A Lifetime Prediction Method for SiC Power Devices Based on Composite Stress Failure Mode

Abstract

This invention discloses a method for predicting the lifetime of SiC power devices based on composite stress failure modes, belonging to the field of power semiconductor device reliability assessment technology. The method applies composite stress aging tests to SiC power devices, collecting drain current transient response sequences and case temperature fluctuation sequences. The drain current transient response sequences are mapped to phase space through chaotic attractor reconstruction, extracting the first topological invariant reflecting failure evolution. The case temperature fluctuation sequence is symbolized to construct a symbolic transfer entropy matrix characterizing the thermal stress cycling mode. A joint state-space model is established based on the first topological invariant and the symbolic transfer entropy matrix. The remaining lifetime range is predicted based on the state transition probability. This method overcomes the limitation of single-stress accelerated testing in reproducing the multi-stress coupled degradation mechanism under real-world conditions, improving the robustness of failure evolution feature extraction.

Description

Technical Field

[0001] This invention relates to the field of power semiconductor device reliability assessment technology, specifically to a method for predicting the lifetime of SiC power devices based on composite stress failure modes. Background Technology

[0002] Silicon carbide power devices (SiCs) are subjected to combined stresses from DC bias, high-frequency switching, and temperature fluctuations in high-frequency, high-voltage applications such as new energy power generation and electric vehicles. Their degradation paths are complex, and failure modes are highly coupled. Existing device lifetime prediction methods often employ single-stress accelerated aging tests to construct empirical degradation models, such as isothermal reverse bias tests or power cycling tests. The stress forms applied in these tests differ significantly from actual operating conditions, failing to accurately reproduce the degradation mechanisms under multi-stress coupling. When devices are simultaneously subjected to electrical and thermal stresses, there are competing and synergistic effects between different failure modes. Single-stress accelerated models struggle to capture this interaction, leading to predictions that deviate from actual lifetimes. Physical model-based lifetime prediction schemes rely on precise modeling of internal carrier transport and trap dynamics. However, the charging and discharging behavior of interface states and near-interface oxide layer traps under combined stress exhibits nonlinear coupling characteristics, making physical parameter extraction difficult and drastically increasing model complexity. While data-driven methods can directly learn failure evolution patterns from degradation monitoring data, existing schemes typically use time-domain or frequency-domain statistical characteristics as health factors, which are sensitive to noise and lack the ability to characterize early, subtle degradation signs. In describing thermal stress state, conventional methods only focus on the amplitude statistics of case temperature, ignoring the temporal transfer mode and local dynamic characteristics of case temperature fluctuations, and cannot fully characterize the cumulative damage effect of thermal cycling on device packaging and bonding structures. Summary of the Invention

[0003] The purpose of this invention is to provide a lifetime prediction method for SiC power devices based on composite stress failure modes. This method addresses the problem that conventional lifetime prediction methods cannot effectively characterize the electrothermal stress coupling degradation mechanism under composite stress. It extracts robust invariant features sensitive to failure evolution from the chaotic dynamics of the transient response of drain current, and mines the transfer mode information of thermal stress cycling from the symbolic representation of case temperature fluctuations. By fusing the two methods, a degradation state space is constructed, enabling probabilistic prediction of the remaining lifetime range.

[0004] The objective of this invention can be achieved through the following technical solutions: This invention provides a method for predicting the lifetime of SiC power devices based on composite stress failure modes, comprising the following steps: The transient response sequence of drain current and the case temperature fluctuation sequence of SiC power devices were collected during a composite stress aging test. This composite stress aging test included DC bias stress, switching cycle stress, and high-temperature storage stress, applied in a preset time-alternating pattern. The switching frequency of the switching cycle stress increased stepwise with aging time to more realistically simulate the complex multi-stress coupling conditions experienced by power devices in practical applications such as electric vehicle drive and new energy grid connection. The drain current waveform of the SiC power device was continuously acquired at a fixed sampling frequency within a preset time window after each switching action. Peak hold processing was performed on the drain current waveform to extract the peak drain current value within each switching cycle. The peak drain current values ​​from multiple consecutive switching cycles were arranged chronologically to generate the transient response sequence of drain current. Simultaneously, the case temperature value of the SiC power device was continuously acquired at the same sampling frequency. Sliding median filtering was performed on the case temperature value to remove measurement noise, resulting in the case temperature fluctuation sequence. This acquisition method yielded high-quality raw data that sensitively reflects major failure mechanisms such as bond wire breakage and solder layer fatigue.

[0005] The drain current transient response sequence is mapped to phase space using chaotic attractor reconstruction to extract the first topological invariant reflecting the failure evolution. Specifically, a delayed embedding operation is performed on the drain current transient response sequence, setting the embedding dimension value and the delay time value to construct multiple phase point vectors; each phase point vector consists of multiple data points selected from the drain current transient response sequence at delay time intervals. Preferably, the embedding dimension value m and the delay time value... A combination of the spurious nearest neighbor method and the mutual information method is used to determine the optimal unfolding of the system dynamic trajectory when reconstructing the phase space. The multiple phase point vectors are arranged in the phase space to form a chaotic attractor trajectory; persistent cohomology analysis is performed on the chaotic attractor trajectory to calculate the persistence graph of the Betti number as a function of the scale parameter in different dimensions. Cohomology class feature values ​​corresponding to all generating loops with the maximum lifetime period are extracted from the persistence graph and used as the first topological invariant. The first topological invariant essentially characterizes the geometric and topological changes of the nonlinear degenerate attractor under multi-stress coupling in the transient response of the drain current. It is highly sensitive to the degenerate state and has good robustness to noise interference.

