A GaN device current collapse suppression method and system

Abstract

The present invention relates to the field of electronic technology, and specifically to a method and system for suppressing the current collapse of a GaN device. The method comprises: obtaining real-time monitoring data of a GaN device; constructing a risk level prediction model, and inputting the real-time monitoring data into the risk level prediction model to obtain a risk level; constructing a parameter compensation prediction model, and inputting the real-time monitoring data into the parameter compensation prediction model based on the risk level to obtain a compensation parameter; and adjusting the gate voltage slope, gate voltage step amplitude, and dead time of the next switching cycle of the GaN device based on the compensation parameter. The present invention solves the problem of a single method for suppressing the current collapse of gallium nitride devices and excessive device load in the prior art. By predicting the risk level, a one-size-fits-all adjustment method is avoided in non-high-risk situations, and a more gentle gradual gate voltage adjustment method is adopted, which can reduce electromagnetic interference, reduce switching losses, suppress current spikes, and improve the reliability of GaN devices.

Description

Technical Field

[0001] The present invention relates to the field of electronic technology, and in particular to a method and system for suppressing current collapse of a GaN device. Background Art

[0002] Gallium nitride (GaN) power devices are widely used in new energy vehicles, renewable energy conversion, and other fields due to their advantages such as high frequency, high efficiency, and high power density. However, when GaN devices operate under harsh operating conditions such as high voltage and high temperature, the accumulation of trapped charge at the heterojunction interface can cause current collapse, which manifests as a sudden increase in on-resistance, degradation of drain current, and in severe cases, device failure. The core cause of current collapse is: when the device is turned off, the trap-assisted tunneling effect caused by the high drain voltage and the surge in trap activation rate caused by the increase in junction temperature. The two work together to trap channel electrons, destroying the normal conductivity of the device.

[0003] Existing technologies mostly use instantaneous step gate voltage boost to suppress current collapse. However, long-term step gate voltage boost can easily cause device overload, and fixed boost may cause the gate voltage to exceed the oxide layer's withstand voltage limit. Under low-temperature conditions, excessive boost will cause a sudden increase in channel electron concentration, causing an abnormal increase in conduction loss and reducing energy efficiency. At the same time, the "one-size-fits-all" control method ignores the degree of current collapse and will also introduce additional losses due to excessive boost. Therefore, a more flexible current collapse suppression method is urgently needed to promote the development of gallium nitride devices. Summary of the Invention

[0004] In response to the defects in the prior art, the present invention provides a method and system for suppressing the current collapse of GaN devices, which solves the problem that the prior art has a single method for suppressing the current collapse of GaN devices and the device load is too large.

[0005] In order to achieve the above-mentioned purpose, one aspect of the present invention provides a method for suppressing current collapse of a GaN device, comprising: obtaining real-time monitoring data of a GaN device; constructing a risk level prediction model, and inputting the real-time monitoring data into the risk level prediction model to obtain a risk level; constructing a parameter compensation prediction model, and inputting the real-time monitoring data into the parameter compensation prediction model based on the risk level to obtain a compensation parameter; and adjusting the gate voltage slope, gate voltage step amplitude and dead time of the next switching cycle of the GaN device based on the compensation parameter.

[0006] The present invention can accurately capture the device operating status by real-time monitoring of GaN device data. The risk level prediction model combined with physical constraints can accurately identify the current collapse risk parameter compensation prediction model. After training with historical data, it can output the optimal compensation parameters and dynamically adjust parameters such as the gate voltage slope, which can effectively suppress current collapse and improve device reliability.

[0007] Optionally, obtaining real-time monitoring data of the GaN device includes: monitoring the GaN device to obtain the on-resistance, off-state drain voltage, junction temperature and turn-on delay of the GaN device in one switching cycle; normalizing the on-resistance, the off-state drain voltage, the junction temperature and the turn-on delay to obtain real-time monitoring data.

[0008] The present invention can accurately capture the operating status of the GaN device by monitoring key parameters such as the on-resistance, off-state drain voltage, junction temperature, and turn-on delay in real time within a switching cycle of the device. By normalizing these parameters, the influence of differences in the dimensions and numerical ranges of different data types on subsequent calculations can be eliminated, thereby improving the reliability of real-time monitoring.

[0009] Optionally, constructing a risk level prediction model includes: obtaining a tunneling activation voltage threshold, and using the tunneling activation voltage threshold to set a high drain voltage penalty item; obtaining a trap activation temperature threshold, and using the trap activation temperature threshold to set a high temperature penalty item; using the high drain voltage penalty item and the high temperature penalty item to set a comprehensive loss function; and constructing a risk level prediction model based on polynomial logistic regression using the comprehensive loss function.

[0010] The present invention obtains the tunneling activation voltage threshold and the trap activation temperature threshold, sets the high drain voltage penalty term and the high temperature penalty term, and can accurately identify the working conditions such as high voltage and high temperature that are likely to cause current collapse. By combining the two to set a comprehensive loss function, and then constructing a risk level prediction model based on polynomial logistic regression, it can take into account both classification accuracy and physical constraints, realize the accurate identification of the current collapse risk of GaN devices, and improve the accuracy of risk level judgment.

[0011] Optionally, constructing the parameter compensation prediction model includes: acquiring historical detection data and constructing training samples based on the historical detection data; constructing the parameter compensation prediction model, and using the training samples to train and evaluate the parameter compensation prediction model.

[0012] By acquiring historical detection data and constructing training samples, we provide a rich source of learning material for the parameter compensation prediction model. Using these training samples to train and evaluate the model enables it to accurately learn the nonlinear mapping relationship between historical monitoring characteristics and optimal compensation parameters, thereby accurately predicting the appropriate compensation parameters and improving the scientific nature and accuracy of the compensation parameters.

[0013] Optionally, obtaining historical detection data and constructing training samples based on the historical detection data includes: obtaining historical on-resistance, historical off-state drain voltage, historical junction temperature and historical turn-on delay; determining the adjustment range of GaN device parameters, and traversing sampling with a predetermined step size according to the adjustment range to obtain multiple historical gate voltage slopes, historical gate voltage step amplitudes and historical dead times; measuring the historical resistance recovery time and historical switching energy consumption of the GaN device according to the historical gate voltage slope, the historical gate voltage step amplitude and the historical dead time; determining the optimal historical gate voltage slope, the optimal historical gate voltage step amplitude and the optimal historical dead time based on the historical gate voltage slope, the historical gate voltage step amplitude and the historical dead time using the historical resistance recovery time and the historical switching energy consumption; and constructing training samples using the historical on-resistance, the historical off-state drain voltage, the historical junction temperature, the historical turn-on delay, the optimal historical gate voltage slope, the optimal historical gate voltage step amplitude and the optimal historical dead time.

[0014] By acquiring historical monitoring data such as on-resistance, we provide realistic device operating characteristics for training samples. We then determine the compensation parameter adjustment range and perform traversal sampling, ensuring comprehensive and representative parameter coverage and providing ample candidate parameter combinations for subsequent optimization. By measuring historical resistance recovery time and switching energy consumption, we identify key metrics for evaluating parameter quality, making optimization more targeted. Determining optimal parameters by minimizing these two metrics effectively improves device performance, reduces energy loss and recovery time, and ultimately constructs training samples using historical monitoring data and optimal compensation parameters. These samples encompass both input characteristics and desired outputs, providing high-quality learning data for the parameter compensation prediction model and improving the accuracy of the training samples.

[0015] Optionally, determining the optimal historical gate voltage slope, the optimal historical gate voltage step amplitude and the optimal historical dead time based on the historical gate voltage slope, the historical gate voltage step amplitude and the historical dead time using the historical resistance recovery time and the historical switching energy consumption includes: obtaining an indicator weight based on the historical resistance recovery time and the historical switching energy consumption; introducing a particle swarm optimization algorithm, and setting an objective function for the particle swarm optimization algorithm in combination with the indicator weight with the goal of minimizing the resistance recovery time and the switching energy consumption; and calculating the optimal historical gate voltage slope, the optimal historical gate voltage step amplitude and the optimal historical dead time based on the historical gate voltage slope, the historical gate voltage step amplitude, the historical dead time, the historical resistance recovery time and the historical switching energy consumption using the particle swarm optimization algorithm.

