A Method and System for Reverse Recovery Turn-On Path Identification and Dynamic Characteristic Prediction of SiC MOSFETs

Abstract

This invention relates to a method and system for identifying reverse recovery conduction paths and predicting dynamic characteristics of SiC MOSFETs, belonging to the field of power semiconductor device testing and reliability analysis. The method includes: Step 1: Constructing a dual-pulse test circuit and acquiring waveform data; Step 2: Preprocessing the data; Step 3: Extracting feature parameters and constructing feature vectors; Step 4: Outputting the conduction path type and device degradation state; Step 5: Based on the conduction path type, combined with the total current decomposition relationship and the dynamic response characteristics of the device under different test conditions, establishing a conduction path decomposition model, and calculating the current proportions of the MOS channel conduction path, body diode conduction path, and Schottky diode conduction path; Step 6: Predicting reverse recovery parameters; Step 7: Assessing device status, predicting remaining lifetime, and assessing failure risk. This invention automates the processing of dual-pulse test data and, for the first time, achieves quantitative identification of conduction paths.

Description

Technical Field

[0001] This invention belongs to the field of power semiconductor device testing and reliability analysis, specifically involving a method and system for identifying the reverse recovery conduction path and predicting the dynamic characteristics of SiC MOSFETs. Background Technology

[0002] Silicon carbide metal-oxide-semiconductor field-effect transistors (SiC MOSFETs) have been widely used in new energy vehicles, power electronics, and photovoltaic fields due to their advantages such as high voltage resistance, high frequency, and low loss. During switching, the device exhibits reverse recovery behavior, and its dynamic characteristics have a significant impact on system efficiency and reliability. For SiC MOSFETs, during the third quadrant conduction and reverse recovery phases, current may be conducted through multiple paths, including the MOS channel, body diode, and Schottky diode. The coupling effect of different conduction paths significantly affects the reverse recovery characteristics and is modulated by factors such as gate turn-off voltage and temperature, causing the device's dynamic behavior to exhibit obvious nonlinear characteristics.

[0003] Currently, double-pulse testing, as a standard method for evaluating the dynamic performance of power devices, mainly extracts parameters such as reverse recovery charge, peak current, and reverse recovery time from oscilloscope waveforms. However, this method relies on manual analysis, resulting in low efficiency, and only obtains total current information, failing to distinguish the contributions of different conduction paths. Furthermore, the numerous dynamic features contained in the waveform are not fully utilized, limiting in-depth analysis of the device's physical mechanisms. In addition, existing methods are typically based on single test results, lacking the ability to model device parameter changes with temperature and stress, making it difficult to achieve degradation assessment and lifetime prediction. Therefore, there is an urgent need for an intelligent analysis method that can fully utilize double-pulse test waveform information and achieve conduction path identification and dynamic characteristic prediction. Summary of the Invention

[0004] To address the problems of indistinguishable conduction paths, insufficient utilization of waveform information, and reliance on manual analysis in existing double-pulse testing methods, this invention proposes a method for identifying the reverse recovery conduction path and predicting dynamic characteristics of SiC MOSFETs based on double-pulse testing. This invention acquires voltage, current, and gate signal waveforms, extracts time-domain, electrical, and waveform features, constructs feature vectors, and inputs them into a machine learning model to map reverse recovery parameters to the conduction path. Furthermore, it uses a conduction path decomposition model to identify the current proportions of the MOS channel, body diode, and Schottky diode. Simultaneously, it introduces parameters such as gate turn-off voltage and temperature to establish a multi-parameter coupled model, combining multiple test data to achieve device degradation assessment and lifetime prediction. This invention achieves quantitative identification of the conduction path and accurate modeling of dynamic characteristics, offering advantages such as high automation and strong physical interpretability.

[0005] This invention also proposes a SiC MOSFET reverse recovery conduction path identification and dynamic characteristic prediction system based on dual-pulse testing.

[0006] The technical solution of the present invention is as follows: A method for identifying the reverse recovery conduction path and predicting the dynamic characteristics of SiC MOSFETs based on dual-pulse testing includes: Step 1: Build a dual-pulse test circuit to collect the drain-source voltage Vds(t), drain current Ids(t), external gate drive resistance, load inductance, temperature, load current, bus voltage, and gate drive voltage of the SiC MOSFET under test during the switching and reverse recovery processes. Step 2: Preprocess the waveform data acquired in Step 1, including: noise filtering, time axis alignment, amplitude normalization, and outlier data removal. Step 3: Extract feature parameters from the waveform data preprocessed in Step 2 and construct a feature vector; the feature parameters include time-domain features, electrical features, waveform morphology features and test condition features, wherein the time-domain features are used to characterize the time response of the device during turn-on, turn-off and reverse recovery processes, the electrical features are used to characterize the reverse recovery and switching loss characteristics of the device, the waveform morphology features are used to characterize the dynamic features, and the test condition features include gate turn-off voltage, bus voltage, load current, external gate resistance and temperature; Step 4: Input the feature vector into the trained machine learning model; Establish the mapping relationship between the input feature vector and the key parameters of reverse recovery, and output the reverse recovery charge, reverse recovery peak current, reverse recovery time and switching loss parameters; Output the conduction path type and the device degradation status; Step 5: Based on the conduction path type, combined with the total current decomposition relationship and the dynamic response characteristics of the device under different test conditions, establish a conduction path decomposition model. The conduction path decomposition model includes the equivalent current response functions of the MOS channel conduction path, the body diode conduction path, and the Schottky diode conduction path. Calculate the current proportion of the MOS channel conduction path, the body diode conduction path, and the Schottky diode conduction path. Step 6: Introduce gate turn-off voltage and temperature parameters, establish a multi-parameter coupled model of reverse recovery characteristics, and realize reverse recovery parameter prediction; Step 7: Construct a device key parameter evolution model based on multiple double-pulse test data, and realize device status assessment, remaining lifetime prediction and failure risk assessment according to the changing trends of reverse recovery parameters, switching loss parameters and conduction path ratio.

[0007] According to a preferred embodiment of the present invention, in step 3, the time-domain features include: turn-on delay time td(on), turn-off delay time td(off), current rise time tr, voltage fall time tf, and reverse recovery time trr; Electrical characteristics include: reverse recovery charge Qrr, reverse recovery peak current Irrm, turn-on loss Eon, turn-off loss Eoff, current change rate di / dt, and voltage change rate dv / dt. Waveform characteristics include: current second peak characteristics, oscillation frequency and damping characteristics, tail current characteristics, and voltage overshoot amplitude; The characteristics of the second peak current include: the amplitude of the second peak current I. 2nd The time t for the second peak to occur 2nd The ratio of the secondary peak current to the primary peak current Irrm; oscillation frequency and damping characteristics, including: oscillation frequency fosc, oscillation period Tosc, and damping coefficient or attenuation rate; tail current characteristics, including: tail duration t. tail The magnitude of the trailing current and the rate of trailing decay; Test conditions include: gate turn-off voltage Vgs(off), bus voltage, load current, external gate drive resistance and temperature, load inductance, and gate drive voltage test condition parameters. The feature parameters are uniformly encoded to construct the sample feature vector X.

[0008] According to a preferred embodiment of the present invention, step 4, the process of constructing and training the machine learning model includes: Constructing Sample Labels: Based on the collected waveform data and extracted feature parameters, construct the output labels required for supervised learning; among them, reverse recovery charge Qrr, reverse recovery peak current Irrm, reverse recovery time trr, turn-on loss Eon, and turn-off loss Eoff are used as parameter prediction labels; the conduction path type of the samples is labeled to obtain classification labels for MOS channel dominant conduction, body diode dominant conduction, Schottky diode dominant conduction, or multi-path mixed conduction; the samples are labeled with normal, slightly degraded, or significantly degraded states to form degradation state labels; Constructing machine learning models: The machine learning models include a reverse recovery parameter prediction model, a conduction path identification model, and a degradation state assessment model; the reverse recovery parameter prediction model, the conduction path identification model, and the degradation state assessment model are established respectively; the reverse recovery parameter prediction model is a regression model, used to establish the mapping relationship between the input feature vector and the key parameters of reverse recovery, and outputs the reverse recovery charge Qrr, the reverse recovery peak current Irrm, the reverse recovery time trr, and the turn-on loss Eon and the turn-off loss Eoff; The conduction path identification model is used to output the conduction path type of the device under the current test conditions; The degradation state assessment model is used to output the current degradation state of the device; Machine learning model training and optimization: The samples are divided into training set, validation set and test set. The input features are standardized or normalized to complete the training of the machine learning model. The machine learning model is then evaluated according to the indicators and the parameters are optimized to obtain the trained machine learning model.

[0009] More preferably, the reverse recovery parameter prediction model includes an input layer, a hidden layer, and an output layer. The input layer receives the sample feature vector X constructed in step 3. The hidden layer uses a fully connected neural network or a tree-based ensemble method for nonlinear mapping to capture the complex relationship between the input features and the output parameters. The output layer generates continuous value regression results, including the reverse recovery charge Qrr, the reverse recovery peak current Irrm, the reverse recovery time trr, and the turn-on loss Eon and turn-off loss Eoff, thereby enabling the prediction of key reverse recovery parameters. The conduction path identification model is a classification model. The input layer takes the sample feature vector X as input, and the hidden layer uses a fully connected neural network or a convolutional neural network to extract features, or uses support vector machine or gradient boosting tree classification methods for discrimination. The output layer predicts the conduction path type, including MOS channel dominant conduction, body diode dominant conduction, Schottky diode dominant conduction, or multi-path hybrid conduction. The degradation status assessment model is a classification or scoring model based on the evolution sequence of key parameters. The input layer inputs include the current ratio and reverse recovery parameters of the MOS channel conduction path, body diode conduction path, and Schottky diode conduction path obtained from steps 5 and 6. The hidden layer uses a fully connected neural network to process multidimensional inputs or timing features. The output layer generates the device degradation status and risk level, including normal, mild degradation, or severe degradation, or outputs a degradation score or remaining lifetime.

[0010] According to a preferred embodiment of the present invention, the machine learning model employs at least one of random forest, support vector machine, gradient boosting tree, and fully connected neural network; When the original time-series waveform is used directly as input, a long short-term memory network, a convolutional neural network, or a combination of both is also used for modeling.

