Flywheel motor driving si c power tube gate oxide degradation monitoring method and system

Abstract

The application relates to a flywheel motor driving SiC power tube gate oxide degradation monitoring method and system, and relates to the field of power semiconductor device health monitoring. The application is to solve the problem that the existing gate oxide degradation monitoring method is difficult to separate and dynamically track the oxide trap charge and interface state trap charge contribution from the switch waveform coupled with the working condition. The application collects the transient window signal and working condition information in the SiC MOSFET on-off process; the oxide trap charge and the interface state trap charge are defined as the hidden state of the recursive neural network, the oxide trap charge and the interface state trap charge of the current on-off process are updated through the gate recurrent mechanism based on the transient window signal, the working condition information and the hidden state of the last on-off process; and the gate oxide degradation result of the SiC MOSFET is obtained based on the physical relationship between the trap charge and the threshold voltage drift according to the updated oxide trap charge and interface state trap charge.

Description

Technical Field

[0001] This application belongs to the field of health monitoring of power semiconductor devices, and in particular relates to a method and system for monitoring the gate oxide degradation of SiC power transistors driven by flywheel motors. Background Technology

[0002] SiC MOSFETs (Silicon Carbide Metal-Oxide-Semiconductor Field-Effect Transistors) have been widely used in high-power-density power conversion, new energy converters, aerospace power systems, and high-reliability power electronic equipment due to their advantages such as high breakdown electric field, high switching speed, high-temperature operation capability, and low conduction loss. However, the gate oxide layer and its interface are critical weak points of SiC MOSFETs. Under long-term gate bias, electrothermal stress, power cycling, irradiation stress, and high-frequency switching stress, oxide trap charges and interface state trap charges will gradually be generated and evolve, causing threshold voltage drift, channel mobility degradation, and parasitic capacitance changes, ultimately reducing the device drive margin and system safety margin. Therefore, online monitoring of the gate oxide degradation state is crucial.

[0003] Currently, existing methods for monitoring grid oxide degradation can be mainly classified into the following categories:

[0004] (1) Static electrical parameter method

[0005] This method reflects the degree of degradation by measuring static parameters such as threshold voltage, on-state resistance, and gate leakage current in offline or shutdown states, with clear physical meaning. However, its shortcomings stem from the measurement principle itself: static parameter measurements usually require isolating the device under test from the operating circuit and applying specific test voltages or currents, thus necessitating the interruption of the converter's normal operation (such as shutdown or bypass), making it impossible to continuously track the degradation trajectory while the device is subjected to actual operating stresses (such as high-frequency switching, temperature changes, and load fluctuations). This "offline" measurement method not only leads to monitoring lag but also loses crucial information on the dynamic evolution of trapped charges under stress.

[0006] (2) Dynamic switching characteristic method

[0007] This method indirectly characterizes degradation using online waveform features such as turn-on delay, turn-off delay, Miller plateau time, drain current rate of change, or Rogowski coil signal. Its main drawback stems from the mixed sensitivity of these features: these switching characteristics are not only affected by threshold voltage and capacitance changes caused by gate oxide degradation, but are also highly susceptible to disturbances from bus voltage, load current, gate resistance, junction temperature, and PCB parasitic parameters. Due to the lack of a mechanism to separate the effects of these operating conditions from the degradation effect, the correspondence between single waveform features under different operating conditions drifts significantly, making it difficult to stably and uniquely map them to the gate oxide degradation state. This often requires complex calibration or compensation in engineering, limiting its versatility.

[0008] (3) Capacitance or gate charge measurement method

[0009] This method infers degradation by monitoring changes in input capacitance, gate-source capacitance, gate-drain capacitance, or gate charge. However, its main drawback stems from the engineering complexity of the online measurement chain: to accurately capture minute changes in capacitance or charge during normal device switching, it typically requires injecting a high-frequency small signal, configuring high-bandwidth sampling circuits, analog switches, comparators, or latches. This not only increases system cost and PCB area, but the injected signal may also conflict with the converter's electromagnetic compatibility requirements, making it difficult to deploy in practical compact power electronic equipment and hindering its widespread adoption.

[0010] (4) Deep learning methods based on pure data-driven approaches

[0011] This type of method can handle high-dimensional noisy waveforms and, theoretically, can learn a nonlinear mapping from operating conditions to degradation. However, its fundamental flaw stems from neglecting the physical structure: the model learns the statistical relationship between the input waveform and the degradation label entirely from the data, lacking causal chain constraints between gate dynamics, trap charge evolution, and threshold voltage drift. Therefore, the model easily degenerates into a "black box" regressor, extremely sensitive to the label density and operating condition distribution in the training data; when the device experiences new operating conditions or stress combinations not present in the training set, the model predictions are highly prone to failure. Furthermore, purely data-driven methods struggle to explain the contributions of oxide trap charges and interface state trap charges, failing to meet the needs of mechanistic analysis and reliable evaluation.

