Sic device life prediction method and system based on multi-source information fusion

Abstract

This invention discloses a method and system for predicting the lifetime of SiC devices based on multi-source information fusion, belonging to the field of power electronic equipment health management technology. It addresses the problem of low lifetime prediction accuracy under variable temperature conditions. First, quasi-static parameters and high-frequency oscillation waveforms are simultaneously acquired. The spectral characteristics are spatially mapped and converted using real-time junction temperature to construct a temperature-insensitive package aging index. Then, the threshold voltage recovery trajectory is analyzed to filter out fast-changing charge interference and lock the slow-changing oxide layer damage component to characterize the chip's intrinsic failure. Finally, the state deviation is calculated by combining physical prior trends and multi-source measured posterior correction factors, dynamically narrowing the prediction interval and outputting the remaining lifetime. This invention effectively achieves thermal effect decoupling and true / false aging identification, significantly improving the robustness and accuracy of the prediction.

Description

Technical Field

[0001] This invention relates to the field of power electronic equipment health management technology, specifically to a method and system for predicting the lifespan of SiC devices based on multi-source information fusion. Background Technology

[0002] With the rapid development of new energy vehicles, photovoltaic power generation, and smart grids, silicon carbide power devices, due to their superior material properties such as high voltage, high frequency, and high temperature resistance, are gradually replacing traditional silicon-based devices in core power conversion systems. These devices typically operate under extreme electrothermal stress environments, and their operational reliability directly affects the safety and stability of the entire system. Sudden device failure during operation can lead to system shutdown or even serious safety accidents. Therefore, accurate health monitoring and remaining life prediction of silicon carbide devices, without affecting normal equipment operation, and realizing a shift from reactive maintenance to condition-based maintenance, has become a critical need that urgently needs to be addressed in the industry.

[0003] Existing lifetime prediction methods are mainly divided into single-parameter monitoring methods and pure physical model methods, but both have substantial drawbacks. Single-parameter monitoring methods typically select on-resistance or threshold voltage as aging indicators. However, in actual variable operating conditions, these electrical parameters are highly susceptible to junction temperature fluctuations, making it difficult to distinguish whether parameter drift is caused by temperature changes or device aging, leading to misjudgments in monitoring results. While pure physical model methods have a clear theoretical basis, they rely on idealized failure mechanism assumptions and cannot adapt to the discreteness of actual manufacturing processes and the dynamic complexity of operating conditions. Furthermore, existing technologies generally ignore the bias temperature instability effect unique to silicon carbide devices, meaning that devices will experience recoverable parameter drift under electrical stress. This reversible drift is often misjudged as permanent failure, severely affecting the accuracy of remaining lifetime prediction. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for predicting the lifetime of SiC devices based on multi-source information fusion, thus solving the problems mentioned in the background.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a SiC device lifetime prediction method based on multi-source information fusion, comprising the following steps: S1. Acquiring quasi-static operating parameters during the operation of the SiC device, and simultaneously locking the high-frequency oscillation range during the device turn-off process using a high-frequency sampling circuit, capturing the underdamped oscillation waveform containing parasitic inductance degradation information, and synchronously latching the body diode voltage drop across the source and drain terminals during the dead time; S2. Calculating the real-time junction temperature of the device based on the body diode voltage drop, using a preset reference temperature as the target, spatially mapping and converting the spectral features extracted from the underdamped oscillation waveform using the real-time junction temperature, constructing a feature drift vector of the current state point relative to the healthy reference point, and calculating the magnitude of the feature drift vector as the temperature at which the device package degradation degree is quantified. S3. Monitor the drift recovery trajectory of the threshold voltage during the device's non-operating intervals. Deconstruct the recovery trajectory using a multi-scale time constant separation algorithm, filter out the interface state charge trapping component dominated by the fast time constant, and lock the oxide layer irreversible damage component dominated by the slow time constant as the cumulative damage vector characterizing the intrinsic failure of the chip; S4. Calculate the theoretical thermal stress accumulation degree using quasi-static operating parameters to establish the a priori trend of device lifetime degradation. Use the temperature-insensitive aging index and the cumulative damage vector as posterior correction factors reflecting the actual degradation degree. Calculate the state deviation between the posterior correction factor and the a priori trend. Dynamically shrink and correct the remaining lifetime prediction range of the device using the state deviation, and output the remaining lifetime prediction value with converged confidence.

[0006] Further, step S1 includes the following steps: configuring a high-frequency analog comparator circuit to monitor the gate-source voltage signal in real time, and marking the inflection point when the falling edge drops to the Miller plateau region as the turn-off oscillation trigger anchor point; after the trigger anchor point, starting a high-bandwidth acquisition channel to record the entire process of underdamped oscillation decay caused by the resonance of the package parasitic inductance and junction capacitance in the gate-source circuit until the oscillation amplitude converges to the background noise baseline; logic locking the dead time window of the drive signal, and after the turn-off transient ends and the anti-parallel diode enters the steady-state freewheeling stage, opening the differential sampling channel to capture the forward conduction voltage drop across the source and drain terminals, synchronously recording the quasi-static load current corresponding to the sampling time, and generating a timestamp-aligned multidimensional raw dataset.

[0007] Furthermore, based on the real-time junction temperature of the device calculated from the body diode voltage drop, and with a preset reference temperature as the target, the specific process of spatial mapping and conversion of the spectral features extracted from the underdamped oscillation waveform using the real-time junction temperature is as follows: Based on the wide bandgap thermistor characteristics of SiC, the multi-dimensional electrothermal coupling calibration relationship between the pre-stored body diode forward voltage drop, load current, and junction temperature is used to substitute the body diode voltage drop and synchronous load current into the calibration relationship to inversely solve for the real-time junction temperature inside the device; the time-frequency domain joint transformation of the underdamped oscillation waveform is performed to extract the oscillation natural frequency characterizing the circuit inductance attribute and the oscillation attenuation coefficient characterizing the circuit resistance attribute; a temperature sensitivity compensation matrix is ​​introduced to quantify the inherent drift rate of the oscillation natural frequency and the oscillation attenuation coefficient as a function of temperature; using the temperature difference between the real-time junction temperature and the preset reference temperature, the natural frequency and attenuation coefficient under high-temperature conditions are inversely mapped back to the reference temperature coordinate system through the compensation matrix to obtain the equivalent spectral features that eliminate thermally induced drift components.

[0008] Furthermore, the specific process of constructing the characteristic drift vector of the current state point relative to the health reference point and calculating the magnitude of the characteristic drift vector as a temperature-insensitive aging index for quantifying the degree of device package degradation is as follows: Construct a two-dimensional package impedance characteristic plane with the natural oscillation frequency as the horizontal axis and the oscillation attenuation coefficient as the vertical axis, and calibrate the characteristic coordinates of the device in its factory healthy state and at the reference temperature as the origin reference coordinates; use the equivalent spectrum characteristics as the current aging state coordinates, and establish a directed degradation vector on the two-dimensional package impedance characteristic plane pointing from the origin reference coordinates to the current aging state coordinates; calculate the geometric magnitude of the directed degradation vector using the Euclidean distance algorithm, and define the geometric magnitude as a package health degradation index that reflects the combined effect of inductance changes caused by bond wire peeling and resistance changes caused by solder layer fatigue. When the magnitude value increases monotonically with running time, it is determined that the physical impedance characteristics of the package structure have undergone irreversible degradation.

