept_j represents the calculated intercept value, β′ represents the global shape factor, and η_j represents the characteristic lifetime value under the j-th stress condition.

[0027] In one specific implementation, the step of constructing a deep neural network model and training the deep neural network model using the feature lifetime values ​​obtained from the fitting calculation to generate an accelerated lifetime assessment model includes:

[0028] Construct an input layer, hidden layer, and output layer for a deep neural network model, and set activation functions, weights, and biases in the hidden layer;

[0029] The boundary point of the electrical stress condition is determined based on the steep increase point of the gate leakage current, and low electrical stress condition and high electrical stress condition are generated.

[0030] Determine the range of electrical stress conditions corresponding to the characteristic lifetime value, and perform format conversion on the characteristic lifetime value according to different electrical stress conditions;

[0031] If the electrical stress condition corresponding to the characteristic lifetime value is a low electrical stress condition, then the characteristic lifetime value is converted into a real number format and a real number characteristic lifetime value is generated.

[0032] If the electrical stress condition corresponding to the characteristic lifetime value is a high electrical stress condition, then the characteristic lifetime value is converted to dB unit format and a logarithmic characteristic lifetime value is generated.

[0033] The real-valued feature lifetime and the logarithmic feature lifetime are normalized to generate real-valued normalized feature lifetime and logarithmic normalized feature lifetime.

[0034] Determine the range of electrical stress conditions corresponding to the feature lifetime value, and train the deep neural network model according to different electrical stress conditions;

[0035] If the electrical stress condition corresponding to the feature lifetime value is a low electrical stress condition, then the deep neural network model is trained using the real-number normalized feature lifetime value, and a real-number neural network model is generated.

[0036] If the electrical stress condition corresponding to the feature lifetime value is a high electrical stress condition, then the deep neural network model is trained using the log-normalized feature lifetime value, and a log-normalized neural network model is generated.

[0037] The real-number neural network model and the logarithmic neural network model are set as the accelerated life assessment model.

[0038] In one specific implementation, the step of constructing a deep neural network model and training the deep neural network model using the feature lifetime values ​​obtained from the fitting calculation to generate an accelerated lifetime assessment model includes:

[0039] Construct an input layer, hidden layer, and output layer for a deep neural network model, and set activation functions, weights, and biases in the hidden layer;

[0040] The characteristic lifetime values ​​are converted to real number format and dB unit format respectively, and the format-converted real number characteristic lifetime value and logarithmic characteristic lifetime value are generated.

[0041] The real number feature lifetime and the logarithmic feature lifetime after format conversion are normalized to generate real number normalized feature lifetime and logarithmic normalized feature lifetime.

[0042] The deep neural network model is trained using the real-number normalized feature lifetime value to generate a real-number neural network model;

[0043] The deep neural network model is trained using the log-normalized feature lifetime values ​​to generate a log-normalized neural network model;

[0044] Based on the training results of the real neural network model and the logarithmic neural network model, the electrical stress conditions are divided into low electrical stress conditions and high electrical stress conditions.

[0045] The real neural network model is used under the low electrical stress condition, and the logarithmic neural network model is used under the high electrical stress condition, and the final accelerated life assessment model is generated.

[0046] In one specific implementation, after generating the accelerated life assessment model, the method further includes:

[0047] The results output by the accelerated life assessment model are inversely normalized.

[0048] Secondly, this application provides a life assessment model building device, which adopts the following technical solution: the device is applied to a life assessment model building system, the life assessment model building system includes power devices, and the device includes:

[0049] The failure time calculation module is used to determine the electrical stress condition of the power device based on the gate leakage current of the power device, and to measure the failure time of the power device under the electrical stress condition.

[0050] The distribution function calculation module is used to calculate the Vickers cumulative distribution function of the power device at the failure time.

[0051] The characteristic lifetime calculation module is used to calculate the characteristic lifetime value of the power device under the electrical stress condition based on the Vickers cumulative distribution function and the preset global shape factor.

[0052] The model building and training module is used to build a deep neural network model and train the deep neural network model using the feature lifetime values ​​obtained by fitting calculation to generate an accelerated lifetime assessment model, wherein the accelerated lifetime assessment model is a partitioned deep neural network model.

[0053] Thirdly, this application provides a computer device that adopts the following technical solution: it includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as any of the above-described life assessment model establishment methods.

[0054] Fourthly, this application provides a computer-readable storage medium that stores a computer program capable of being loaded by a processor and executing any of the above-mentioned life assessment model establishment methods.

[0055] In summary, this application has the following beneficial technical effects:

[0056] First, the basic parameters of the power device are measured. Then, a deep neural network model is trained using these parameters to obtain a model capable of evaluating the power device's lifetime. Even for power devices with complex fabrication structures, this model can be used for accurate lifetime assessment and reliability testing, addressing the shortcomings of existing GaN power device lifetime models. Secondly, the constructed deep neural network model improves the accuracy of fitting the GaN power device lifetime curve, predicts reliability characteristics before and after the sharp increase in gate leakage current, and identifies the formation of gate leakage current channels, providing a basis for reliability prediction and design. Attached Figure Description

[0057] Figure 1 This is a flowchart of the life assessment model establishment method in the embodiments of this application;

[0058] Figure 2 This is a schematic diagram illustrating the sharp voltage increase of the device in the 7V-8V range in the embodiments of this application;

[0059] Figure 3 This is a schematic diagram showing the results of the medium breakdown full accelerated lifetime test in the embodiments of this application;

[0060] Figure 4 This is a schematic diagram illustrating the Vickers distribution fitting of power device lifetime results under different electrical stress conditions in the embodiments of this application;

[0061] Figure 5 This is a schematic diagram of the deep neural network model structure in the embodiments of this application;

[0062] Figure 6 This is a schematic diagram showing the results of fitting power and device lifetime using the traditional exponential function model and power function model in the embodiments of this application;

[0063] Figure 7 This is a schematic diagram showing the results of the approximate power and device lifetime of the evaluation model constructed in this embodiment of the present application;

[0064] Figure 8 This is a schematic diagram of the life assessment model establishment device in the embodiments of this application;

[0065] Figure 9 This is a schematic diagram used to illustrate a computer device in the embodiments of this application.