[0011] According to a preferred embodiment of the present invention, step 5 includes the following specific implementation process: Step 5.1: Establish the equivalent current response functions I_MOS(t), I_body(t), and I_SBD(t) for the MOS channel conduction path, body diode conduction path, and Schottky diode conduction path, respectively, as follows: I_MOS(t)=fMOS (Vds(t),Vgs(t),T,Ɵ MOS ); I_body(t) = f body (Vds(t),Vgs(t),T,Ɵ body ); I_SBD(t)=f SBD (Vds(t),Vgs(t),T,Ɵ SBD ); Among them, Ɵ MOS , Ɵ body , Ɵ SBD These are the model parameters for the MOS channel conduction path, the body diode conduction path, and the Schottky diode conduction path, respectively, used to characterize the conduction response characteristics of each path under different test conditions; f MOS () represents the MOS channel current response function; f body () represents the body diode current response function; f SBD () represents the current response function of a Schottky diode; Vds(t) is the curve of drain-source voltage changing with time, i.e., the transient waveform of the voltage between the drain and source of a MOSFET over time; Vgs(t) is the curve of gate-source voltage changing with time, i.e., the transient waveform of the voltage between the gate and source of a MOSFET over time, used to control the on or off state of the device; T refers to the temperature under test or device operating conditions. Apply constraints: When the identification result is a MOS channel conduction path, set I_MOS (i.e., I_MOS(t)) as the dominant component; when the identification result is a body diode conduction path, set I_body (i.e., I_body(t)) as the dominant component; when the identification result is a Schottky diode conduction path, set I_SBD (i.e., I_SBD(t)) as the dominant component; when the identification result is a multi-path mixed conduction, allow at least two conduction paths to participate significantly in conduction simultaneously. Step 5.2: Using the measured drain-source voltage Vds(t), drain current Ids(t), gate turn-off voltage Vgs(off), temperature T, reverse recovery charge Qrr, reverse recovery peak current Irrm, reverse recovery time trr, and waveform characteristics as constraints or optimization objectives, the current components I_MOS(t), I_body(t), and I_SBD(t) of each conduction path are obtained through parameter fitting, least squares optimization, or iterative solution; including: First, based on the equivalent current response functions of each conduction path established in step 5.1, the expression for the total current is constructed: I_total=I MOS (t,Ɵ MOS )+Ibody(t,Ɵbody)+I SBD(t,Ɵ SBD ); Using the measured drain current Ids(t) as a reference target, the error function, i.e. the objective function, is constructed as follows: 2 ; Meanwhile, the reverse recovery charge Qrr, peak current Irrm, reverse recovery time trr, and waveform morphology characteristics are transformed into additional constraints and added to the objective function; Subsequently, the least squares method or iterative optimization algorithm is used to optimize the model parameters Ɵ. MOS Ɵbody、Ɵ SBD Solve the problem to minimize the objective function; After the parameters converge, the obtained optimal parameters are substituted into the response function of each conduction path to obtain the corresponding time-series current components. I_body(t) and This allows for the decomposition of the total current; Step 5.3: Calculate the proportion of each conduction path based on the proportional relationship between the current components of each conduction path and the total current: ; ; ; in, , and These represent the instantaneous proportions of the MOS channel conduction path, the body diode conduction path, and the Schottky diode conduction path in the total current, respectively; that is, the current proportions of the MOS channel conduction path, the body diode conduction path, and the Schottky diode conduction path. I_total = + + , This refers to the total current, i.e., the measured drain current Ids(t).

[0012] According to a preferred embodiment of the present invention, step 6 includes the following specific implementation process: Based on test data under different gate turn-off voltages Vgs(off) and temperature T, a mapping model between the reverse recovery parameters and external conditions is established, i.e., a multi-parameter coupling model of the reverse recovery characteristics: Qrr = f(Vgs(off), T, X); Where f represents the mapping relationship of the function, Qrr represents the value of the reverse recovery charge, and Qrr = f(Vgs(off),T, X) means that the reverse recovery charge is expressed as a function of the gate turn-off voltage Vgs(off), temperature T, and eigenvector X. Inverse recovery parameter prediction is achieved through a multi-parameter coupling model of inverse recovery characteristics.

[0013] According to a preferred embodiment of the present invention, step 7 is specifically implemented as follows: Based on multiple sets of dual-pulse test data acquired at different times, the change sequences of reverse recovery charge Qrr, reverse recovery peak current Irrm, reverse recovery time trr, switching loss Eon / Eoff, and conduction path ratio are extracted to establish an evolution model of key device parameters with time, thermal stress, electrical stress, or switching cycle number, i.e., key parameter evolution model. Assess the current degradation status of the device based on the drift amplitude of key parameters, the changing trend of the conduction path ratio, and the frequency of abnormal waveforms; including: The drift amplitude of key parameters, changes in the proportion of conduction paths, and the frequency of abnormal waveforms are combined and analyzed comprehensively using a weighted scoring or classification model. Based on the comprehensive results, the current state of the device is classified as follows: Key parameters show minimal drift, the proportion of conduction paths is stable, there are few abnormal waveforms, and the current state of the device is normal. Some parameters have drifted slightly, the proportion of conduction paths has changed slightly, and abnormal waveforms occasionally appear. The current state of the device is mild degradation. Key parameters have drifted significantly, the proportion of conduction paths has changed markedly, abnormal waveforms are frequent, and the device is currently in a state of severe degradation. Set the rate of change D of the key parameter relative to the initial value. P for: D P = Where P is the current test parameter and P0 is the initial health status parameter, including Qrr, Irrm, trr, Eon, and Eoff; All key parameters D P When the percentage is less than 5%, the drift of the key parameter is small; at least one key parameter 5% ≤ D P <15%, indicating slight drift in some parameters; at least one critical parameter D P ≥15% indicates a significant drift in key parameters; Set the change D of the proportion of the conducting path relative to the initial state. η :D η =|η-η0|, where η is the current percentage of a certain conductive path, and η0 is the percentage of that path in the initial state; The change in the proportion of each path is less than 5%, indicating a stable proportion of conductive paths; the change in the proportion of at least one path is 5%-15%, indicating a slight change in the proportion of conductive paths; the change in the proportion of at least one path is D. η A value of 15% or higher indicates a significant change in the proportion of the conductive path. Set the frequency F of abnormal waveform occurrenceabn :F abn = Where, N abn N represents the number of abnormal waveforms. total This represents the total number of test waveforms; An abnormal waveform frequency of less than 5% indicates few abnormal waveforms; an abnormal waveform frequency of 5%-20% indicates occasional abnormal waveforms; and an abnormal waveform frequency of 20% or more indicates frequent abnormal waveforms.

[0014] Based on the degradation state assessment results and parameter evolution trends, the remaining usable lifetime of the device is predicted, and the failure risk level is output; including: Remaining usable lifetime prediction, including: Based on the key parameter evolution model, the trend curves of key parameters changing with time, number of cycles, or stress conditions are obtained; Set the device failure threshold; Extrapolating parameter changes using a key parameter evolution model, the time or number of switching cycles required to reach the failure threshold is calculated: t fail or n fail =f −1 ; The difference between the current time or number of cycles and the predicted failure time / number of cycles is the remaining usable lifetime (RUL): RUL = t fail -t current Or RUL=n fail -n current ; Where f represents the key parameter evolution model; f −1 t represents finding the inverse function; fail This represents the time required to predict and reach the failure parameter threshold; t current This represents the actual operating time of the device, that is, the time elapsed from the start of use or testing to the current moment; n fail This represents the number of switching cycles required to predict and reach the failure parameter threshold; n current This indicates the number of switching cycles the device has undergone so far, i.e., the cumulative number of switching cycles from the start of its use to the present. Failure risk level output includes: The degree to which the set parameter R approaches the failure threshold P :R P = ×100%, where, The failure threshold is defined by P0 as the initial value and P as the current value; when all key parameters R... P Less than 70%, abnormal waveform frequency F abn Less than 5%, and the change in the proportion of the conduction path D η Less than 5% is considered low risk. When any key parameter 70%≤R P <90%, or abnormal waveform frequency F abn The percentage is 5% to 20%, or the proportion of the conduction path changes by D. η The risk level is 5% to 15%, which is considered medium risk. When any key parameter R P ≥90% or has exceeded the failure threshold, or abnormal waveform frequency F abn ≥20%, or change in the proportion of conduction path D η ≥15% indicates high risk.

[0015] A system for identifying the reverse recovery conduction path and predicting the dynamic characteristics of SiC MOSFETs based on dual-pulse testing includes: The waveform data acquisition module is configured to: build a dual-pulse test circuit to acquire the drain-source voltage Vds(t), drain current Ids(t), external gate drive resistance, load inductance, temperature, load current, bus voltage, and gate drive voltage of the SiC MOSFET under test during the switching process and reverse recovery process. The preprocessing module is configured to preprocess the acquired waveform data, including noise filtering, time axis alignment, amplitude normalization, and outlier removal. The feature parameter extraction module is configured to extract feature parameters from the preprocessed waveform data and construct a feature vector. The feature parameters include time-domain features, electrical features, waveform morphology features, and test condition features. The time-domain features are used to characterize the time response of the device during turn-on, turn-off, and reverse recovery. The electrical features are used to characterize the reverse recovery and switching loss characteristics of the device. The waveform morphology features are used to characterize the dynamic features. The test condition features include gate turn-off voltage, bus voltage, load current, external gate resistance, and temperature. The machine learning module is configured to input the feature vector into the trained machine learning model; Establish the mapping relationship between the input feature vector and the key parameters of reverse recovery, and output the reverse recovery charge, reverse recovery peak current, reverse recovery time and switching loss parameters; output the conduction path type and output the device degradation status; The current percentage calculation module is configured to: establish a conduction path decomposition model based on the conduction path type, combined with the total current decomposition relationship and the dynamic response characteristics of the device under different test conditions. The conduction path decomposition model includes the equivalent current response functions of the MOS channel conduction path, the body diode conduction path and the Schottky diode conduction path, and calculate the current percentage of the MOS channel conduction path, the body diode conduction path and the Schottky diode conduction path. The reverse recovery parameter prediction module is configured to: introduce gate turn-off voltage and temperature parameters, establish a multi-parameter coupled model of reverse recovery characteristics, and realize reverse recovery parameter prediction. The device status assessment, remaining lifetime prediction, and failure risk assessment module is configured to: construct a device key parameter evolution model based on multiple double-pulse test data, and realize device status assessment, remaining lifetime prediction, and failure risk assessment based on the changing trends of reverse recovery parameters, switching loss parameters, and conduction path ratio.