[0006] The shell temperature fluctuation sequence is symbolized to construct a symbolic transfer entropy matrix to characterize the thermal stress cycling mode. The shell temperature fluctuation sequence is divided into multiple symbol intervals according to its amplitude range, and each symbol interval is assigned a discrete symbol. Each shell temperature value in the sequence is replaced with the discrete symbol corresponding to its respective symbol interval, resulting in a symbolic sequence. Preferably, the boundaries of each interval are dynamically determined using an equal probability quantile method, ensuring that each symbol interval contains approximately the same number of shell temperature data points, effectively avoiding symbol representation bias caused by uneven shell temperature distribution. The symbol sequence... The combination of two adjacent symbols is defined as a symbol transition state. The frequency of occurrence of all possible symbol transition states is counted, and a symbol transition probability matrix P is constructed, where the matrix elements... This represents the probability estimate of transitioning from the current symbol i to the next symbol j. Each element in the symbol transition probability matrix P is multiplied by the self-information of the corresponding symbol transition state to obtain the weighted transition probability matrix. Then, each row of the weighted transition probability matrix is ​​normalized to generate the symbol transition entropy matrix H. The row index of the symbol transition entropy matrix H corresponds to the current symbol, and the column index corresponds to the next symbol. Its elements... It is a normalized weighted transition probability that contains the uncertainty of the amplitude pattern and temporal correlation of temperature changes, and can accurately describe the change in the rhythm of shell temperature fluctuations caused by solder layer aging.

[0007] According to the first topological invariant Using the symbolic transfer entropy matrix H, a joint state-space model describing the degradation process of the SIC power device is established. The first dimension of the observation feature vector is used as the first dimension; all matrix elements of the symbolic transfer entropy matrix H are expanded row-wise and concatenated into a feature row vector, which serves as the second dimension of the observation feature vector, thus constructing a joint observation feature vector x that integrates electrical nonlinearity features and thermal mode features. The observation feature vector x is then input into a sparse Bayesian learning framework, with the kernel function defined as the Laplace kernel. The degradation rate latent variable corresponding to each observation feature vector is calculated using a correlation vector machine regression algorithm. Preferably, the latent variable of the obtained degradation rate... An adaptive Kalman filter is introduced for smoothing to eliminate jitter in the degradation rate estimation caused by local measurement perturbations. The smoothed degradation rate latent variable is then processed. By comparing the SiC power device at each time step with multiple preset degradation state thresholds, it is discretized into one of several degradation states, and the sequence of degradation states on the continuous time axis is used as the hidden state chain of the Hidden Markov Model. The parameters of the hidden state chain are estimated using the Viterbi algorithm to obtain the state transition probability matrix A and the state emission probability matrix B. Combining these two matrices forms the joint state space model. This model unifies the multi-source characteristics of electrothermal processes within the Hidden Markov Framework, achieving a compact expression and dynamic tracking of the composite stress failure evolution process.

[0008] Based on the state transition probabilities in the joint state-space model, the remaining service life of the SiC power device is predicted. The state transition probability distribution corresponding to the current degradation state is extracted from the state transition probability matrix A of the joint state-space model. This distribution contains the probability values ​​for transitioning from the current degradation state to all other degradation states. Monte Carlo random sampling is performed based on this distribution to generate multiple degradation paths from the current time until a preset failure state is reached. Each degradation path contains a series of sequentially passed degradation states. The number of time steps taken by each degradation path from the current degradation state to the preset failure state is calculated, and the time steps of all degradation paths are combined into a time step set. The tenth percentile of this time step set is taken as the lower limit of the remaining service life. The 90th percentile of the set of time steps is taken as the upper limit of the remaining lifespan. The output is from and The remaining service life range constituted is calculated. This range-based prediction result quantifies the uncertainty of life prediction under combined stress, providing a more reliable decision-making basis for condition-based maintenance and dynamic reliability assessment of power devices.

[0009] The beneficial effects of this invention are: The transient response sequence of drain current is reconstructed into a chaotic attractor trajectory through delayed embedding operations, and the first topological invariant is extracted as a degradation feature using persistent homology analysis. Conventional time-domain statistical features change gradually in the early degradation stage, resulting in a high detection delay for weak failure signals. Chaotic attractors are extremely sensitive to small changes in the system's dynamic behavior; when the density of states at the device interface changes slightly due to accumulated electrical stress, the attractor topology undergoes quantifiable deformation. Homology-type features with the maximum lifetime period are extracted using the persistence graph of the Betti number as a function of scale parameters. This filters out transient topological noise introduced by measurement noise and preserves stable geometric structures reflecting permanent damage. Therefore, it captures early signs of internal defect accumulation before the drain current waveform shows significant amplitude drift, improving the timeliness of remaining lifetime prediction. Symbolic processing based on dynamic partitioning of equal probability quantiles is used to convert the case temperature fluctuation sequence into a symbolic sequence, and a symbolic transfer entropy matrix is ​​constructed. Conventional case temperature analysis, relying on single statistical quantities such as mean or extreme values, cannot reflect the temporal transfer law of temperature fluctuations. Symbolization maps continuous temperature values ​​to a discrete symbol space, eliminating the impact of small temperature perturbations on state partitioning. Simultaneously, the equal probability quantile boundaries ensure that each symbol interval contains equal information, avoiding the problem of uneven information density in high and low temperature regions with fixed amplitude thresholds. By statistically analyzing the transition probabilities of adjacent symbols and introducing self-information weighting, the resulting transfer entropy matrix can simultaneously characterize the fluctuation amplitude distribution of thermal stress cycles and the uncertainty of the transfer path. When the device's heat dissipation performance deteriorates due to solder layer voids or thermal grease degradation, the case temperature fluctuation pattern changes accordingly and is significantly amplified in the transfer entropy matrix. Attached Figure Description

[0010] The invention will now be further described with reference to the accompanying drawings.