[0066] Figure reference numerals: 801, Failure time calculation module; 802, Distribution function calculation module; 803, Characteristic lifetime calculation module; 804, Model building and training module. Detailed Implementation

[0067] The following combination Figures 1-9 This application will be described in further detail.

[0068] This application discloses a method for establishing a life assessment model. This method can construct a life assessment model with a wide range of applications, enabling the use of this model for failure analysis and life prediction of power devices. With the continuous advancement of semiconductor technology, solid-state power electronics technology has experienced rapid and sustained development. The development of solid-state devices is not only driving the technological revolution in the information industry but also prompting people to think and live in new ways. Semiconductor materials, as a fundamental discipline of semiconductor technology, play a crucial role in the development of energy and information technology.

[0069] The evolution of semiconductor technology can be summarized into three important stages: first-generation semiconductor materials represented by silicon (Si), second-generation semiconductor materials represented by gallium arsenide (GaAs), and third-generation semiconductor materials represented by gallium nitride (GaN) and silicon carbide (SiC). Compared with other materials, GaN is more suitable for high-frequency, high-power applications, especially in the MHz frequency range. Due to its higher electron saturation drift velocity and electron mobility, GaN is the best choice for high-frequency, high-efficiency power modules.

[0070] Epitaxial GaN power devices on silicon (Si) substrates suffer from lattice mismatch and thermal mismatch, which can lead to fatal defects in the GaN layer. Therefore, complex buffer layers and epitaxial layers are required. In addition, natural GaN is a depletion-mode device, and the realization of enhancement-mode GaN usually involves realizing metal-insulator-semiconductor (MIS) structures and P-type gate structures, which also bring additional reliability issues to the device.

[0071] In realizing metal-insulator-semiconductor (MIS) structures, the long plasma irradiation time and the heavy ultraviolet radiation from the plasma severely damage the semiconductor surface. This surface damage leads to increased leakage current, decreased threshold voltage (Vth) stability, and current collapse (increased dynamic on-resistance). Furthermore, p-type gate gallium nitride (p-GaN) exhibits forward gate overvoltage failure and time-dependent breakdown. When a high electrical stress voltage is applied to the gate, a large voltage drop and electric field appear in the p-GaN depletion region near the metal interface, promoting the formation of penetration paths. This degradation mechanism is consistent with time-dependent dielectric breakdown.

[0072] Time-dependent dielectric breakdown (TDDB) has been extensively studied in power devices manufactured using silicon and other processes, and it has become an essential evaluation method for transistor reliability. However, during dielectric breakdown, dielectric layer breakdown can lead to device destruction and loss of switching function, resulting in power module failure. Therefore, accurately assessing the lifetime of power devices and analyzing the failure mechanisms of new structures are crucial for improving device reliability and module quality. Currently, commonly used accelerated lifetime assessment models for power devices include exponential function models and power function models. By fitting the characteristic lifetimes to exponential and power function models, the lifetime at the rated operating voltage or the typical operating voltage at the rated lifetime can be obtained. For traditional devices, three characteristic lifetimes are typically obtained through full-life accelerated testing, and these three characteristic lifetimes are basically on the same linear distribution in the log coordinate system. Therefore, exponential and power function models can achieve good fitting results and minimize error functions.

[0073] However, traditional exponential and power function models can only fit the characteristic lifetime under a fixed distribution pattern. As system performance improves and breakdown voltage increases, the failure mechanisms and lifetime characteristics of newly developed device processes and structures become more complex, making it difficult to accurately assess failure mechanisms and lifetimes using traditional models. Furthermore, traditional models have low accuracy in fitting accelerated lifetimes under high electrical stress conditions, resulting in poor fitting performance. To address the problems of traditional models in calculating the lifetime of power devices, this application provides a method for establishing a lifetime assessment model.

[0074] Reference Figure 1 The method includes the following steps:

[0075] S10, determine the electrical stress conditions of the power device based on the gate leakage current of the power device, and measure the failure time of the power device under electrical stress conditions.

[0076] Specifically, when establishing a lifetime assessment model, it is necessary to first obtain the relevant parameters of the power device to facilitate model learning and training. Accelerated lifecycle testing of the power device can obtain the failure time parameter, that is, the time required for the power device to completely fail under certain stress conditions. This parameter can then be used to build and train the model. The failure time of the device varies under different electrical stress conditions.

[0077] Specific methods for measuring the failure time of power devices may include: First, measuring the steep increase point of the gate leakage current of the power device, and determining the steep increase stress range of the gate leakage current based on the steep increase point; specifically, the Kelvin test method can be used, and the gate DC voltage of the device can be measured in steps within a preset voltage range using a semiconductor parameter analyzer. For example, if the source DC voltage Vs = 0V, the drain DC voltage Vd = 0.1V, and the gate DC voltage Vg is swept from 0V to 15V in steps of 0.05V, referencing... Figure 2 The leakage current increased sharply in the range of 7V-8V gate voltage stress, indicating that there is a different gate dielectric degradation mechanism at this point. Therefore, the range can be set to [7V, 8.3V].