[0009] Furthermore, the drift recovery trajectory of the threshold voltage is monitored during the device's non-operating interval. The recovery trajectory is deconstructed using a multi-scale time constant separation algorithm to filter out the interface state charge trapping component dominated by the fast time constant. The specific process is as follows: A weak constant measurement current is applied through the natural cooling relaxation window after the device is turned off, and a bias scanning signal is applied to the gate. The relaxation decay curve of the threshold voltage changing with time is continuously recorded. The relaxation decay curve is analyzed by multi-exponential fitting using a hybrid decay mathematical model that includes the interface state electron emission mechanism and the oxide layer trap tunneling mechanism. The fast-changing component corresponding to the short relaxation time in the model is identified and determined to be a recoverable drift caused by the escape of electrons trapped by shallow energy level traps at the SiC / SiO2 interface. The fast-changing component is numerically separated from the total recovery trajectory, and the remaining residual trajectory is retained.

[0010] Furthermore, the specific process of locking the irreversible damage component of the oxide layer dominated by the slow time constant as the cumulative damage vector characterizing the intrinsic failure of the chip is as follows: perform secondary analysis on the remaining residual trajectory to extract the slow variable component corresponding to the long relaxation time, and determine it as an irreversible drift caused by deep energy level defects in the near-interface oxide layer or permanent lattice damage in the gate oxide layer; calculate the residual amplitude of the slow variable component in steady state and its long-term growth rate over service time; combine the residual amplitude and the long-term growth rate to generate a two-dimensional vector, which is defined as the cumulative damage vector characterizing the degradation of the insulation performance of the chip's gate oxide layer, and is used to quantify the intrinsic physical damage inside the chip.

[0011] Furthermore, by calculating the theoretical thermal stress accumulation degree through quasi-static operating condition parameters, a priori benchmark trend for device lifetime degradation is established. Temperature-insensitive aging index and cumulative damage vector are used as posterior correction factors reflecting the actual degradation degree. The specific process of calculating the state deviation between the posterior correction factor and the priori benchmark trend is as follows: Based on the quasi-static operating condition parameters, the thermal cycling load spectrum is extracted, and the theoretical cumulative fatigue damage degree of the device is calculated by inputting a preset fatigue failure model, generating a priori benchmark trend curve that monotonically increases with time; the package health degradation index and cumulative damage vector are normalized and fused and mapped to the same health metric space as the priori benchmark trend curve to form a multidimensional measured state point; the weighted Euclidean distance between the multidimensional measured state point and the priori benchmark trend curve at the current moment is calculated, and the distance is defined as the state deviation characterizing the inconsistency between the physical model prediction and the actual device state.

[0012] Furthermore, the specific process of dynamically shrinking and correcting the remaining lifetime prediction interval of the device through state deviation, and outputting the remaining lifetime prediction value with converged confidence, is as follows: Construct a state observer based on particle filtering, and use state deviation as a likelihood evaluation index to update the normalized weights of the particle swarm; when the state deviation is less than a preset deviation threshold, it is determined that the degradation state is within the coverage of the prior model, the weights of particles located in the high likelihood region of the prior baseline trend are increased, and the variance of the posterior probability density distribution is reduced to narrow the confidence interval; when the state deviation exceeds the preset deviation threshold, it is determined that a nonlinear abrupt degradation has occurred, triggering the particle resampling mechanism to migrate the particle swarm towards the multidimensional measured state points, reconstructing the mathematical expectation of the remaining lifetime prediction value based on the measured state points and adjusting the prediction interval width to cover the uncertainty, and outputting the remaining lifetime probability distribution after converged iteration.

[0013] The SiC device lifetime prediction system based on multi-source information fusion includes the following modules: a multi-source acquisition module, used to acquire quasi-static operating parameters during SiC device operation, and simultaneously lock the high-frequency oscillation range during device turn-off through a high-frequency sampling circuit, capturing the underdamped oscillation waveform containing parasitic inductance degradation information, and synchronously latching the body diode voltage drop across the source and drain during the dead time; a feature decoupling module, used to calculate the real-time junction temperature of the device based on the body diode voltage drop, and using a preset reference temperature as the target, spatially mapping and converting the spectral features extracted from the underdamped oscillation waveform through the real-time junction temperature to construct the feature drift vector of the current state point relative to the health reference point, and calculating the magnitude of the feature drift vector as a temperature-insensitive aging index for quantifying the degree of device package degradation; and damage separation. The first module monitors the drift recovery trajectory of the threshold voltage during device non-operational intervals. It deconstructs the recovery trajectory using a multi-scale time constant separation algorithm, filtering out the interface state charge trapping component dominated by the fast time constant and locking in the oxide layer irreversible damage component dominated by the slow time constant, which serves as the cumulative damage vector characterizing intrinsic chip failure. The second module calculates the theoretical thermal stress accumulation using quasi-static operating parameters, establishing a priori trend for device lifetime degradation. It uses temperature-insensitive aging indicators and the cumulative damage vector as posterior correction factors reflecting the actual degradation level, calculating the state deviation between the posterior correction factor and the priori trend. This state deviation dynamically shrinks and corrects the device's remaining lifetime prediction range, outputting a confidently converged remaining lifetime prediction value.

[0014] The present invention has the following beneficial effects:

[0015] (1) A multi-source information fusion-based method for predicting the lifetime of SiC devices achieves accurate extraction and decoupling of device aging characteristics under variable temperature conditions. By locking the turn-off oscillation interval through high-frequency sampling and simultaneously acquiring the body diode voltage drop, the method can simultaneously capture transient characteristics reflecting the package state and static parameters reflecting the thermal state. More importantly, by using real-time junction temperature to perform spatial mapping and conversion of spectral characteristics, the interference of temperature fluctuations on the oscillation frequency and attenuation coefficient is effectively eliminated, and a temperature-insensitive aging index is constructed. This enables the system to accurately distinguish the real physical degradation caused by bond wire breakage or solder layer fatigue in complex dynamic temperature environments, avoid misjudgments caused by temperature effects, and significantly improve the robustness of aging monitoring.

[0016] (2) The SIC device lifetime prediction system based on multi-source information fusion improves the confidence and accuracy of remaining lifetime prediction. By analyzing the recovery trajectory of the threshold voltage during non-operating intervals, the method innovatively separates the fast time constant reversible drift caused by interface state charge trapping from the slow time constant irreversible drift caused by oxide layer damage, thereby accurately locking the intrinsic failure of the chip and eliminating the spurious aging interference unique to silicon carbide devices. At the same time, combined with the Bayesian inference framework, the prior trend of the physical model is combined with the posterior correction of the measured data, and the prediction interval is dynamically narrowed by utilizing state bias. This dual-driven mechanism of physics and data not only solves the problem of poor generalization ability of a single model, but also outputs convergent and reliable remaining lifetime prediction values ​​under small sample data.