[0016] The beneficial effects of this invention are as follows: 1. This invention enables automated processing of dual-pulse test data: avoiding errors from manual reading and improving test efficiency and consistency.

[0017] 2. This invention achieves quantitative identification of conduction paths for the first time: it can distinguish the conduction paths of MOS channels, body diodes and Schottky diodes, and calculate their current proportions.

[0018] 3. This invention makes full use of waveform information to improve analysis accuracy: it not only uses traditional parameters, but also introduces waveform morphology features to improve the model's expressive ability.

[0019] 4. This invention establishes a multi-parameter coupling relationship between Qrr, Vgs(off), and temperature, revealing the influence mechanism of gate modulation and temperature effect on reverse recovery characteristics.

[0020] 5. This invention enables device degradation and lifetime prediction: breaking through the limitation of traditional methods that can only perform single analysis.

[0021] 6. This invention is applicable to various SiC power device structures, including conventional MOSFETs and devices with built-in SBD MOSFETs. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the dual-pulse test circuit structure used in this invention; Figure 2 This is a schematic diagram of the data processing flow of the present invention; Figure 3 This is a schematic diagram of the machine learning model structure of the present invention; Figure 4 This is a schematic diagram of the conduction path identification of the present invention; Figure 5 This is a schematic diagram illustrating the multi-parameter coupling relationship between the reverse recovery parameters and the gate turn-off voltage and temperature in this invention. Figure 6 This is a waveform diagram of the switch. Figure 7 This is a waveform diagram for the reverse recovery test; Figure 8This is a schematic diagram showing the curves of reverse recovery charge Qrr as a function of gate turn-off voltage Vgs(off) at different temperatures. Figure 9 This is a schematic diagram showing the curves of reverse recovery charge Qrr as a function of temperature under different gate turn-off voltages.

[0023] Figure 10 This is a diagram of the reverse recovery parameter prediction model architecture of the present invention.

[0024] Figure 11 This is a diagram of the conduction path identification model architecture of the present invention.

[0025] Figure 12 This is a diagram of the degradation state assessment model architecture of the present invention. Detailed Implementation

[0026] The present invention will be further described below with reference to the embodiments and accompanying drawings, but is not limited thereto.

[0027] Terminology Explanation: 1. SiC MOSFET: A metal-oxide-semiconductor field-effect transistor based on silicon carbide (SiC) material.

[0028] 2. Double Pulse Test (DPT): A test method used to characterize the dynamic performance of power devices. It obtains the voltage and current waveforms of the device during the turn-on, turn-off, and reverse recovery processes by applying two pulse signals.

[0029] 3. Reverse recovery characteristics: The dynamic behavior of reverse current appearing and gradually decaying due to the carrier storage effect during the transition of power devices from the on state to the off state. It is usually characterized by reverse recovery charge (Qrr), peak current (Irrm) and reverse recovery time (trr).

[0030] 4. Reverse Recovery Charge (Qrr): During reverse recovery, due to the carrier storage effect, the amount of charge released by the MOSFET when it transitions from the on state to the off state after being turned off. Specifically, Qrr represents the total charge consumed as the reverse current decreases from its peak value to zero. The formula is expressed as follows: , where I(t) is the curve of the reverse recovery current changing with time, and t0 and t1 are the start and end times when the current drops to zero.

[0031] 5. Peak current (Irrm): The maximum reverse current that the MOSFET withstands during reverse recovery. It usually occurs at the start of reverse recovery and represents the maximum instantaneous current amplitude.

[0032] 6. Reverse recovery time (trr): The time elapsed from the peak current of the reverse recovery current until the current decays to a specified value (e.g., 10% of the peak current).

[0033] 7. Turn-on delay time td(on): The time from when the gate voltage (Vgs) rises to 10% to when the drain-source voltage (Vds) drops to 90%.

[0034] 8. Turn-off delay time td(off): refers to the time from when the gate voltage (Vgs) drops to 90% to when the drain-source voltage (Vds) drops to 10%.

[0035] 9. Current rise time tr: The time it takes for the drain current (Id) to rise from 10% to 90%.

[0036] 10. Voltage drop time tf: The time it takes for the drain-source voltage (Vds) to drop from 90% to 10%.

[0037] 11. Turn-on loss Eon: The energy loss of a MOSFET when it transitions from the off state to the on state. That is, the energy loss that occurs when the current begins to increase during the turn-on process.

[0038] 12. Turn-off loss Eoff: The energy loss when a MOSFET switches from the on state to the off state, which mainly occurs during the current decrease process.

[0039] 13. Rate of change of current di / dt: The rate at which current changes with time, expressed as the ratio of the change in current to the change in time.

[0040] 14. Voltage change rate dv / dt: The rate at which voltage changes with time, expressed as the ratio of the voltage change to the time change.

[0041] 15. Conduction path: The current conduction path inside the device during the third quadrant conduction or reverse recovery process, including the MOS channel conduction path, the body diode conduction path, and the Schottky diode conduction path.

[0042] 16. Proportion of conduction paths: The proportion of current carried by different conduction paths in the total current, used to characterize the contribution of each path to the overall conduction behavior.

[0043] 17. Feature Extraction: The process of extracting key parameters reflecting the dynamic characteristics of a device from a double-pulse test waveform, including time-domain features, electrical features, and waveform morphology features.

[0044] 18. Machine learning models: Data-driven models used to establish the mapping relationship between input features and output results, including random forests, support vector machines, neural networks, etc.

[0045] 19. Multi-parameter coupling model: A mathematical model that describes the relationship between reverse recovery characteristics and multiple variables such as gate turn-off voltage, temperature, and device state.

[0046] 20. Degradation state: The state in which the performance of a device gradually deteriorates under long-term operation or stress, manifested by phenomena such as increased reverse recovery charge and increased switching losses.

[0047] 21. Lifetime prediction: A method for estimating the future failure time or remaining usable time of a device based on the changing trends of its parameters.

[0048] 22. Random Forest: An ensemble learning-based model that makes a final prediction by building multiple decision trees and voting on or averaging their results. Each tree is trained by randomly selecting samples and features to avoid overfitting, thus improving the model's stability and accuracy. Random forests can be used for regression and classification tasks.

[0049] 23. Support Vector Machine (SVM): A supervised learning model widely used in classification and regression problems. SVM constructs an optimal hyperplane in a high-dimensional space to maximize the margin between different classes.

[0050] 24. Gradient Boosting Tree (GBT): An ensemble learning method that progressively combines multiple weak learners (usually decision trees) into a strong learner. The training objective of each tree is to minimize the loss function, and the training of each tree is adjusted based on the residuals of previous trees, thereby improving the model's prediction accuracy. GBT is suitable for regression and classification problems and has excellent modeling capabilities for nonlinear relationships.

[0051] 25. Fully Connected Neural Network (FCNN): A basic form of neural network where each neuron in each layer is connected to all neurons in the previous layer. FCNN is suitable for handling non-linear relationships and can extract deep features from data through learning from multiple layers of neurons. FCNN can be used for tasks such as regression and classification, and is typically used for modeling data with complex patterns.

[0052] 26. Long Short-Term Memory (LSTM): A variant of Recurrent Neural Network (RNN) specifically designed to address the vanishing gradient problem encountered by standard RNNs when dealing with long-term dependencies. LSTM effectively captures long-term dependencies in sequential data and is widely used in tasks such as time series forecasting and natural language processing. In this patent, LSTM can be used to process time series data in the backpropagation process.

[0053] 27. Convolutional Neural Network (CNN): A type of feedforward neural network widely used in image processing and pattern recognition. CNNs extract local features through convolutional layers, reduce the dimensionality of features through pooling layers, and finally make predictions through fully connected layers. CNNs can automatically learn the spatial structure features in the input data, making them very effective in processing image and waveform data.

[0054] Example 1 A method for identifying the reverse recovery conduction path and predicting the dynamic characteristics of SiC MOSFETs based on dual-pulse testing includes: Step 1: Build a double-pulse test circuit, such as... Figure 1 As shown, the drain-source voltage Vds(t), drain current Ids(t), external gate drive resistance, load inductance, temperature, load current, bus voltage, and gate drive voltage (turn-on voltage Vgs(on) and turn-off voltage Vgs(off)) of the tested SiC MOSFET are collected during the switching and reverse recovery processes. The existing dual-pulse test system (DPT equipment) can directly complete the circuit construction and the acquisition of waveforms and test condition parameters such as Vds, Id, and Vgs.

[0055] Under the set bus voltage (e.g., 600V) and load conditions, the device under test (DUT) is placed in the SWD position, a double-pulse signal is applied, and switching waveform data—drain-source voltage Vds(t) and drain current Ids(t)—is acquired using a high-speed data acquisition system. Temperature and gate drive parameters are also recorded. Besides acquiring the switching waveform, measuring the reverse recovery waveform is also crucial in this embodiment. During the reverse recovery test, the DUT is placed in the FWD position, i.e., in freewheeling mode, and the reverse recovery waveform is acquired. The drain-source voltage Vds(t) and drain current Ids(t) are monitored synchronously using current and voltage probes. Based on the voltage and current waveform data during the reverse recovery process, key parameters such as reverse recovery time (trr), reverse recovery charge (Qrr), and reverse recovery peak current (Irrm) are further analyzed and extracted. Figure 6 This is a waveform diagram of the switch. Figure 6 In the diagram, a represents the waveform of the drain-source voltage and drain current of the device under test during the switching process, and b ... Figure 7 This is a waveform diagram for the reverse recovery test; Figure 7 The changes of drain current Ids(t) and drain-source voltage Vds(t) over time are shown. The peak current (Irrm), second peak, current tail, and voltage overshoot phenomenon during the reverse recovery stage can be clearly observed, providing a direct basis for extracting reverse recovery parameters (Qrr, trr, etc.).