[0078] Next, several different electrical stress conditions are determined based on the range of steeply increasing stress. Several sets of stress conditions are set, and the specific number of sets can be set by the user; there is no limitation here. For example, within the range of [7V, 8.3V], nine sets of stress conditions are selected, and at least eight identical devices are tested under each set of stress conditions. Finally, the failure time of the power devices under several electrical stress conditions is measured using a dielectric breakdown full-life accelerated test. For example, referring to… Figure 3 The power device has Vg = 7.1V and Vd = 0.1V. The gate leakage current changes with time. As time increases, the current increases slowly, and noise appears. When the time is greater than 324s, the current increases sharply, indicating the formation of a penetration channel in the dielectric. Hard breakdown leads to device failure. The breakdown time at this time is recorded. The breakdown time of each power device under the corresponding stress condition is calculated using the same method and sorted and numbered. The sorting method is not limited. In this application, the calculated failure time is sorted and numbered in ascending order.

[0079] S20, calculate the Wechsler cumulative distribution function of the power device at the failure time.

[0080] Specifically, multiple sets of electrical stress conditions were set during the experiment to observe the failure time or breakdown time of power devices under different electrical stress conditions. The Weibull distribution, also known as the Weibull distribution, is a continuous probability distribution widely used in reliability analysis and life testing. By calculating the Weibull cumulative distribution function of the power device under different electrical stress conditions, the characteristic lifetime value of the power device under different electrical stress conditions can be calculated. The original measured failure times are randomly ordered. First, the measured failure times are sorted by size. In this embodiment, they are sorted in ascending order of failure time. Each power device has a corresponding measured failure time and a number. When calculating the device lifetime, the shape factor and characteristic lifetime value can be extracted using the cumulative distribution function. The Weibull cumulative distribution function can be expressed as:

[0081]

[0082] Where t represents the failure time, β represents the shape factor, η represents the scaling factor (which can also be understood as the failure rate), γ represents the time delay, and F(t) represents the Wechsler cumulative distribution function.

[0083] Among them, there are many ways to calculate the value F(t) of the Wechsler cumulative distribution function, and the median rank method has a relatively better calculation effect. In this embodiment of the application, the median rank method is used as an example for illustration. The calculation method of the median rank method can be expressed as follows:

[0084]

[0085] Where i represents the device number after sorting under a certain set of electrical stress conditions, which is also the number of the power device corresponding to the failure time t in the cumulative distribution function above; F(i) represents the value of the Vickers cumulative distribution function of the i-th power device under a certain set of electrical stress conditions, which can also be understood as the probability that the penetration path will form at a certain time under a given stress condition; n represents the total number of power devices under each set of electrical stress conditions. The calculated value of F(i) is the value of F(t).

[0086] S30 calculates the characteristic lifetime value of the power device under electrical stress conditions based on the Widmanstätten cumulative distribution function and the preset global shape factor.

[0087] Specifically, by calculating and fitting the Vickers cumulative distribution function and a preset global shape factor, the characteristic lifetime values ​​of the power device under different electrical stress conditions can be obtained. The specific calculation method may include: first, fitting the Vickers distribution of the failure time based on the Vickers cumulative distribution function; when the time delay parameter γ in the Vickers cumulative distribution function is 0, the Vickers cumulative distribution function can be expressed as:

[0088] ln[-ln(1-F(t))]=βln(t)-βln(η),

[0089] Where t represents the failure time, β represents the shape factor, η represents the scaling factor, and F(t) represents the Wechsler cumulative distribution function.

[0090] By fitting ln[-ln(1-F(t))] and ln(t), a straight line with a slope of β and an intercept of -βln(η) can be generated. Under a semi-logarithmic chart, this fitting result can be called the Wechsler distribution.

[0091] Next, the intercept of the failure time under electrical stress conditions is calculated based on the Wechsler distribution of the failure time. Specifically, the fitting formula for the Wechsler distribution can be expressed as:

[0092] y_j=β·x_j[1:n]+intercept_j,

[0093] Where x_j[1:n] represents the natural logarithm of the failure time from the 1st to the nth device under the j-th stress condition, intercept_j represents the intercept under each stress condition, and y_j represents the value of ln[-ln(1-F(t))]. Assuming the initial value of β is set to 1 and the initial value of intercept_j is set to 0, the global slope β and the intercept values ​​under each stress can be obtained by fitting using the least squares method. The results obtained by fitting using the Weyssee distribution are referenced. Figure 4 As shown, the global shape factor β is set to 1.25 in this application scheme. In practical applications, the global shape factor can be adjusted and determined according to the actual fitting results of the user, and no restrictions are imposed here.

[0094] Finally, the characteristic lifetime value of the power device under electrical stress conditions is calculated based on the preset global shape factor and intercept value.

[0095] The methods for calculating characteristic lifetime values ​​include:

[0096]

[0097] Where intercept_j represents the calculated intercept value, β′ represents the global shape factor, and η_j represents the characteristic lifetime value under the j-th stress condition. intercept_j corresponds to -βln(η) in the Wechsler cumulative distribution function, and the characteristic lifetime η_j corresponds to the value when ln[-ln(1-F(t))]=0.

[0098] S40. Construct a deep neural network model and train the deep neural network model using the feature lifetime values ​​obtained from the fitting calculation to generate an accelerated lifetime assessment model. The accelerated lifetime assessment model is a partitioned deep neural network model.

[0099] Specifically, a basic deep neural network model is constructed. Then, based on the characteristic lifetime values ​​calculated by the power device under different electrical stress conditions, the initial deep neural network model is learned and trained to generate an accelerated lifetime assessment model that can be used to evaluate applications. The model is trained separately under different electrical stress conditions. Therefore, the final accelerated lifetime assessment model is a partitioned deep neural network model.