[0012] (5) Existing physical information machine learning methods

[0013] In recent years, some studies have begun to incorporate physical laws into machine learning, but most remain at the level of "physical feature preprocessing," "physical parameters as auxiliary inputs," or "adding soft constraints to the loss function." The shortcomings of these methods stem from the shallow and passive nature of physical embedding: physical laws are only used as input features or regularization terms, and are not truly embedded into the network's recursive state update process. Therefore, the model still cannot utilize the causal recursive relationship between SiC MOSFET gate dynamics, trap charge evolution, parasitic capacitance changes, and threshold voltage drift to constrain state estimation, resulting in difficulties in effectively separating the dynamic evolution laws of oxide trap charges and interface state trap charges in sparse labeling and cross-condition scenarios.

[0014] In summary, how to embed the SiC MOSFET gate dynamic equation, the physical causal chain of trap charge and threshold voltage drift into the recursive state estimation process, so as to achieve online recursive estimation of double trap charge without frequent shutdowns, without complex injection circuits, and only relying on turn-on transient window signals and operating condition information, is a technical problem that urgently needs to be solved. Summary of the Invention

[0015] This application addresses the challenge of existing gate oxide degradation monitoring methods in separating and dynamically tracking the contributions of oxide trap charges and interface state trap charges from strongly coupled switching waveforms under online, low-intrusion conditions. It provides a gate oxide degradation monitoring method and system for flywheel motor-driven SiC power transistors based on a physical recurrent neural network. By acquiring the transient window signal during device turn-on, the method recursively estimates oxide trap charges and interface state trap charges online, and further outputs threshold voltage drift, gate oxide degradation status, and health warning results under reference conditions.

[0016] The first aspect of this application provides a method for monitoring the gate oxide degradation of a flywheel motor-driven SiC power transistor, including:

[0017] Acquire transient window signals and operating condition information during the turn-on process of SiC MOSFETs;

[0018] The oxide trap charge and the interface state trap charge are defined as the hidden states of the recurrent neural network. Based on the transient window signal, operating condition information and the hidden state of the previous opening process, the oxide trap charge and the interface state trap charge of the current opening process are updated respectively through the gating recursion mechanism.

[0019] Based on the updated oxide trap charge and interface state trap charge, the gate oxide degradation result of the SiC MOSFET is obtained according to the physical relationship between trap charge and threshold voltage drift.

[0020] In one possible design, the updating of the oxide trap charge and interface state trap charge of the current start-up process based on the transient window signal, operating condition information, and the hidden state of the previous start-up process via a gated recursion mechanism includes:

[0021] Extract observation features from the transient window signal and operating condition information;

[0022] The equivalent charge increment, double trap allocation coefficient, and double trap charge recovery gating coefficient are calculated based on the observed features and the hidden state of the previous opening process.

[0023] The oxide trap charge and interface state trap charge of the current opening process are updated based on the double trap charge, equivalent charge increment, double trap allocation coefficient, and double trap charge recovery gating coefficient of the previous opening process.

[0024] In one possible design, the extraction of observed features from the transient window signal and operating condition information includes:

[0025] A waveform encoder is used to extract observation features from the transient window signal and operating condition information.

[0026] The waveform encoder is a one-dimensional convolutional network, a gated recurrent unit, a temporal convolutional network, a Transformer encoder, a multilayer perceptron, or a physical feature encoder.

[0027] In one possible design, the calculation of the equivalent charge increment, double-trap allocation coefficient, and double-trap charge recovery gating coefficient based on the observed features and the hidden state of the previous open-loop process includes:

[0028] The equivalent charge increment is calculated according to the following formula:

[0029] ,

[0030] in, For the first The equivalent charge increment during the second opening process. This refers to the softplus function. This represents a function that calculates the equivalent charge increment based on input features. For the first Observational characteristics of the second opening process For the first Operating condition information for the next opening process. For the first The hidden state of the next opening process. For the first The sampling interval of each open process;

[0031] The double-trap allocation coefficient is calculated according to the following formula:

[0032] ,

[0033] ,

[0034] in, and The first The distribution coefficients of oxide traps and interface state traps during the initial commissioning process. This represents the softmax function. A function for calculating trap assignment coefficients based on input features;

[0035] The charge recovery gating coefficient of the double trap is calculated according to the following formula:

[0036] ,

[0037] ,

[0038] in, and The first Charge recovery gating coefficients of oxide traps and interface state traps during the initial turn-on process. This represents the sigmoid function. This is a function for calculating the trap charge recovery gating coefficient based on input characteristics.

[0039] In one possible design, updating the oxide trap charge and interface state trap charge of the current turn-on process based on the double-trap charge, equivalent charge increment, double-trap allocation coefficient, and double-trap charge recovery gating coefficient of the previous turn-on process includes:

[0040] The oxide trap charge and interface state trap charge of the current open process are updated according to the following formula:

[0041] ,

[0042] ,

[0043] in, and The first and Oxide trap charge during the secondary opening process. and For the first and Interface state trap charge during the secondary opening process. and The first Charge recovery gating coefficients of oxide traps and interface state traps during the initial turn-on process. and The first The distribution coefficients of oxide traps and interface state traps during the initial commissioning process. For the first The equivalent charge increment during the second opening process.