[0056] Step 2: Preprocess the waveform data acquired in Step 1 (including drain-source voltage Vds(t), drain current Ids(t), external gate drive resistance, load inductance, temperature, load current, bus voltage, and gate drive voltage of the SiC MOSFET under test during switching and reverse recovery processes), such as... Figure 2 As shown, this includes: noise filtering, time axis alignment, amplitude normalization, and outlier removal; to improve data quality and model stability. Step 3: Extract feature parameters from the preprocessed waveform data in Step 2. The extraction method is an existing method, specifically including: numerical integration (for Qrr), numerical differentiation (for di / dt, dv / dt), peak detection (Irrm, etc.); and construct a feature vector. The feature parameters include time-domain features, electrical features, waveform morphology features, and test condition features. Among them, time-domain features are used to characterize the time response of the device during turn-on, turn-off, and reverse recovery; electrical features are used to characterize the reverse recovery and switching loss characteristics of the device; waveform morphology features are used to characterize the dynamic characteristics such as waveform peak, oscillation, and tail; and test condition features include gate turn-off voltage, bus voltage, load current, external gate resistance, and temperature. Step 4: Input the feature vectors into the trained machine learning model; Establish a mapping relationship between the input feature vector and the key parameters of reverse recovery, and output the reverse recovery charge, reverse recovery peak current, reverse recovery time, and switching loss parameters; including: A training set is built using the sample data from the completed tests. Each sample includes an input sample feature vector X and the corresponding true output parameters, such as Qrr, Irrm, trr, Eon, and Eoff. The true output parameters can be obtained from the measured waveforms through integration, peak extraction, time determination, and energy integration.

[0057] The sample feature vector X is input into the regression model for training, so that the regression model learns the nonlinear mapping relationship between the input features and the key parameters of reverse recovery. After training is complete, for a new test sample, input the feature vector of the new test sample and output the prediction results, including reverse recovery charge, reverse recovery peak current, reverse recovery time and switching loss parameters.

[0058] The system outputs the conduction path type and the device degradation state; the conduction path type output is part of the classification and recognition process. First, based on existing experimental data, third-quadrant conduction characteristics, and differences in reverse recovery waveforms at different gate turn-off voltages and temperatures, samples are labeled, for example, as: MOS channel dominant conduction, body diode dominant conduction, Schottky diode dominant conduction, or multi-path hybrid conduction. Then, the feature vector X is used as input, and the conduction path type is used as the output label to train the classification model. After training, inputting the test sample into the model will output the corresponding conduction path type.

[0059] Device degradation status output can also be achieved through classification or scoring methods. Based on the changing trends of key parameters during repeated tests, aging tests, or stress tests, such as the drift magnitude of parameters like Qrr, Irrm, Eon, Eoff, and conduction path percentage, samples are labeled as normal, slightly degraded, or significantly degraded. After model training is complete, the current degradation status of the device can be output based on the feature vector of the test sample, or the corresponding degradation score and failure risk level can be output.

[0060] Step 5: Based on the conduction path type, combined with the total current decomposition relationship and the dynamic response characteristics of the device under different test conditions, establish a conduction path decomposition model. The conduction path decomposition model includes the equivalent current response functions of the MOS channel conduction path, the body diode conduction path, and the Schottky diode conduction path. Calculate the current proportion of the MOS channel conduction path, the body diode conduction path, and the Schottky diode conduction path. Step 6: Introduce gate turn-off voltage and temperature parameters, establish a multi-parameter coupled model of reverse recovery characteristics, and realize reverse recovery parameter prediction; Step 7: Construct a device key parameter evolution model based on multiple double-pulse test data, and realize device status assessment, remaining lifetime prediction and failure risk assessment according to the changing trends of reverse recovery parameters, switching loss parameters and conduction path ratio.

[0061] Example 2 The difference between the method for identifying the reverse recovery conduction path and predicting the dynamic characteristics of SiC MOSFETs based on dual-pulse testing as described in Example 1 is that: In step 3, the time-domain features include: turn-on delay time td(on), turn-off delay time td(off), current rise time tr, voltage fall time tf, and reverse recovery time trr; Electrical characteristics include: reverse recovery charge Qrr, reverse recovery peak current Irrm, turn-on loss Eon, turn-off loss Eoff, current change rate di / dt, and voltage change rate dv / dt. Waveform characteristics include: current second peak characteristics, oscillation frequency and damping characteristics, tail current characteristics, and voltage overshoot amplitude; The characteristics of the second peak current include: the amplitude of the second peak current I. 2nd The time t for the second peak to occur 2nd And the ratio of the secondary peak current to the primary peak current Irrm (e.g., I 2nd / Irrm); used to characterize the intensity and timing of the secondary current rise during the reverse recovery process. Oscillation frequency and damping characteristics include: oscillation frequency fosc (calculated from the time interval between adjacent peaks or troughs), oscillation period Tosc, and damping coefficient or attenuation rate (calculated from the oscillation envelope or the attenuation ratio of adjacent peak amplitudes, such as logarithmic attenuation rate); used to describe the speed and decay characteristics of current or voltage oscillations. Tail current characteristics include: tail duration t. tail The tail current amplitude (such as the average tail current or the end current value) and the tail decay rate (which can be represented by the rate of decrease of the tail current over time or the exponential decay coefficient) are used to characterize the degree of slow decay of the current in the later stage of reverse recovery.

[0062] Test conditions include: gate turn-off voltage Vgs(off), bus voltage, load current, external gate drive resistance and temperature, load inductance, and gate drive voltage (turn-on voltage Vgs(on), turn-off voltage Vgs(off)) test condition parameters; The feature parameters are uniformly encoded to construct the sample feature vector X. X = [x1, x2, x3, …, xn].

[0063] The feature vector X can be composed of manually extracted feature parameters, or it can be combined with test condition parameters as model input while retaining the original time-series waveform segments, so as to improve the ability to characterize the dynamic behavior of the device.

[0064] Step 4, the process of building and training the machine learning model, includes: Sample Label Construction: Based on the collected waveform data and extracted feature parameters, output labels required for supervised learning are constructed. Specifically, reverse recovery charge Qrr, reverse recovery peak current Irrm, reverse recovery time trr, turn-on loss Eon, and turn-off loss Eoff are used as parameter prediction labels. Combining the differences in reverse recovery waveforms under different gate turn-off voltages, temperatures, and load conditions, the third quadrant conduction characteristics, and preset conduction path discrimination rules, conduction path types are labeled for the samples, resulting in classification labels for MOS channel-dominated conduction, body diode-dominated conduction, Schottky diode-dominated conduction, or multi-path mixed conduction. Based on the drift of key parameters during repeated tests, stress tests, or aging tests, samples are labeled as normal, slightly degraded, or significantly degraded, forming degradation state labels. Building machine learning models: such as Figure 3 As shown, the machine learning model includes a reverse recovery parameter prediction model, a conduction path identification model, and a degradation state assessment model. The reverse recovery parameter prediction model, conduction path identification model, and degradation state assessment model are established respectively. The reverse recovery parameter prediction model is a regression model used to establish the mapping relationship between the input feature vector and the key reverse recovery parameters, outputting the reverse recovery charge Qrr, the reverse recovery peak current Irrm, the reverse recovery time trr, and the turn-on loss Eon and turn-off loss Eoff. The conduction path identification model is used to output the conduction path type of the device under the current test conditions; The degradation state assessment model is used to output the current degradation state of the device; The conduction path type results output by the conduction path identification model are used as prior constraints for the conduction path decomposition model in step 5, to help solve the current components and their proportions in each conduction path.

[0065] Machine learning model training and optimization: The samples are divided into training, validation, and test sets. Input features are standardized or normalized before training the machine learning model. The model is then evaluated based on metrics such as mean squared error, mean absolute error, classification accuracy, recall, and confusion matrix, and parameters are optimized to improve the model's accuracy and generalization ability. A well-trained machine learning model is obtained.

[0066] Machine learning model output: Input the feature vectors corresponding to the test samples in the test set into the training, and output the key parameters of reverse recovery, the type of conduction path and the degradation state results, and provide basic data for subsequent calculation of conduction path ratio and lifetime prediction.

[0067] like Figure 10As shown, the reverse recovery parameter prediction model includes an input layer, a hidden layer, and an output layer, used to predict the dynamic characteristics of the device under different test conditions. The input layer receives the sample feature vector X constructed in step 3; it includes time-domain, electrical, and waveform morphology features extracted from the double-pulse test waveform, and may also include external condition parameters such as gate turn-off voltage Vgs(off) and temperature T. The hidden layer uses a fully connected neural network (FCNN) or a tree-based ensemble method (such as random forest or gradient boosting tree) for nonlinear mapping to capture the complex relationship between input features and output parameters; the output layer generates continuous value regression results, including reverse recovery charge Qrr, reverse recovery peak current Irrm, reverse recovery time trr, and turn-on loss Eon and turn-off loss Eoff, enabling the prediction of key reverse recovery parameters.

[0068] like Figure 4 and Figure 10 As shown, the conduction path identification model is a classification model. The input layer takes the sample feature vector X as input and combines it with external conditions such as gate turn-off voltage Vgs(off) and temperature T to identify the main conduction path of SiC MOSFET during reverse recovery. The hidden layer uses a fully connected neural network or convolutional neural network (CNN) to extract features, or uses classification methods such as support vector machine (SVM) and gradient boosting tree (GBT) for discrimination. The output layer predicts the conduction path type, including MOS channel-dominated conduction, body diode-dominated conduction, Schottky diode-dominated conduction, or multi-path mixed conduction; thus providing a basis for subsequent conduction path decomposition and degradation analysis.