[0016] By assigning weights to historical resistance recovery time and switching energy consumption, this invention allows for differentiated optimization based on their importance, making the optimization target more aligned with actual needs. By introducing a particle swarm optimization algorithm and combining it with weighted objective function settings, this method efficiently searches for the parameter combination that minimizes the objective function within the permitted parameter adjustment range, improving the accuracy of the optimal gate voltage slope, step amplitude, and dead time.

[0017] Optionally, obtaining the indicator weight based on the historical resistance recovery time and the historical switch energy consumption includes: stacking the historical resistance recovery time and the historical switch energy consumption in rows to form an indicator matrix; normalizing the elements in the indicator matrix to obtain a normalized matrix; calculating the proportion of each element in the normalized matrix in the corresponding column, and constructing a probability matrix based on the proportion; calculating the discreteness of the historical resistance recovery time and the historical switch energy consumption according to the probability matrix; and calculating the indicator weight of the historical resistance recovery time and the historical switch energy consumption based on the discreteness.

[0018] The present invention forms an indicator matrix by stacking historical resistance recovery time and switch energy consumption, and can perform structured storage on the performance indicators of multiple groups of compensation parameters. The indicator matrix is ​​normalized to eliminate the influence of differences in the dimensions and numerical ranges of different indicators. A probability matrix is ​​constructed and the degree of dispersion is calculated to quantify the fluctuation differences of different indicators. Indicator data with low dispersion is more stable and more representative for comprehensive evaluation. Finally, weights are determined according to the degree of dispersion, so that the model pays more attention to key indicators and improves the scientific nature and reliability of the weights.

[0019] Optionally, adjusting the gate voltage slope, gate voltage step amplitude and dead time of the next switching cycle of the GaN device based on the compensation parameters includes: obtaining a dead time reference value, a gate voltage slope threshold and a gate voltage step amplitude threshold of the GaN device; correcting the dead time reference value according to the junction temperature to obtain a dynamic dead time threshold; using the gate voltage slope threshold, the gate voltage step amplitude threshold and the dynamic dead time threshold to constrain the compensation parameters; and using the constrained compensation parameters to adjust the gate voltage slope, gate voltage step amplitude and dead time of the next switching cycle of the GaN device.

[0020] The present invention obtains thresholds such as dead-time reference values ​​to provide a reference range for parameter adjustment. The dead-time reference values ​​are then corrected according to junction temperature to obtain dynamic thresholds, adapting to varying temperature conditions and ensuring reliable device operation. Using various threshold constraint compensation parameters prevents parameters from exceeding safety limits, thus avoiding electromagnetic interference, device damage, and other issues. Ultimately, the constrained parameters are used to adjust the relevant parameters for the next switching cycle of the device, achieving real-time closed-loop control. This allows the device to obtain optimal drive parameters during each switching cycle, effectively suppressing current collapse, reducing switching losses, and improving the device's stability and reliability under high-frequency conditions.

[0021] Optionally, the correcting the dead time reference value according to the junction temperature to obtain a dynamic dead time threshold includes: obtaining multiple sets of test junction temperatures and test dead time reference values; constructing a regression model using the multiple sets of test junction temperatures and the test dead time reference values; and inputting the junction temperature into the regression model to obtain a dynamic dead time threshold.

[0022] By acquiring multiple sets of test junction temperature and dead time benchmark values, the relationship between junction temperature and dead time can be captured. A regression model is constructed to quantify this relationship, enabling the dead time to be dynamically adjusted with junction temperature. Inputting real-time junction temperature into the model generates a dynamic dead time threshold, which adapts to varying temperature conditions, avoids unreasonable dead time caused by temperature fluctuations, and improves the rationality of the dead time threshold.

[0023] Another aspect of the present invention provides a GaN device current collapse suppression system, comprising: a processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute a GaN device current collapse suppression method as described in any one of the previous aspects of the present invention.

[0024] The GaN device current collapse suppression system of the present invention has a compact structure, stable performance, high integration and simple composition. It can stably execute the GaN device current collapse suppression method provided in the previous aspect of the present invention, further improving the overall applicability and practical application capabilities of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a flow chart of a method for suppressing current collapse in a GaN device according to an embodiment of the present invention;

[0026] Figure 2 This is a schematic structural diagram of a GaN device current collapse suppression system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0027] Specific embodiments of the present invention will be described in detail below. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the present invention. In the following description, numerous specific details are set forth to provide a thorough understanding of the present invention. However, it will be apparent to one of ordinary skill in the art that these specific details are not necessarily required to practice the present invention. In other instances, well-known circuits, software, or methods are not specifically described to avoid obscuring the present invention.

[0028] Throughout this specification, references to "one embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with the embodiment or example is included in at least one embodiment of the present invention. Therefore, appearances of the phrases "in one embodiment," "in an embodiment," "an example," or "an example" in various places throughout this specification are not necessarily all referring to the same embodiment or example. Furthermore, the particular features, structures, or characteristics may be combined in any suitable combinations and / or subcombinations in one or more embodiments or examples. Furthermore, those of ordinary skill in the art will appreciate that the figures provided herein are for illustrative purposes only and are not necessarily drawn to scale.

[0029] See Figure 1 , in order to solve the defects in the prior art, in an optional embodiment, as Figure 1 A method for suppressing current collapse of a GaN device is shown, comprising the following steps:

[0030] Step S1, obtaining real-time monitoring data of the GaN device.

[0031] Among them, obtaining real-time monitoring data of GaN devices specifically includes:

[0032] Step S101 : monitoring a GaN device to obtain an on-resistance, an off-state drain voltage, a junction temperature, and a turn-on delay of the GaN device in one switching cycle.

[0033] In this embodiment, for the measurement of on-resistance, a pulse current injection method can be used to inject a pulse current of 100mA and 1μs during the on-time of the device. The magnitude of this current needs to be strictly designed to avoid the self-heating effect of the device due to excessive current affecting the measurement accuracy, and to ensure a sufficient signal-to-noise ratio. The drain-source voltage drop is measured using a four-terminal Kelvin connection method, which can effectively eliminate the influence of lead resistance and ensure that the measurement result only reflects the on-resistance of the device itself. The PWM drive needs to be turned off during measurement to prevent the high-frequency noise generated during the switching process from interfering with the voltage sampling. The sampling resistor uses a high-precision resistor with an accuracy of 0.1%, and is matched with an ADC with a sampling rate of 1MS / s to ensure the accuracy of voltage and current data acquisition. Finally, the on-resistance value is calculated by Ohm's law.

[0034] Monitoring the off-state drain voltage presents challenges in signal acquisition under high-voltage conditions. A high-voltage differential probe with a bandwidth of 100MHz and a withstand voltage of 1000V, such as the Tektronix P5200A, is used to accurately capture the voltage waveform at the instant of shutdown. The signal conditioning circuit utilizes a 1000:1 voltage-divider resistor network combined with a non-inverting amplifier circuit. The voltage-divider resistors are high-precision metal film resistors with values ​​of 99kΩ and 1kΩ, respectively, to safely attenuate and linearly amplify the high-voltage signal. The conditioned signal is sampled using a 16-bit ADC with a sampling rate of 500kS / s. At a switching frequency of 100kHz, five data points can be collected during each off-state cycle, ensuring a complete record of the drain voltage evolution. To avoid common-mode interference, the entire measurement circuit requires a well-designed grounding scheme, employing a star grounding configuration to minimize ground loops.

