SiC MOSFET junction temperature mechanical stress wave monitoring method and system

Abstract

The invention discloses a mechanical stress wave monitoring method and system for SiC MOSFET junction temperature, and relates to the technical field of semiconductor device detection. The method comprises the following steps: enabling a device to reach a plurality of preset junction temperatures through a heating device, and obtaining a plurality of turn-off current values through an adjusting circuit at each preset junction temperature; collecting a mechanical stress wave signal excited in the transient state of the switch each time, and synchronously collecting a corresponding turn-off current; extracting a logarithm root mean square, a peak-to-peak value and kurtosis from the mechanical stress wave signal to form an acoustic feature vector, and combining the acoustic feature vector with the turn-off current to form an input feature vector; establishing a junction temperature prediction model by taking the input feature vector as an input and a preset junction temperature as an output; and realizing real-time prediction of the junction temperature of the SiC MOSFET device by using the junction temperature prediction model. According to the invention, non-intrusive and high-precision online monitoring of the junction temperature of the SiC MOSFET device is realized.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor device testing technology, and particularly relates to a method and system for monitoring the mechanical stress wave of SiC MOSFET junction temperature. Background Technology

[0002] Silicon carbide (SiC) MOSFETs, with their superior characteristics such as high switching speed, high voltage withstand capability, and low conduction loss, have become core components in high-power-density power electronic systems such as electric vehicles and smart grids. However, under long-term high-frequency switching and high-voltage, high-current stress, the internal thermal stress concentration of the device can easily lead to material fatigue and aging, resulting in an increased risk of failure. Therefore, accurate online junction temperature monitoring is the technical foundation for implementing active thermal management and ensuring the long-term reliable operation of SiC MOSFETs.

[0003] Currently, power device junction temperature monitoring technologies are mainly divided into three categories:

[0004] (1) Thermal network model method: The junction temperature is estimated by establishing a thermal resistance-capacitance network of the device package. This method is low-cost and easy to implement, but its accuracy depends heavily on the precision of the model parameters and the accuracy of the real-time loss calculation. Since the thermal conductivity of the packaging material is temperature-dependent, the thermal network parameters will fluctuate significantly over a wide temperature range, resulting in significant estimation errors.

[0005] (2) Temperature-Sensitive Electrical Parameter Method (TSEP): This method uses electrical parameters that are strongly correlated with temperature (such as on-state voltage drop, switching delay time, etc.) to infer the junction temperature. This method has high sensitivity and fast response speed and is currently the mainstream technology. However, it has inherent defects: there is electrical coupling between the measurement circuit and the main power circuit, which requires complex electrical isolation design; at the same time, it has high requirements for the bandwidth and accuracy of the sampling circuit, which increases the system cost and implementation difficulty.

[0006] (3) Method based on mechanical stress waves: Recent studies have found that the instantaneous power loss generated by the switching transients of power devices creates a huge temperature gradient inside the chip, which excites mechanical stress waves based on the thermoelastic effect and propagates outward. Theoretically, the junction temperature can be indirectly assessed by analyzing this stress wave signal, which has the natural advantages of being non-invasive and electrically isolated. However, existing monitoring schemes based on mechanical stress waves have two major bottlenecks in practical applications:

[0007] First, there is the issue of thermal degradation at the hardware acquisition end. Traditional solutions directly attach the acoustic emission sensor to the surface of high-temperature devices. When heat is conducted to the piezoelectric element, it induces a thermoelectric effect and causes the interfacial acoustic coupling agent to fail. This results in an irreversible nonlinear decay of the sensor's sensitivity, completely disrupting the stable mapping relationship between stress wave amplitude and junction temperature.

[0008] Secondly, there is the problem of limited data processing features. Mechanical stress waves are the dynamic modulation result of strong coupling between multiple physical fields (electricity, heat, and machinery): the operating current determines the power loss and affects the thermal shock intensity; the junction temperature affects the waveform morphology by changing the Young's modulus and damping coefficient of the material. However, existing technologies mostly follow the traditional acoustic emission detection approach, extracting only a single energy feature (such as signal energy or root mean square value) and combining it with simple polynomial fitting for temperature estimation. This extreme dimensionality reduction process compresses the rich waveform distortion information caused by transient thermal shock gradients and high-temperature softening of materials. As a result, the single-feature model cannot decouple the nonlinear synergistic modulation effect of current and temperature on mechanical stress waves, leading to large junction temperature estimation errors under actual working conditions, which is difficult to meet the needs of engineering applications.