[0069] like Figure 12 As shown, the degradation state assessment model is a classification or scoring model based on the evolution sequence of key parameters. The input layer includes the current proportion and reverse recovery parameters of the MOS channel conduction path, body diode conduction path, and Schottky diode conduction path obtained from steps 5 and 6, i.e., the key parameter sequence (such as Qrr, Irrm, trr, Eon, Eoff) and the changing trend and waveform anomaly characteristics of the conduction path proportion. The hidden layer uses a fully connected neural network (FCNN) to process multi-dimensional input or timing features to capture the patterns of device changes over time, temperature, or cycle count. The output layer generates the device degradation state and risk level, including normal, mild degradation, or severe degradation, or outputs a degradation score or remaining lifetime. The degradation state assessment model can evaluate the current degradation state (normal, mild degradation, severe degradation) and risk level of the device and can be used for remaining lifetime prediction, providing a basis for device reliability assessment.

[0070] The machine learning model employs at least one of the following: random forest, support vector machine, gradient boosting tree, and fully connected neural network. When using raw time-series waveforms directly as input, these waveforms refer to continuous data sequences of voltage, current, etc., that change over time without feature extraction processing. For example, discrete sampling sequences of Vds(t) and Ids(t) within a certain time window, directly acquired, can be represented as a multidimensional time-series input arranged in chronological order. Long Short-Term Memory (LSTM) networks, convolutional neural networks, or a combination of both are also employed for modeling to enhance the ability to extract time-series information from dynamic waveforms.

[0071] The specific implementation process of step 5 includes: Step 5.1: Establish the equivalent current response functions I_MOS(t), I_body(t), and I_SBD(t) for the MOS channel conduction path, body diode conduction path, and Schottky diode conduction path, respectively, as follows: I_MOS(t)=f MOS (Vds(t),Vgs(t),T,Ɵ MOS ); I_body(t) = f body (Vds(t),Vgs(t),T,Ɵ body ); I_SBD(t)=f SBD (Vds(t),Vgs(t),T,Ɵ SBD ); Among them, Ɵ MOS , Ɵ body , Ɵ SBD These are the model parameters for the MOS channel conduction path, the body diode conduction path, and the Schottky diode conduction path, respectively, used to characterize the conduction response characteristics of each path under different test conditions; f MOS () represents the MOS channel current response function; f body () represents the body diode current response function; f SBD () represents the current response function of a Schottky diode; Vds(t) is the curve of drain-source voltage changing with time, i.e., the transient waveform of the voltage between the drain and source of a MOSFET over time; Vgs(t) is the curve of gate-source voltage changing with time, i.e., the transient waveform of the voltage between the gate and source of a MOSFET over time, used to control the on or off state of the device; T refers to the temperature under test or device operating conditions. The gate voltage directly regulates the conduction path of the MOS channel and indirectly couples it to the conduction paths of the body diode and the Schottky diode. Based on this, and combined with the conduction mode output by the conduction path identification model, constraints are applied to each path component. When the identification result is a MOS channel conduction path, I_MOS (i.e., I_MOS(t)) is set as the dominant component; when the identification result is a body diode conduction path, I_body (i.e., I_body(t)) is set as the dominant component; when the identification result is a Schottky diode conduction path, I_SBD (i.e., I_SBD(t)) is set as the dominant component; when the identification result is multi-path mixed conduction, at least two conduction paths are allowed to significantly participate in conduction simultaneously.

[0072] Step 5.2: Using the measured drain-source voltage Vds(t), drain current Ids(t), gate turn-off voltage Vgs(off), temperature T, reverse recovery charge Qrr, reverse recovery peak current Irrm, reverse recovery time trr, and waveform characteristics as constraints or optimization objectives, the current components I_MOS(t), I_body(t), and I_SBD(t) of each conduction path are obtained through parameter fitting, least squares optimization, or iterative solution; including: First, based on the equivalent current response functions of each conduction path established in step 5.1, the expression for the total current is constructed: I_total=I MOS (t,Ɵ MOS )+Ibody(t,Ɵbody)+I SBD (t,Ɵ SBD ); Using the measured drain current Ids(t) as a reference target, the error function, i.e. the objective function, is constructed as follows: 2 ; Simultaneously, the reverse recovery charge Qrr, peak current Irrm, reverse recovery time trr, and waveform morphology characteristics are transformed into additional constraints and added to the objective function to ensure that the decomposition results satisfy the actual physical characteristics. Based on this, prior constraints are applied in conjunction with the conduction path identification results, such as restricting a certain conduction path to be the dominant component or constraining its weight range.

[0073] Subsequently, the least squares method or iterative optimization algorithms (such as gradient descent, Levenberg-Marquardt method, etc.) are used to optimize the model parameters Ɵ. MOS Ɵbody、Ɵ SBD Solve the problem to minimize the objective function; After the parameters converge, the obtained optimal parameters are substituted into the response function of each conduction path to obtain the corresponding time-series current components. I_body(t) and This allows for the decomposition of the total current; Step 5.3: Calculate the proportion of each conduction path based on the proportional relationship between the current components of each conduction path and the total current: ; ; ; in, , and These represent the instantaneous proportions of the MOS channel conduction path, the body diode conduction path, and the Schottky diode conduction path in the total current, respectively; that is, the current proportions of the MOS channel conduction path, the body diode conduction path, and the Schottky diode conduction path. I_total = + + , This refers to the total current, i.e., the measured drain current Ids(t).

[0074] By averaging, integrating, or calculating the peak value of the instantaneous proportion over a specified time interval, the overall contribution ratio of each conduction path during the reverse recovery phase can be obtained, thereby enabling quantitative identification of the contribution of different conduction paths.

[0075] The specific implementation process of step 6 includes: Based on test data under different gate turn-off voltages Vgs(off) and temperature T, a mapping model between the reverse recovery parameters and external conditions is established, i.e., a multi-parameter coupling model of the reverse recovery characteristics: Qrr = f(Vgs(off), T, X); Where f represents the mapping relationship of the function, Qrr represents the value of the reverse recovery charge, and Qrr = f(Vgs(off),T, X) means that the reverse recovery charge is expressed as a function of the gate turn-off voltage Vgs(off), temperature T, and eigenvector X. Inverse recovery parameter prediction is achieved through a multi-parameter coupled model based on inverse recovery characteristics. This includes:

[0076] First, a training dataset is constructed based on a large number of sample data obtained under different gate turn-off voltages Vgs(off), temperatures T, and corresponding test conditions. Each sample includes input variables (Vgs(off), T, X) and corresponding output parameters Qrr, where X is the feature vector obtained in step 3. Secondly, a suitable model form is selected to establish a multi-parameter coupling function f(.). This function can be a machine learning model (such as random forest, gradient boosting tree, neural network, etc.) or a parameterized fitting function (such as multinomial regression, exponential function, or hybrid model). By fitting the training data, the multi-parameter coupling function f(.) learns the nonlinear mapping relationship between the input variables and the back-recovery charge.

[0077] During model training, the measured Qrr is used as the supervision signal. The model parameters are optimized by minimizing the error (e.g., mean square error) between the predicted and true values ​​to obtain the optimal mapping function f(.). After model training is complete, in the prediction phase, the Vgs(off), temperature T, and feature vector X of the sample to be tested are input into the model to obtain the corresponding predicted reverse recovery charge value Qrr. For other reverse recovery parameters (such as Irrm and trr), corresponding multi-parameter coupled models can be established in the same way to achieve joint or independent prediction of multiple parameters. Through the above process, the gate turn-off voltage, temperature, and device dynamic characteristics are coupled and modeled, thereby enabling the prediction of reverse recovery parameters under different operating conditions. Figure 5 This is a schematic diagram illustrating the multi-parameter coupling relationship between the reverse recovery parameters and the gate turn-off voltage and temperature in this invention. Figure 8 This is a schematic diagram showing the curves of reverse recovery charge Qrr as a function of gate turn-off voltage Vgs(off) at different temperatures. Figure 8 The data reflects the variation of Qrr as the gate turn-off voltage increases or decreases, as well as the moderating effect of temperature on this relationship, and is used to analyze the influence of gate turn-off voltage and temperature on reverse recovery behavior. Figure 9 This is a schematic diagram showing the curves of reverse recovery charge Qrr as a function of temperature under different gate turn-off voltages. Figure 9 The data reflects the change characteristics of Qrr as the temperature increases, as well as the differences in temperature sensitivity under different Vgs(off) conditions, providing a reference for studying the dynamic characteristics of devices under high and low temperature environments.

[0078] The specific implementation process of step 7 includes: Based on multiple sets of dual-pulse test data acquired at different times, the variation sequences of reverse recovery charge Qrr, reverse recovery peak current Irrm, reverse recovery time trr, switching loss Eon / Eoff, and conduction path ratio are extracted. An evolution model of the key parameters of the device as a function of time, thermal stress, electrical stress, or switching cycle number is established, namely the key parameter evolution model. Its essence is to express the parameters (such as Qrr, Irrm, trr, Eon / Eoff, and conduction path ratio) as functions of time or stress variables.

[0079] The key parameter evolution model includes the following: The power-law or linear extrapolation model of the change of key device parameters with the number of switching cycles n is: Qrr(n) = Qrr0 + a•n b Where Qrr0 is the initial value, a and b are the fitting parameters, and n is the number of switching cycles; The stress-driven model with respect to time t and temperature T is: Qrr(t,T) = Qrr0 + k•t• Where k is the rate coefficient, Ea is the activation energy, and k B Boltzmann's constant; The evolution model of the proportion of conduction paths is: αMOS(t) = f(t,T,stress condition), which is used to describe the changing trend of each path with time or stress.

[0080] Assess the current degradation status of the device based on the drift amplitude of key parameters, the changing trend of the conduction path ratio, and the frequency of abnormal waveforms; including: Key parameter drift magnitude; including: calculating the drift magnitude by comparing the changes in the currently tested key parameters (such as Qrr, Irrm, trr, Eon, Eoff) with the changes in the device's initial state or reference value: ΔQrr When the drift amplitude exceeds the set threshold, it indicates that the device performance has degraded; The trend of the proportion of current in the conduction path; including: analyzing the trend of the current proportion of different conduction paths (MOS channel, body diode, SBD) with time, number of cycles or test conditions; if the proportion of a certain path continues to decrease or increases abnormally, it indicates that the conduction mechanism may have changed, which is a signal of device degradation; Frequency of abnormal waveform occurrences; including: statistically analyzing the frequency of abnormal features in the reverse recovery waveform, such as: excessively large current second peak, increased oscillation amplitude, and prolonged tail current. Frequent occurrence of abnormal waveforms indicates abnormal carrier distribution or parasitic effects within the device, suggesting degradation.