[0035] Real-time monitoring of junction temperature can be achieved through direct measurement and indirect measurement. The direct measurement method uses the built-in thermistor of the GaN device to inject a constant current of 1mA into the gate-source diode during the device shutdown period. At this time, the diode is in a forward biased state, and its forward voltage is linearly related to temperature, approximately -2mV / °C. The forward voltage is measured by a high-precision voltmeter, and combined with the reference voltage value at 25°C, the current junction temperature can be calculated. To improve measurement reliability, the temperature coefficient of the thermistor needs to be calibrated regularly in a constant temperature chamber to ensure measurement accuracy in different temperature ranges. The indirect measurement method uses an infrared thermal imager, and selects a model with a wavelength range of 8-14μm and a resolution of 640×512. During measurement, the field of view is precisely focused on the device chip area, and the junction temperature data is obtained through the temperature inversion algorithm of the thermal imager. This method can avoid the impact of contact measurement on the working state of the device and is suitable for real-time monitoring in high-temperature environments.

[0036] Measuring turn-on delay requires precise capture of the timing relationship between the drive voltage and drain current, using a dual-channel oscilloscope for simultaneous monitoring. One channel is connected to the drive voltage signal, with a probe bandwidth set to 100MHz. The other channel is connected to the drain current loop via a current probe, a TCP0030 model, also guaranteeing a 100MHz bandwidth. The trigger condition is set to the rising edge of the drive voltage, with a trigger level of 10% of the rated value, ensuring a consistent starting point for each measurement. The oscilloscope's edge detection function automatically identifies the moment when the drive voltage rises to 10% of the rated value and the moment when the drain current rises to 10% of the rated value. The time difference between the two moments is the turn-on delay. To reduce measurement errors, the oscilloscope must be regularly calibrated to ensure that the timebase accuracy is within ±0.1%. At the same time, the shielding performance of the signal transmission cable should be optimized to prevent external electromagnetic interference from affecting timing measurements.

[0037] The data acquisition system's hardware architecture is based on the STM32H743 as the primary control chip. Its ARM Cortex-M7 core operates at up to 480MHz, meeting the requirements of multi-channel, high-speed data acquisition and processing. The ADC is configured in 4-channel simultaneous sampling mode, with 16-bit resolution and a sampling rate of 1MS / s. The built-in programmable gain amplifier (PGA) dynamically adjusts the gain (1-16x) based on the input signal amplitude, improving the accuracy of small signal acquisition. To achieve synchronous acquisition of various parameters, the system uses a 10MHz crystal oscillator as a unified clock source, ensuring that the ADC sampling clock and the PWM drive clock are coherent, with time deviation within 100ns. The data cache utilizes a dual-buffer mechanism. When one buffer stores data for 100 switching cycles, it triggers a DMA transfer to the main memory, while the other buffer simultaneously begins new data acquisition, ensuring continuous data recording.

[0038] To improve the data acquisition system's interference resistance, LC low-pass filters with a 10MHz cutoff frequency are added to each signal channel to effectively filter out high-frequency noise. A decoupling network consisting of a 10μF tantalum capacitor and a 0.1μF ceramic capacitor in parallel at the power input suppresses the impact of power supply ripple on the analog circuits. Digital and analog circuits are strictly separated during PCB layout and isolated by a ground plane to reduce mutual interference.

[0039] To improve the accuracy of long-term measurements, the data acquisition system can also establish a calibration mechanism. Temperature calibration is performed once a week. The voltage-temperature coefficient of the thermistor diode is calibrated in a constant temperature chamber at 25°C and 100°C to correct the coefficient deviation. For on-resistance calibration, a 10mΩ standard resistor with an accuracy of 0.01% is used. After connecting to the measurement circuit, the signal conditioning parameters are adjusted to ensure that the error between the measured value and the standard value is within ±1%. Through the above comprehensive design and measures, the real-time monitoring system can accurately obtain the key parameters of GaN devices and provide reliable data support for current collapse suppression. It is suitable for 650V / 10A GaN HEMT devices. The data acquisition delay at a switching frequency of 1MHz can be controlled within 2μs, meeting the real-time monitoring needs of high-frequency power electronic systems.

[0040] Step S102 , normalizing the on-resistance, the off-state drain voltage, the junction temperature, and the turn-on delay to obtain real-time monitoring data.

[0041] When normalizing real-time monitoring data, first determine the historical maximum and minimum values ​​of the on-resistance, off-state drain voltage, junction temperature, and turn-on delay. Then, for each real-time monitoring data point, use the normalization formula to calculate and uniformly map all types of data to the [0,1] interval to obtain normalized data, eliminating the impact of differences in the dimensions and numerical ranges of different data types on subsequent calculations.

[0042] The normalization formula is:

[0043] ,

[0044] in, is the normalized data, is the original data, is the minimum value, is the maximum value.

[0045] Step S2: construct a risk level prediction model, and input the real-time monitoring data into the risk level prediction model to obtain the risk level.

[0046] Among them, building a risk level prediction model specifically includes the following sub-steps:

[0047] Step S201 : obtaining a tunneling activation voltage threshold, and using the tunneling activation voltage threshold to set a high drain voltage penalty term.

[0048] To obtain the tunneling activation voltage threshold, Silvaco ATLAS was used to construct a two-dimensional heterojunction model of a GaN power device. The geometric parameters of the AlGaN barrier layer, GaN buffer layer, and electrodes were reproduced, and physical properties such as the breakdown electric field and trap energy levels were configured. The Shockley-Read-Hall model was introduced to describe charge transport. A DC steady-state simulation was performed using a gradient drain voltage of 400V to 600V (with a step size of 20V) to extract the electric field distribution at the drain-buffer interface (this region is the electric field concentration zone and directly determines the probability of trap-assisted tunneling). The critical electric field for trap-assisted tunneling was derived based on the trap charge density formula. When the interface electric field first exceeds the critical electric field for trap-assisted tunneling, the corresponding drain voltage is the tunneling activation voltage threshold. Simultaneously, temperature-varying simulations from 300K to 400K were performed to verify the temperature robustness of the model.

[0049] The high drain voltage penalty term satisfies the following formula:

[0050] ,

[0051] in, is the high drain voltage penalty term, is the sample size, is the sigmoid function, For the The off-state drain voltage of each sample, is the tunneling activation voltage threshold, For the The predicted probability that a sample is high risk.

[0052] When the drain voltage exceeds the threshold value, the GaN device is prone to activating trapped charges due to the tunneling effect, resulting in an exponential increase in the risk of current collapse. Therefore, the model needs to be forced to learn the high-risk characteristics under this working condition. Judge the The off-state drain voltage of the sample Whether the tunneling activation voltage threshold is exceeded ,like , the sigmoid function output approaches 1, activating the penalty mechanism. If , the output of the sigmoid function approaches 0, the penalty mechanism is not activated, and at the same time, Measure how well the model predicts high risk, is the predicted probability that the sample is high risk, The larger the value, the more serious the model's missed high-risk (safe and medium-risk probability ratio) is. Finally, the product of the tunneling activation voltage threshold activation result and the missed judgment is calculated for each sample. The product of all samples is summed up and then divided by the number of samples to obtain the high drain voltage penalty term. The model minimizes it through back propagation. ,when When the threshold is significantly exceeded, If the drain voltage is too low (missed), the penalty term will increase sharply, forcing the model to adjust the parameters. Finally, it learns that the higher the drain voltage, the higher the risk probability of the output. The higher the rule, the greater the penalty. This is achieved when the off-state drain voltage exceeds the tunneling activation voltage threshold and the model misses a high risk. The greater the off-state drain voltage exceeds the threshold and the more serious the miss, the greater the penalty. This constrains the model to accurately identify high-risk features (i.e., a high-risk probability that the output matches the drain voltage) when the drain voltage exceeds the threshold.

[0053] It should be noted that As the first The predicted probability that a sample is high risk is derived from the multinomial logistic regression model combined with the Softmax probability transformation under the comprehensive loss function constraint.