[0011] Figure 1 This is a flowchart of the SiC power device lifetime prediction method based on composite stress failure mode; Figure 2 This is a flowchart of delayed embedding and extraction of the first topological invariant from the drain current transient response sequence; Figure 3 This is a flowchart of the generation process of the symbolic transfer entropy matrix of the shell temperature fluctuation sequence. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] See Figure 1This invention provides a method for predicting the lifetime of SiC power devices based on composite stress failure modes, comprising the following steps: acquiring the transient response sequence of drain current and the case temperature fluctuation sequence of the SiC power device in a composite stress aging test; mapping the transient response sequence of drain current to phase space through chaotic attractor reconstruction, and extracting the first topological invariant reflecting the failure evolution; symbolizing the case temperature fluctuation sequence to construct a symbolic transfer entropy matrix for characterizing the thermal stress cycling mode; establishing a joint state-space model describing the degradation process of the SiC power device based on the first topological invariant and the symbolic transfer entropy matrix; and predicting the remaining lifetime range of the SiC power device based on the state transition probabilities in the joint state-space model.

[0014] In practice, a composite stress aging test is conducted on SiC power devices by simultaneously applying DC bias stress, switching cycle stress, and high-temperature storage stress. This composite stress aging test is performed using a programmable power supply, an arbitrary waveform generator, a power switch drive circuit, and a temperature-controlled chamber. The DC bias stress is applied between the drain and source of the SiC power device by a constant drain-source voltage output from the programmable power supply, with the DC bias stress amplitude set to 80% of the rated drain-source voltage of the SiC power device. The switching cycle stress is generated by an arbitrary waveform generator producing a pulse-width modulation signal, which is converted into a gate drive voltage by the power switch drive circuit and applied to the gate of the SiC power device, causing the SiC power device to cycle between on and off states. The initial switching frequency of the switching cycle stress is set to 50% of the typical application frequency in the SiC power device datasheet. The high-temperature storage stress is provided by the temperature-controlled chamber, which maintains the ambient temperature of the SiC power device at a constant 85% of the highest operating junction temperature of the SiC power device.

[0015] DC bias stress, switching cycle stress, and high-temperature storage stress are applied according to a preset time alternation pattern. The preset time alternation pattern means that a complete alternation cycle includes a DC bias plus high-temperature storage phase, a switching cycle plus high-temperature storage phase, and a pure high-temperature storage phase. The duration of the DC bias plus high-temperature storage phase is a first preset duration. During this phase, DC bias stress and high-temperature storage stress act simultaneously on the SiC power device, while the switching cycle stress is disconnected. The duration of the switching cycle plus high-temperature storage phase is a second preset duration. During this phase, switching cycle stress and high-temperature storage stress act simultaneously on the SiC power device, while the DC bias stress is disconnected. The duration of the pure high-temperature storage phase is a third preset duration. During this phase, only high-temperature storage stress acts on the SiC power device, while both DC bias stress and switching cycle stress are disconnected. The first, second, and third set durations are determined based on the time proportions of the on-state, switching state, and standby state in a typical task profile of the SiC power device, ensuring that the ratio among the three is equal to the ratio of the on-state time proportion, switching state time proportion, and standby state time proportion of the SiC power device in the target application. The composite stress aging test is performed cyclically according to the above alternating cycle until the test ends.

[0016] The switching frequency under cyclic stress increases in a stepwise manner with aging time. The total aging time of the composite stress aging test is divided into multiple steps. The switching frequency at the start of the q-th step is adjusted to the initial switching frequency multiplied by an increasing coefficient κ raised to the power of (q-1), where κ ranges from 1.1 to 1.3, and the specific value of κ is set according to the accelerated degradation requirements of the gate oxide layer of the SiC power device. Within the same step, the switching frequency remains constant. The duration of each step is the same, and the number of steps is no less than five.

[0017] During the composite stress aging test, the drain current waveform of the SiC power device was continuously acquired at a fixed sampling frequency within a preset time window after each switching action. The fixed sampling frequency was set to be at least one hundred times the current switching frequency of the switching cycle stress. The preset time window started at the moment when the gate drive voltage rose to the threshold voltage of the SiC power device, and the duration of the preset time window was set to 20% of the conduction time in the current switching cycle to ensure that the preset time window was completely within the stable conduction range after the Miller plateau ended. The drain current waveform was acquired using a high-precision current sensor connected in series in the drain circuit of the SiC power device. The output signal of the current sensor was sampled and stored at a fixed sampling frequency by an analog-to-digital converter.

[0018] Peak hold processing is performed on the drain current waveform to extract the peak drain current value within each switching cycle. Peak hold processing is implemented using a real-time comparator circuit. The non-inverting input of the comparator receives the digital value of the drain current waveform after analog-to-digital conversion, while the inverting input is connected to the output of a peak register. The comparator output controls the update of the peak register: when the digital value of the drain current waveform is greater than the current value stored in the peak register, the comparator outputs a high level, enabling the peak register to write the digital value of the drain current waveform; when the digital value of the drain current waveform is less than or equal to the current value stored in the peak register, the comparator outputs a low level, and the peak register retains its original value. At the end of each preset time window, the value latched in the peak register is the peak drain current value for the current switching cycle. Subsequently, the peak register is reset to zero to prepare for peak detection in the next switching cycle.

[0019] The peak drain current values ​​from multiple consecutive switching cycles are arranged in time sequence to generate a drain current transient response sequence. The drain current transient response sequence is represented as follows: Where n is the switching cycle number, N represents the total number of switching cycles accumulated during the composite stress aging test. This represents the peak drain current collected during the nth switching cycle.