[0044] In one possible design, obtaining the gate oxide degradation result of the SiC MOSFET based on the updated oxide trap charge and interface state trap charge, and on the physical relationship between trap charge and threshold voltage drift, includes:

[0045] Based on the physical relationship between the trap charge and the threshold voltage drift, a reference threshold voltage is output;

[0046] The gate oxide health index is calculated based on the double-trap charge state and the reference threshold voltage, and a status warning is determined based on the gate oxide health index.

[0047] In one possible design, the output of the reference threshold voltage based on the physical relationship between the trap charge and the threshold voltage drift includes:

[0048] The reference threshold voltage is calculated using the following formula:

[0049] ,

[0050] or, ,

[0051] Or perhaps, ,

[0052] in, As a reference threshold voltage, The initial threshold voltage, For reference threshold voltage drift, and These are the charge polarity coefficients of oxide traps and interface state traps, respectively. For the first Oxide trap charge during the secondary opening process. For the first Interface state trap charge during the secondary opening process. This is the equivalent gate oxide capacitance. For temperature compensation, This is the operating condition offset compensation item;

[0053] The expression for the reference threshold voltage drift is:

[0054] .

[0055] In one possible design, the calculation of the gate oxide health index based on the double-trap charge state and the reference threshold voltage, and the determination of whether a state warning is needed based on the gate oxide health index, includes:

[0056] The oxygen health index is calculated using the following formula:

[0057] ,

[0058] in, For the first The oxygen health indicators during the second opening process. For the first Oxide trap charge during the secondary opening process. For the first Interface state trap charge during the secondary opening process. As a reference threshold voltage, The initial threshold voltage, This represents the limiting value of the oxide trap charge. This represents the limiting value of the interface state trap charge. This represents the limit value of the threshold voltage drift. , and All are weighted coefficients;

[0059] or In this case, a status alert is required. The preset warning threshold is used.

[0060] In one possible design, the method for monitoring gate oxide degradation of flywheel motor-driven SiC power transistors also includes: constructing at least one auxiliary physical constraint to enhance the identifiability of oxide trap charges and interface state trap charges;

[0061] The auxiliary physical constraints include one or more of the following: mobility degradation constraints, input capacitance constraints, gate charge response constraints, and identifiable constraints based on the sensitivity matrix.

[0062] The second aspect of this application provides a gate oxide degradation monitoring system for a flywheel motor-driven SiC power transistor, comprising:

[0063] Acquisition unit: Used to acquire transient window signals and operating condition information during the turn-on process of SiC MOSFET;

[0064] Gated recursive unit: used to define oxide trap charge and interface state trap charge as the hidden state of recurrent neural network, and update oxide trap charge and interface state trap charge of the current opening process respectively through gated recursive mechanism based on the transient window signal, operating condition information and the hidden state of the previous opening process.

[0065] Physical mapping unit: used to obtain the gate oxide degradation result of the SiC MOSFET based on the physical relationship between trap charge and threshold voltage drift, according to the updated oxide trap charge and interface state trap charge.

[0066] A third aspect of this application provides a gate oxide degradation monitoring device for a flywheel motor driven SiC power transistor. The flywheel motor driven SiC power transistor gate oxide degradation monitoring device includes a processor and a memory. The memory stores at least one instruction, which is loaded and executed by the processor to implement the above-described flywheel motor driven SiC power transistor gate oxide degradation monitoring method.

[0067] A fourth aspect of this application provides a computer storage medium storing at least one instruction, which is loaded and executed by a processor to implement the above-described method for monitoring gate oxide degradation of a SiC power transistor driven by a flywheel motor.

[0068] The beneficial effects of this application are:

[0069] 1. Low-intrusion online monitoring. This application only requires a transient window signal and a small amount of operating condition information to recursively estimate the gate oxide degradation state, reducing reliance on frequent shutdown threshold voltage testing and complex capacitance measurement circuits, and reducing the demand for offline tags by 50%.

[0070] 2. High interpretability. The model's hidden states are directly defined as oxide trap charges and interface state trap charges, and the output threshold voltage drift is obtained through physical mapping of the gate oxide capacitance per unit area, rather than a simple black-box waveform regression.

[0071] 3. Capable of separating dual-trap contributions. By employing charge distribution gating, recovery gating, stress prior constraints, and multi-physics auxiliary constraints, the equivalent oxide trap charge and the equivalent interface state trap charge are updated separately, solving the problem that a single equivalent charge cannot distinguish the source of degradation.