[0017] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0018] Figure 1 This is a flowchart of the SiC device lifetime prediction method based on multi-source information fusion according to the present invention.

[0019] Figure 2 This is a flowchart of the SiC device lifetime prediction system based on multi-source information fusion according to the present invention. Detailed Implementation

[0020] This application's embodiments address the problems of low lifetime prediction accuracy and high false alarm rate in existing technologies, which cannot distinguish between temperature interference and actual aging under complex temperature variations, and are unable to eliminate the reversible parameter drift unique to silicon carbide devices, through a multi-source information fusion-based method and system for predicting the lifetime of SiC devices.

[0021] The overall concept of the solution in this application embodiment is as follows:

[0022] First, operating parameters are collected at the macroscopic level, and high-frequency oscillation waveforms containing package parasitic parameter information are captured at the microscopic level. Real-time temperature is obtained using the body diode voltage drop. Second, transient characteristics severely affected by temperature are mapped back to the reference temperature space using physical laws, removing thermal noise and extracting pure package aging indicators. Third, the unique relaxation characteristics of the device are used to separate recoverable charge trap effects and unrecoverable lattice damage in the time dimension, extracting pure chip damage vectors. Finally, the theoretical lifetime trend calculated based on the physical fatigue model is used as prior knowledge, and the extracted dual aging characteristics of the package and chip are used as experimental evidence. Probabilistic statistical methods are used to dynamically correct and narrow the prediction results, ultimately outputting a high-confidence remaining lifetime prediction result.

[0023] Please see Figure 1This invention provides a technical solution: a SiC device lifetime prediction method based on multi-source information fusion, comprising the following steps: S1. Acquiring quasi-static operating parameters during SiC device operation, and simultaneously locking the high-frequency oscillation range during device turn-off using a high-frequency sampling circuit, capturing the underdamped oscillation waveform containing parasitic inductance degradation information, and synchronously latching the body diode voltage drop across the source and drain during the dead time; S2. Calculating the real-time junction temperature of the device based on the body diode voltage drop, using a preset reference temperature as the target, spatially mapping and converting the spectral features extracted from the underdamped oscillation waveform using the real-time junction temperature, constructing a feature drift vector of the current state point relative to the healthy reference point, and calculating the magnitude of the feature drift vector as a temperature-insensitive aging quantifier of the device package degradation degree. S3. Monitor the drift recovery trajectory of the threshold voltage during the device's non-operating interval. Deconstruct the recovery trajectory using a multi-scale time constant separation algorithm, filter out the interface state charge trapping component dominated by the fast time constant, and lock the oxide layer irreversible damage component dominated by the slow time constant as the cumulative damage vector characterizing the intrinsic failure of the chip; S4. Calculate the theoretical thermal stress accumulation degree through quasi-static operating parameters, establish the a priori trend of device lifetime degradation, use the temperature-insensitive aging index and the cumulative damage vector as the posterior correction factor reflecting the actual degradation degree, calculate the state deviation between the posterior correction factor and the a priori trend, and dynamically shrink and correct the remaining lifetime prediction range of the device through the state deviation, outputting the remaining lifetime prediction value with converged confidence.

[0024] In this implementation scheme, step S1 is mainly used to establish a multi-dimensional sensing dataset reflecting the comprehensive health status of the device, aiming to solve the problem of insufficient information from a single data source. This step, while collecting slowly changing quasi-static operating condition parameters such as load current, uses high-frequency sampling technology to capture the underdamped oscillation waveform at the moment of device turn-off. The underdamped oscillation waveform refers to the decaying voltage waveform generated by LC resonance between the parasitic inductance and junction capacitance inside the package at the moment of switching. Its frequency and decay rate directly correspond to the physical connection state between the bond wire and the solder layer. Furthermore, the dead time mentioned in the step refers to the small time window during which the upper and lower diodes are simultaneously turned off to prevent bridge arm shoot-through. During this period, the voltage drop of the body diode is collected using the freewheeling characteristics of the anti-parallel diode, obtaining pure temperature-sensitive electrical parameters unaffected by switching noise. This process achieves simultaneous acquisition from macroscopic operating conditions to microscopic transient characteristics, providing a complete data foundation for subsequent decoupling analysis. The core function of step S2 is to perform temperature decoupling of physical characteristics, aiming to eliminate the interference of ambient temperature fluctuations on aging monitoring and prevent misjudgments caused by temperature rise. This step first calculates the real-time junction temperature inside the device using the body diode voltage drop. Then, it uses this temperature data to perform spatial mapping and conversion of the spectral characteristics of the oscillation waveform. That is, through mathematical transformation, the frequency and damping parameters measured at high temperature are projected back to the coordinate space at a standard reference temperature. Based on this, a feature drift vector is constructed. This vector, in a two-dimensional or multi-dimensional feature space, describes the direction and distance of the current state point from the initial healthy state point, with its magnitude representing the degree of deviation. By calculating this magnitude as a temperature-insensitive aging index, the effective separation of thermally induced parameter drift and physical aging parameter drift is technically achieved, ensuring that the monitored index only reflects the true degradation of the package structure. Step S3 mainly targets the bias temperature instability phenomenon unique to silicon carbide devices to identify true and false aging, aiming to eliminate recoverable false aging characteristics. The drift recovery trajectory monitored in this step refers to the curve of the threshold voltage gradually recovering to its initial value over time after the device stops operating. Through a multi-scale time constant separation algorithm, that is, utilizing the difference in time response speed of different physical mechanisms, the recovery trajectory is decomposed into two parts: fast change and slow change. The fast time constant component corresponds to interface state charge trapping, a reversible physical process in which electrons are temporarily trapped and then escape at the SiC-oxide interface, and is considered system noise that needs to be filtered out. The slow time constant component corresponds to irreversible oxide damage, representing permanent lattice defects within the gate oxide layer. Locking this component as the cumulative damage vector allows for precise quantification of the chip's intrinsic failure level, significantly improving the prediction specificity. Step S4 achieves a deep fusion of physical mechanisms and data-driven approaches, aiming to output high-confidence remaining lifetime prediction results.This step first uses quasi-static operating parameters combined with a fatigue model to calculate the theoretical thermal stress accumulation, thereby establishing a priori trend for device lifetime degradation—that is, the theoretical aging path derived from physical laws. Subsequently, the temperature-insensitive aging index and cumulative damage vector extracted in steps S2 and S3 are used as posterior correction factors, representing measured evidence reflecting the device's current actual health status. By calculating the state deviation between the two, and using this deviation to dynamically shrink and correct the prediction interval, the probability distribution range of the lifetime prediction results is adaptively adjusted based on the degree of agreement between measured data and theoretical expectations. This process ensures that the prediction results conform to physical degradation laws while using measured data to correct errors caused by individual differences and sudden degradation, achieving convergence in prediction accuracy.