[0020] Simultaneously, the case temperature values ​​of the SiC power device are continuously acquired at the same sampling frequency. The case temperature values ​​are obtained through a thin-film thermocouple sensor mounted on the surface of the SiC power device near the chip. The output signal of the thin-film thermocouple sensor, after cold junction compensation and analog-to-digital conversion, is sampled at a fixed sampling frequency identical to that of the drain current waveform acquisition, resulting in the original case temperature sampling sequence. The timestamp of each sampling point in the original case temperature sampling sequence is strictly aligned with the timestamp of the drain current waveform sampling point.

[0021] The original shell temperature sampling sequence is subjected to sliding median filtering to remove measurement noise, resulting in a shell temperature fluctuation sequence. The sliding median filtering is implemented as follows: a sliding window length L is set, where L is an odd number and ranges from 11 to 21. The specific value of L is determined by a trade-off between the shell temperature sampling rate and the temperature change rate. For the k-th original shell temperature sample value... Take the included and before and after All values ​​within a window consisting of adjacent sampled values ​​are sorted according to their amplitude, and the median value after sorting is taken as the filtered shell temperature value. The process of sliding median filtering is expressed by the following formula: ; in, This represents the shell temperature value output after sliding median filtering at time k. This represents the original shell temperature sample value, where L is the sliding window length, and L is an odd number. This operation retrieves the median of the values ​​in the set, where k is the index of the sampling point and the range of k is [value missing]. to , where k is the total length of the original shell temperature sampling sequence. For sampling points at the beginning and end of the sequence that are less than a full window length, edge interpolation is used to supplement virtual sampling values, with the nearest valid original shell temperature sampling value being the virtual sampling value. All time points are... Arranged chronologically, the shell temperature fluctuation sequence is obtained.

[0022] In practice, a delayed embedding operation is performed on the transient response sequence of the drain current. (See also...) Figure 2 The transient response sequence of the drain current is denoted as ,in N represents the total number of switching cycles accumulated during the composite stress aging test. The parameters for the delayed embedding operation include the embedding dimension value m and the delay time value. Embedded dimension value m and delay time value The method of determining the spurious nearest neighbor and the mutual information method are used together.

[0023] The implementation of the spurious nearest neighbor method is as follows: For a given candidate embedding dimension value m, construct all phase point vectors from the drain current transient response sequence. The phase point vectors are constructed as follows: , , where T denotes the vector transpose. For each phase point vector Find the nearest neighboring phase vector in m-dimensional phase space by Euclidean distance. Calculate the m-dimensional phase point vector and Euclidean distance between Then add one dimension to get 3D phase point vector and ,calculate Euclidean distance under dimension Define the spurious nearest neighbor criterion. ,in For unit step function, For distance variation tolerance, Set to 10. When The first time it was less than five percent or When the embedding dimension value no longer decreases significantly with increasing m, the corresponding minimum m is taken as the embedding dimension value. Delay time value. The mutual information method is used to determine the transient response sequence of the drain current. Rather than delay Sequence after step Mutual information between Mutual information is calculated based on the joint distribution histogram and marginal distribution histogram of the sequences. Follow Increases while decreasing. Take Decay to the point corresponding to the first local minimum The delay time value is determined jointly by the false nearest neighbor method and the mutual information method. The process is as follows: first, the mutual information method is used to determine the delay time value. In the delay time value Based on this, the candidate embedding dimension value m is obtained by traversing the candidate embedding dimension value m using the spurious nearest neighbor method.

[0024] Set embedding dimension value and delay time value Then, multiple phase point vectors are constructed. The phase point vectors are constructed as follows: ; in, Let represent the i-th phase point vector, where i is the index of the phase point vector, and the value of i ranges from 1 to 2. , M is the total number of phase point vectors; These are elements in the transient response sequence of the drain current; is the delay time value for the delayed embedding operation; m is the embedding dimension value for the delayed embedding operation; T represents the vector transpose operation. Each phase vector The transient response sequence of drain current is ordered by the delay time value It consists of m data points selected at intervals, and the dimension of the point vector is m.

[0025] Arrange M phase point vectors in an m-dimensional phase space, with each phase point vector representing a point in the m-dimensional phase space. Connect the phase point vectors sequentially in ascending order of index i to form a continuous trajectory, which is the chaotic attractor trajectory. The chaotic attractor trajectory reflects the dynamic evolution structure of the drain current transient response sequence in the phase space.

[0026] Continuous cohomology analysis was performed on the chaotic attractor trajectory. The Vietoris-Rips complex construction method was employed, using the M phase point vectors from the chaotic attractor trajectory as point cloud data, and a scaling parameter was set. When the Euclidean distance between two phase vectors is less than or equal to At that time, add an edge connecting the two phase point vectors in the Vietoris-Rips complex. This is done as the scale parameter... As the value is gradually increased from zero to a maximum value covering the entire point cloud, the topology of the Vietoris-Rips complex changes, with generating loops constantly appearing and disappearing. Parameters at each scale are calculated. The zero Vietoris-Rips complex under the given value And one Vibetty number Zero Vibeti number The number of connected components, a one-dimensional Betty number. This represents the number of one-dimensional holes or generating loops. The scale parameter value for the occurrence of each generating loop is also included. and disappearance scale parameter value Record as coordinate pairs The coordinate pairs of all generated loops form a persistence graph in a plane with the occurrence scale parameter as the horizontal axis and the disappearance scale parameter as the vertical axis. In the persistence graph, each generated loop corresponds to a point, and the lifetime of a generated loop is defined as... .