[0100] In this application, the basic parameters of the power device are first measured. These parameters are then used to train a deep neural network model to obtain a model capable of evaluating the power device's lifetime. Even for power devices with complex fabrication structures, this model can be used for accurate lifetime assessment and reliability testing, thus addressing the shortcomings of existing GaN power device lifetime models. Secondly, the constructed deep neural network model improves the accuracy of fitting the GaN power device lifetime curve, predicts reliability characteristics before and after the sharp increase in gate leakage current, and identifies the formation of gate leakage current channels, providing a basis for reliability prediction and design.

[0101] In one embodiment, constructing a deep neural network model and training it with the feature lifetime values ​​obtained from the fitting calculation to generate an accelerated lifetime assessment model can be specifically performed as follows:

[0102] First, we construct the input layer, hidden layers, and output layer of the deep neural network model, and set the activation function in the hidden layer. A deep neural network is a complex network structure composed of multiple layers, typically including an input layer, hidden layers, and an output layer. The hidden layer is a crucial component of a deep neural network. The hidden layer is located between the input and output layers. A deep neural network can contain multiple hidden layers, which are stacked sequentially, with the output of each layer serving as the input to the next layer. (Refer to...) Figure 5In this embodiment, the constructed deep neural network model includes an input layer, an output layer, and three hidden layers. Each hidden layer contains 16 neurons. The main function of the hidden layers is to extract features from the input data. The number of hidden layers and the number of neurons in each hidden layer can be set by the user according to actual modeling needs. The number of neurons should be set appropriately for the model's operation. If the number of neurons is too small, the accuracy of the fitting will be low and the fitting effect will be poor. If the number of neurons is too large, the fitting accuracy will be higher, but the network will become more complex and the computation time will be longer. Overfitting may also occur. Therefore, the number of neurons should not be too large or too small. This application uses 16 neurons. The three hidden layers and the 16 neurons in each hidden layer in this embodiment are only examples for illustration and are not intended to be limiting. Each hidden layer is equipped with an activation function, weights, and biases. By using different weights and activation functions, nonlinear transformations are applied to the input data to capture different levels of abstract features of the input data. The activation function can be selected and adjusted by the user according to the actual computational needs and the actual modeling effect. In this embodiment, the Sigmoid function is used as an example for illustration. The calculation method of the Sigmoid function can be expressed as follows:

[0103]

[0104] Where z represents the weighting function, and the weighting function z can be expressed as:

[0105] z=Σw i x i +b,

[0106] Where, x i In this embodiment of the application, x represents the independent variable of the neural network, specifically the first hidden layer. i This refers to the electrical stress condition input to the input layer, and the x value of the second hidden layer. i That is, z is the output of the first hidden layer, and x is the output of the third hidden layer. i That is, the z and w values ​​output by the second hidden layer. i 'b' represents the weight, and 'b' represents the bias.

[0107] After constructing the basic neural network model of God, the input training data is processed to distinguish between high and low electrical stress conditions. Specifically, this distinction is made by determining the boundary point of the electrical stress conditions based on the steep increase point of the gate leakage current, thus obtaining the low and high electrical stress conditions. (Refer to...) Figure 6Traditional exponential and power function models, when fitting accelerated lifetime under high electrical stress conditions, suffer from poor fitting accuracy and dynamic range due to the fact that more error weights are allocated to low stress conditions. Therefore, by distinguishing between high and low electrical stress conditions, the fitting accuracy of the overall accelerated lifetime of power devices can be improved, thereby increasing the accuracy of determining the failure mechanism of power devices.

[0108] Next, the range of electrical stress conditions corresponding to the characteristic lifetime value is determined, and the format of the characteristic lifetime value is converted according to different electrical stress conditions. If the electrical stress condition corresponding to the characteristic lifetime value is a low electrical stress condition, the characteristic lifetime value is converted into a real number format, and a real number characteristic lifetime value is generated. The conversion method can be expressed as follows:

[0109] η_j_s=η_j,

[0110] Where η_j represents the characteristic lifetime value calculated using the Wechsler distribution, and η_j_s represents the real characteristic lifetime value after real number transformation.

[0111] If the electrical stress condition corresponding to the characteristic lifetime value is a high electrical stress condition, then the characteristic lifetime value is converted to dB unit format, and a logarithmic characteristic lifetime value is generated. The conversion method can be expressed as follows:

[0112] η_j_dB = 20log(abs(η_j)),

[0113] Where η_j represents the characteristic lifetime value calculated using the Wechsler distribution, and η_j_dB represents the logarithmic characteristic lifetime value after format conversion in dB units.

[0114] Referring to Table 1, the characteristic lifetime differs by three orders of magnitude under high and low stress conditions. This leads to a larger contribution of the high-order-of-magnitude characteristic lifetime to the weight of the error function during training, resulting in a decrease in the fitting accuracy of the low-order-of-magnitude characteristic lifetime. Therefore, under low stress conditions, training is performed using a real number format, while under high stress conditions, training is performed using an algorithm format in dB units.

[0115] Table 1. Characteristic lifetimes under different stress conditions

[0116]

[0117] Next, the feature lifetime values ​​after format conversion are normalized, that is, the real feature lifetime values ​​and logarithmic feature lifetime values ​​are normalized, and real-normalized feature lifetime values ​​and logarithmic-normalized feature lifetime values ​​are generated. The normalization process can be expressed as follows:

[0118]

[0119] Where η_jx represents the real-valued characteristic lifetime and the logarithmic characteristic lifetime before normalization, η_x min η_x represents the minimum characteristic lifetime. max Represents the maximum characteristic lifetime, η_jx norm This represents the normalized real-valued eigenlife and the log-normalized eigenlife.