[0072] 4. Suitable for sparse label scenarios. Threshold voltage can usually only be measured offline at aging nodes. When only 10% to 30% of the threshold voltage labels are used, the model can still maintain threshold voltage drift estimation accuracy similar to that of full-label training. Compared with ordinary neural networks trained only with labeled samples, the threshold voltage drift estimation error can be reduced by about 15% to 50%.

[0073] 5. Strong noise resistance. The recursive state update merges the fast variable observations of each activation into the slow variable degradation trajectory, reducing the jitter amplitude of the threshold voltage drift estimation sequence by 30% and improving the continuity and availability of the online monitoring curve.

[0074] 6. Strong adaptability across operating conditions. The operating condition vector and compensation term can explicitly handle mixed factors such as temperature, load, bus voltage, and drive parameters, reducing the threshold voltage drift estimation error by 15% under conditions of cross-temperature, cross-load, or cross-bus voltage.

[0075] 7. Facilitates engineering deployment. After model training is completed, the online inference process only includes window sampling, encoding, gating recursion, and physical mapping. It can be deployed on edge computing units, controllers, or host computer health management software, with online inference time controlled in the millisecond to hundreds of milliseconds range. Attached Figure Description

[0076] Figure 1 This is a schematic diagram of the physical recurrent neural network method.

[0077] Figure 2 This is an overall block diagram of the flywheel motor-driven SiC power transistor gate oxide degradation monitoring system in the embodiment;

[0078] Figure 3 This is a schematic diagram of the physical link from the turn-on transient waveform to the threshold voltage drift in the embodiment;

[0079] Figure 4 This is a flowchart of the training process in the embodiment;

[0080] Figure 5 This is a flowchart of the online deployment process in the example. Detailed Implementation

[0081] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0082] Specific Implementation Method 1: This implementation method describes a gate oxide degradation monitoring method for SiC power transistors driven by a flywheel motor based on a double-trap charge physical recurrent neural network, comprising:

[0083] 1. Signal Acquisition Stage: Acquire a transient window signal during the SiC MOSFET turn-on process. Preferably, the first... During the initial turn-on process, the Rogowski coil output voltage It can also be the gate current, the rate of change of drain current, the gate-source voltage, the drain-source voltage, or other equivalent switching transient signals.

[0084] Specifically, during the turn-on transient phase of the SiC MOSFET, the Rogowski coil output voltage... This can be expressed as a function of the rate of change of gate current or associated switching current. This embodiment uses the following observation relationship:

[0085] (1),

[0086] in, The mutual inductance or equivalent sensitivity coefficient of the Rogowski coil. This refers to the gate current or the current equivalent to the gate charging process. For measuring noise.

[0087] 2. Working condition vector Input stage: Synchronous input gate drive voltage Gate resistor Bus voltage Load current Junction temperature or shell temperature Sampling interval Operating condition information, etc.

[0088] In the gate drive circuit, the gate-source voltage during the initial turn-on phase can be described using an equivalent RC charging relationship. :

[0089] (2),

[0090] (3),

[0091] (4).

[0092] When the gate-source voltage reaches the threshold voltage, the turn-on delay or threshold crossover time occurs. satisfy:

[0093] (5).

[0094] The above formula enables the switching transient window signal Gate dynamic parameters, input capacitance With threshold voltage Establish physical connections to provide a physical backbone for subsequent inversion of hidden degradation states from observable waveforms.

[0095] 3. Waveform preprocessing stage: [This refers to the preprocessing of waveforms.] Window truncation, time alignment, amplitude normalization, baseline correction, filtering or integral transform are performed to obtain the preprocessed standardized observation window. .

[0096] 4. To address the issues of how to distinguish the proportions of oxide trap charges and interface state trap charges, and how to update them separately, this embodiment proposes a dual-trap charge allocation gating recursive formula.

[0097] First, by waveform encoder Extracting observation features :

[0098] (6),

[0099] in, This can be a one-dimensional convolutional network, a gated recurrent unit, a temporal convolutional network, a Transformer encoder, a multilayer perceptron, or a physical feature encoder.

[0100] Then, based on the observed features Calculate the equivalent charge increment and double trap allocation coefficient :

[0101] (7),

[0102] (8),

[0103] (9),

[0104] in, For the first The equivalent charge increment during the second opening process. This refers to the softplus function. This represents a function that calculates the equivalent charge increment based on input features. For the first Observational characteristics of the second opening process For the first Operating condition information for the next opening process. For the first The hidden state of the next opening process. For the first The sampling interval of each open process, and The first The distribution coefficients of oxide traps and interface state traps during the initial commissioning process. This represents the softmax function. This is a function for calculating trap assignment coefficients based on input features.

[0105] Furthermore, a charge recovery gating coefficient is introduced. Used to describe trap-to-capture, recovery effect, or short-term hysteresis:

[0106] (10)

[0107] (11),

[0108] in, and The first Charge recovery gating coefficients of oxide traps and interface state traps during the initial turn-on process. This represents the sigmoid function. This is a function for calculating the trap charge recovery gating coefficient based on input characteristics.