[0009] In summary, existing technologies cannot solve the problems of sensor thermal degradation and the inability to decouple the current-temperature nonlinear relationship due to single features. There is an urgent need for an online monitoring method for SiC MOSFET junction temperature that combines high accuracy, high anti-interference capability, and complete electrical isolation. Summary of the Invention

[0010] To address the aforementioned deficiencies in the existing technology, the present invention aims to provide a method and system for monitoring the mechanical stress wave of SiC MOSFET junction temperature, thereby solving at least one of the problems in existing mechanical stress wave monitoring schemes: signal distortion caused by high-temperature thermal degradation of the sensor and large junction temperature estimation error due to the inability of a single feature model to decouple the nonlinear coupling effect of current and temperature.

[0011] This invention solves the above-mentioned technical problems through the following technical solution: a method for monitoring the mechanical stress wave of SiC MOSFET junction temperature, comprising:

[0012] The SiC MOSFET device is heated to multiple preset junction temperatures. At each preset junction temperature, the SiC MOSFET device is adjusted to obtain multiple different turn-off current values ​​during multiple switching transients, with each switching transient corresponding to one of the turn-off current values. The mechanical stress wave signal excited by each switching transient is collected by an acoustic sensor, and the turn-off current corresponding to that switching transient is collected synchronously by a current acquisition unit to obtain the mechanical stress wave signal under different preset junction temperatures and different turn-off currents.

[0013] The acoustic feature vectors, including at least the root mean square, peak-to-peak value, and kurtosis, are extracted from the mechanical stress wave signals under different preset junction temperatures and different turn-off currents. The acoustic feature vectors are then combined with the corresponding turn-off currents to form the input feature vectors.

[0014] Using the input feature vector as input and the corresponding preset junction temperature as output, a mapping relationship between the input features and the junction temperature is established to obtain a junction temperature prediction model;

[0015] The current mechanical stress wave signal and current turn-off current of the SiC MOSFET device under test are acquired in real time. The acoustic feature vector of the current mechanical stress wave signal is extracted and combined with the current turn-off current and input into the junction temperature prediction model. The junction temperature prediction model outputs the current junction temperature prediction value of the SiC MOSFET device under test.

[0016] This invention constructs a complete sample dataset covering different junction temperatures and different current combinations. The input (mechanical stress wave signal, turn-off current) and output (preset junction temperature) of each sample correspond one-to-one, providing high-quality training data for subsequent machine learning modeling. This enables the trained junction temperature prediction model to have good generalization ability and be applicable to complex working conditions such as load changes.

[0017] This invention departs from the traditional approach of extracting only a single energy feature (such as the root mean square value), and instead extracts a three-dimensional acoustic feature vector from the acquired mechanical stress wave signal, including at least the logarithmic root mean square, peak-to-peak value, and kurtosis. The logarithmic root mean square characterizes and linearizes the nonlinear growth trend of macroscopic acoustic wave energy with temperature; the peak-to-peak value captures the maximum instantaneous thermal shock stress intensity generated by switching transients; and the kurtosis quantifies the steepness and tail thickness of the signal probability density distribution, sensitively capturing waveform distortion caused by high-temperature softening of the material. This multi-dimensional feature extraction strategy comprehensively characterizes the mechanical stress wave signal from three dimensions: energy, impact intensity, and waveform morphology, fully preserving the rich waveform distortion information caused by transient thermal shock gradients and high-temperature softening of the material's elastic modulus. Compared to single features, the three-dimensional feature vector provides richer input information for subsequent models, making it possible to distinguish the different effects of current changes and temperature changes on the signal.

[0018] The extracted three-dimensional acoustic feature vector is combined with the corresponding turn-off current to form the input feature vector, and the corresponding preset junction temperature is used as the output to establish a nonlinear mapping relationship between the input features and the junction temperature. This data-driven modeling strategy fully utilizes the ability of machine learning algorithms such as Gaussian process regression to handle high-dimensional, nonlinear, and small-sample problems. By constructing a modeling sample set covering different preset junction temperatures and different turn-off currents, the model can learn the variation law of mechanical stress wave signal under different operating conditions and automatically separate the complex nonlinear modulation effect of current change and temperature change on the signal.