[0025] Specifically, step S1 includes the following steps: configuring a high-frequency analog comparator circuit to monitor the gate-source voltage signal in real time, and marking the inflection point when the falling edge drops to the Miller plateau region as the turn-off oscillation trigger anchor point; after the trigger anchor point, starting a high-bandwidth acquisition channel to record the entire process of underdamped oscillation decay caused by the resonance of the package parasitic inductance and junction capacitance in the gate-source circuit until the oscillation amplitude converges to the background noise baseline; locking the dead time window of the drive signal with logic, and after the turn-off transient ends and the anti-parallel diode enters the steady-state freewheeling stage, opening the differential sampling channel to capture the forward conduction voltage drop across the source and drain terminals, synchronously recording the quasi-static load current corresponding to the sampling time, and generating a timestamp-aligned multidimensional raw dataset.

[0026] In this implementation scheme, step S1 achieves synchronous capture of the macroscopic and microscopic characteristics of the SiC device through high-precision timing control. First, a high-frequency analog comparator circuit is used to monitor the dynamic changes of the gate-source voltage. The inflection point when the falling edge drops to the Miller plateau region is selected as the trigger anchor point. This is because the Miller plateau effect corresponds to the stage where the drain-source voltage changes drastically during the device's turn-on or turn-off process. At this time, the gate-source voltage is relatively stable, providing extremely high time synchronization accuracy and avoiding false triggering caused by switching noise. The entire underdamped oscillation process recorded after the trigger anchor point is physically the natural response of a second-order LC oscillation circuit composed of stray inductance in the circuit and parasitic capacitance of the device under turn-off impact. This waveform directly carries the fingerprint information of the bond wire connection status and package insulation characteristics. Subsequently, the dead time window of the logic lock-in drive signal, i.e. the tiny time period during which the upper and lower transistors are simultaneously turned off, is sampled using the freewheeling characteristics of the anti-parallel diode. This effectively avoids high-frequency interference from switching transients and ensures the purity and alignment of the forward conduction voltage drop and quasi-static load current data across the source and drain terminals, providing accurate raw data support for subsequent decoupling analysis based on multiphysics.

[0027] Specifically, the process of spatially mapping and converting the spectral features extracted from the underdamped oscillation waveform using the real-time junction temperature is as follows: Based on the wide bandgap thermistor characteristics of SiC, the multi-dimensional electrothermal coupling calibration relationship between the pre-stored forward voltage drop of the body diode, load current, and junction temperature is used to substitute the body diode voltage drop and synchronous load current into the calibration relationship to solve for the real-time junction temperature inside the device; the time-frequency domain joint transformation of the underdamped oscillation waveform is performed to extract the oscillation natural frequency characterizing the inductance attribute of the circuit and the oscillation attenuation coefficient characterizing the resistance attribute of the circuit; a temperature sensitivity compensation matrix is ​​introduced to quantify the inherent drift rate of the oscillation natural frequency and the oscillation attenuation coefficient as a function of temperature; using the temperature difference between the real-time junction temperature and the preset reference temperature, the natural frequency and attenuation coefficient under high-temperature conditions are inversely mapped back to the reference temperature coordinate system through the compensation matrix to obtain the equivalent spectral features that eliminate thermally induced drift components.

[0028] In this implementation scheme, the core step in eliminating thermal interference is to inversely calculate the junction temperature based on the body diode voltage drop and perform spatial mapping conversion of the spectral characteristics. First, based on the inherent temperature-sensitive characteristics of SiC wide-bandgap semiconductor materials, there is a nonlinear coupling relationship between the body diode forward voltage drop, load current, and junction temperature. To accurately inversely calculate the real-time junction temperature inside the device, the following polynomial surface fitting model is used: In the formula, The calculated real-time junction temperature inside the device; to : The multidimensional electrothermal coupling coefficient determined in advance through a constant temperature oil bath calibration experiment; The forward voltage drop of the body diode was collected. The quasi-static load current is recorded synchronously. After obtaining the real-time junction temperature, a temperature sensitivity compensation matrix is ​​introduced to eliminate the inherent influence of temperature on high-frequency oscillation characteristics, thus inversely restoring the features extracted under high-temperature conditions to the reference temperature. The formula for calculating the equivalent spectral characteristics is as follows: In the formula, The equivalent natural frequency of the oscillation after eliminating the thermally induced drift component; The equivalent oscillation damping coefficient after eliminating thermally induced drift components; The measured natural frequency of the oscillation is extracted from the underdamped oscillation waveform using the fast Fourier transform. The measured oscillation attenuation coefficient was extracted using the envelope fitting method; : The linear drift rate of oscillation frequency as a function of temperature; : The linear drift rate of the attenuation coefficient as a function of temperature; The preset reference temperature is typically set to 25 degrees Celsius. The linear drift rate is also specified. and The method for determining the slope is as follows: offline testing of the new device is conducted across the entire temperature range, eigenvalues ​​are recorded at different temperatures, and linear regression analysis is performed to obtain the slope. This step ensures that subsequent degradation analysis focuses only on physical changes in the package structure itself, rather than parameter shifts caused by temperature fluctuations.

[0029] Specifically, the process of constructing a feature drift vector of the current state point relative to the health reference point and calculating the magnitude of the feature drift vector as a temperature-insensitive aging index for quantifying the degree of device package degradation is as follows: A two-dimensional package impedance characteristic plane is constructed with the natural oscillation frequency as the horizontal axis and the oscillation attenuation coefficient as the vertical axis. The characteristic coordinates of the device in its factory healthy state and at the reference temperature are calibrated as the origin reference coordinates. The equivalent spectral characteristics are used as the coordinates of the current aging state. A directed degradation vector is established on the two-dimensional package impedance characteristic plane, pointing from the origin reference coordinates to the current aging state coordinates. The geometric magnitude of the directed degradation vector is calculated using the Euclidean distance algorithm. The geometric magnitude is defined as a package health degradation index reflecting the combined effect of inductance changes caused by bond wire stripping and resistance changes caused by solder layer fatigue. When the magnitude value monotonically increases with operating time, it is determined that the physical impedance characteristics of the package structure have undergone irreversible degradation.

[0030] In this implementation scheme, a feature drift vector is constructed and an aging index is calculated to reduce the multi-dimensional and complex degradation information into a single quantifiable value, facilitating a more intuitive assessment of package health. The constructed two-dimensional package impedance characteristic plane unifies changes in inductance and resistance properties within the same geometric space, where the feature drift vector characterizes the direction and distance of deviation from the initial factory state. To comprehensively reflect the synergistic effect of bond wire peeling and solder layer fatigue, a weighted Euclidean distance algorithm is used to calculate a temperature-insensitive aging index, the formula of which is as follows: In the formula, Temperature-insensitive aging index for quantifying the degree of degradation of device packaging; : Weighting coefficients for changes in oscillation frequency, used to adjust the sensitivity to parasitic inductance degradation; The initial oscillation frequency of the device when it is in a healthy state at the factory and at the reference temperature; : A weighted coefficient for changes in the attenuation coefficient, used to adjust the sensitivity to loop resistance degradation; : The initial oscillation attenuation coefficient of the device under factory healthy conditions and at the reference temperature. Weighting coefficients in the formula. and The method for determining the weight is as follows: Based on accelerated aging test data, principal component analysis (PCA) is used to calculate the contribution ratio of the frequency change rate and the decay change rate to the total degradation variance, and these values ​​are normalized and used as weight values. This index, through the calculation of the modulus, physically represents the irreversible migration of the device package parasitic impedance in the characteristic space. When this value increases monotonically, it intuitively reflects the cumulative process of physical damage such as the decrease in inductance caused by broken bond wires inside the package or the increase in resistance caused by voids in the solder layer.