[0109] Finally, the recursive update formula for the double-trap charge is obtained:

[0110] (12),

[0111] (13)

[0112] in, For the first Oxide trap charge during the secondary opening process. For the first Interface state trap charge during the secondary opening process.

[0113] Formulas (6) to (13) decompose the total equivalent charge change into oxide trap charge increment and interface state trap charge increment, while allowing the two to have different recovery coefficients, enabling the model to recursively track the dual-trap degradation state across different aging stresses, different operating conditions and different devices.

[0114] by As a recursive hidden state, the oxide trap charge and the interface state trap charge are updated by the double trap charge recursion formula.

[0115] To enhance the identifiability of double-trap charge decomposition, this embodiment introduces stress prior constraints. For different aging stresses such as positive high-temperature gate bias, negative high-temperature gate bias, high-temperature gate switching, power cycling, and irradiation, a soft prior vector is set. :

[0116] (14)

[0117] in, and These are the prior values ​​for oxide traps and interface state traps, respectively.

[0118] This prior is not used as a hard label, but as a physical guide for charge separation:

[0119] (15)

[0120] in, For stress a priori separation loss, For the first Weighting coefficients during the activation process.

[0121] Positive high-temperature gate bias conditions can increase the prior weight of oxide trap charge distribution, while negative high-temperature gate bias conditions can increase the prior weight of interface state trap charge distribution. High-temperature gate switching conditions can be set as a hybrid prior. Therefore, the model does not require direct measurement. and The labels can also be used to learn the relative contributions of the two through waveform response and threshold drift trend under different stress conditions.

[0122] Oxide trap charges and interface state trap charges enhance channel carrier scattering, leading to a decrease in channel mobility. In this embodiment, the mobility degradation is expressed as:

[0123] (16)

[0124] in, For the first The estimated channel mobility during the opening process. The initial channel mobility. For elementary charge, and These are the influence coefficients of oxide trap charge and interface state trap charge on mobility scattering, respectively. and The first Equivalent oxide trap estimated charge and equivalent interface state trap estimated charge corresponding to the next opening process.

[0125] Gate estimation input capacitance of SiC MOSFET It can be represented as:

[0126] (17)

[0127] in, No. The gate-source capacitance estimated during the turn-on process. No. The gate-drain capacitance is estimated during the opening process.

[0128] Trapped charges alter the electric field and surface potential distribution of the gate oxide layer, thus affecting the gate equivalent capacitance. The equivalent gate oxide charge can be defined to estimate the charge. :

[0129] (18)

[0130] in, and These are all equivalent weighting coefficients, used to describe the degree of influence of different trap charges on the gate oxide electric field and equivalent capacitance. Correspondingly, the gate-source junction estimated capacitance... Estimated capacitance of gate-drain junction It can be represented as:

[0131] (19)

[0132] (20)

[0133] in, and All are doping concentrations in the drift region. Gate oxide capacitance per unit area The dielectric constant of silicon carbide material is... and The first Gate-source voltage and gate-drain voltage during the turn-on process For the first Equivalent gate oxide charge estimated by an opening process.

[0134] Taking into account parasitic capacitance and geometric capacitance, it can be further expressed as:

[0135] (twenty one),

[0136] (twenty two),

[0137] in, For the first The total gate-source capacitance estimated by the turn-on process. For the first The total gate-drain capacitance estimated by the opening process. For gate-drain parasitic capacitance, Gate-source geometric capacitance, Gate-source parasitic capacitance, For gate-drain oxide layer capacitance, For the first The gate-drain junction capacitance is estimated during the turn-on process.

[0138] The above formula will , It is associated with the gate input capacitance and is used to constrain the estimation of the network's dynamic response to the input capacitance and gate during online switching.

[0139] Under the action of the gate drive pulse, the gate source voltage The dynamic can be written as:

[0140] (twenty three),

[0141] Corresponding gate charging current It can be represented as:

[0142] (twenty four),

[0143] Gate charging time constant for:

[0144] (25),

[0145] In the formula, For the first Gate drive voltage amplitude during the next turn-on event. For time variables, No. Gate resistance during the next turn-on event.

[0146] Equations (23) to (25) can serve as auxiliary constraints for physical interpretation and parameter inversion of online switching waveforms in this embodiment.

[0147] Estimated solely by threshold voltage drift and There may be issues with identification. Therefore, this embodiment can construct multi-physics observation pairs. and Sensitivity matrix:

[0148] (26)

[0149] when When the column vectors are linearly independent, it indicates that different physical observations are independent of each other. and They have different sensitivities, and both possess the condition for separability. In this embodiment, the identifiable constraint is defined as:

[0150] (27)

[0151] in, A small positive constant. By minimizing This can improve the network's... and This enhances the model's ability to differentiate between different voltages, preventing it from simply fitting the same threshold voltage drift using multiple charge combinations.