[0054] The risk level prediction model is based on the multinomial logistic regression framework. It maps the input features into a three-category space (safe = 0, medium risk = 1, high risk = 2) through linear transformation, and then converts it into a probability distribution through the softmax function. Specifically, for the first The linear transformation output of the samples is (corresponding to the linear scores of the three risk levels), the predicted probability satisfies the following formula:

[0055] ,

[0056] For samples Risk Level The probability of For samples Risk Level The linear score of For samples Risk Level The linear score of For samples The sum of all scores.

[0057] in, Specifically 、 and , Represents a sample The level is the predicted probability of safety, Represents a sample The predicted probability of being classified as medium risk.

[0058] This step is the mathematical basis for generating prediction probabilities, ensuring that the probability values ​​are in the interval [0,1] and sum to 1, which conforms to the basic characteristics of classification probability in the subsequent training stage. Forced promotion Linear score (allowing high-risk conditions to naturally larger).

[0059] Step S202, obtaining a trap activation temperature threshold, and using the trap activation temperature threshold to set a high temperature penalty item;

[0060] To obtain the trap activation temperature threshold, Silvaco ATLAS was used to construct a GaN device trap dynamics model, integrating the Arrhenius equation (activation energy 0.8 eV, consistent with the trap energy level depth) and the Shockley-Read-Hall model. The device structural parameters and measured trap concentration were input, and the trap activation rate and leakage current degradation rate were simulated from 100°C to 150°C (with a step size of 5°C). When the trap activation rate is greater than 50% and the leakage current degradation conforms to the Arrhenius exponential law, the theoretical critical temperature range (120℃-130℃) is extracted, and high-temperature reverse-bias accelerated aging experiments are carried out simultaneously: the drain voltage is fixed as the tunneling activation voltage threshold, and gradient junction temperatures of 110℃, 120℃, 125℃, 130℃, and 135℃ are applied to multiple groups of devices. The failure time is recorded with the leakage current increasing by 100% relative to the initial value as the failure criterion. The experimental failure time drops sharply from the exponential period, and the temperature is determined as the trap activation temperature threshold by combining the cross-validation of the trap activation rate of 55% and the Arrhenius fit greater than 95% in the TCAD simulation.

[0061] The high temperature penalty term satisfies the following formula:

[0062] ,

[0063] in, For high temperature punishment, is the sample size, is the sigmoid function, For the The junction temperature of each sample, The temperature threshold for trap activation, For the The predicted probability that a sample is high risk.

[0064] The same principle as the high drain voltage penalty term is used in the above formula. Judge the The junction temperature of the sample Whether the trap activation temperature threshold is exceeded ,like , the sigmoid function output approaches 1, activating the penalty mechanism. If , the output of the sigmoid function approaches 0, the penalty mechanism is not activated, and at the same time, Measure how well the model predicts high risk, is the predicted probability that the sample is high risk, The larger it is, the more serious the model's missed high-risk (safe and medium-risk probabilities) judgment is. Finally, the product of the junction temperature exceeding the threshold activation result and the missed judgment is calculated for each sample. The product of all samples is summed up and then divided by the number of samples to obtain the high-temperature penalty term. When the junction temperature exceeds the threshold and the model misses high-risk judgment, the greater the junction temperature exceeds and the more serious the missed judgment, the greater the high-temperature penalty. This constrains the model to accurately identify high-temperature and high-risk working conditions.

[0065] The source is the same as that in step S201.

[0066] Step S203 : Setting a comprehensive loss function using the high drain voltage penalty term and the high temperature penalty term.

[0067] The comprehensive loss function satisfies the following formula:

[0068] ,

[0069] ,

[0070] in, is the loss function, is the cross entropy loss term, is the high drain voltage penalty term, is the high temperature penalty term, For the samples belong to the risk category The probability of For the samples belong to the risk category The predicted probability of .

[0071] The source is the same as that in step S201.

[0072] The cross entropy loss term is used to measure the probability of the model predicting the risk level and the actual risk probability The model also introduces high drain voltage and high temperature penalties. When the off-state drain voltage exceeds the tunneling activation voltage threshold or the junction temperature exceeds the trap activation temperature threshold, a sigmoid function activates the penalty mechanism. Combined with the model's predicted probability of high risk, an additional penalty is imposed for missed high-risk scenarios, forcing the model to focus on conditions prone to current collapse, such as high voltage and high temperature. Ultimately, a weighted summation of these three factors creates a comprehensive optimization objective that balances classification accuracy and physical constraints, guiding the risk level prediction model to accurately identify current collapse risks in GaN devices.

[0073] Will 、 and Substituting into the comprehensive loss function, the comprehensive loss function satisfies the following formula:

[0074] ,

[0075] Step S204: constructing a risk level prediction model based on polynomial logistic regression using the comprehensive loss function.

[0076] When acquiring training data, in addition to measuring the on-resistance, off-state drain voltage, junction temperature, and turn-on delay, the accumulated charge must also be measured synchronously during the device's shutdown period: the drain current in the shutdown phase is collected in real time using a high-precision current probe, and the current signal is discretized at a sampling frequency of no less than 10 MHz. During the shutdown transient (the time period from when the gate voltage drops below the threshold to when the drain current stabilizes, typically 100 ns-1 μs), the product of current and time is numerically integrated to obtain the accumulated charge value. This value directly reflects the degree of trapped charge accumulation, because when the drain is turned off, the charging and discharging process of the trapped charge modulates the drain current. The larger the accumulated charge value obtained by integration, the higher the trap density at the gate-drain interface and the more significant the risk of current collapse.

[0077] Based on the operating requirements of GaN devices and expert experience, three critical thresholds (corresponding to cumulative charge values) are set to distinguish the threshold intervals corresponding to levels of safety, medium risk, and high risk. The risk level is determined for each group of samples and the output label is 0 for safety, 1 for medium risk, and 2 for high risk.

[0078] The samples are integrated into structured samples. First, each feature is normalized to the interval [0, 1] to eliminate the dimension effect. Then, the samples are grouped according to the time series or working condition type. Each group of feature vectors is bound to a unique risk level label to form a training set format of feature matrix plus label vector.

[0079] To train a risk level prediction model using polynomial logistic regression, the preprocessed training data must first be input into the model and mapped to a three-category space through a linear layer. The predicted probability of each risk level is output through the softmax function. The cross-entropy term of the comprehensive loss function optimizes the classification accuracy. The physical penalty term uses the sigmoid function to impose a penalty on voltage / temperature features that exceed the threshold, forcing the model to focus on high-risk conditions. The Adam optimizer is used to minimize the loss function. During training, the training set, validation set, and test set are divided into training, validation, and test sets in an 8:1:1 ratio. The learning rate and penalty weight are regularly adjusted through the validation set. Finally, the test set is used to evaluate the model's prediction accuracy for safe, medium-risk, and high-risk levels, as well as its robustness under extreme conditions.

[0080] Finally, the normalized real-time monitoring data is input into the trained risk level prediction model. The risk level prediction model maps the features to a three-category space through linear transformation, and the softmax function is used to calculate the predicted probability distribution of safe (0), medium risk (1), and high risk (2). If the off-state drain voltage exceeds the tunneling activation voltage threshold or the junction temperature exceeds the trap activation temperature threshold, the model is constrained by the physical penalty term in the training phase, which will significantly improve the predicted probability of the high risk level. The level corresponding to the maximum probability is finally taken as the output. At the same time, the probability value of each level can also be output for risk confidence assessment, realizing accurate risk level determination of the current operating condition of the GaN device.

[0081] It should be noted that when the risk level is safe, no parameter compensation is performed. When the risk level is high, instantaneous step gate voltage boost should be used. When the risk level is medium, gradual gate voltage boost should be used. The compensation parameters are predicted by the parameter compensation prediction model.