[0120] After normalization, the range of electrical stress conditions corresponding to the feature lifetime values ​​is determined, and the deep neural network model is trained according to different electrical stress conditions. If the electrical stress condition corresponding to the feature lifetime value is a low electrical stress condition, the deep neural network model is trained using real-number normalized feature lifetime values, generating a real-number neural network model. If the electrical stress condition corresponding to the feature lifetime value is a high electrical stress condition, the deep neural network model is trained using log-normalized feature lifetime values, generating a logarithmic neural network model. Finally, the real-number neural network model and the logarithmic neural network model are set as the accelerated lifetime assessment model. The final accelerated lifetime assessment model can be expressed as:

[0121]

[0122] Where Vkink represents the boundary between high and low electrical stress conditions, Dnn_dB represents the logarithmic neural network model trained under high electrical stress conditions, Dnn_s represents the real number neural network model trained under low electrical stress conditions, and Vgs_j represents the actual applied electrical stress conditions.

[0123] When building the model, normalized data is used for learning and training. However, considering that normalized data may not accurately reflect the true lifetime value, after the model is built, in practical applications, the actual output data needs to be denormalized and converted to real number format to obtain the true characteristic lifetime value of the power device. Real number format conversion means that real characteristic lifetime values ​​do not require additional processing, while logarithmic characteristic lifetime values, i.e., values ​​in dB unit format, are converted to real values.

[0124] In one embodiment, constructing a deep neural network model and training it with the feature lifetime values ​​obtained from the fitting calculation to generate an accelerated lifetime assessment model can be specifically performed as follows:

[0125] First, we construct the input layer, hidden layers, and output layer of the deep neural network model, and set the activation function in the hidden layer. A deep neural network is a complex network structure composed of multiple layers, typically including an input layer, hidden layers, and an output layer. The hidden layer is a crucial component of a deep neural network. The hidden layer is located between the input and output layers. A deep neural network can contain multiple hidden layers, which are stacked sequentially, with the output of each layer serving as the input to the next layer. (Refer to...) Figure 5 In this embodiment, the constructed deep neural network model includes an input layer, an output layer, and three hidden layers. Each hidden layer contains 16 neurons. The main function of the hidden layers is to extract features from the input data. The number of hidden layers and the number of neurons in each hidden layer can be set by the user according to actual modeling needs. Too few or too many neurons will affect the fitting accuracy of the network. Therefore, this application uses 16 neurons. The three hidden layers and the number of neurons in each hidden layer in this embodiment are only for illustrative purposes and are not intended to be limiting. An activation function is set in each hidden layer. By using different weights and activation functions, the input data is nonlinearly transformed to capture different levels of abstract features of the input data. The user can select and adjust the activation function according to actual computational needs and actual modeling effects. In this embodiment, the Sigmoid function is used as an example. The Sigmoid function can be calculated as follows:

[0126]

[0127] Where z represents the weighting function, and the weighting function z can be expressed as:

[0128] Z = Σw i x i +b,

[0129] Where, x i In this embodiment of the application, x represents the independent variable of the neural network, specifically the first hidden layer. i This refers to the electrical stress condition input to the input layer, and the x value of the second hidden layer. i That is, z is the output of the first hidden layer, and x is the output of the third hidden layer. i That is, the z and w values ​​output by the second hidden layer. i 'b' represents the weight, and 'b' represents the bias.

[0130] Reference Figure 6Considering that traditional exponential and power function models have low fitting accuracy when fitting accelerated lifetime under high electrical stress conditions, and that high stress fitting accuracy is low and dynamic range is small due to a large allocation of error weights to low stress conditions, resulting in poor fitting performance, distinguishing between high and low electrical stress conditions can improve the overall fitting accuracy of accelerated lifetime for power devices, thereby improving the accuracy of determining the failure mechanism of power devices. Besides determining the division between high and low electrical stress conditions by the steep increase point of gate leakage current, a method can also be to first perform global model training, and then determine the dividing point based on the model training results.

[0131] Specifically, referring to Table 1 above, the characteristic lifetime under high stress and low stress conditions differs by three orders of magnitude. This leads to a larger contribution of the high-order-of-magnitude characteristic lifetime to the weight of the error function during training, resulting in a decrease in the fitting accuracy of the low-order-of-magnitude characteristic lifetime. At the same time, the algorithm format using real numbers and dB units is used for training, and the high-low boundary point can be distinguished based on the training results.

[0132] First, the characteristic lifetime values ​​are converted to real number format and dB unit format respectively, and then the converted real number characteristic lifetime values ​​and logarithmic characteristic lifetime values ​​are generated; the real number format conversion method can be expressed as:

[0133] η_j_s=η_j,

[0134] Where η_j represents the characteristic lifetime value calculated using the Wechsler distribution, and η_j_s represents the real characteristic lifetime value after real-number conversion. In this case, a real-number format conversion is performed on all electrical stress conditions, without distinguishing between high and low values;

[0135] The conversion method for logarithmic formats can be represented as:

[0136] η_j_dB = 20log(abs(η_j)),

[0137] Where η_j represents the characteristic lifetime value calculated using the Wechsler distribution, and η_j_dB represents the logarithmic characteristic lifetime value after format conversion to dB units. Similarly, logarithmic format conversion is performed on all electrical stress conditions, without distinguishing between high and low stress levels.