[0081] Comprehensive evaluation; including: combining key parameter drift amplitude, conduction path ratio changes, and abnormal waveform frequency, and conducting comprehensive analysis through weighted scoring or classification models; based on the comprehensive results, classifying the current state of the device as follows: Key parameters show minimal drift, the proportion of conduction paths is stable, there are few abnormal waveforms, and the current state of the device is normal. Some parameters have drifted slightly, the proportion of conduction paths has changed slightly, and abnormal waveforms occasionally appear. The current state of the device is mild degradation. Key parameters have drifted significantly, the proportion of conduction paths has changed markedly, abnormal waveforms are frequent, and the device is currently in a state of severe degradation. "Small drift, slight drift, significant drift" – set the rate of change D of the key parameter relative to its initial value. P for: D P = Where P is the current test parameter and P0 is the initial health status parameter, including Qrr, Irrm, trr, Eon, and Eoff; All key parameters D P When the percentage is less than 5%, the drift of the key parameter is small; at least one key parameter 5% ≤ D P <15%, indicating slight drift in some parameters; at least one critical parameter D P ≥15% indicates a significant drift in key parameters; "The proportion of conductive paths is stable, slightly changing, or significantly changing," let's define the change D of the proportion of conductive paths relative to the initial state. η :D η =|η-η0|, where η is the current percentage of a certain conductive path, and η0 is the percentage of that path in the initial state; The change in the proportion of each path is less than 5%, indicating a stable proportion of conductive paths; the change in the proportion of at least one path is 5%-15%, indicating a slight change in the proportion of conductive paths; the change in the proportion of at least one path is D. η A value greater than or equal to 15% indicates a significant change in the proportion of the conduction path; "abnormal waveforms appearing infrequently, occasionally, or frequently" sets the frequency F of abnormal waveform occurrence. abn :F abn = Where, N abn N represents the number of abnormal waveforms. total This represents the total number of test waveforms; abnormal waveforms include abnormally increased secondary current peaks, abnormally increased oscillation amplitudes, and prolonged tailing currents.

[0082] An abnormal waveform frequency of less than 5% indicates few abnormal waveforms; an abnormal waveform frequency of 5%-20% indicates occasional abnormal waveforms; and an abnormal waveform frequency of 20% or more indicates frequent abnormal waveforms.

[0083] Status determination: (1) Normal: Key parameter drift amplitude <5%, conduction path ratio change less than 5%, abnormal waveform frequency <5%; (2) Mild degradation: Any indicator reaches the range of mild degradation, but does not reach the range of severe degradation; (3) Severe degradation: any key parameter drift ≥15%, or any conduction path percentage change ≥15%, or abnormal waveform frequency ≥20%.

[0084] This assessment method uses quantitative indicators combined with historical and real-time test data to determine the degree of device degradation.

[0085] Based on the degradation state assessment results and parameter evolution trends, the remaining usable lifetime of the device is predicted, and the failure risk level is output; including: Remaining usable lifetime prediction, including: Based on the evolution model of key parameters, the trend curves of key parameters changing with time, number of cycles or stress conditions are obtained; for example, Qrr(t), Irrm(t), trr(t), etc.

[0086] Set device failure thresholds (e.g., Qrr exceeds the specified range, or trr reaches the safety limit). Extrapolating parameter changes using a key parameter evolution model, the time or number of switching cycles required to reach the failure threshold is calculated: t fail or n fail =f −1 (Parameter threshold); The difference between the current time or number of cycles and the predicted failure time / number of cycles is the remaining usable lifetime (RUL): RUL = t fail -t current Or RUL=n fail -n current ; Among them, t fail With n fail =f −1 Meaning: f represents the critical parameter evolution model; it describes the relationship between critical device parameters and time, cycle number, or stress conditions. For example: Qrr=f(t) or Qrr=f(n), where Qrr is the reverse recovery charge, t is time, and n is the number of switching cycles. −1 This represents the inverse function; that is, given a key parameter value (such as a failure threshold), it calculates the corresponding time or number of iterations. Parameter threshold: A predefined device failure standard, for example: Qrr > Q. threshold 、trr>t threshold ;t fail This represents the time required to predict and reach the failure parameter threshold; t current This represents the actual operating time of the device, that is, the time elapsed from the start of use or testing to the current moment; n fail This represents the number of switching cycles required to predict and reach the failure parameter threshold; n current This indicates the number of switching cycles the device has undergone so far, i.e., the cumulative number of switching cycles from the start of its use to the present.

[0087] Failure risk level output includes: Risk levels are determined based on the following indicators: the proximity of current key parameters to failure thresholds, parameter change rate (drift rate), trend of conduction path proportion changes, and frequency of abnormal waveform occurrences. Devices are then classified into risk levels according to a comprehensive scoring or classification model.

[0088] When the parameters are stable and far from the threshold, the risk is low. When the parameter deviation is significant and close to the threshold, it is considered medium risk; When parameters approach or exceed the threshold, frequent abnormal waveforms indicate high risk. The terms "parameter stable, far from the threshold, significant deviation, and close to the threshold" can be defined by the degree to which the parameter approaches the failure threshold. The degree to which the set parameter R approaches the failure threshold P :R P = ×100%, where, The failure threshold is defined by P0 as the initial value and P as the current value; R P =0% indicates a near-initial healthy state; R P <70% is considered far from the threshold; 70% ≤ R P <90% is close to the threshold; R P ≥90% is close to the threshold; R P =100% indicates that the failure threshold has been reached.

[0089] When the parameter change rate is less than 5% in multiple consecutive tests, it indicates that the parameter is stable; when the parameter change rate reaches or exceeds 10%, or R... P A value of ≥70% indicates a significant parameter offset.

[0090] When all key parameters R P Less than 70%, abnormal waveform frequency F abn Less than 5%, and the change in the proportion of the conduction path D η Less than 5% is considered low risk. When any key parameter 70%≤R P <90%, or abnormal waveform frequency F abn The percentage is 5% to 20%, or the proportion of the conduction path changes by D. η The risk level is 5% to 15%, which is considered medium risk. When any key parameter R P ≥90% or has exceeded the failure threshold, or abnormal waveform frequency F abn ≥20%, or change in the proportion of conduction path D η ≥15% indicates high risk.

[0091] The implementation method is as follows: Extrapolation of fitted functions (regression / exponential models): By using data on the changes of key parameters over time, cycle number, or stress conditions, these curves are fitted with mathematical functions, such as linear regression (assuming that the parameters change linearly over time) and exponential models (assuming that the parameters show an accelerating or decaying trend over time). After fitting, the function is extrapolated (predicting future time or cycle number) to calculate the time or number of cycles required for the parameters to reach the failure threshold, thereby obtaining the remaining lifetime (RUL).

[0092] Prediction based on machine learning models (regression + classification): ① Train a machine learning model using historical test data and feature vectors: Regression models predict future values ​​of key parameters (such as Qrr, Irrm, trr); classification models predict degradation status or risk level (such as normal, mild degradation, severe degradation). ② For new devices under test or new test conditions, input the feature vector, and the model can output the prediction results.

[0093] The output includes the remaining lifespan value and the corresponding risk level label; ① Remaining lifetime value: The estimated time or number of switching cycles that a device can safely operate under current conditions (RUL). ② Risk Level Labeling: Based on the degree of parameter deviation, conduction path changes, and abnormal waveform frequencies, assess the degradation risk of the device, for example: Low risk: parameters are stable and far from the threshold; Medium risk: Parameter deviation is significant and close to the threshold; High risk: Parameters are close to or exceed the threshold, and abnormal waveforms occur frequently; These outputs can be used for equipment maintenance decisions and reliability assessments, helping to predict the remaining lifespan of devices and schedule maintenance or replacement in advance.

[0094] This invention does not directly output the proportion of each path from the conduction path identification model. Instead, it uses the conduction path identification results as a priori constraints, combines the total current decomposition relationship and measured dynamic waveform characteristics, and fits and solves the current components of each conduction path to achieve quantitative extraction of the contribution ratio of each conduction path.

[0095] Example 3 A system for identifying the reverse recovery conduction path and predicting the dynamic characteristics of SiC MOSFETs based on dual-pulse testing includes: The waveform data acquisition module is configured to: build a dual-pulse test circuit to acquire the drain-source voltage Vds(t), drain current Ids(t), external gate drive resistance, load inductance, temperature, load current, bus voltage, and gate drive voltage (turn-on voltage Vgs(on) and turn-off voltage Vgs(off)) of the SiC MOSFET under test during the switching and reverse recovery processes; the existing dual-pulse test system (DPT equipment) can directly complete the circuit construction and the acquisition of waveforms and test condition parameters such as Vds, Id, and Vgs.

[0096] The preprocessing module is configured to preprocess the acquired waveform data (including the drain-source voltage Vds(t), drain current Ids(t), external gate drive resistance, load inductance, temperature, load current, bus voltage, and gate drive voltage of the tested SiC MOSFET during the switching and reverse recovery processes) by performing noise filtering, time axis alignment, amplitude normalization, and outlier removal to improve data quality and model stability. The feature parameter extraction module is configured to extract feature parameters from the preprocessed waveform data using existing methods, including numerical integration (for Qrr), numerical differentiation (for di / dt, dv / dt), and peak detection (Irrm, etc.); and construct feature vectors. The feature parameters include time-domain features, electrical features, waveform morphology features, and test condition features. The time-domain features characterize the time response of the device during turn-on, turn-off, and reverse recovery; the electrical features characterize the reverse recovery and switching loss characteristics of the device; the waveform morphology features characterize dynamic features such as waveform peak value, oscillation, and tailing; and the test condition features include gate turn-off voltage, bus voltage, load current, external gate resistance, and temperature. The machine learning module is configured to input feature vectors into the trained machine learning model; Establish a mapping relationship between the input feature vector and the key parameters of reverse recovery, and output the reverse recovery charge, reverse recovery peak current, reverse recovery time, and switching loss parameters; including: A training set is built using the sample data from the completed tests. Each sample includes an input sample feature vector X and the corresponding true output parameters, such as Qrr, Irrm, trr, Eon, and Eoff. The true output parameters can be obtained from the measured waveforms through integration, peak extraction, time determination, and energy integration.