[0027] Extract the homology class feature values ​​corresponding to all generating cycles with the maximum lifetime from the persistence graph, and use them as the first topological invariant. The extraction method is: index all points in the persistence graph by their lifetime. Sort the generated cycles from largest to smallest, and select the top q generating cycles with the longest lifetimes. The value of q is set to 3. The rationale for setting q to 3 is that the number of feature points significantly deviating from the diagonal in the univariate persistence graph usually corresponds to the main periodic structure of the system. Selecting the top 3 longest-lived generated cycles can capture the most important topological features. Extract the individual features of these q generated cycles. The coordinate values, arranged sequentially with the occurrence and disappearance scale parameters of the q generating cycles, form an eigenvector, which is the first topological invariant. The first topological invariant is based on the coordinates of the q generating cycles. Represented in the form of .

[0028] In practice, the shell temperature fluctuation sequence is divided into multiple sign intervals according to its amplitude range, and the boundaries of each interval are dynamically determined using the equal probability quantile method. (See also...) Figure 3 The shell temperature fluctuation sequence is denoted as , K represents the total length of the shell temperature fluctuation sequence. The number of symbol intervals is set to S, with S set to 8. The reason for choosing S as 8 is that when S is less than 8, the ability to distinguish the details of temperature fluctuations is insufficient; when S is greater than 8, the statistical samples of symbol transition states are too sparse. Choosing S as 8 strikes a balance between the ability to resolve temperature dynamics and the statistical robustness of the symbol transition probability matrix. All K shell temperature values ​​in the shell temperature fluctuation sequence are sorted in ascending order to obtain the sorted sequence. Calculate the indices of the S-1 dividing points. ,in , This indicates rounding up. It retrieves the sorted shell temperature value corresponding to the dividing point. As boundary values ​​of the symbol interval, S-1 boundary values ​​are obtained. And supplement , This forms S symbol intervals: Each symbol interval is assigned a discrete symbol, denoted as . The discrete symbol set is Each shell temperature value in the shell temperature fluctuation sequence Replace it with the discrete symbol corresponding to its symbol interval: when When it falls into the a-th symbol interval, Mapping to discrete symbols , The discrete symbols obtained after mapping are arranged in their original time order to obtain a symbol sequence, denoted as . ,in Represents the discrete symbol at time t.

[0029] A symbol transition state is defined as a combination of two adjacent symbols in a symbol sequence. The symbol sequence has K-1 pairs of adjacent symbols. , Count the frequency of all possible symbol transition states: Construct an S×S counting matrix F, where the elements are... Indicates from symbol Transfer to symbol The total number of occurrences in the symbol sequence, where 'a' is the current symbol index and 'b' is the symbol index at the next time step. Iterate through t from 1 to K-1, if the current symbol... And the symbol of the next moment Then The value increases by 1. The symbol transition probability matrix is ​​obtained by row-by-row normalization of the counting matrix F: for the current symbol For the corresponding row a, calculate the sum of the column counts for that row. Where c is the index of the symbol at the next time step, then the element in the a-th row and b-th column of the symbol transition probability matrix is... The calculation method is as follows .when When, define .

[0030] Multiplying each element in the symbol transition probability matrix by the self-information of the symbol transition state corresponding to that element yields the weighted transition probability matrix. For a symbol transition from... Transfer to symbol The symbol transition state, its self-information is defined as ,when When, define To avoid undefined logarithmic operations, the element in the a-th row and b-th column of the weighted transition probability matrix. The calculation method is as follows .

[0031] Normalize each row of the weighted transition probability matrix to generate the symbolic transition entropy matrix. The element in the a-th row and b-th column of the symbolic transition entropy matrix... The calculation formula is as follows: ; in, The element values ​​corresponding to row index a and column index b in the symbolic transfer entropy matrix represent the values ​​at the current symbol level. Under the condition that the next time step symbol is Normalized transfer entropy; 'a' is the current symbol index. b is the symbol index for the next time step. S represents the number of symbol intervals, and S takes a value of 8. Let be the element in the a-th row and b-th column of the symbol transition probability matrix, representing the transition from symbol to symbol . Transfer to symbol The probability of; Let be the element in the a-th row and c-th column of the symbol transition probability matrix, where c is the summation index for traversing the symbols at the next time step; This represents a logarithmic operation with base 2. When At that time, it was stipulated If the denominator of a certain row Then all of the rows The value is 1 / S. The symbol transition entropy matrix H is calculated according to the above formula. The row index of the symbol transition entropy matrix is ​​the current symbol, and the column index is the symbol at the next time step.

[0032] In practice, the first topological invariant is used as the first dimension of the observed feature vector. The first topological invariant is the homology class feature value corresponding to the generating cycle with the maximum lifetime extracted from the persistence graph, represented as a vector composed of the appearance scale parameter values ​​and disappearance scale parameter values ​​of the q generating cycles arranged in order. q takes the value 3. Let the first topological invariant vector be denoted as... , The dimension is 2q=6.

[0033] The symbolic transfer entropy matrix is ​​expanded row-wise and concatenated to form a feature row vector, which serves as the second dimension of the observed feature vector. The symbolic transfer entropy matrix is ​​denoted as H, with dimensions S×S, where S is the number of symbol intervals, and S takes the value 8. The row-wise expansion is performed by expanding the first row of the symbolic transfer entropy matrix H... Arrange in sequence, then add the elements in the second row. Following the elements in the first row, and so on, until all elements in the S-th row are concatenated, resulting in a string of length . The feature row vector, denoted as . , The dimension is .

[0034] The first topological invariant vector The eigenrow of the sign transition entropy matrix By piecing together the beginning and end, a dimension is formed. The observed feature vector, denoted as The semicolon indicates a vector concatenation operation. During the composite stress aging test, an observation feature vector is generated in the above manner at each preset degradation state sampling time. The interval between degradation state sampling times is set to perform sampling once after each complete alternation cycle, generating an observation feature vector sequence. T represents the total number of sampling times for the degenerate state.