[0082] When the risk level reaches high, due to the extremely rapid risk evolution and model prediction delays, a hardware-level instantaneous step gate voltage boost is directly triggered. The gate voltage step amplitude is immediately forced to a safe redundancy value based on the gate voltage slope upper threshold (e.g., a +1V instantaneous compensation added to the rated gate voltage of +6V, without exceeding the gate voltage step upper threshold). This burst gate voltage boost within a single switching cycle rapidly injects channel charge to offset trapping effects. The dead time can be optionally locked to the upper limit of the dynamic dead time threshold to prevent direct conduction between the upper and lower bridge arms. This layered design ensures precise control of medium-risk operating conditions through model prediction, while addressing the urgent suppression needs of high-risk conditions with a hardware-driven instantaneous boost mechanism, achieving a balance between timeliness and accuracy.

[0083] Step S3: constructing a parameter compensation prediction model, and inputting the real-time monitoring data into the parameter compensation prediction model based on the risk level to obtain compensation parameters.

[0084] The construction of the parameter compensation prediction model specifically includes the following sub-steps:

[0085] Step S301: Acquire historical detection data and construct training samples based on the historical detection data.

[0086] The process of obtaining historical detection data and constructing training samples based on the historical detection data includes:

[0087] Step S30101, obtaining historical on-resistance, historical off-state drain voltage, historical junction temperature and historical turn-on delay.

[0088] In this embodiment, historical on-resistance, historical off-state drain voltage, historical junction temperature, and historical turn-on delay are obtained through historical monitoring of GaN devices. The method is consistent with step S101. The historical monitoring data should be screened once to exclude safe and high-risk data groups as much as possible and retain the characteristics of medium-risk sample data.

[0089] Step S30102: determining the adjustment range of the GaN device parameters, and obtaining a plurality of historical gate voltage slopes, historical gate voltage step amplitudes, and historical dead time by traversing sampling with a predetermined step length according to the adjustment range.

[0090] In this embodiment, when determining the adjustment range of GaN device parameters, the physical boundaries of the gate voltage slope, gate voltage step amplitude, and dead time must first be determined based on the device data sheet and the safe operating area: the gate voltage slope is generally limited to 10V / ns to 50V / ns (to avoid electromagnetic interference caused by excessive gate voltage slope), the gate voltage step amplitude does not exceed ±10% of the rated gate voltage, and the dead time is controlled to 20ns to 100ns (to prevent direct conduction of the upper and lower bridge arms). Within this range, a predetermined step size is set based on engineering experience (e.g., a gate voltage slope step size of 5V / ns, a step amplitude step size of 0.2V, and a dead time step size of 5ns), and all parameter combinations are generated through Cartesian product traversal. Each set of parameters must be hardware-verified to ensure that it is within the safety threshold (e.g., the gate voltage after the step does not exceed the absolute maximum rating), ultimately forming a parameter library containing historical gate voltage slopes, step amplitudes, and dead times.

[0091] Step S30103 , measuring the historical resistance recovery time and historical switching energy consumption of the GaN device according to the historical gate voltage slope, the historical gate voltage step amplitude, and the historical dead time.

[0092] Measuring the historical resistance recovery time and switching energy consumption of GaN devices based on historical gate voltage slopes, historical gate voltage step amplitudes, and historical dead time requires a pulse test platform. A programmable drive signal is generated using a Keysight B1505A power device analyzer. The GaN device is driven according to preset parameters (e.g., a gate voltage slope of 20V / ns, a step amplitude of +6.2V, and a dead time of 30ns). The device is operated at a switching frequency of 100kHz for 500 cycles to induce current collapse. A four-terminal Kelvin connection is used with a 0.1% high-precision sampling resistor. A Tektronix DPO7254C oscilloscope (2.5GHz bandwidth) is used to acquire the on-resistance waveform in real time. After the drive signal returns to standard parameters, the time it takes for the on-resistance to drop from its peak value to 1.1 times its initial value is measured as the resistance recovery time. Switching energy consumption is calculated using a power analyzer (such as the Yokogawa WT3000) as the integrated sum of gate drive power and drain loss in each switching cycle. The average value over 100 consecutive cycles is taken to eliminate random fluctuations. The junction temperature (via a built-in thermal diode) and drain voltage waveform are simultaneously recorded to ensure strict alignment of the measured data with the compensation parameter combination.

[0093] Step S30104, determining an optimal historical gate voltage slope, an optimal historical gate voltage step amplitude, and an optimal historical dead time according to the historical gate voltage slope, the historical gate voltage step amplitude, and the historical dead time using the historical resistance recovery time and the historical switching energy consumption.

[0094] The method of determining the optimal historical gate voltage slope, the optimal historical gate voltage step amplitude, and the optimal historical dead time according to the historical gate voltage slope, the historical gate voltage step amplitude, and the historical dead time using the historical resistance recovery time and the historical switching energy consumption specifically includes the following steps:

[0095] Step S3010401: Obtain an indicator weight according to the historical resistance recovery time and the historical switch energy consumption.

[0096] Wherein, obtaining the indicator weight according to the historical resistance recovery time and the historical switch energy consumption specifically includes the following sub-steps:

[0097] Step S301040101: stack the historical resistance recovery time and the historical switch energy consumption by row to form an indicator matrix.

[0098] In this embodiment, for each parameter combination of historical gate voltage slope, historical gate voltage step amplitude and historical dead time, the corresponding resistance recovery time is used as the first row element of the matrix, and the switching energy consumption is used as the second row element of the matrix. They are arranged in sequence according to the test order of the parameter combination, and finally a 2-row and N-column indicator matrix is ​​constructed, where N is the total number of parameter combinations, to achieve structured storage of multiple groups of compensation parameter performance indicators.

[0099] Step S301040102: normalize the elements in the indicator matrix to obtain a normalized matrix.

[0100] The normalized matrix satisfies the following formula:

[0101] ,

[0102] in, is the first in the normalized matrix OK The value of the column, is the first OK The value of the column, and The first The maximum and minimum values ​​of the column.

[0103] The above formula, for each column of the indicator matrix, first calculates the difference between the current element and the minimum value in that column, reflecting the relative offset of the element within the column. It then divides the difference by the difference between the maximum and minimum values ​​in that column (i.e., the column range). The offset is then scaled by the column's fluctuation range, ultimately mapping each column of data uniformly to the interval [0,1]. This eliminates the impact of differences in dimensionality or numerical ranges among different indicators.

[0104] Step S301040103, calculate the proportion of each element in the normalized matrix in the corresponding column, and construct a probability matrix according to the proportion.

[0105] The probability matrix satisfies the following formula:

[0106] ,

[0107] in, is the probability matrix OK The value of the column, is the first in the normalized matrix OK The value of the column, is the sample size, is the first in the normalized matrix OK The value of the column.

[0108] For the normalized matrix All sample data in the column (a total of , corresponding to Go to row), first calculate the sum of all normalized values ​​in the column; then sum the current element Divide by the total of the column, and the resulting ratio is the probability matrix of the corresponding position After this processing, all the The sum is always 1. The normalized value of the same indicator is converted into a probability distribution in the form of column proportions so that the contribution of different indicators can be quantitatively compared through probability proportions.

[0109] Step S301040104, calculating the historical resistance recovery time and the degree of discreteness of the historical switch energy consumption respectively according to the probability matrix.

[0110] The degree of discreteness satisfies the following formula:

[0111] ,

[0112] in, For the The degree of dispersion of the indicators, is the sample size, is the probability matrix OK The value of the column.

[0113] For the probability matrix The column corresponding to the indicator (that is, the probability distribution of the indicator , first calculate the sum of the products of the probability of each sample and its own logarithm (essentially the reverse expression of negative information entropy), and then divide by Achieve normalization. When the probability distribution of the indicator is more uniform (such as the proportion of each sample is close to equal, showing a uniform distribution), the absolute value of the sum is larger, and after normalization The closer it is to 1, the higher the degree of dispersion of the indicator (the more dispersed the data fluctuations). Conversely, if the probability is concentrated in a few samples (for example, the proportion of a certain sample approaches 1, and the rest approaches 0), the summation approaches 0. The closer it is to 0, the lower the degree of dispersion (the more concentrated the data). In this way, the discrete characteristics of the indicator are mapped to the interval [0,1], making it easier to quantitatively compare the fluctuation differences of different indicators.