[0031] Specifically, the process of monitoring the drift recovery trajectory of the threshold voltage during the device's non-operational interval, and deconstructing the recovery trajectory using a multi-scale time constant separation algorithm to filter out the interface state charge trapping component dominated by the fast time constant is as follows: A weak, constant measurement current is applied through the natural cooling relaxation window after device turn-off, and a bias scanning signal is applied to the gate to continuously record the relaxation decay curve of the threshold voltage over time; a hybrid decay mathematical model incorporating the interface state electron emission mechanism and the oxide layer trap tunneling mechanism is used to perform multi-exponential fitting analysis on the relaxation decay curve, identifying the fast-changing component corresponding to the short relaxation time in the model, which is determined to be a recoverable drift caused by the escape of electrons trapped in shallow energy level traps at the SiC / SiO2 interface. The fast-changing component is numerically separated from the total recovery trajectory, and the remaining residual trajectory is retained.

[0032] In this implementation scheme, monitoring is performed through the natural cooling relaxation window after the device is turned off. This utilizes the physical characteristic of silicon carbide MOSFETs where trapped carriers are gradually released from energy level traps after the gate bias is cut off. It should be noted that the device's non-operating gap or natural cooling relaxation window does not refer to the microsecond-level switching off time of the device during high-frequency PWM pulse width modulation, but rather to the system-level macroscopic standby, shutdown, or sleep periods that occur in the actual operation of power electronic equipment. For example, in the application of electric drive systems for new energy vehicles, this non-operating gap can be the second-level shutdown when the vehicle is waiting at a traffic light, the parking standby stage, the system self-test period before charging, or the natural cooling stage after the vehicle is powered off; in the application of photovoltaic inverters, this gap can be the inverter's sleep period at night when there is no sunlight or the standby window when switching to grid-connected status. Within these system-level macroscopic gaps, the main power output of the SiC device is paused, thus providing a sufficient time window (covering milliseconds to several seconds or even several minutes) for the lifetime prediction system, fully meeting the requirements for extracting slow time constants. After establishing the aforementioned macroscopic non-operating gap, a weak, constant measurement current is applied to accurately read the turn-on threshold voltage without inducing device self-heating. The continuously recorded relaxation decay curves visually reflect the dynamic recovery process of the device's internal charge state over time. To accurately distinguish drift caused by different physical mechanisms, a hybrid decay mathematical model is used to analyze the curves. The core of this step lies in using mathematical fitting to decouple macroscopic voltage changes into microscopic physical processes. The specific fitting analytical model is as follows: In the formula, : The theoretical curve of the threshold voltage changing over time obtained by fitting; : The voltage drift amplitude corresponding to the fast-changing component, characterizing the intensity of charge trapping at the interface state; The fast time constant of charge emission from interface states, typically in the range of milliseconds to seconds, reflects the electron escape velocity from shallow level traps. : The voltage drift amplitude corresponding to the slowly varying component, characterizing the trapping strength of the oxide layer trap; The slow time constant of oxide trap tunneling is usually much larger than the fast time constant, reflecting the charge retention capability of deep-level defects. : The steady-state threshold voltage after the device has fully recovered. Wherein, for fast time constants... With slow time constant The method for determining the discrimination threshold is as follows: The characteristic frequency response of the SiC / SiO2 interface is extracted through broadband dielectric spectroscopy, and the upper limit of the characteristic time constant corresponding to the interface state density is set as the boundary threshold. After fitting, the calculation process for numerically separating the rapidly changing components from the total recovery trajectory aims to filter out noise; the calculation formula is as follows. , here The residual trajectory obtained from the measured voltage data This refers to the clean data after removing interference from reversible interface states.

[0033] Specifically, the process of locking the irreversible damage component of the oxide layer dominated by the slow time constant as the cumulative damage vector characterizing the intrinsic failure of the chip is as follows: perform secondary analysis on the remaining residual trajectory to extract the slow variable component corresponding to the long relaxation time, and determine it as an irreversible drift caused by deep level defects in the near-interface oxide layer or permanent lattice damage in the gate oxide layer; calculate the residual amplitude of the slow variable component in steady state and its long-term growth rate over service time; combine the residual amplitude and the long-term growth rate to generate a two-dimensional vector, which is defined as the cumulative damage vector characterizing the degradation of the insulation performance of the chip's gate oxide layer, and is used to quantify the intrinsic physical damage inside the chip.

[0034] In this implementation, the remaining residual trajectory undergoes secondary analysis to lock the slowly varying components, aiming to extract features that truly reflect the health status of the gate oxide layer. The slowly varying components retained in the residual trajectory physically correspond to deep-level defects in the near-interface oxide layer or permanent lattice damage in the gate oxide layer. These damages are typically caused by hot electron injection or hole trapping under high electric field stress and are irreversible. To quantify this damage, a two-dimensional cumulative damage vector is constructed, combining the steady-state residual magnitude with its growth rate over service time, achieving a dual assessment of both the degree of damage and the rate of degradation. The formula for constructing this cumulative damage vector is as follows: In the formula, : The cumulative damage vector characterizing the degradation of the gate oxide insulation performance of the chip; The amplitude of the slowly varying component extracted from the model fitted in step S3 represents the oxide layer defect density at the current moment. : The health threshold voltage of the device in its initial state, used for normalization processing; The increment of the amplitude of the slow variable component during the most recent monitoring period; The ratio of the two values ​​represents the long-term growth rate of damage. Normalized weighting coefficients for amplitude residuals; Normalized weighting coefficients for the long-term growth rate. Wherein, the weighting coefficients... and The method for determining the vector is as follows: a failure probability model is established based on accelerated life test data, and the contribution of amplitude and rate indices to gate oxide breakdown failure is calculated using the maximum likelihood estimation method, thereby setting the weight ratio. This vector, through multi-dimensional description, can more comprehensively capture the evolution process of the device from defect accumulation to final breakdown, and has higher predictive sensitivity compared to a single parameter.

[0035] Specifically, the theoretical thermal stress accumulation is calculated using quasi-static operating condition parameters to establish a priori trend for device lifetime degradation. Temperature-insensitive aging indices and cumulative damage vectors are used as posterior correction factors reflecting the actual degradation level. The specific process for calculating the state deviation between the posterior correction factors and the priori trend is as follows: Based on quasi-static operating condition parameters, the thermal cycling load spectrum is extracted, and a preset fatigue failure model is input to calculate the theoretical cumulative fatigue damage of the device, generating a priori trend curve that monotonically increases with time; the package health degradation index and cumulative damage vector are normalized and fused and mapped to the same health dimension space as the priori trend curve to form multidimensional measured state points; the weighted Euclidean distance between the multidimensional measured state points and the priori trend curve at the current moment is calculated, and the distance is defined as the state deviation characterizing the inconsistency between the physical model prediction and the actual device state.