[0035] Observation feature vector sequence The input is fed into the sparse Bayesian learning framework. The core architecture of the sparse Bayesian learning framework is a correlation vector machine regression model, which consists of an input layer, a kernel function mapping layer, and a linear output layer. The input layer receives observation feature vectors with a dimension of 70. The kernel function mapping layer uses a Laplacian kernel for feature mapping. The Laplacian kernel is defined as follows: ,in and Given two observed feature vectors, This represents the sum of the absolute values ​​of the elements of the vector. Norm, For kernel width parameter, The values ​​are obtained from the set through cross-validation. The value that minimizes the leave-one-out verification error is selected. The linear output layer performs a weighted summation of the kernel function mapping results and outputs the degradation rate hidden variable. The predicted value. The mathematical form of the correlation vector machine regression model is: ,in The weight coefficients of the j-th related vector are... This is a bias term.

[0036] The training process of the correlation vector machine regression model is as follows: the observed feature vector sequence is... As input to the model, and simultaneously for each observed feature vector A target degradation rate value is set as the training label. The target degradation rate value is calculated from the relative change rate of the average peak drain current between adjacent sampling times, denoted as the average peak drain current over r consecutive switching cycles near the t-th sampling time. If r is taken as half of the total number of switching cycles within a complete alternating cycle of the switching cyclic stress, then the target degradation rate is defined as follows: The model is trained using a sparse Bayesian learning method, with each weight coefficient... Introduce an independent Gaussian prior distribution with zero mean and prior variance. Hyperparameters are defined and automatically determined from the training data. They are learned automatically by maximizing the marginal likelihood function on the training data. and noise variance noise variance The initial value is set to the degradation rate of the observed target. One-tenth of the sample variance. After training converges, most hyperparameters... Approaching infinity, corresponding to the weighting coefficient The posterior distribution is concentrated at zero, with only a few related vectors corresponding to non-zero weight coefficients, forming a sparse solution.

[0037] The trained Relevance Vector Machine (RVM) regression model receives an observed feature vector. As input, the output degradation rate is a hidden variable. Point estimates of degradation rate latent variables The expression is as follows: ; in, Let R be the latent variable representing the degradation rate at sampling time t; R is the set of relevance vector indices, containing the weight coefficients after training. The training sample index corresponding to the basis function whose posterior mean is not zero; Let be the posterior mean of the weight coefficients of the j-th correlation vector in the correlation vector machine regression model; This represents the observed feature vector at the current sampling time; This is the observed feature vector corresponding to the j-th correlation vector; Represents the observed feature vector With observed feature vector Between Norm, which is the sum of the absolute differences between corresponding elements of two vectors; The kernel width parameter of the Laplace nucleus. Cross-validation from the candidate value set Selected from; This represents the posterior mean of the bias term. The degradation rate latent variable sequence is obtained by calculating this at each sampling time point in the observed feature vector sequence in the manner described above. .

[0038] The degradation rate latent variable sequence is smoothed by introducing an adaptive Kalman filter. The state-space model of the adaptive Kalman filter consists of state equations and observation equations. The state equations are defined as follows: ,in The noise of the state transition process follows a zero-mean Gaussian distribution, and the noise variance is denoted as . The observation equation is defined as follows: ,in These are the latent variable observations of the degradation rate directly output by the relevant vector machine regression model. The observed noise follows a zero-mean Gaussian distribution, and the variance of the observed noise is denoted as . Process noise variance and observation noise variance All methods employ an adaptive estimation method based on innovation for dynamic updates: at the t-th filtering step, the innovation is calculated. ,in This is a single-step prediction of the state at time t based on the state at time t-1; the length is taken as... The sample variance of the news sequence within the sliding window is used as the observation noise variance. The estimated value, The value is 20; process noise variance The latent degradation rate is estimated using the novel correlation method. Smoothing is performed using a fixed-interval smoothing algorithm. After obtaining filtered estimates for all time steps through forward filtering, the smoothed degradation rate latent variable is obtained by recursively calculating from the last time step. .

[0039] The smoothed degradation rate hidden variable The values ​​are compared with a set of preset degradation state thresholds. The preset degradation state thresholds are a monotonically increasing numerical sequence. The degradation state threshold is preset to , , , The degradation state thresholds are set based on the following: when the degradation rate is below 0.05, the device is in the early stage of near-linear degradation; when the degradation rate exceeds 0.50, the device is about to fail; the two intermediate thresholds, 0.15 and 0.30, further subdivide the intermediate degradation stage into a slow-accelerated degradation stage and a fast-accelerated degradation stage. Based on these degradation state thresholds, the SiC power device is discretized into five degradation states. :when At that time, the current degradation state is ;when At that time, the current degradation state is ;when At that time, the current degradation state is ;when At that time, the current degradation state is ;when At that time, the current degradation state is ,in The default failure state is set. The degradation state at each moment is arranged in chronological order, and the degradation state number sequence is denoted as follows: , This sequence is the hidden state chain of the Hidden Markov Model.