[0097] The sample feature vector X is input into the regression model for training, so that the regression model learns the nonlinear mapping relationship between the input features and the key parameters of reverse recovery. After training is complete, for a new test sample, input the feature vector of the new test sample and output the prediction results, including reverse recovery charge, reverse recovery peak current, reverse recovery time and switching loss parameters.

[0098] The system outputs the conduction path type and the device degradation state; the conduction path type output is part of the classification and recognition process. First, based on existing experimental data, third-quadrant conduction characteristics, and differences in reverse recovery waveforms at different gate turn-off voltages and temperatures, samples are labeled, for example, as: MOS channel dominant conduction, body diode dominant conduction, Schottky diode dominant conduction, or multi-path hybrid conduction. Then, the feature vector X is used as input, and the conduction path type is used as the output label to train the classification model. After training, inputting the test sample into the model will output the corresponding conduction path type.

[0099] Device degradation status output can also be achieved through classification or scoring methods. Based on the changing trends of key parameters during repeated tests, aging tests, or stress tests, such as the drift magnitude of parameters like Qrr, Irrm, Eon, Eoff, and conduction path percentage, samples are labeled as normal, slightly degraded, or significantly degraded. After model training is complete, the current degradation status of the device can be output based on the feature vector of the test sample, or the corresponding degradation score and failure risk level can be output.

[0100] The current percentage calculation module is configured to: establish a conduction path decomposition model based on the conduction path type, combined with the total current decomposition relationship and the dynamic response characteristics of the device under different test conditions. The conduction path decomposition model includes the equivalent current response functions of the MOS channel conduction path, the body diode conduction path and the Schottky diode conduction path, and calculate the current percentage of the MOS channel conduction path, the body diode conduction path and the Schottky diode conduction path. The reverse recovery parameter prediction module is configured to: introduce gate turn-off voltage and temperature parameters, establish a multi-parameter coupled model of reverse recovery characteristics, and realize reverse recovery parameter prediction. The device status assessment, remaining lifetime prediction, and failure risk assessment module is configured to: construct a device key parameter evolution model based on multiple double-pulse test data, and realize device status assessment, remaining lifetime prediction, and failure risk assessment based on the changing trends of reverse recovery parameters, switching loss parameters, and conduction path ratio.

Claims

1. A method for identifying the reverse recovery conduction path and predicting the dynamic characteristics of SiC MOSFETs based on dual-pulse testing, characterized in that, include: Step 1: Build a dual-pulse test circuit to collect the drain-source voltage Vds(t), drain current Ids(t), external gate drive resistance, load inductance, temperature, load current, bus voltage, and gate drive voltage of the SiC MOSFET under test during the switching and reverse recovery processes. Step 2: Preprocess the waveform data acquired in Step 1, including: noise filtering, time axis alignment, amplitude normalization, and outlier data removal. Step 3: Extract feature parameters from the waveform data preprocessed in Step 2 and construct a feature vector; the feature parameters include time-domain features, electrical features, waveform morphology features and test condition features, wherein the time-domain features are used to characterize the time response of the device during turn-on, turn-off and reverse recovery processes, the electrical features are used to characterize the reverse recovery and switching loss characteristics of the device, the waveform morphology features are used to characterize the dynamic features, and the test condition features include gate turn-off voltage, bus voltage, load current, external gate resistance and temperature; Step 4: Input the feature vector into the trained machine learning model; Establish the mapping relationship between the input feature vector and the key parameters of reverse recovery, and output the reverse recovery charge, reverse recovery peak current, reverse recovery time and switching loss parameters; Output the conduction path type and the device degradation status; Step 5: Based on the conduction path type, combined with the total current decomposition relationship and the dynamic response characteristics of the device under different test conditions, establish a conduction path decomposition model. The conduction path decomposition model includes the equivalent current response functions of the MOS channel conduction path, the body diode conduction path, and the Schottky diode conduction path. Calculate the current proportion of the MOS channel conduction path, the body diode conduction path, and the Schottky diode conduction path. Step 6: Introduce gate turn-off voltage and temperature parameters, establish a multi-parameter coupled model of reverse recovery characteristics, and realize reverse recovery parameter prediction; Step 7: Construct a device key parameter evolution model based on multiple double-pulse test data, and realize device status assessment, remaining lifetime prediction and failure risk assessment according to the changing trends of reverse recovery parameters, switching loss parameters and conduction path ratio.

2. The method for identifying the reverse recovery conduction path and predicting the dynamic characteristics of SiC MOSFETs based on dual-pulse testing according to claim 1, characterized in that, In step 3, the time-domain features include: turn-on delay time td(on), turn-off delay time td(off), current rise time tr, voltage fall time tf, and reverse recovery time trr; Electrical characteristics include: reverse recovery charge Qrr, reverse recovery peak current Irrm, turn-on loss Eon, turn-off loss Eoff, current change rate di / dt, and voltage change rate dv / dt. Waveform characteristics include: current second peak characteristics, oscillation frequency and damping characteristics, tail current characteristics, and voltage overshoot amplitude; The characteristics of the second peak current include: the amplitude of the second peak current I. 2nd The time t for the second peak to occur 2nd The ratio of the secondary peak current to the primary peak current Irrm; oscillation frequency and damping characteristics, including: oscillation frequency fosc, oscillation period Tosc, and damping coefficient or attenuation rate; tail current characteristics, including: tail duration t. tail The magnitude of the trailing current and the rate of trailing decay; Test conditions include: gate turn-off voltage Vgs(off), bus voltage, load current, external gate drive resistance and temperature, load inductance, and gate drive voltage test condition parameters. The feature parameters are uniformly encoded to construct the sample feature vector X.

3. The method for identifying the reverse recovery conduction path and predicting the dynamic characteristics of SiC MOSFETs based on dual-pulse testing according to claim 2, characterized in that, Step 4, the process of building and training the machine learning model, includes: Constructing Sample Labels: Based on the collected waveform data and extracted feature parameters, construct the output labels required for supervised learning; among them, reverse recovery charge Qrr, reverse recovery peak current Irrm, reverse recovery time trr, turn-on loss Eon, and turn-off loss Eoff are used as parameter prediction labels; the conduction path type of the samples is labeled to obtain classification labels for MOS channel dominant conduction, body diode dominant conduction, Schottky diode dominant conduction, or multi-path mixed conduction; the samples are labeled with normal, slightly degraded, or significantly degraded states to form degradation state labels; Constructing machine learning models: The machine learning models include a reverse recovery parameter prediction model, a conduction path identification model, and a degradation state assessment model; the reverse recovery parameter prediction model, the conduction path identification model, and the degradation state assessment model are established respectively; the reverse recovery parameter prediction model is a regression model, used to establish the mapping relationship between the input feature vector and the key parameters of reverse recovery, and outputs the reverse recovery charge Qrr, the reverse recovery peak current Irrm, the reverse recovery time trr, and the turn-on loss Eon and the turn-off loss Eoff; The conduction path identification model is used to output the conduction path type of the device under the current test conditions; The degradation state assessment model is used to output the current degradation state of the device; Machine learning model training and optimization: The samples are divided into training set, validation set and test set. The input features are standardized or normalized to complete the training of the machine learning model. The machine learning model is then evaluated according to the indicators and the parameters are optimized to obtain the trained machine learning model.

4. The method for identifying the reverse recovery conduction path and predicting the dynamic characteristics of SiC MOSFETs based on dual-pulse testing according to claim 3, characterized in that, The reverse recovery parameter prediction model includes an input layer, a hidden layer, and an output layer. The input layer receives the sample feature vector X constructed in step 3. The hidden layer uses a fully connected neural network or a tree-based ensemble method for nonlinear mapping to capture the complex relationship between the input features and the output parameters. The output layer generates continuous value regression results, including the reverse recovery charge Qrr, the reverse recovery peak current Irrm, the reverse recovery time trr, and the turn-on loss Eon and turn-off loss Eoff, thereby enabling the prediction of key reverse recovery parameters. The conduction path identification model is a classification model. The input layer takes the sample feature vector X as input, and the hidden layer uses a fully connected neural network or a convolutional neural network to extract features, or uses support vector machine or gradient boosting tree classification methods for discrimination. The output layer predicts the conduction path type, including MOS channel dominant conduction, body diode dominant conduction, Schottky diode dominant conduction, or multi-path hybrid conduction. The degradation status assessment model is a classification or scoring model based on the evolution sequence of key parameters. The input layer inputs include the current ratio and reverse recovery parameters of the MOS channel conduction path, body diode conduction path, and Schottky diode conduction path obtained from steps 5 and 6. The hidden layer uses a fully connected neural network to process multidimensional inputs or timing features. The output layer generates the device degradation status and risk level, including normal, mild degradation, or severe degradation, or outputs a degradation score or remaining lifetime.

5. The method for identifying the reverse recovery conduction path and predicting the dynamic characteristics of SiC MOSFETs based on dual-pulse testing according to claim 1, characterized in that, The machine learning model employs at least one of the following: random forest, support vector machine, gradient boosting tree, and fully connected neural network.