[0036] In this implementation scheme, the theoretical thermal stress accumulation is calculated and a priori trend is established through quasi-static operating condition parameters. Its core logic lies in using a mature physical model to provide a baseline reference for lifetime prediction, thereby solving the problem of inaccurate predictions in pure data-driven methods when initial data is scarce. First, a thermal cycling load spectrum is extracted based on the quasi-static operating condition parameters. This step uses the rainflow counting method to process the collected irregular load current and junction temperature data, transforming them into a series of stress cycles with specific amplitudes and mean values, thus simulating each thermal shock experienced by the device during actual operation. Next, the extracted load spectrum is input into the fatigue failure model to calculate the theoretical cumulative fatigue damage of the device. This calculation process is based on the linear damage accumulation theory, and the specific calculation formula is as follows: In the formula, Theoretical cumulative fatigue damage up to the current moment; The total number of arrays of different types of thermal cycles identified; The number of times the k-th type of thermal cycle occurs in actual operating conditions; : The theoretical maximum number of cycles that the k-th thermal cycle type can withstand before failure. Fatigue coefficient related to SiC module packaging material properties The junction temperature fluctuation amplitude corresponding to the kth thermal cycle; The fatigue index typically ranges from 3 to 5. Among these, the parameters... and The determination method is as follows: it is obtained by consulting the power cycle reliability manual provided by the device manufacturer or by fitting the data through offline accelerated aging tests on the same model of device. The result calculated by this formula increases monotonically with time, forming a priori trend curve reflecting the ideal degradation trajectory of the device. Subsequently, the process of calculating the state deviation is actually a multi-dimensional distance measurement process, which aims to quantify the difference between the physical model prediction and the measured data. To this end, the temperature-insensitive aging index obtained in step S2 and the cumulative damage vector obtained in step S3 are first normalized to map them to the same dimension space, and then the state deviation is calculated, as follows: In the formula, : Characterizes the state deviation between the physical model prediction and the actual device state; : The reliability weight of the packaging degradation index in the deviation calculation; The temperature-insensitive aging index calculated in step S2; : Historical statistical standard deviation of packaging aging index, used to normalize the denominator; : The confidence weight of the chip cumulative damage vector in the deviation calculation; : The magnitude of the cumulative damage vector obtained in step S3; : Historical statistical standard deviation of chip damage vector. Where, weights and The method for determining this deviation is as follows: The Analytic Hierarchy Process (AHP) is used to determine the deviation based on the current operating condition. For example, under high-frequency vibration conditions, the proportion of package weight is increased; under high-temperature steady-state conditions, the proportion of chip weight is increased. This deviation value directly reflects the degree to which the actual health status of the device deviates from the theoretical expectation at the current moment.

[0037] Specifically, the process of dynamically shrinking and correcting the remaining lifetime prediction interval of the device through state deviation, and outputting a confidence-converged remaining lifetime prediction value is as follows: A state observer based on particle filtering is constructed, and the normalized weights of the particle swarm are updated using state deviation as a likelihood evaluation index; when the state deviation is less than a preset deviation threshold, the degradation state is determined to be within the coverage of the prior model, the weights of particles located in the high-likelihood region of the prior baseline trend are increased, and the variance of the posterior probability density distribution is reduced to narrow the confidence interval; when the state deviation exceeds the preset deviation threshold, a nonlinear abrupt degradation is determined, triggering a particle resampling mechanism to migrate the particle swarm towards the multidimensional measured state points, reconstructing the mathematical expectation of the remaining lifetime prediction value based on the measured state points, and adjusting the prediction interval width to cover uncertainty, outputting the remaining lifetime probability distribution after convergence iteration.

[0038] In this implementation scheme, constructing a state observer based on particle filtering and performing dynamic shrinkage and correction is the core step in using Bayesian recursive estimation theory to handle nonlinear non-Gaussian systems. First, a particle swarm is initialized to simulate the probability distribution of remaining lifetimes. The calculated state deviation is used as observational evidence to update the importance weights of the particles. The likelihood update formula is as follows: In the formula, : The normalized weight of the i-th particle after the update; The weight of the i-th particle before the update; The covariance of the observation noise reflects the uncertainty of the measurement system. The instantaneous state deviation calculated in the preceding steps. The purpose of this formula is that when the state deviation is small, the exponent term is larger, assigning higher weights to particles close to the physical model's predictions. This causes the particle swarm to concentrate towards the prior baseline trend, reducing the variance of the posterior probability density distribution and narrowing the prediction interval, thus enhancing confidence in the physical model. Conversely, when the state deviation exceeds a preset threshold, it indicates a sudden change in the device not covered by the physical model, triggering a resampling mechanism. This mechanism discards low-weight particles and replicates high-weight particles, shifting the center of the particle swarm towards the multidimensional measured state point. Finally, the output remaining lifetime prediction is based on the weighted sum of all particle weights, calculated as follows: In the formula, : The expected value of the remaining lifetime prediction after convergence iteration; : Total number of particles in the particle filter; : The assumed remaining lifetime represented by the i-th particle. This process not only corrects the expected center of the prediction, but also dynamically changes the width of the prediction interval by adjusting the dispersion of the particle distribution, thereby ensuring that the prediction results closely follow the actual degradation trajectory while reasonably covering the uncertainty of the system.

[0039] Please see Figure 2The SiC device lifetime prediction system based on multi-source information fusion includes the following modules: a multi-source acquisition module, used to acquire quasi-static operating parameters during SiC device operation, and simultaneously lock the high-frequency oscillation range during device turn-off through a high-frequency sampling circuit, capturing the underdamped oscillation waveform containing parasitic inductance degradation information, and synchronously latching the body diode voltage drop across the source and drain during the dead time; a feature decoupling module, used to calculate the real-time junction temperature of the device based on the body diode voltage drop, and using a preset reference temperature as the target, spatially mapping and converting the spectral features extracted from the underdamped oscillation waveform through the real-time junction temperature to construct the feature drift vector of the current state point relative to the health reference point, and calculating the magnitude of the feature drift vector as a temperature-insensitive aging index for quantifying the degree of device package degradation; damage analysis; and other modules. The isolation module monitors the drift recovery trajectory of the threshold voltage during device non-operational intervals. It deconstructs the recovery trajectory using a multi-scale time constant separation algorithm, filtering out the interface state charge trapping component dominated by the fast time constant and locking the oxide layer irreversible damage component dominated by the slow time constant as the cumulative damage vector characterizing the intrinsic failure of the chip. The fusion prediction module calculates the theoretical thermal stress accumulation degree using quasi-static operating parameters, establishes the a priori trend of device lifetime degradation, uses temperature-insensitive aging indicators and cumulative damage vector as posterior correction factors reflecting the actual degradation degree, calculates the state deviation between the posterior correction factor and the a priori trend, and dynamically shrinks and corrects the remaining lifetime prediction range of the device through the state deviation, outputting a remaining lifetime prediction value with converged confidence.