[0152] 5. Physical Output Mapping: Based on the physical relationship between trap charge and threshold voltage drift, the reference threshold voltage drift is output. and reference threshold voltage .

[0153] The gate oxide degradation state is decomposed into oxide trap charge and interface state trap charge, and the degradation index is output through threshold voltage physical mapping. Preferably, the threshold voltage drift under the reference condition is expressed as:

[0154] (28)

[0155] (29)

[0156] in, The gate oxide capacitance per unit area or equivalent. This is the initial threshold voltage.

[0157] Considering different device polarities, test directions, or symbol conventions, refer to the threshold voltage. It can be further written as:

[0158] (30)

[0159] in, and These are the charge polarity coefficients of oxide traps and interface state traps, respectively, which can be determined through initial calibration or device type.

[0160] If we consider the mixed temperature and operating conditions during online measurement, we can further introduce an operating condition compensation term:

[0161] (31),

[0162] in, For temperature compensation, This is the operating condition bias compensation term learned by the neural network. If the output target is the threshold drift under the reference condition, then temperature and operating condition compensation are used to recover the degradation state from online observation to the reference condition, avoiding misjudging operating condition disturbances as gate oxide degradation.

[0163] 6. Output the grid oxygen health index, trap contribution ratio, and early warning results according to equations (32) to (34), and then... Save it as the recursive state for the next activation event.

[0164] Constructing a gated oxygen health index based on double-trap charge states :

[0165] (32),

[0166] The system outputs a gate oxide degradation warning when one of the following conditions is met:

[0167] if or ,but (33),

[0168] in, This represents the limiting value of the oxide trap charge. This represents the limiting value of the interface state trap charge. This represents the limit value of the threshold voltage drift. The preset warning threshold is used. , , and It can be set according to device specifications, reliability test results, or engineering safety margins. , and All are weighted coefficients.

[0169] For the first Early warning signs of grid oxide degradation during the second opening process. This indicates that an alert has been triggered. This indicates that the current health indicators and threshold voltage drift are both within a safe range.

[0170] Simultaneously, the proportion of oxide traps and the proportion of interface state traps can be defined:

[0171] (34),

[0172] in, To prevent small constants with denominators of zero. When When the value is higher, it indicates that the oxide trap charge contributes more to the current threshold drift; when... When the value is higher, it indicates that the interface state trap charge contributes more significantly to mobility degradation and interface scattering.

[0173] 7. When new offline threshold voltage test data or a small amount of real operating condition calibration data are available, the model is incrementally updated or corrected using the following formula and optional auxiliary physical constraints.

[0174] During model training, sparse threshold voltage labels, switching waveform observations, and physical consistency can be used as constraints. The total loss function is:

[0175] (35),

[0176] The definitions of each item are as follows:

[0177] Threshold voltage tag loss: (36)

[0178] Waveform reconstruction loss: (37)

[0179] Physical consistency loss: (38)

[0180] State smoothing loss: (39)

[0181] in, ~ These are the weighting coefficients for each item. Indicates the first Does each window have an offline threshold voltage label? This represents the reconstructed turn-on transient waveform obtained from the gate dynamic equation and charge state; Optional auxiliary physical constraints may include input capacitance constraints, mobility constraints, gate charge response constraints, and identifiability constraints.

[0182] 8. Training and online deployment phase: The model is trained using sparse offline threshold voltage labels, switching transient waveform reconstruction constraints, physical consistency constraints, and optional identifiable constraints. After training, the degradation state is recursively updated after each or several activation events.

[0183] To enable the model to adapt to different degradation environments such as high-temperature gate bias, high-temperature gate switching, power cycling, and irradiation, this application can construct electric field, temperature, dose, and switching stress as degradation excitation terms:

[0184] (40),

[0185] (41),

[0186] in, For the gate oxide electric field, For reference electric field, , The electric field acceleration coefficient is... and To activate energy, Boltzmann's constant, For temperature.

[0187] When the test includes cumulative radiation dose In this case, the degradation incentive term can be expanded to:

[0188] (42),

[0189] (43),

[0190] in, and These are the influence coefficients of irradiation on the formation of oxide traps and interface state traps, respectively. The above stress excitation term can be used as... , or The input can also be used as a physical prior factor in the charge recursion formula.

[0191] In this embodiment, several SiC MOSFETs of the same model were selected and subjected to aging tests under positive high-temperature gate bias (P-HTGB, VGS=+30V, 175℃), negative high-temperature gate bias (N-HTGB, VGS=-20V, 175℃), and high-temperature gate switching (HTGS, VGS=+25 / -15V, 100kHz, 175℃). Stress was paused at each aging node, and after cooling to specified reference conditions, the threshold voltage was measured, and the aging time, number of cycles, temperature, and stress type were recorded. The devices were placed on a dual-pulse test platform, and the turn-on transient window signal was acquired under fixed operating conditions (load inductance 200μH, bus voltage 300V, load current 15A, drive voltage +15 / -5V, gate resistance 10Ω). The operating condition vectors of each aging node were used to form training samples. For those without... The label's on-state window is trained using only waveform reconstruction loss, smoothing loss, stress prior loss, and physical consistency loss. After training, the model can recursively estimate the hidden degenerate states based on the new on-state transient window and output the threshold voltage drift under reference conditions.