[0138] Next, the real and logarithmic feature lifetimes after format conversion are normalized to generate real-normalized feature lifetimes and logarithmic-normalized feature lifetimes. The normalization method can be expressed as follows:

[0139]

[0140] Where η_jx represents the real-valued characteristic lifetime and the logarithmic characteristic lifetime before normalization, η_x minη_x represents the minimum characteristic lifetime. max Represents the maximum characteristic lifetime, η_jx norm This represents the normalized real-valued eigenlife and the log-normalized eigenlife.

[0141] After the data processing is completed, the deep neural network model is trained using real-number normalized feature lifetime values ​​to generate a real-number neural network model; at the same time, the deep neural network model is trained using log-normalized feature lifetime values ​​to generate a log-number neural network model.

[0142] Subsequently, based on the training results of the real-valued neural network model and the logarithmic neural network model, the electrical stress conditions are divided into low-stress and high-stress conditions. After the two training methods are completed, there will be an overlapping region between the two models, which can be used as the distinguishing point between high and low electrical stress conditions. Finally, the real-valued neural network model is used for low-stress conditions, and the logarithmic neural network model is used for high-stress conditions, generating the final accelerated life assessment model. The final accelerated life assessment model can be expressed as:

[0143]

[0144] Where Vkink represents the boundary between high and low electrical stress conditions, Dnn_dB represents the logarithmic neural network model trained under high electrical stress conditions, Dnn_s represents the real number neural network model trained under low electrical stress conditions, and Vgs_j represents the actual applied electrical stress conditions.

[0145] When building the model, normalized data is used for learning and training. However, considering that normalized data may not accurately reflect the true lifetime value, after the model is built, in practical applications, the actual output data needs to be denormalized and converted to real number format to obtain the true characteristic lifetime value of the power device. Real number format conversion means that real characteristic lifetime values ​​do not require additional processing, while logarithmic characteristic lifetime values, i.e., values ​​in dB unit format, are converted to real values.

[0146] In summary, model training under high and low electrical stress conditions can be conducted by first dividing the model into high and low electrical stress conditions and then training the model separately, or by using both methods for global model training. Afterwards, the regions under high and low electrical stress conditions are distinguished based on the training results. Simulation experiments show that the lifespans of the power devices simulated by both methods are basically consistent with minimal differences, and both methods can accurately fit the lifespan of the power devices. (Refer to...) Figure 7The lifetime of the power device evaluated by the model constructed in the embodiments of this application can be seen to fit the actual lifetime value well under both high and low electrical stress conditions.

[0147] In this application, a deep neural network model is used to train the model for high and low stress conditions separately. Considering that traditional prediction and evaluation models do not distinguish between high and low stress, the fitting effect is better under low electrical stress conditions, while the fitting effect is poor and the fitting error is large under high electrical stress conditions. Therefore, by determining the high and low boundary point, the model is divided into two parts, and fitting calculations are performed separately for different electrical stress conditions, thereby improving the fitting accuracy.

[0148] The model constructed using the modeling method of this application can perform failure analysis and lifetime prediction on the new failure mechanisms and lifetime characteristics exhibited by newly developed GaN power devices. It can identify the failure models of new power devices in advance and determine the failure mechanisms of new structures, thereby overcoming the fixed requirements of traditional exponential function models and power function models on characteristic lifetimes and providing a theoretical basis for device reliability design.

[0149] Figure 1 This is a flowchart illustrating a method for establishing a life assessment model in one embodiment. It should be understood that, although... Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows; unless explicitly stated otherwise, there is no strict order requirement for the execution of these steps, and they can be executed in other orders; and Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0150] Based on the above method, this application also discloses a life assessment model establishment apparatus.

[0151] Reference Figure 8 The device includes the following modules:

[0152] The failure time calculation module 801 is used to determine the electrical stress conditions of the power device based on the gate leakage current of the power device, and to measure the failure time of the power device under electrical stress conditions.

[0153] The distribution function calculation module 802 is used to calculate the Vickers cumulative distribution function of the power device at the failure time.

[0154] The characteristic lifetime calculation module 803 is used to calculate the characteristic lifetime value of the power device under electrical stress conditions based on the Vickers cumulative distribution function and the preset global shape factor.

[0155] The model building and training module 804 is used to build a deep neural network model and train the deep neural network model using the feature lifetime values ​​obtained by fitting calculation to generate an accelerated lifetime assessment model, which is a partitioned deep neural network model.

[0156] In one embodiment, the failure time calculation module 801 is specifically used to measure the steep increase point of the gate leakage current of the power device, and determine the steep increase stress range of the gate leakage current based on the steep increase point; determine several different electrical stress conditions based on the steep increase stress range; and measure the failure time of the power device under several electrical stress conditions using a dielectric breakdown full-life accelerated test.

[0157] In one embodiment, the distribution function calculation module 802 calculates the Vickers cumulative distribution function in the following ways:

[0158]

[0159] Where t represents the failure time, β represents the shape factor, η represents the scaling factor, γ represents the time delay, and F(t) represents the Wechsler cumulative distribution function.

[0160] In one embodiment, the characteristic lifetime calculation module 803 is specifically used to fit a Widman distribution of the failure time according to the Widman cumulative distribution function; calculate the intercept value of the failure time under electrical stress conditions according to the Widman distribution of the failure time; and calculate the characteristic lifetime value of the power device under electrical stress conditions according to a preset global shape factor and intercept value. The calculation method of the characteristic lifetime value includes:

[0161]

[0162] Where, intercept_j represents the calculated intercept value, β′ represents the global shape factor, and η_j represents the characteristic lifetime value under the j-th stress condition.