[0114] Step S301040105: Calculate the indicator weights of the historical resistance recovery time and the historical switch energy consumption according to the discrete degree.

[0115] The weights of historical resistance recovery time and historical switching energy consumption satisfy the following formula:

[0116] ,

[0117] in, For the The weight of the indicator, For the The degree of dispersion of the indicators, For the The degree of dispersion of an indicator.

[0118] Subtract the discreteness from 1 , the lower the degree of dispersion (the more concentrated the data, the smaller the fluctuation), the greater the contribution item, and then the sum of the dispersion of the two indicators is calculated as the denominator to achieve normalization. Finally, the indicator with lower dispersion, The larger the weight The higher the weight (because the data is more stable, the characteristics are more consistent, and the more representative the comprehensive assessment is), the lower the weight of the indicator with high dispersion (large fluctuation). Through this formula, the discrete characteristics of the two indicators are mapped to normalized weights, ensuring that the sum of the weights is 1.

[0119] Step S3010402: introducing a particle swarm optimization algorithm, and setting an objective function for the particle swarm optimization algorithm based on the indicator weights with the goal of minimizing resistance recovery time and switch energy consumption.

[0120] In this embodiment, a particle swarm optimization algorithm is introduced. A swarm of particles is randomly initialized within the allowable compensation parameter adjustment range (e.g., gate voltage slope 10-50V / ns, step amplitude ±10% of rated value, and dead time 20-100ns). Each particle represents a set of compensation parameter combinations to be optimized (gate voltage slope, step amplitude, and dead time). The resistance recovery time and switching energy consumption weights, calculated using the discreteness calculation, are then combined with the historically measured resistance recovery time and switching energy consumption data for each parameter set to construct an objective function. Using the weights as coefficients, the weighted sum of the resistance recovery time and switching energy consumption is calculated, with the goal of minimizing this weighted sum. During the algorithm iterations, each particle updates its movement speed and position based on its own historical optimal position and the global optimal position found by the entire particle swarm. During the update, nonlinear mapping is used to constrain the parameters within a safe threshold to avoid exceeding the physical limits of the device. The particle swarm continuously searches for the parameter combination that minimizes the objective function value until the preset maximum number of iterations is reached or the objective function value converges, and finally determines the optimal gate voltage slope, step amplitude and dead time, providing high-quality training samples for the parameter compensation prediction model.

[0121] The objective function satisfies the following formula:

[0122] ,

[0123] in, is the objective function, is the historical resistance recovery time weight, is the historical switching energy consumption weight, is the resistance recovery time corresponding to the iteration of the particle swarm optimization algorithm, is the minimum value of the historical resistance recovery time, is the switching energy consumption corresponding to the iteration of the particle swarm optimization algorithm, is the minimum value of historical switching energy consumption.

[0124] In the above formula, the minimum historical switching energy consumption and the minimum historical resistance recovery time are derived from historical detection data. Specifically, when constructing the training sample, the historical on-resistance, off-state drain voltage, junction temperature, and turn-on delay are first obtained. Then, the compensation parameter adjustment range is determined and the samples are traversed at a predetermined step size to obtain multiple historical gate voltage slopes, step amplitudes, and dead times. The historical resistance recovery time and historical switching energy consumption of the GaN device corresponding to each set of parameters are then measured. Finally, the minimum value is directly extracted from the resistance recovery time and switching energy consumption data obtained from these historical measurements.

[0125] Step S3010403, using the particle swarm optimization algorithm to calculate the optimal historical gate voltage slope, the optimal historical gate voltage step amplitude, and the optimal historical dead time according to the historical gate voltage slope, the historical gate voltage step amplitude, the historical dead time, the historical resistance recovery time, and the historical switching energy consumption.

[0126] In this embodiment, a parameter index mapping model is constructed based on the historical gate voltage slope, the historical gate voltage step amplitude, the historical dead time, the historical resistance recovery time and the historical switching energy consumption; the historical dead time and the historical resistance recovery time are weightedly summed using the index weight to obtain a comprehensive index set; the historical gate voltage slope, the historical gate voltage step amplitude and the historical dead time corresponding to the minimum value in the comprehensive index set are selected as the initial population; the objective function value of the objective function is calculated based on the initial population using the parameter index mapping model; the particle swarm optimization algorithm is iterated based on the objective function value; and the optimal historical gate voltage slope, the optimal historical gate voltage step amplitude and the optimal historical dead time are obtained based on the iteration.

[0127] Based on the historical gate voltage slope, historical gate voltage step amplitude, and historical dead time as input features, and the historical resistance recovery time and historical switching energy consumption as output labels, a neural network is used to construct a parameter indicator mapping model.

[0128] A weighted sum is performed on the historical resistance recovery time and historical switching energy consumption corresponding to each set of historical gate voltage slope, historical gate voltage step amplitude, and historical dead time, that is, the comprehensive index = historical resistance recovery time × corresponding weight + historical switching energy consumption × corresponding weight. This results in a comprehensive index set containing the comprehensive evaluation values ​​of all parameter combinations. This index set is used to quantify the overall performance of each parameter combination.

[0129] The corresponding weights come from the weight calculation of the historical resistance recovery time and the historical switch energy consumption in step S301040105.

[0130] The element with the smallest value in the comprehensive index set is selected. This element corresponds to a set of historical gate voltage slope, historical gate voltage step amplitude, and historical dead time. This parameter combination is added to the initial population of the particle swarm optimization algorithm. This operation uses the minimization of the comprehensive index as the basis for selecting the initial population, ensuring that the particle swarm begins its iterative search from the point with the best comprehensive performance among the historical parameters, providing a reasonable starting point for the subsequent optimization process.

[0131] Each set of historical gate voltage slope, historical gate voltage step amplitude and historical dead time in the initial population is taken as input and input into the constructed parameter index mapping model to obtain the corresponding predicted historical resistance recovery time and predicted historical switching energy consumption. Then, the objective function value of each set of parameter combinations is calculated according to the objective function formula, providing a quantitative optimization basis for the iteration of the particle swarm optimization algorithm.

[0132] Based on the iterative process of the particle swarm optimization algorithm, each particle represents a set of gate voltage slope, step amplitude, and dead time parameter combinations. A parameter-index mapping model is used to predict the resistance recovery time and switching energy consumption corresponding to each parameter combination. The objective function value is calculated and the individual optimal position of the particle and the global optimal position of the swarm are updated. After a preset number of iterations or after the objective function convergence condition is reached, the parameter combination corresponding to the global optimal position is selected to obtain the optimal historical gate voltage slope, optimal historical gate voltage step amplitude, and optimal historical dead time that minimizes the objective function.

[0133] It should be noted that the population value range of the particle swarm optimization algorithm needs to be reasonably determined according to the adjustment range allowed by the parameters.

[0134] Step S30105 , constructing a training sample using the historical on-resistance, the historical off-state drain voltage, the historical junction temperature, the historical turn-on delay, the optimal historical gate voltage slope, the optimal historical gate voltage step amplitude, and the optimal historical dead time.

[0135] In this embodiment, the four types of monitoring data, namely historical on-resistance, historical off-state drain voltage, historical junction temperature, and historical turn-on delay, are first associated one-to-one with the three types of compensation parameters, namely the optimal historical gate voltage slope, the optimal historical gate voltage step amplitude, and the optimal historical dead time, obtained by the particle swarm optimization algorithm. For each set of historical operating condition data (i.e., each set of historical on-resistance, off-state drain voltage, junction temperature, and turn-on delay), the corresponding optimal compensation parameters are matched to form a complete sample record. Then, the feature data in all sample records are normalized, and data of different dimensions and numerical ranges are uniformly mapped to the [0,1] interval to eliminate the impact of dimensional differences on model training. Finally, all processed sample records are organized in a certain format to form a structured training sample set for training the parameter compensation prediction model, in which each sample contains the historical monitoring feature as the input label and the corresponding optimal compensation parameter as the output label.