[0040] The Viterbi algorithm is used to estimate the parameters of the hidden state chain. The parameters of the Hidden Markov Model include the initial state probability vector. The state transition probability matrix A and the state emission probability matrix B. The initial state probability vector. The i-th element This indicates that the degenerate state sequence begins at... The probability, The latent variable of the degradation rate after smoothing from the first sampling time. The indicator function for the corresponding degenerate state is determined by comparing it with the degenerate state threshold. The element in the i-th row and j-th column of the state transition probability matrix A... Indicates a state of degradation Transition to a degenerate state The probability, The initial value is set as follows: when j=i When j = i + 1 When j=i-1 Other cases ,in Furthermore, the boundary states are truncated so that the sum of the probabilities in each row is 1. The element in the i-th row and k-th column of the state emission probability matrix B... This indicates that when the hidden state is The probability of emitting the observed degenerate state number k at any time, the observed degenerate state number The Viterbi algorithm, given a hidden state chain observation sequence... and current parameter estimation Given conditions A and B, iteratively execute the E-step and M-step. The E-step calculates the posterior probability distribution of the hidden state at each time step using a forward-backward algorithm; the M-step updates the estimated values ​​of each element in parameter matrices A and B based on the posterior probability distribution, specifically by... The estimated value is equal to that from the degenerate state Transition to a degenerate state The expected frequency divided by the frequency from the degenerate state The total expected frequency transferred out The estimated value is equal to that in the degenerate state. The expected frequency of the observed degradation state number k is divided by the degradation state. The total expected frequency of occurrence. The convergence condition for the iteration is that the maximum absolute value of the change in each element of the state transition probability matrix A between two consecutive iterations is less than 10. -4 The combined state transition probability matrix A and state emission probability matrix B, after convergence of the combined iterations, constitute a joint state-space model.

[0041] In practical implementation, the state transition probability distribution corresponding to the current degenerate state is extracted from the state transition probability matrix of the joint state-space model. The state transition probability matrix of the joint state-space model is denoted as... , The dimensions are 5 rows and 5 columns, and the five degradation states are as follows: , , , , ,in This is the preset failure state. The current degradation state is based on the smoothed degradation rate latent variable at the time of the most recent degradation state sampling. The current degradation state is determined by comparing it with the degradation state threshold and recorded as follows. , Current degradation state In the state transition probability matrix The corresponding row index is denoted as s. The correspondence between row indexes and degradation states is as follows: Corresponding to line 1, Corresponding to line 2, Corresponding to line 3, Corresponding to line 4, Corresponding to line 5. From the state transition probability matrix. Extracting all column elements from the s-th row yields a row vector of length 5. This row vector represents the state transition probability distribution corresponding to the current degenerate state, denoted as . ,in Indicates the current degenerate state Transition to a degenerate state The probability, And satisfy .

[0042] According to the state transition probability distribution Monte Carlo random sampling is performed to generate multiple degradation paths from the current moment until a preset failure state is reached. The total number of Monte Carlo random sampling paths is set to [value missing]. , The value is 10000. The value of 10000 is chosen because, at a confidence level of 95%, the standard error of the percentiles obtained from 10000 samplings is less than 5% of the width of the remaining service life interval, meeting the accuracy requirements for engineered life prediction. The generation method for a single degradation path is as follows: [The text abruptly shifts to a different topic] As the starting state of the degradation path, initialize the degradation path state sequence. During each subsequent state transition, the current degenerate state is considered. Corresponding state transition probability distribution Generate a random number u, which follows a uniform distribution on the interval (0,1). Use the inverse transform sampling method to determine the degradation state at the next time step. The calculation formula for the inverse transform sampling method is: ; Where b represents the degenerate state at the next moment. The index; k is an intermediate variable for traversing the indices of the degenerate states; To transition from a degenerate state Transition to a degenerate state The probability is given by m, where m is the summation index, and u is a uniformly distributed random number drawn independently from the interval (0,1) during each state transition. The determined degenerate state is then... Append to the end of the degenerate path state sequence P, and update the current degenerate state to... Repeat the state transition steps until a preset failure state appears in the degradation path state sequence P. At this point, the generation of the degradation path ends. Each degradation path contains a series of degradation states that are passed through in sequence.

[0043] For each generated degradation path, calculate the degradation path from the current degradation state. To the preset failure state The number of time steps elapsed. The number of time steps is defined as the number of steps in the degenerate path state sequence P from the initial state. The next position begins until the first occurrence. The number of state transition steps so far, with each state transition counted as one time step. All The set of time steps for each degradation path is denoted as ,in Let be the number of time steps for the i-th degradation path.

[0044] The 100th percentile of the time step set L is taken as the lower limit of the remaining lifetime, and the 90th percentile of the time step set L is taken as the upper limit of the remaining lifetime. The quantiles are calculated by sorting all elements in the time step set L in ascending order of value, resulting in an ordered sequence. The index of the tenth percentile in the ordered sequence is... The value, the ninetieth percentile corresponds to the index in the ordered sequence of . The value of, among which This indicates rounding up. The tenth percentile is denoted as... Lower limit of remaining service life The number of steps in the degenerate state transition is used as the unit; the ninetieth percentile is denoted as Maximum remaining service life The output is measured in steps of degradation state transition. The output is determined by the lower limit of the remaining lifetime. and remaining service life limit The remaining service life range constituted .

[0045] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for predicting the lifetime of SiC power devices based on composite stress failure modes, characterized in that, Includes the following steps: Collect the transient response sequence of drain current and the case temperature fluctuation sequence of SiC power devices during composite stress aging test; The transient response sequence of the drain current is mapped to the phase space by chaotic attractor reconstruction, and the first topological invariant reflecting the failure evolution is extracted. The shell temperature fluctuation sequence is symbolized to construct a symbolic transfer entropy matrix for characterizing the thermal stress cycling mode; Based on the first topological invariant and the symbolic transfer entropy matrix, a joint state-space model describing the degradation process of SIC power devices is established. Based on the state transition probabilities in the joint state-space model, the remaining lifetime range of the SiC power device is predicted.