[0040] In this implementation scheme, the multi-source acquisition module, as the system's bottom-level sensing unit, is primarily responsible for achieving high-fidelity acquisition and timing synchronization of multi-dimensional physical signals in complex electromagnetic interference environments. Through the coordinated operation of hardware circuits, this module can not only continuously record quasi-static parameters reflecting macroscopic operating loads, but also accurately capture transient waveforms and dead-zone voltages containing rich high-frequency characteristics at microsecond-level switching moments using precise timing logic. Its core function is to construct a raw database with strictly aligned time references, ensuring that the data sources used for subsequent analysis cover the electrical, thermal, and packaging parasitic parameter information of the devices, thus solving the problem of insufficient information from a single signal source. The feature decoupling module, as the system's signal processing and feature extraction unit, is mainly responsible for removing ambient temperature noise interference and quantifying the degree of packaging degradation. This module uses thermistor electrical parameter inversion technology to obtain the internal real-time temperature, and uses this as a reference to perform mathematical transformations and spatial projections on the temperature-sensitive spectral characteristics, uniformly restoring the measured values ​​at different temperatures to a standard reference plane. Its core function is to extract parameter changes purely caused by physical structural damage from mixed signals through vector construction and modulus calculation, outputting stable aging indicators unaffected by operating condition fluctuations, thereby preventing misjudgments of lifespan due to temperature rise. The damage separation module focuses on in-depth diagnosis of the intrinsic health state of the chip, mainly used to distinguish reversible charge trapping effects from irreversible material damage in the time domain. This module utilizes the voltage relaxation characteristics after the device stops working, and decomposes the mixed drift trajectory into components of different time scales through algorithm analysis, accurately filtering out fast-changing components representing temporary performance fluctuations, while locking in slow-changing components representing permanent defects in the gate oxide layer. Its core function is to extract vector features that can truly reflect the cumulative damage of the chip's insulating layer, ensuring that the lifespan prediction model can identify intrinsic failures caused by material aging, rather than being misled by recoverable electrical parameter drift. The fusion prediction module, as the decision-making and calculation center of the system, mainly performs deep fusion calculations of physical mechanism models and multi-source measured data. This module first calculates the ideal degradation path of the device based on classical fatigue failure theory as a reference benchmark. Then, it uses the dual aging characteristics of the package and chip extracted by the preceding module as empirical correction evidence. The deviation between theoretical expectations and actual conditions is calculated to drive the update of the probabilistic prediction algorithm. Its core function is to dynamically correct the prior trends of the physical model using measured data. By continuously narrowing the confidence interval of the prediction results, it ultimately outputs a high-precision remaining lifetime prediction value that conforms to physical laws and possesses individual adaptability.

[0041] In summary, this application has at least the following effects:

[0042] The SiC device lifetime prediction method and system based on multi-source information fusion simultaneously acquires macroscopic quasi-static parameters and microscopic high-frequency transient waveforms. Utilizing a spatial mapping conversion technique based on real-time junction temperature, it effectively eliminates the thermal interference of ambient temperature fluctuations on package aging characteristics, achieving precise decoupling of "temperature sensitivity-aging." Simultaneously, addressing the unique bias temperature instability effect of SiC devices, it innovatively separates reversible interface state charge trapping from irreversible oxide layer lattice damage, eliminating spurious aging noise. Finally, through a Bayesian fusion mechanism guided by a physical failure model and corrected by multi-source measured data, it achieves dynamic contraction and correction of the remaining lifetime prediction range, significantly improving the robustness and prediction accuracy of device health status assessment under complex temperature variations. This effectively solves the technical challenges of existing technologies being easily misled by environmental noise and lacking generalization ability.

[0043] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0044] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0045] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0046] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0047] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0048] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting the lifetime of SiC devices based on multi-source information fusion, characterized in that, Includes the following steps: S1. During the operation of the SiC device, quasi-static operating parameters are collected. At the same time, the high-frequency oscillation range during the device turn-off process is locked by the high-frequency sampling circuit, the underdamped oscillation waveform containing parasitic inductance degradation information is captured, and the body diode voltage drop across the source and drain is synchronously latched during the dead time. S2. Calculate the real-time junction temperature of the device based on the voltage drop of the body diode. With the preset reference temperature as the target, perform spatial mapping and conversion on the spectral features extracted from the underdamped oscillation waveform through the real-time junction temperature to construct the feature drift vector of the current state point relative to the health reference point. Calculate the magnitude of the feature drift vector as a temperature-insensitive aging index for quantifying the degree of device package degradation. S3. Monitor the drift recovery trajectory of the threshold voltage during the device's non-working interval, deconstruct the recovery trajectory using a multi-scale time constant separation algorithm, filter out the interface state charge trapping component dominated by the fast time constant, and lock the oxide layer irreversible damage component dominated by the slow time constant as the cumulative damage vector characterizing the chip's intrinsic failure. S4. Calculate the theoretical thermal stress accumulation degree through quasi-static operating condition parameters, establish the a priori benchmark trend of device life degradation, use temperature-insensitive aging index and cumulative damage vector as posterior correction factors to reflect the actual degree of degradation, calculate the state deviation between the posterior correction factor and the a priori benchmark trend, dynamically shrink and correct the remaining life prediction range of the device through the state deviation, and output the remaining life prediction value with confidence convergence.

2. The method for predicting the lifetime of SiC devices based on multi-source information fusion according to claim 1, characterized in that: Step S1 includes the following steps: A high-frequency analog comparator circuit is configured to monitor the gate-source voltage signal in real time, and the inflection point when the falling edge drops to the Miller plateau region is marked as the turn-off oscillation trigger anchor point. After triggering the anchor point, a high-bandwidth acquisition channel is started to record the entire process of the decay of the underdamped oscillation caused by the resonance of the package parasitic inductance and junction capacitance in the gate-source circuit until the oscillation amplitude converges to the background noise baseline. The logic locks the dead time window of the drive signal. After the turn-off transient ends and the anti-parallel body diode enters the steady-state freewheeling stage, the differential sampling channel is turned on to capture the forward conduction voltage drop across the source and drain terminals. The quasi-static load current corresponding to the sampling time is recorded synchronously to generate a timestamp-aligned multidimensional raw dataset.

3. The method for predicting the lifetime of SiC devices based on multi-source information fusion according to claim 1, characterized in that: The real-time junction temperature of the device is calculated based on the body diode voltage drop. Using a preset reference temperature as the target, the specific process of spatially mapping and converting the spectral features extracted from the underdamped oscillation waveform using the real-time junction temperature is as follows: Based on the wide bandgap thermistor characteristics of SiC, the multidimensional electrothermal coupling calibration relationship between the pre-stored forward voltage drop of the body diode, the load current and the junction temperature is used to substitute the body diode voltage drop and the synchronous load current into the calibration relationship to solve for the real-time junction temperature inside the device. A joint time-frequency domain transformation is performed on the underdamped oscillation waveform to extract the oscillation natural frequency characterizing the inductance property of the circuit and the oscillation attenuation coefficient characterizing the resistance property of the circuit. A temperature sensitivity compensation matrix is ​​introduced to quantify the inherent drift rate of the oscillation natural frequency and oscillation attenuation coefficient as a function of temperature. By utilizing the temperature difference between the real-time junction temperature and the preset reference temperature, the natural frequency and attenuation coefficient under high-temperature conditions are inversely mapped back to the reference temperature coordinate system through the compensation matrix, thereby obtaining the equivalent spectral characteristics that eliminate thermal drift components.