[0192] The trained PI-RNN is deployed to a host computer, edge computing unit, DSP, or MCU-assisted monitoring module. The SiC MOSFET under test operates in a Buck converter or other power electronic converter, with the gate drive circuit applying periodic drive pulses to the device. During normal converter operation, the turn-on window is periodically sampled. Simultaneously, it reads estimated values ​​for bus voltage, load current, drive voltage, gate resistance, switching frequency, and temperature. The online data is then input into the model for recursive updates. and It outputs the threshold voltage drift under reference conditions. When health indicators When the preset threshold is exceeded, the system outputs a warning signal; the estimation results can also be used to adjust the drive strategy, reduce the switching frequency, reduce the throttle operation, or arrange maintenance.

[0193] This embodiment can be used for gate oxide degradation analysis of SiC MOSFETs after reliability tests such as total dose irradiation, high-temperature gate bias, or high-temperature reverse bias. During the test, gate bias, electrothermal stress, or irradiation stress are applied to the SiC MOSFET, and parameters such as threshold voltage, gate leakage current, on-resistance, input capacitance, and gate charging curve are collected at different aging time points. These parameters are input into the physical information recursive neural network of this application to obtain the results at different stress time points. and When the experiment involves radiation dose When this occurs, the extended formula for the degradation excitation term can be used for calculation. This implementation method can be used to analyze whether the threshold voltage drift of SiC MOSFETs mainly originates from oxide trap charges or interface state trap charges under different irradiation doses, different gate bias conditions, and different temperatures.

[0194] Specific Implementation Method Two: The flywheel motor-driven SiC power transistor gate oxide degradation monitoring system described in this implementation method includes:

[0195] Acquisition unit: Used to acquire transient window signals and operating condition information during the turn-on process of SiC MOSFET;

[0196] Gated recursive unit: used to define oxide trap charge and interface state trap charge as the hidden state of recurrent neural network, and update oxide trap charge and interface state trap charge of the current opening process respectively through gated recursive mechanism based on the transient window signal, operating condition information and the hidden state of the previous opening process.

[0197] Physical mapping unit: used to obtain the gate oxide degradation result of the SiC MOSFET based on the physical relationship between trap charge and threshold voltage drift, according to the updated oxide trap charge and interface state trap charge.

[0198] Specific Implementation Method 3: The flywheel motor driven SiC power transistor gate oxide degradation monitoring device described in this embodiment includes a processor and a memory. The memory stores at least one instruction, which is loaded and executed by the processor to implement the flywheel motor driven SiC power transistor gate oxide degradation monitoring method as described in Specific Implementation Method 1.

[0199] Specific Implementation Method Four: A computer storage medium as described in this embodiment stores at least one instruction, which is loaded and executed by a processor to implement the gate oxide degradation monitoring method for flywheel motor-driven SiC power transistors as described in Specific Implementation Method One.

[0200] While specific embodiments of this application have been described herein with reference to them, it should be understood that these embodiments are merely examples of the principles and applications of this application. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of this application as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A method for monitoring gate oxide degradation of SiC power transistors driven by flywheel motors, characterized in that, include: Acquire transient window signals and operating condition information during the turn-on process of SiC MOSFETs; The oxide trap charge and the interface state trap charge are defined as the hidden states of the recurrent neural network. Based on the transient window signal, operating condition information and the hidden state of the previous opening process, the oxide trap charge and the interface state trap charge of the current opening process are updated respectively through the gating recursion mechanism. Based on the updated oxide trap charge and interface state trap charge, the gate oxide degradation result of the SiC MOSFET is obtained according to the physical relationship between trap charge and threshold voltage drift; The step of updating the oxide trap charge and interface state trap charge of the current opening process based on the transient window signal, operating condition information, and the hidden state of the previous opening process through a gating recursion mechanism includes: Extract observation features from the transient window signal and operating condition information; The equivalent charge increment, double trap allocation coefficient, and double trap charge recovery gating coefficient are calculated based on the observed features and the hidden state of the previous opening process. The oxide trap charge and interface state trap charge of the current opening process are updated based on the double trap charge, equivalent charge increment, double trap allocation coefficient, and double trap charge recovery gating coefficient of the previous opening process, respectively. The calculation of the equivalent charge increment, double-trap allocation coefficient, and double-trap charge recovery gating coefficient based on the observed features and the hidden state of the previous opening process includes: The equivalent charge increment is calculated according to the following formula: ， in, For the first The equivalent charge increment during the second opening process. This refers to the softplus function. This represents a function that calculates the equivalent charge increment based on input features. For the first Observational characteristics of the second opening process For the first Operating condition information for the next opening process. For the first The hidden state of the next opening process. For the first The sampling interval of each open process; The double-trap allocation coefficient is calculated according to the following formula: ， ， in, and The first The distribution coefficients of oxide traps and interface state traps during the initial commissioning process. This represents the softmax function. A function for calculating trap assignment coefficients based on input features; The charge recovery gating coefficient of the double trap is calculated according to the following formula: ， ， in, and The first Charge recovery gating coefficients of oxide traps and interface state traps during the initial turn-on process. This represents the sigmoid function. This is a function for calculating the trap charge recovery gating coefficient based on input features; The step of updating the oxide trap charge and interface state trap charge of the current turn-on process based on the double trap charge, equivalent charge increment, double trap allocation coefficient, and double trap charge recovery gating coefficient of the previous turn-on process includes: The oxide trap charge and interface state trap charge of the current open process are updated according to the following formula: ， ， in, and The first and Oxide trap charge during the secondary opening process. and For the first and Interface state trap charge during the secondary opening process. and The first Charge recovery gating coefficients of oxide traps and interface state traps during the initial turn-on process. and The first The distribution coefficients of oxide traps and interface state traps during the initial commissioning process. For the first The equivalent charge increment during the second opening process.