[0163] In one embodiment, the model building and training module 804 is specifically used to construct the input layer, hidden layer, and output layer of a deep neural network model, and to set activation functions, weights, and biases in the hidden layer; determine the boundary point of the electrical stress condition based on the steep increase point of the gate leakage current, and generate low electrical stress conditions and high electrical stress conditions; determine the range to which the electrical stress condition corresponding to the feature lifetime value belongs, and perform format conversion on the feature lifetime value according to different electrical stress conditions; if the electrical stress condition corresponding to the feature lifetime value is a low electrical stress condition, then the feature lifetime value is converted to a real number format and a real number feature lifetime value is generated; if the electrical stress condition corresponding to the feature lifetime value is a high electrical stress condition, then the feature lifetime value is converted to a dB unit format and a logarithmic feature lifetime value is generated. The lifespan assessment process involves: normalizing the real and logarithmic characteristic lifetime values ​​to generate real-normalized and logarithmic-normalized characteristic lifetime values; determining the range of electrical stress conditions corresponding to the characteristic lifetime values ​​and training the deep neural network model based on different electrical stress conditions; if the electrical stress condition corresponding to the characteristic lifetime value is a low electrical stress condition, then the deep neural network model is trained using the real-normalized characteristic lifetime value, generating a real neural network model; if the electrical stress condition corresponding to the characteristic lifetime value is a high electrical stress condition, then the deep neural network model is trained using the logarithmic-normalized characteristic lifetime value, generating a logarithmic neural network model; and finally, setting the real neural network model and the logarithmic neural network model as accelerated lifespan assessment models.

[0164] In one embodiment, the model building and training module 804 is specifically used to construct the input layer, hidden layer, and output layer of a deep neural network model, and set activation functions, weights, and biases in the hidden layer; convert the feature lifetime values ​​to real number format and dB unit format respectively, and generate real number feature lifetime values ​​and logarithmic feature lifetime values ​​after format conversion; normalize the real number feature lifetime values ​​and logarithmic feature lifetime values ​​after format conversion, and generate real number normalized feature lifetime values ​​and logarithmic normalized feature lifetime values; train the deep neural network model using the real number normalized feature lifetime values ​​and generate a real number neural network model; train the deep neural network model using the logarithmic normalized feature lifetime values ​​and generate a logarithmic neural network model; based on the training results of the real number neural network model and the logarithmic neural network model, divide the electrical stress conditions into low electrical stress conditions and high electrical stress conditions; use the real number neural network model corresponding to the low electrical stress conditions, and use the logarithmic neural network model corresponding to the high electrical stress conditions, and generate the final accelerated lifetime assessment model.

[0165] In one embodiment, the model building and training module 804 is also used to perform inverse normalization processing on the results output by the accelerated lifetime assessment model.

[0166] The life assessment model establishment apparatus provided in this application embodiment can be applied to the life assessment model establishment method provided in the above embodiment. For relevant details, please refer to the above method embodiment. Its implementation principle and technical effect are similar, and will not be repeated here.

[0167] It should be noted that the life assessment model building device provided in this embodiment is only illustrated by the above-described division of functional modules / units when building a life assessment model. In practical applications, the above functions can be assigned to different functional modules / units as needed, that is, the internal structure of the life assessment model building device can be divided into different functional modules / units to complete all or part of the functions described above. Furthermore, the implementation method of the life assessment model building method provided in the above method embodiment and the implementation method of the life assessment model building device provided in this embodiment belong to the same concept. The specific implementation process of the life assessment model building device provided in this embodiment is detailed in the above method embodiment and will not be repeated here.

[0168] This application also discloses a computer device.

[0169] Specifically, such as Figure 9 As shown, the computer device can be a desktop computer, laptop computer, handheld computer, or cloud server, etc. The computer device may include, but is not limited to, a processor and memory. The processor and memory can be connected via a bus or other means. The processor can be a Central Processing Unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, graphics processing units (GPUs), embedded neural network processing units (NPUs) or other dedicated deep learning coprocessors, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0170] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the methods in the above embodiments of this application. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions, and modules stored in the memory, thereby implementing the methods in the above embodiments. The memory may include a program storage area and a data storage area, wherein the program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0171] This application also discloses a computer-readable storage medium.

[0172] Specifically, the computer-readable storage medium is used to store a computer program, which, when executed by a processor, implements the methods described in the above-described method embodiments. Those skilled in the art will understand that implementing all or part of the processes in the methods described in the above-described embodiments of this application can be accomplished by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium can also include combinations of the above types of memory.

[0173] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.

Claims

1. A method for establishing a life assessment model, characterized in that: The method is applied to a life assessment model building system, the life assessment model building system including power devices, and the method includes: The electrical stress condition of the power device is determined based on the gate leakage current of the power device, and the failure time of the power device under the electrical stress condition is measured. Calculate the Wechsler cumulative distribution function of the power device at the failure time; The characteristic lifetime value of the power device under the electrical stress condition is calculated by fitting the Vickers cumulative distribution function and the preset global shape factor. A deep neural network model is constructed, and the feature lifetime values ​​obtained by fitting calculation are used to train the deep neural network model to generate an accelerated lifetime assessment model, wherein the accelerated lifetime assessment model is a partitioned deep neural network model. The process of constructing a deep neural network model and training the deep neural network model using the feature lifetime values ​​obtained from the fitting calculation to generate an accelerated lifetime assessment model includes: Construct an input layer, hidden layer, and output layer for a deep neural network model, and set activation functions, weights, and biases in the hidden layer; The characteristic lifetime values ​​are converted to real number format and dB unit format respectively, and the format-converted real number characteristic lifetime value and logarithmic characteristic lifetime value are generated. The real number feature lifetime and the logarithmic feature lifetime after format conversion are normalized to generate real number normalized feature lifetime and logarithmic normalized feature lifetime. The deep neural network model is trained using the real-number normalized feature lifetime value to generate a real-number neural network model; The deep neural network model is trained using the log-normalized feature lifetime values ​​to generate a log-normalized neural network model; Based on the training results of the real neural network model and the logarithmic neural network model, the electrical stress conditions are divided into low electrical stress conditions and high electrical stress conditions. The real neural network model is used under the low electrical stress condition, and the logarithmic neural network model is used under the high electrical stress condition, and the final accelerated life assessment model is generated.