[0136] Step S302: construct a parameter compensation prediction model, and use the training samples to train and evaluate the parameter compensation prediction model.

[0137] A parameter compensation prediction model was constructed using a deep neural network, employing a three-layer, fully connected network architecture. The number of input layer nodes matched the dimensions of the historical monitoring features of the training samples (e.g., on-resistance, off-state drain voltage, junction temperature, and turn-on delay, a total of four features). The hidden layer had 128 nodes and employed a Relu (Reinforced Luminance) activation function. The number of output layer nodes corresponded to the dimensions of the compensation parameters (e.g., gate voltage slope, step amplitude, and dead time, a total of three parameters). During training, the constructed training samples were split into training, validation, and test sets in an 8:1:1 ratio. The Adam optimizer was used to minimize the mean squared error loss function, with a batch size of 32 and an initial learning rate of 0.001, which was dynamically adjusted based on the validation set loss. During the evaluation phase, the average absolute error between the predicted compensation parameters and the actual optimal parameters (e.g., gate voltage slope error ≤ 2V / ns, step amplitude error ≤ 0.3V, dead time error ≤ 5ns) was calculated using the test set. At the same time, a scatter plot and confusion matrix of the predicted and true values ​​were drawn to verify the generalization ability of the model and ensure that the deep neural network can accurately learn the nonlinear mapping relationship between historical monitoring features and the optimal compensation parameters.

[0138] Based on the risk level, the real-time monitoring data is input into the parameter compensation prediction model to obtain compensation parameters. The risk level here refers to medium risk. That is, when the risk level prediction model outputs medium risk, the compensation parameters are determined according to the real-time monitoring data using the parameter compensation prediction model, and then the adjustment method of the gradient gate voltage is determined using the compensation parameters.

[0139] Step S4: adjusting the gate voltage slope, gate voltage step amplitude, and dead time of the next switching cycle of the GaN device based on the compensation parameters.

[0140] The step of adjusting the gate voltage slope, gate voltage step amplitude, and dead time of the next switching cycle of the GaN device based on the compensation parameters specifically includes the following sub-steps:

[0141] Step S401 : obtaining a dead time reference value, a gate voltage slope threshold, and a gate voltage step amplitude threshold of a GaN device.

[0142] In this embodiment, the dead time reference value, gate voltage slope threshold, and gate voltage step amplitude threshold of the GaN device need to be determined based on the device data sheet and safe operating area (SOA).

[0143] The dead time reference value is determined by measuring the shoot-through critical time of the upper and lower bridge arms through hardware circuitry. At a junction temperature of 25°C, the device is driven at a switching frequency of 100kHz, and the dead time is gradually reduced until shoot-through current occurs. A reference value (typically 50ns) is set at 1.5 times the critical value.

[0144] The upper limit of the gate voltage slope threshold is determined by electromagnetic interference (EMI) limits. By measuring the drive signal harmonics with a spectrum analyzer, the minimum slope that causes EMI to exceed the standard (such as exceeding the CISPR 25 Class 5 standard) is set as the upper limit (generally 50V / ns). The lower limit is set at 10V / ns to avoid excessive switching losses.

[0145] The gate voltage step amplitude threshold is based on the rated gate voltage. Positive steps must not exceed +10% of the rated value (e.g., the upper limit is +6.6V when the rated voltage is +6V), and negative steps must not fall below -10% of the rated value (e.g., the lower limit is -1.8V when the rated voltage is -2V), ensuring that the gate oxide breakdown threshold is not exceeded. All thresholds must be verified through reliability testing in high-temperature (150°C) and low-temperature (-40°C) environments. Ultimately, a threshold table with temperature coefficients is generated to provide a benchmark for dynamic parameter adjustment.

[0146] Step S402 : Correcting the dead time reference value according to the junction temperature to obtain a dynamic dead time threshold.

[0147] The step of correcting the dead time reference value according to the junction temperature to obtain the dynamic dead time threshold specifically includes the following sub-steps:

[0148] Step S40201, obtaining multiple groups of test junction temperatures and test dead time reference values.

[0149] When experimentally obtaining test dead-time benchmark values ​​corresponding to multiple sets of different junction temperatures, junction temperature test points were set in 5°C increments from 25°C to 150°C. The GaN device was driven at a switching frequency of 100kHz for each set of junction temperatures, and the dead-time was gradually reduced until the oscilloscope detected shoot-through currents in the upper and lower bridge arms (abnormal drain current spikes). The minimum dead-time before shoot-through was multiplied by 1.5 as the test dead-time benchmark value for that junction temperature. The junction temperature and the corresponding benchmark value were simultaneously recorded to form a junction temperature dead-time benchmark data set.

[0150] Step S40202: construct a regression model using the multiple groups of test junction temperatures and the test dead time reference value.

[0151] In this embodiment, a quadratic polynomial regression model is used, and coefficients are calculated using the least squares method to minimize the mean squared error between the model's predicted values ​​and the measured reference values. To verify the model's generalization capability, the dataset is split into training and test sets at an 8:2 ratio. Regularization parameters are adjusted through cross-validation to ensure that the model's prediction error is less than 3ns across the entire temperature range. Ultimately, a regression model is constructed that accurately reflects the mapping between junction temperature and dead time.

[0152] Step S40203: input the junction temperature into the regression model to obtain a dynamic dead time threshold.

[0153] In this embodiment, the regression model has fully learned how junction temperature changes affect dead time requirements. When a dynamic dead time threshold is needed, the current GaN device junction temperature data, obtained through real-time monitoring or calculation, is accurately input into the constructed regression model. Based on the learned mapping, the model rapidly calculates and outputs a dynamic dead time threshold adapted to the current junction temperature operating conditions. This threshold changes in real time with the junction temperature, effectively ensuring reliable device operation under varying temperature conditions and avoiding faults such as upper and lower bridge arm shoot-through caused by improper dead time settings.

[0154] Step S403 : Constraining the compensation parameters by using the gate voltage slope threshold, the gate voltage step amplitude threshold, and the dynamic dead time threshold.

[0155] In this embodiment, once the model outputs the three compensation parameters—gate voltage slope, gate voltage step amplitude, and dead time—the system immediately initiates a threshold verification mechanism. For the gate voltage slope parameter, the predicted value is compared with the gate voltage slope threshold (including upper and lower limits). If the predicted gate voltage slope exceeds the set upper slope threshold, the compensation parameter is corrected to the upper slope threshold to avoid severe electromagnetic interference caused by an excessively high slope, increase signal harmonics, and affect the system's electromagnetic compatibility, while also preventing unnecessary stress on devices caused by excessively high slopes.

[0156] At the same time, we can also consider the situation where the gate voltage slope is too small. If the predicted value is less than the slope lower limit threshold, the parameter is adjusted to the slope lower limit threshold because the too small slope will cause the switching loss to increase sharply and reduce the working efficiency of the device.

[0157] For the gate voltage step amplitude parameter, it is also compared with the corresponding amplitude threshold (upper and lower limits). If the predicted step amplitude exceeds the upper threshold, it needs to be corrected to the upper amplitude threshold because the excessively high amplitude may damage the gate oxide layer and threaten the reliability of the device.

[0158] At the same time, we can also consider the situation where the gate voltage step amplitude is too small. If it is lower than the lower limit threshold, the device switching process cannot be effectively controlled due to the low amplitude and the current collapse suppression effect is affected, so it is adjusted to the lower limit threshold.

[0159] For the dead time parameter, the predicted dead time is compared with the dynamic dead time threshold. If the predicted value is less than the threshold, it is easy to cause the upper and lower bridge arms to pass through, causing a short circuit fault and damaging the device. Therefore, it is also adjusted to the dynamic dead time threshold.