4. The method for predicting the lifetime of SiC devices based on multi-source information fusion according to claim 3, characterized in that: The specific process of constructing the feature drift vector of the current state point relative to the health baseline point and calculating the magnitude of the feature drift vector as a temperature-insensitive aging index for quantifying the degree of device package degradation is as follows: A two-dimensional package impedance characteristic plane is constructed with the natural frequency of oscillation as the horizontal axis and the oscillation attenuation coefficient as the vertical axis. The characteristic coordinates of the device under factory health conditions and at the reference temperature are used as the origin reference coordinates. Using the equivalent spectral characteristics as the coordinates of the current aging state, a directed degradation vector is established on the two-dimensional package impedance characteristic plane, pointing from the origin reference coordinates to the coordinates of the current aging state. The geometric modulus of the directed degradation vector is calculated using the Euclidean distance algorithm. The geometric modulus is defined as a package health degradation index that reflects the combined effect of inductance changes caused by bond wire stripping and resistance changes caused by solder layer fatigue. When the modulus value increases monotonically with operating time, it is determined that the physical impedance characteristics of the package structure have undergone irreversible degradation.

5. The method for predicting the lifetime of SiC devices based on multi-source information fusion according to claim 1, characterized in that: The drift recovery trajectory of the threshold voltage is monitored during the device's non-operating interval. The recovery trajectory is deconstructed using a multi-scale time constant separation algorithm to filter out the interface state charge trapping component dominated by the fast time constant. The specific process is as follows: By applying a weak, constant measurement current and a bias scan signal to the gate during the natural cooling relaxation window after device turn-off, the relaxation decay curve of the threshold voltage over time is continuously recorded. By using a hybrid decay mathematical model that incorporates the interface state electron emission mechanism and the oxide layer trap tunneling mechanism, the relaxation decay curve is analyzed by multi-exponential fitting. The fast-changing component corresponding to the short relaxation time in the model is identified and determined to be a recoverable drift caused by the escape of electrons trapped by shallow energy level traps at the SiC / SiO2 interface. The fast-changing component is numerically separated from the total recovery trajectory, and the remaining residual trajectory is retained.

6. The method for predicting the lifetime of SiC devices based on multi-source information fusion according to claim 5, characterized in that: The specific process of locking the irreversible damage component of the oxide layer, dominated by a slow time constant, as the cumulative damage vector characterizing the intrinsic failure of the chip is as follows: The remaining residual trajectory is analyzed a second time to extract the slow-varying component corresponding to the long relaxation time, which is determined to be an irreversible drift caused by deep level defects in the near-interface oxide layer or permanent lattice damage to the gate oxide layer. Calculate the residual amplitude of the slowly varying component in steady state and its long-term growth rate over the service life. The combined magnitude residual and long-term growth rate generate a two-dimensional vector, defined as the cumulative damage vector characterizing the degradation of the chip's gate oxide insulation performance, which is used to quantify intrinsic physical damage inside the chip.

7. The method for predicting the lifetime of SiC devices based on multi-source information fusion according to claim 1, characterized in that: The theoretical thermal stress accumulation is calculated using quasi-static operating parameters to establish a priori trend for device lifetime degradation. Temperature-insensitive aging indices and cumulative damage vectors are used as posterior correction factors reflecting the actual degree of degradation. The specific process for calculating the state deviation between the posterior correction factors and the priori trend is as follows: The thermal cycling load spectrum is extracted based on quasi-static operating condition parameters. The theoretical cumulative fatigue damage of the device is calculated by inputting a preset fatigue failure model and generating a priori trend curve that monotonically increases with time. The health degradation index and cumulative damage vector are normalized and fused and mapped to the same health metric space as the prior baseline trend curve to form multidimensional measured state points. Calculate the weighted Euclidean distance between the multidimensional measured state point and the prior baseline trend curve at the current moment. Define the distance as the state deviation that characterizes the inconsistency between the physical model prediction and the actual device state.

8. The method for predicting the lifetime of SiC devices based on multi-source information fusion according to claim 7, characterized in that: The specific process of dynamically narrowing and correcting the remaining lifetime prediction range of the device through state deviation, and outputting the remaining lifetime prediction value with converged confidence, is as follows: Construct a state observer based on particle filtering, and use state deviation as a likelihood evaluation index to update the normalized weights of the particle swarm. When the state deviation is less than the preset deviation threshold, the degenerate state is determined to be within the coverage of the prior model. The weight of particles located in the high likelihood region of the prior baseline trend is increased, and the variance of the posterior probability density distribution is reduced to narrow the confidence interval. When the state deviation exceeds the preset deviation threshold, a nonlinear abrupt degradation is determined, triggering the particle resampling mechanism to migrate the particle swarm towards the multidimensional measured state point. The mathematical expectation of the remaining lifetime prediction is reconstructed based on the measured state point, and the prediction interval width is adjusted to cover the uncertainty. The remaining lifetime probability distribution after convergence iteration is output.

9. A SiC device lifetime prediction system based on multi-source information fusion, applied to the SiC device lifetime prediction method based on multi-source information fusion as described in any one of claims 1-8, characterized in that, Includes the following modules: The multi-source acquisition module is used to acquire quasi-static operating parameters during the operation of SiC devices. At the same time, it locks the high-frequency oscillation range during the device turn-off process through a high-frequency sampling circuit, captures the underdamped oscillation waveform containing parasitic inductance degradation information, and synchronously latches the body diode voltage drop across the source and drain during the dead time. The feature decoupling module is used to calculate the real-time junction temperature of the device based on the body diode voltage drop. With a preset reference temperature as the target, the spectral features extracted from the underdamped oscillation waveform are spatially mapped and converted through the real-time junction temperature to construct the feature drift vector of the current state point relative to the health reference point. The magnitude of the feature drift vector is calculated as a temperature-insensitive aging index for quantifying the degree of device package degradation. The damage separation module is used to monitor the drift recovery trajectory of the threshold voltage during the device's non-working interval. It deconstructs the recovery trajectory through a multi-scale time constant separation algorithm, filters out the interface state charge trapping component dominated by the fast time constant, and locks the oxide layer irreversible damage component dominated by the slow time constant as the cumulative damage vector characterizing the chip's intrinsic failure. The fusion prediction module is used to calculate the theoretical thermal stress accumulation degree through quasi-static operating condition parameters, establish the prior benchmark trend of device lifetime degradation, use temperature-insensitive aging index and cumulative damage vector as posterior correction factors to reflect the actual degradation degree, calculate the state deviation between the posterior correction factor and the prior benchmark trend, dynamically shrink and correct the remaining lifetime prediction range of the device through the state deviation, and output the remaining lifetime prediction value with confidence convergence.

Images