[0019] By acquiring mechanical stress wave signals using acoustic sensors, non-invasive monitoring with complete electrical isolation from the main power circuit is achieved. This eliminates the need for complex electrical isolation designs and avoids the electrical coupling problems inherent in the Thermo-Sensitive Electrical Parameter (TSEP) method. During real-time monitoring, the current mechanical stress wave signal and current turn-off current of the device under test are simultaneously acquired. Acoustic feature vectors are extracted and input into the junction temperature prediction model, allowing for real-time output of the current junction temperature estimate. This process does not affect the normal operation of the device and is suitable for online monitoring scenarios. Furthermore, because the input features include turn-off current information, the model can adapt to complex operating conditions such as varying load conditions.

[0020] Furthermore, the acoustic sensor is indirectly connected to the SiC MOSFET device via a thermal barrier, specifically configured as follows:

[0021] The first surface of the heat transfer barrier is dry-contacted to the first surface of the SiC MOSFET device, the acoustic sensor is abutted to the second surface of the heat transfer barrier, and an acoustic coupling agent is coated between the acoustic sensor and the second surface of the heat transfer barrier; the heating device is disposed on the second surface side of the SiC MOSFET device opposite to the first surface.

[0022] This invention completely solves the thermal degradation problem caused by directly attaching traditional acoustic emission sensors to the surface of high-temperature devices by introducing a heat transfer barrier and employing a combination of dry contact and localized heating. Specifically, the first surface of the heat transfer barrier abuts against the first surface of the SiC MOSFET device in a dry contact manner, blocking the heat conduction path from the device to the sensor. The heating device is located on the second surface of the device opposite to the first surface, providing localized heating only, keeping both the heat transfer barrier and the acoustic sensor at ambient temperature. This design avoids the pyroelectric effect induced by heat conduction to the piezoelectric crystal and eliminates the risk of interface acoustic coupling agent failure due to high temperatures. An acoustic coupling agent is coated between the sensor and the second surface of the heat transfer barrier, ensuring efficient transmission of sound waves from the heat transfer barrier to the sensor. Through the above hardware design, the signal acquisition front end of this invention can maintain stable sensing sensitivity over a long period, ensuring that the mapping relationship between the amplitude of the mechanical stress wave signal and the chip junction temperature does not undergo nonlinear attenuation, laying a reliable signal foundation for subsequent high-precision junction temperature inversion.

[0023] Furthermore, the heat transfer barrier is a waveguide rod made of polyimide; the acoustic coupling agent is a silicon-based coupling agent.

[0024] Polyimide has an extremely low thermal conductivity (approximately 0.2 W / (m·K)), which is much lower than that of metals and general polymers. This can block the conduction of heat from high-temperature devices to acoustic sensors to the greatest extent, ensuring that the sensors are always at ambient temperature and completely eliminating the problem of thermal degradation. At the same time, polyimide, as a waveguide material, has good mechanical strength and processing performance, and can stably transmit mechanical stress waves without introducing additional signal attenuation or distortion.

[0025] Silicon-based coupling agents possess excellent high-temperature stability and acoustic transmittance, forming a good acoustic matching layer between the sensor and the waveguide rod. This ensures efficient transmission of mechanical stress waves to the sensor while preventing coupling agent failure or performance degradation due to high temperatures, further guaranteeing the stability and fidelity of signal acquisition.

[0026] Furthermore, before extracting the acoustic feature vector, the method also includes filtering the mechanical stress wave signal; the frequency band of the filtering is set according to the noise distribution range determined by the lead breakage test, so as to filter out low-frequency mechanical background noise and retain the effective signal.

[0027] Lead breakage testing can accurately identify the structural resonant frequency and frequency band distribution of mechanical background noise of the measurement platform. Based on this, the set filter frequency band can selectively filter out low-frequency mechanical background noise (such as 0~50kHz) while retaining the high-frequency components of the effective signal excited by thermal stress of the device (such as 50~150kHz), thereby significantly improving the signal-to-noise ratio. The noise-filtered signal makes the subsequently extracted logarithmic root mean square, peak-to-peak value, kurtosis and other features more realistically reflect the modulation effect of device junction temperature and turn-off current on mechanical stress wave, avoiding feature distortion caused by noise interference, and laying the foundation for high-precision modeling.

[0028] Furthermore, the root mean square logarithm is calculated according to the following formula:

[0029] ;

[0030] Where F1 represents the logarithmic root mean square; RMS represents the root mean square; and N represents the total number of sampling points. This represents the mechanical stress wave signal at the nth sampling point;

[0031] The peak value is calculated using the following formula:

[0032] ;

[0033] Where F2 represents the peak-to-peak value;

[0034] The kurtosis is calculated using the following formula:

[0035] ;

[0036] Where F3 represents kurtosis; This represents the mean value of the mechanical stress wave signal; This represents the standard deviation of the mechanical stress wave signal.