2. The method for predicting the lifetime of SiC power devices based on composite stress failure modes according to claim 1, characterized in that, The acquisition of the drain current transient response sequence and case temperature fluctuation sequence of the SiC power device during the composite stress aging test is specifically as follows: The composite stress aging test is constituted by simultaneously applying DC bias stress, switching cycle stress and high temperature storage stress to SiC power devices. During the composite stress aging test, the drain current waveform of the SIC power device is continuously acquired at a fixed sampling frequency within a preset time window after each switching action. The drain current waveform is processed to maintain the peak value of the drain current in each switching cycle. The drain current peak values ​​of multiple consecutive switching cycles are arranged in time sequence to generate the drain current transient response sequence. During the composite stress aging test, the case temperature values ​​of the SiC power device are continuously collected at the same sampling frequency. The case temperature values ​​are then subjected to sliding median filtering to remove measurement noise, thereby obtaining the case temperature fluctuation sequence.

3. The method for predicting the lifetime of SiC power devices based on composite stress failure modes according to claim 1, characterized in that, The process of mapping the drain current transient response sequence to phase space through chaotic attractor reconstruction and extracting the first topological invariant reflecting the failure evolution is specifically as follows: A delay embedding operation is performed on the drain current transient response sequence. The embedding dimension value and the delay time value are set to construct multiple phase point vectors. Each phase point vector consists of multiple data points selected from the drain current transient response sequence at intervals of the delay time value. The multiple phase point vectors are arranged in phase space to form a chaotic attractor trajectory. A persistent cohomology analysis was performed on the chaotic attractor trajectory to calculate the persistence plot of the Betti number as a function of the scale parameter in different dimensions. Extract the homology class feature values ​​corresponding to all generating cycles with the maximum lifetime from the persistence graph, and use them as the first topological invariant.

4. The method for predicting the lifetime of SiC power devices based on composite stress failure modes according to claim 3, characterized in that, The embedding dimension value and the delay time value are determined jointly using the false nearest neighbor method and the mutual information method.

5. The method for predicting the lifetime of SiC power devices based on composite stress failure modes according to claim 1, characterized in that, The shell temperature fluctuation sequence is symbolized to construct a symbolic transfer entropy matrix to characterize the thermal stress cycling mode, specifically: The shell temperature fluctuation sequence is divided into multiple symbol intervals according to the amplitude range. Each symbol interval is assigned a discrete symbol. Each shell temperature value in the shell temperature fluctuation sequence is replaced with the discrete symbol corresponding to its symbol interval to obtain a symbol sequence. The combination of two adjacent symbols in the symbol sequence is defined as a symbol transition state. The frequency of occurrence of all possible symbol transition states is counted to construct a symbol transition probability matrix. Each element in the symbol transition probability matrix is ​​multiplied by the self-information of the symbol transition state corresponding to that element to obtain a weighted transition probability matrix. Then, each row of the weighted transition probability matrix is ​​normalized to generate the symbol transition entropy matrix, where the row index of the symbol transition entropy matrix is ​​the current symbol and the column index is the symbol at the next time step.

6. The method for predicting the lifetime of SiC power devices based on composite stress failure modes according to claim 5, characterized in that, When the shell temperature fluctuation sequence is divided into multiple symbol intervals according to the amplitude range, the boundaries of each interval are dynamically determined using the equal probability quantile method.

7. The method for predicting the lifetime of SiC power devices based on composite stress failure modes according to claim 1, characterized in that, Based on the first topological invariant and the symbolic transfer entropy matrix, a joint state-space model describing the degradation process of the SIC power device is established, specifically as follows: The first topological invariant is used as the first dimension of the observed feature vector, and all matrix elements of the symbolic transfer entropy matrix are expanded row by row and concatenated to form a feature row vector, which is used as the second dimension of the observed feature vector. The observed feature vectors are input into a sparse Bayesian learning framework, the kernel function is defined as the Laplace kernel, and the degradation rate latent variable corresponding to each observed feature vector is calculated by the correlation vector machine regression algorithm. The degradation rate hidden variable is compared with multiple preset degradation state thresholds, the SIC power device at each time moment is discretized into one of multiple degradation states, and the degradation state sequence on the continuous time axis is used as the hidden state chain of the hidden Markov model. The hidden state chain is parameterized using the Viterbi algorithm to obtain the state transition probability matrix and the state emission probability matrix. The state transition probability matrix and the state emission probability matrix are combined to form the joint state space model.

8. The method for predicting the lifetime of SiC power devices based on composite stress failure modes according to claim 7, characterized in that, The latent variable of degradation rate is smoothed by introducing an adaptive Kalman filter and then compared with multiple preset degradation state thresholds.

9. The method for predicting the lifetime of SiC power devices based on composite stress failure modes according to claim 1, characterized in that, Based on the state transition probabilities in the joint state-space model, the remaining lifetime of the SiC power device is predicted, specifically as follows: Extract the state transition probability distribution corresponding to the current degenerate state from the state transition probability matrix of the joint state space model. The state transition probability distribution contains the probability values ​​of transitioning from the current degenerate state to all other degenerate states. Monte Carlo random sampling is performed based on the state transition probability distribution to generate multiple degradation paths from the current moment until a preset failure state is reached. Each degradation path contains a series of degradation states passed through in sequence. Calculate the number of time steps for each degradation path from the current degradation state to the preset failure state, and form a time step set for all degradation paths; Take the tenth percentile of the set of time steps as the lower limit of the remaining lifespan, and take the ninetieth percentile of the set of time steps as the upper limit of the remaining lifespan. Output the interval formed by the lower limit of the remaining lifespan and the upper limit of the remaining lifespan.

10. The method for predicting the lifetime of SiC power devices based on composite stress failure modes according to claim 1, characterized in that, In the composite stress aging test, DC bias stress, switching cycle stress and high temperature storage stress are applied in a preset time alternation pattern, and the switching frequency of the switching cycle stress increases stepwise with the aging time.

Images