[0160] Through such a constraint mechanism, the compensation parameters ultimately acting on the GaN device are strictly limited to the parameter range for safe and stable operation of the device, thereby achieving effective optimization and reliable protection of device performance.

[0161] At the same time, the situation where the dead time is too large can also be considered. For example, if the predicted dead time is more than twice the dynamic dead time threshold, it is corrected to twice the dynamic dead time threshold to avoid excessive increase in switching loss.

[0162] Step S404 : Using the constrained compensation parameters, adjust the gate voltage slope, gate voltage step amplitude, and dead time of the next switching cycle of the GaN device.

[0163] In this embodiment, since the GaN device operates in switching mode, in order to promptly respond to dynamic changes in device operating conditions (such as junction temperature and drain voltage), it is necessary to use the next switching cycle as an adjustment cycle to achieve real-time closed-loop control. Compensation parameters are used to ensure that the GaN device obtains optimal drive parameters based on the latest operating conditions in each new switching cycle, effectively suppressing current collapse, reducing switching losses, and ensuring stable operation under high-frequency conditions.

[0164] It should be noted that gradual gate voltage boost optimizes the switching process by adjusting the gate voltage slope and step amplitude. However, changes in device junction temperature can affect carrier mobility and switching speed. If the dead time is fixed, increased switching delays may cause shoot-through in the upper and lower bridge arms at high temperatures, while increased switching speeds at low temperatures may lead to redundant dead time and increase losses. Dead time adjustment dynamically adjusts to the junction temperature, extending the dead time at high temperatures to avoid shoot-through and shortening it at low temperatures to reduce losses. This, combined with the gate voltage parameters, can match the actual device switching timing during the gradual boost process, further suppressing current collapse and optimizing system efficiency.

[0165] like Figure 2 As shown, a GaN device current collapse suppression system is also provided, including: a processor, an input device, an output device and a memory, wherein the processor, the input device, the output device and the memory are interconnected, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute relevant steps of a relevant embodiment of a GaN device current collapse suppression method of the present invention.

[0166] The present invention provides a GaN device current collapse suppression system. Each functional component can be integrated into a single processing unit, each component can exist physically separately, or two or more components can be integrated into a single unit. These integrated components can be implemented as either hardware or software functions.

[0167] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and description of the present invention.

Claims

1. A method for suppressing current collapse of a GaN device, characterized in that: The method comprises: Obtain real-time monitoring data of GaN devices; Constructing a risk level prediction model, and inputting the real-time monitoring data into the risk level prediction model to obtain a risk level; Constructing a parameter compensation prediction model, and inputting the real-time monitoring data into the parameter compensation prediction model to obtain compensation parameters based on the risk level being medium risk; The constructing of the risk level prediction model comprises: Obtaining a tunneling activation voltage threshold, and using the tunneling activation voltage threshold to set a high drain voltage penalty term; Obtaining a trap activation temperature threshold, and using the trap activation temperature threshold to set a high temperature penalty item; Setting a comprehensive loss function using the high drain voltage penalty term and the high temperature penalty term; Constructing a risk level prediction model based on multinomial logistic regression using the comprehensive loss function; The gate voltage slope, gate voltage step amplitude and dead time of the next switching cycle of the GaN device are adjusted based on the compensation parameters.

2. A method for suppressing current collapse of a GaN device according to claim 1, characterized in that: The obtaining of real-time monitoring data of the GaN device includes: Monitoring the GaN device to obtain the on-resistance, off-state drain voltage, junction temperature, and turn-on delay of the GaN device during one switching cycle; The on-resistance, the off-state drain voltage, the junction temperature and the turn-on delay are normalized to obtain real-time monitoring data.

3. The method for suppressing current collapse of a GaN device according to claim 1, wherein: The constructing parameter compensation prediction model comprises: Acquire historical test data and construct training samples based on the historical test data; A parameter compensation prediction model is constructed, and the parameter compensation prediction model is trained and evaluated using the training samples.

4. A method for suppressing current collapse of a GaN device according to claim 3, characterized in that: The acquiring of historical detection data and constructing training samples based on the historical detection data includes: Get historical on-resistance, historical off-state drain voltage, historical junction temperature and historical turn-on delay; Determining an adjustment range of GaN device parameters, and obtaining a plurality of historical gate voltage slopes, historical gate voltage step amplitudes, and historical dead time by traversing sampling with a predetermined step size according to the adjustment range; Measuring the historical resistance recovery time and historical switching energy consumption of the GaN device according to the historical gate voltage slope, the historical gate voltage step amplitude and the historical dead time; Determine an optimal historical gate voltage slope, an optimal historical gate voltage step amplitude, and an optimal historical dead time according to the historical gate voltage slope, the historical gate voltage step amplitude, and the historical dead time using the historical resistance recovery time and the historical switching energy consumption; A training sample is constructed using the historical on-resistance, the historical off-state drain voltage, the historical junction temperature, the historical turn-on delay, the optimal historical gate voltage slope, the optimal historical gate voltage step amplitude, and the optimal historical dead time.

5. A method for suppressing current collapse of a GaN device according to claim 4, characterized in that: The determining of the optimal historical gate voltage slope, the optimal historical gate voltage step amplitude, and the optimal historical dead time by using the historical resistance recovery time and the historical switch energy consumption according to the historical gate voltage slope, the historical gate voltage step amplitude, and the historical dead time includes: Obtaining an indicator weight according to the historical resistance recovery time and the historical switch energy consumption; Introducing a particle swarm optimization algorithm, and combining the indicator weights to set an objective function for the particle swarm optimization algorithm with the goal of minimizing resistance recovery time and switching energy consumption; The particle swarm optimization algorithm is used to calculate the optimal historical gate voltage slope, the optimal historical gate voltage step amplitude, and the optimal historical dead time according to the historical gate voltage slope, the historical gate voltage step amplitude, the historical dead time, the historical resistance recovery time, and the historical switching energy consumption.

6. A method for suppressing current collapse of a GaN device according to claim 5, characterized in that: Obtaining the indicator weight according to the historical resistance recovery time and the historical switch energy consumption includes: Stacking the historical resistance recovery time and the historical switch energy consumption in rows to form an indicator matrix; Normalizing the elements in the indicator matrix to obtain a normalized matrix; Calculating the proportion of each element in the normalized matrix in the corresponding column, and constructing a probability matrix based on the proportion; Calculate the historical resistance recovery time and the discrete degree of the historical switch energy consumption respectively according to the probability matrix; The indicator weights of the historical resistance recovery time and the historical switch energy consumption are calculated according to the discrete degree.

7. A method for suppressing current collapse of a GaN device according to claim 2, characterized in that: The adjusting the gate voltage slope, gate voltage step amplitude and dead time of the next switching cycle of the GaN device based on the compensation parameters includes: Obtaining a dead time reference value, a gate voltage slope threshold, and a gate voltage step amplitude threshold of a GaN device; Correcting the dead time reference value according to the junction temperature to obtain a dynamic dead time threshold; Constraining the compensation parameters by using the gate voltage slope threshold, the gate voltage step amplitude threshold, and the dynamic dead time threshold; The constrained compensation parameters are used to adjust the gate voltage slope, gate voltage step amplitude and dead time of the next switching cycle of the GaN device.

8. A method for suppressing current collapse of a GaN device according to claim 7, characterized in that: The step of correcting the dead time reference value according to the junction temperature to obtain a dynamic dead time threshold value includes: Obtain multiple groups of test junction temperatures and test dead time reference values; Constructing a regression model using the multiple groups of test junction temperatures and the test dead time reference value; The junction temperature is input into the regression model to obtain a dynamic dead time threshold.

9. A GaN device current collapse suppression system, characterized in that: include: A processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute a GaN device current collapse suppression method according to any one of claims 1 to 8.

Images