[0037] The logarithmic root mean square formula takes the logarithm of the root mean square value of the mechanical stress wave signal. On the one hand, it preserves the trend information of signal energy change with temperature, and on the other hand, it linearizes the originally nonlinear growth trend, which is convenient for subsequent model learning. The peak-to-peak value formula directly calculates the difference between the maximum and minimum values ​​of the signal, accurately reflecting the maximum instantaneous thermal shock stress intensity generated by the switching transient, which is a key information that cannot be characterized by a single energy feature. The kurtosis formula calculates fourth-order statistics based on the mean and standard deviation of the signal, quantifies the steepness and tail thickness of the signal probability density distribution, and can sensitively capture the small waveform distortion caused by the high-temperature softening of materials.

[0038] Furthermore, the acoustic feature vector also includes frequency domain features or time-frequency domain features; the frequency domain features include frequency domain peaks or signal energy in a specific frequency band.

[0039] Frequency domain features can reveal the energy distribution of mechanical stress wave signals at different frequency components, while time-frequency domain features can simultaneously reflect the changes in the signal in both time and frequency dimensions, providing additional information dimensions for junction temperature monitoring. Frequency domain peak values ​​reflect the frequency components with the most concentrated energy in the signal, while specific frequency band signal energy characterizes the energy changes within a specific frequency band related to material properties or structural modes. These features complement the time-domain three-dimensional acoustic features, enabling a more comprehensive characterization of the multi-dimensional properties of mechanical stress wave signals and further improving the prediction accuracy and robustness of the model under complex working conditions.

[0040] Furthermore, machine learning algorithms are used to establish a mapping relationship between input features and junction temperature.

[0041] The introduction of machine learning algorithms enables the model to automatically learn the complex nonlinear relationship between input features and junction temperature, without relying on the precise parameters of the physical model, thus overcoming the dependence of traditional thermal network models on parameter accuracy.

[0042] Furthermore, the machine learning algorithm is a Gaussian process regression algorithm, which establishes the mapping relationship between input features and junction temperature, including:

[0043] Define a kernel function; the kernel function is an additive composite kernel function, used to measure the covariance between input feature vectors;

[0044] Initialize the hyperparameters of the kernel function, including signal variance, length scale, and noise variance;

[0045] Based on the defined kernel function and the input feature vectors, construct the covariance matrix between the input feature vectors;

[0046] The logarithmic marginal likelihood is calculated based on the covariance matrix, and the optimal hyperparameters of the kernel function are obtained by optimizing the algorithm to maximize the logarithmic marginal likelihood, thereby obtaining the junction temperature prediction model.

[0047] Gaussian process regression, as a nonparametric Bayesian method, can not only establish a nonlinear mapping between input features and junction temperature, but also output prediction variance to characterize the uncertainty of the prediction results, providing confidence interval information for junction temperature monitoring and enhancing the reliability of the monitoring results. The introduction of additive composite kernel function enables the model to capture the contribution of different dimensions of input features (such as different components of acoustic features and turn-off current) to junction temperature, effectively decoupling the complex nonlinear modulation effect of current and temperature on mechanical stress waves. By maximizing the logarithmic marginal likelihood to optimize the hyperparameters, the optimal parameter configuration of the kernel function can be automatically determined, enabling the model to achieve the best fitting effect and generalization ability.

[0048] Furthermore, the input feature vector consists only of the acoustic feature vector and does not include the turn-off current; the junction temperature prediction model is a pure acoustic prediction model used for junction temperature monitoring without electrical measurement assistance.

[0049] Even in scenarios without electrical measurement assistance or with severe load fluctuations, the pure acoustic prediction model can still independently monitor junction temperature even if the current acquisition unit fails or cannot obtain an accurate shutdown current value.

[0050] Based on the same concept, the present invention also provides a mechanical stress wave monitoring system for SiC MOSFET junction temperature, comprising:

[0051] A heating device, located on the surface of the SiC MOSFET device, is used to bring the SiC MOSFET device to multiple different preset junction temperatures;

[0052] The adjustment circuit is used to adjust parameters at each preset junction temperature so that the SiC MOSFET device obtains multiple different turn-off current values ​​in multiple switching transients, and each switching transient corresponds to one of the turn-off current values.

[0053] Acoustic sensors are used to acquire mechanical stress wave signals generated during each switching transient.