2. The method for monitoring gate oxide degradation of a flywheel motor-driven SiC power transistor according to claim 1, characterized in that, The extraction of observation features from the transient window signal and operating condition information includes: A waveform encoder is used to extract observation features from the transient window signal and operating condition information. The waveform encoder is a one-dimensional convolutional network, a gated recurrent unit, a temporal convolutional network, a Transformer encoder, a multilayer perceptron, or a physical feature encoder.

3. The method for monitoring gate oxide degradation of a flywheel motor-driven SiC power transistor according to claim 1, characterized in that, The step of obtaining the gate oxide degradation result of the SiC MOSFET based on the updated oxide trap charge and interface state trap charge, and on the physical relationship between trap charge and threshold voltage drift, includes: Based on the physical relationship between the trap charge and the threshold voltage drift, a reference threshold voltage is output; The gate oxide health index is calculated based on the double-trap charge state and the reference threshold voltage, and a status warning is determined based on the gate oxide health index.

4. The method for monitoring gate oxide degradation of a flywheel motor-driven SiC power transistor according to claim 3, characterized in that, The step of outputting a reference threshold voltage based on the physical relationship between the trap charge and the threshold voltage drift includes: The reference threshold voltage is calculated using the following formula: ， or, , Or perhaps, , in, As a reference threshold voltage, The initial threshold voltage, For reference threshold voltage drift, and These are the charge polarity coefficients of oxide traps and interface state traps, respectively. For the first Oxide trap charge during the secondary opening process. For the first Interface state trap charge during the secondary opening process. This is the equivalent gate oxide capacitance. For temperature compensation, This is the operating condition offset compensation item; The expression for the reference threshold voltage drift is: 。 5. The method for monitoring gate oxide degradation of a flywheel motor-driven SiC power transistor according to claim 3, characterized in that, The process of calculating the gate oxide health index based on the double-trap charge state and the reference threshold voltage, and determining whether a state warning is needed based on the gate oxide health index, includes: The oxygen health index is calculated using the following formula: ， in, For the first The oxygen health indicators during the second opening process. For the first Oxide trap charge during the secondary opening process. For the first Interface state trap charge during the secondary opening process. As a reference threshold voltage, The initial threshold voltage, This represents the limiting value of the oxide trap charge. This represents the limiting value of the interface state trap charge. This represents the limit value of the threshold voltage drift. , and All are weighted coefficients; or In this case, a status alert is required. The preset warning threshold is used.

6. The method for monitoring gate oxide degradation of a flywheel motor-driven SiC power transistor according to claim 1, characterized in that, Also includes: Construct at least one auxiliary physical constraint to enhance the identifiability of oxide trap charges and interface state trap charges; The auxiliary physical constraints include one or more of the following: mobility degradation constraints, input capacitance constraints, gate charge response constraints, and identifiable constraints based on the sensitivity matrix.

7. A gate oxide degradation monitoring system for a flywheel motor-driven SiC power transistor used to perform the method of any one of claims 1 to 6, characterized in that, include: Acquisition unit: Used to acquire transient window signals and operating condition information during the turn-on process of SiC MOSFET; Gated recursive unit: used to define oxide trap charge and interface state trap charge as the hidden state of recurrent neural network, and update oxide trap charge and interface state trap charge of the current opening process respectively through gated recursive mechanism based on the transient window signal, operating condition information and the hidden state of the previous opening process. Physical mapping unit: used to obtain the gate oxide degradation result of the SiC MOSFET based on the physical relationship between trap charge and threshold voltage drift, according to the updated oxide trap charge and interface state trap charge.

Images