2. The method according to claim 1, characterized in that: The step of determining the electrical stress condition of the power device based on the gate leakage current of the power device, and measuring the failure time of the power device under the electrical stress condition, includes: Measure the steep increase point of the gate leakage current of the power device, and determine the steep increase stress range of the gate leakage current based on the steep increase point; Several different electrical stress conditions are determined based on the steeply increasing stress range; The failure time of the power device under several electrical stress conditions was measured using a dielectric breakdown full-life accelerated test.

3. The method according to claim 1, characterized in that: The calculation method of the Vickers cumulative distribution function includes: ； Where t represents the failure time. Represents the shape factor. Indicates the scaling factor. Let F(t) represent the time delay, and let F(t) represent the Wechsler cumulative distribution function.

4. The method according to claim 1, characterized in that: The step of calculating the characteristic lifetime value of the power device under the electrical stress condition by fitting the Vickers cumulative distribution function and the preset global shape factor includes: The failure time is fitted with a Wechsler distribution based on the Wechsler cumulative distribution function; The intercept value of the failure time under the electrical stress condition is calculated based on the Wechsler distribution of the failure time. The characteristic lifetime value of the power device under the electrical stress condition is calculated based on the preset global shape factor and the intercept value. The calculation method for the characteristic lifetime value includes: ； in, This represents the calculated intercept value. Represents the global shape factor. This represents the characteristic lifetime value under the j-th stress condition.

5. The method according to claim 2, characterized in that: The construction of a deep neural network model, and the training of the deep neural network model using the feature lifetime values ​​obtained from the fitting calculation, to generate an accelerated lifetime assessment model, is replaced by: Construct an input layer, hidden layer, and output layer for a deep neural network model, and set activation functions, weights, and biases in the hidden layer; The boundary point of the electrical stress condition is determined based on the steep increase point of the gate leakage current, and low electrical stress condition and high electrical stress condition are generated. Determine the range of electrical stress conditions corresponding to the characteristic lifetime value, and perform format conversion on the characteristic lifetime value according to different electrical stress conditions; If the electrical stress condition corresponding to the characteristic lifetime value is a low electrical stress condition, then the characteristic lifetime value is converted into a real number format and a real number characteristic lifetime value is generated. If the electrical stress condition corresponding to the characteristic lifetime value is a high electrical stress condition, then the characteristic lifetime value is converted to dB unit format and a logarithmic characteristic lifetime value is generated. The real-valued feature lifetime and the logarithmic feature lifetime are normalized to generate real-valued normalized feature lifetime and logarithmic normalized feature lifetime. Determine the range of electrical stress conditions corresponding to the feature lifetime value, and train the deep neural network model according to different electrical stress conditions; If the electrical stress condition corresponding to the feature lifetime value is a low electrical stress condition, then the deep neural network model is trained using the real-number normalized feature lifetime value, and a real-number neural network model is generated. If the electrical stress condition corresponding to the feature lifetime value is a high electrical stress condition, then the deep neural network model is trained using the log-normalized feature lifetime value, and a log-normalized neural network model is generated. The real-number neural network model and the logarithmic neural network model are set as the accelerated life assessment model.

6. The method according to claim 1 or 5, characterized in that: After generating the accelerated life assessment model, the method further includes: performing inverse normalization on the output of the accelerated life assessment model.

7. A life assessment model establishment device, characterized in that: The device is used in a life assessment model building system, the life assessment model building system including power devices, and the device includes: The failure time calculation module (801) is used to determine the electrical stress condition of the power device based on the gate leakage current of the power device, and to measure the failure time of the power device under the electrical stress condition. The distribution function calculation module (802) is used to calculate the Widman cumulative distribution function of the power device at the failure time; The characteristic lifetime calculation module (803) is used to calculate the characteristic lifetime value of the power device under the electrical stress condition based on the Vickers cumulative distribution function and the preset global shape factor. The model building and training module (804) is used to build a deep neural network model and train the deep neural network model using the feature lifetime values ​​obtained by fitting calculation to generate an accelerated lifetime assessment model, wherein the accelerated lifetime assessment model is a partitioned deep neural network model. The model construction and training module 804 is specifically used to construct the input layer, hidden layer, and output layer of the deep neural network model, and to set the activation function, weights, and biases in the hidden layer; to convert the feature lifetime values ​​to real number format and dB unit format respectively, and to generate real number feature lifetime values ​​and logarithmic feature lifetime values ​​after format conversion; to normalize the format-converted real number feature lifetime values ​​and logarithmic feature lifetime values; to train the deep neural network model using the real number normalized feature lifetime values ​​and to generate a real number neural network model; to train the deep neural network model using the logarithmic normalized feature lifetime values ​​and to generate a logarithmic neural network model; based on the training results of the real number neural network model and the logarithmic neural network model, the electrical stress conditions are divided into low electrical stress conditions and high electrical stress conditions; a real number neural network model is used under low electrical stress conditions, and a logarithmic neural network model is used under high electrical stress conditions, and the final accelerated lifetime assessment model is generated.

8. A computer device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and executed according to any one of claims 1 to 6.

Images