6. The method for identifying the reverse recovery conduction path and predicting the dynamic characteristics of SiC MOSFETs based on dual-pulse testing according to claim 1, characterized in that, The specific implementation process of step 5 includes: Step 5.1: Establish the equivalent current response functions I_MOS(t), I_body(t), and I_SBD(t) for the MOS channel conduction path, body diode conduction path, and Schottky diode conduction path, respectively, as follows: I_MOS(t)=f MOS (Vds(t),Vgs(t),T,Ɵ MOS ); I_body(t)=f body (Vds(t),Vgs(t),T,Ɵ body ); I_SBD(t)=f SBD (Vds(t),Vgs(t),T,Ɵ SBD ); Among them, Ɵ MOS , Ɵ body , Ɵ SBD These are the model parameters for the MOS channel conduction path, the body diode conduction path, and the Schottky diode conduction path, respectively, used to characterize the conduction response characteristics of each path under different test conditions; f MOS () represents the MOS channel current response function; f body () represents the body diode current response function; f SBD () represents the current response function of a Schottky diode; Vds(t) is the curve of drain-source voltage changing with time, i.e., the transient waveform of the voltage between the drain and source of a MOSFET over time; Vgs(t) is the curve of gate-source voltage changing with time, i.e., the transient waveform of the voltage between the gate and source of a MOSFET over time, used to control the on or off state of the device; T refers to the temperature under test or device operating conditions. Apply constraints: When the identification result is a MOS channel conduction path, set I_MOS (i.e., I_MOS(t)) as the dominant component; when the identification result is a body diode conduction path, set I_body (i.e., I_body(t)) as the dominant component; when the identification result is a Schottky diode conduction path, set I_SBD (i.e., I_SBD(t)) as the dominant component; when the identification result is a multi-path mixed conduction, allow at least two conduction paths to participate significantly in conduction simultaneously. Step 5.2: Using the measured drain-source voltage Vds(t), drain current Ids(t), gate turn-off voltage Vgs(off), temperature T, reverse recovery charge Qrr, reverse recovery peak current Irrm, reverse recovery time trr, and waveform characteristics as constraints or optimization objectives, the current components I_MOS(t), I_body(t), and I_SBD(t) of each conduction path are obtained through parameter fitting, least squares optimization, or iterative solution; including: First, based on the equivalent current response functions of each conduction path established in step 5.1, the expression for the total current is constructed: I_total=I MOS (t,Ɵ MOS )+Ibody(t,Ɵbody)+I SBD (t,Ɵ SBD ); Using the measured drain current Ids(t) as a reference target, the error function, i.e. the objective function, is constructed as follows: 2 ； Meanwhile, the reverse recovery charge Qrr, peak current Irrm, reverse recovery time trr, and waveform morphology characteristics are transformed into additional constraints and added to the objective function; Subsequently, the least squares method or iterative optimization algorithm is used to optimize the model parameters Ɵ. MOS ,Ɵbody,Ɵ SBD Solve the problem to minimize the objective function; After the parameters converge, the obtained optimal parameters are substituted into the response function of each conduction path to obtain the corresponding time-series current components. I_body(t) and This allows for the decomposition of the total current; Step 5.3: Calculate the proportion of each conduction path based on the proportional relationship between the current components of each conduction path and the total current: ; ； ； in, , and These represent the instantaneous proportions of the MOS channel conduction path, the body diode conduction path, and the Schottky diode conduction path in the total current, respectively; that is, the current proportions of the MOS channel conduction path, the body diode conduction path, and the Schottky diode conduction path. I_total = + + , This refers to the total current, i.e., the measured drain current Ids(t).

7. The method for identifying the reverse recovery conduction path and predicting the dynamic characteristics of SiC MOSFETs based on dual-pulse testing according to claim 1, characterized in that, The specific implementation process of step 6 includes: Based on test data under different gate turn-off voltages Vgs(off) and temperature T, a mapping model between the reverse recovery parameters and external conditions is established, i.e., a multi-parameter coupling model of the reverse recovery characteristics: Qrr = f(Vgs(off), T, X); Where f represents the mapping relationship of the function, Qrr represents the value of the reverse recovery charge, and Qrr = f(Vgs(off), T, X) means that the reverse recovery charge is expressed as a function of the gate turn-off voltage Vgs(off), temperature T, and eigenvector X. Inverse recovery parameter prediction is achieved through a multi-parameter coupling model of inverse recovery characteristics.

8. The method for identifying the reverse recovery conduction path and predicting the dynamic characteristics of SiC MOSFETs based on dual-pulse testing according to claim 1, characterized in that, In step 7, based on multiple sets of double-pulse test data acquired at different times, the change sequences of reverse recovery charge Qrr, reverse recovery peak current Irrm, reverse recovery time trr, switching loss Eon / Eoff, and conduction path ratio are extracted to establish an evolution model of the device's key parameters as a function of time, thermal stress, electrical stress, or switching cycle number, i.e., the key parameter evolution model. Assess the current degradation status of the device based on the drift amplitude of key parameters, the changing trend of the conduction path ratio, and the frequency of abnormal waveforms; including: The drift amplitude of key parameters, changes in the proportion of conduction paths, and the frequency of abnormal waveforms are combined and analyzed comprehensively using a weighted scoring or classification model. Based on the comprehensive results, the current state of the device is classified as follows: Key parameters show minimal drift, the proportion of conduction paths is stable, there are few abnormal waveforms, and the current state of the device is normal. Some parameters have drifted slightly, the proportion of conduction paths has changed slightly, and abnormal waveforms occasionally appear. The current state of the device is mild degradation. Key parameters have drifted significantly, the proportion of conduction paths has changed markedly, abnormal waveforms are frequent, and the device is currently in a state of severe degradation. Set the rate of change D of the key parameter relative to the initial value. P For: D P = Where P is the current test parameter and P0 is the initial health status parameter, including Qrr, Irrm, trr, Eon, and Eoff; All key parameters D P When the percentage is less than 5%, the drift of the key parameter is small; at least one key parameter 5% ≤ D P <15%, indicating slight drift in some parameters; at least one critical parameter D P ≥15% indicates a significant drift in key parameters; Set the change D of the proportion of the conducting path relative to the initial state. η :D η =|η-η0|, where η is the current percentage of a certain conductive path, and η0 is the percentage of that path in the initial state; The change in the proportion of each path is less than 5%, indicating a stable proportion of conductive paths; the change in the proportion of at least one path is 5%-15%, indicating a slight change in the proportion of conductive paths; the change in the proportion of at least one path is D. η A value of 15% or higher indicates a significant change in the proportion of the conductive path. Set the frequency F of abnormal waveform occurrence abn :F abn = Where, N abn N represents the number of abnormal waveforms. total This represents the total number of test waveforms. An abnormal waveform frequency of less than 5% indicates few abnormal waveforms; an abnormal waveform frequency of 5%-20% indicates occasional abnormal waveforms; and an abnormal waveform frequency of 20% or more indicates frequent abnormal waveforms.

9. The method for identifying the reverse recovery conduction path and predicting the dynamic characteristics of SiC MOSFETs based on dual-pulse testing according to claim 8, characterized in that, In step 7, based on the degradation state assessment results and parameter evolution trends, the remaining usable lifetime of the device is predicted, and the failure risk level is output; including: Remaining usable lifetime prediction, including: Based on the key parameter evolution model, the trend curves of key parameters changing with time, number of cycles, or stress conditions are obtained; Set the device failure threshold; Extrapolating parameter changes using a key parameter evolution model, the time or number of switching cycles required to reach the failure threshold is calculated: t fail or n fail =f −1 ; The difference between the current time or number of cycles and the predicted failure time / number of cycles is the remaining usable lifetime (RUL): RUL = t fail -t current Or RUL=n fail -n current ; Where f represents the key parameter evolution model; f −1 t represents finding the inverse function; fail This represents the time required to predict and reach the failure parameter threshold; t current This represents the actual operating time of the device, that is, the time elapsed from the start of use or testing to the current moment; n fail This represents the number of switching cycles required to predict and reach the failure parameter threshold; n current This indicates the number of switching cycles the device has undergone so far, i.e., the cumulative number of switching cycles from the start of its use to the present. Failure risk level output includes: The degree to which the set parameter R approaches the failure threshold P :R P = ×100%, where, The failure threshold is defined by P0 as the initial value and P as the current value; when all key parameters R... P Less than 70%, abnormal waveform frequency F abn Less than 5%, and the change in the proportion of the conduction path D η Less than 5% is considered low risk. When any key parameter 70%≤R P <90%, or abnormal waveform frequency F abn The percentage is 5% to 20%, or the proportion of the conduction path changes by D. η The risk level is 5% to 15%, which is considered medium risk. When any key parameter R P ≥90% or has exceeded the failure threshold, or abnormal waveform frequency F abn ≥20%, or change in the proportion of conduction path D η ≥15% indicates high risk.

10. A system for identifying the reverse recovery conduction path and predicting the dynamic characteristics of SiC MOSFETs based on dual-pulse testing, characterized in that, include: The waveform data acquisition module is configured to: build a dual-pulse test circuit to acquire the drain-source voltage Vds(t), drain current Ids(t), external gate drive resistance, load inductance, temperature, load current, bus voltage, and gate drive voltage of the SiC MOSFET under test during the switching process and reverse recovery process. The preprocessing module is configured to preprocess the acquired waveform data, including noise filtering, time axis alignment, amplitude normalization, and outlier removal. The feature parameter extraction module is configured to extract feature parameters from the preprocessed waveform data and construct a feature vector. The feature parameters include time-domain features, electrical features, waveform morphology features, and test condition features. The time-domain features are used to characterize the time response of the device during turn-on, turn-off, and reverse recovery. The electrical features are used to characterize the reverse recovery and switching loss characteristics of the device. The waveform morphology features are used to characterize the dynamic features. The test condition features include gate turn-off voltage, bus voltage, load current, external gate resistance, and temperature. The machine learning module is configured to input the feature vector into the trained machine learning model; Establish the mapping relationship between the input feature vector and the key parameters of reverse recovery, and output the reverse recovery charge, reverse recovery peak current, reverse recovery time and switching loss parameters; output the conduction path type and output the device degradation status; The current percentage calculation module is configured to: establish a conduction path decomposition model based on the conduction path type, combined with the total current decomposition relationship and the dynamic response characteristics of the device under different test conditions. The conduction path decomposition model includes the equivalent current response functions of the MOS channel conduction path, the body diode conduction path and the Schottky diode conduction path, and calculate the current percentage of the MOS channel conduction path, the body diode conduction path and the Schottky diode conduction path. The reverse recovery parameter prediction module is configured to: introduce gate turn-off voltage and temperature parameters, establish a multi-parameter coupled model of reverse recovery characteristics, and realize reverse recovery parameter prediction. The device status assessment, remaining lifetime prediction, and failure risk assessment module is configured to: construct a device key parameter evolution model based on multiple double-pulse test data, and realize device status assessment, remaining lifetime prediction, and failure risk assessment based on the changing trends of reverse recovery parameters, switching loss parameters, and conduction path ratio.

Images