[0054] The current acquisition unit is used to synchronously acquire the turn-off current corresponding to each switching transient.

[0055] The feature extraction unit is used to extract acoustic feature vectors, including at least logarithmic root mean square, peak-to-peak value and kurtosis, from mechanical stress wave signals under different preset junction temperatures and different turn-off currents, and combine the acoustic feature vectors with the corresponding turn-off currents to form an input feature vector.

[0056] The model building unit is used to establish a mapping relationship between the input features and the junction temperature by taking the input feature vector as input and the corresponding preset junction temperature as output, so as to obtain a junction temperature prediction model.

[0057] The real-time monitoring unit is used to acquire the current mechanical stress wave signal and current turn-off current of the SiC MOSFET device under test through the acoustic sensor and current acquisition unit during the real-time monitoring phase, and call the feature extraction unit to extract the current acoustic feature vector, which is then combined with the current turn-off current and input into the junction temperature prediction model. The junction temperature prediction model outputs the current junction temperature prediction value of the SiC MOSFET device under test.

[0058] By adjusting parameters such as pulse width or bus voltage, the circuit can flexibly cover a wide range of turn-off current values ​​from low to high. Combined with multiple preset junction temperatures set by the heating device, a complete sample dataset covering different junction temperatures and different current combinations is constructed. The signal and current of each switching transient are collected synchronously, ensuring a one-to-one correspondence between input features and output labels. This provides high-quality training data for machine learning modeling, enabling the trained junction temperature prediction model to have good generalization ability and be applicable to complex working conditions such as load changes.

[0059] The three-dimensional acoustic features comprehensively characterize the mechanical stress wave signal from three dimensions: energy trend, instantaneous impact intensity, and waveform morphology. They fully preserve the rich information caused by transient thermal shock gradient and high-temperature softening of materials, providing a feature basis for decoupling the nonlinear coupling between current and temperature. Data-driven modeling methods (such as Gaussian process regression) can automatically learn the complex mapping relationship between input features and junction temperature, successfully separating the different effects of current and temperature changes on the signal, thereby improving the accuracy of junction temperature prediction.

[0060] It enables online, non-invasive monitoring of the junction temperature of SiC MOSFET devices without affecting the normal operation of the devices and without requiring complex electrical isolation design. The real-time monitoring unit automatically calls on existing hardware resources and functional modules without manual intervention and can continuously output junction temperature information, providing real-time data support for active thermal management and health status assessment.

[0061] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0062] This invention addresses the technical problem in existing technologies where a single feature cannot decouple the nonlinear coupling of current and temperature, by combining multi-condition data acquisition, multi-dimensional feature extraction, data-driven modeling, and real-time online prediction. This results in a comprehensive improvement in prediction accuracy, natural electrical isolation, and strong adaptability to a wide range of operating conditions.

[0063] This invention forms a complete technical solution from hardware thermal isolation, signal preprocessing, feature extraction to intelligent modeling, effectively ensuring the safe and reliable operation of SiC MOSFET devices. Attached Figure Description

[0064] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1 This is a flowchart of the mechanical stress wave monitoring method for SiC MOSFET junction temperature in Embodiment 1 of the present invention;

[0066] Figure 2 This is a schematic diagram of the data acquisition platform structure in Embodiment 1 of the present invention;

[0067] Figure 3 This is a schematic diagram of the spectrum analysis of the interrupted lead test and single pulse test in Embodiment 1 of the present invention.

[0068] Figure labeling: 1-Acoustic sensor, 2-Acoustic coupling agent, 3-Heat transfer barrier, 4-SiC MOSFET device. Detailed Implementation

[0069] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0071] Example 1

[0072] This embodiment provides a method for monitoring the mechanical stress wave of SiC MOSFET junction temperature. This method constructs a thermally isolated acoustic signal acquisition platform to collect mechanical stress wave signals under different operating conditions, extracts multi-dimensional acoustic features, and combines them with machine learning algorithms to establish a junction temperature prediction model, ultimately achieving online, non-invasive monitoring of the junction temperature of SiC MOSFET devices. Figure 1 As shown, the monitoring method includes the following steps:

[0073] Step S1: Construct a mechanical stress wave signal acquisition platform and collect sample data for modeling.

[0074] like Figure 2 As shown, a SiC MOSFET device 4 with a TO-247-3L package is selected as the test object. A heating device (a heating stage is used in this embodiment) is placed on the lower surface of the SiC MOSFET device 4 (i.e., the second surface side opposite to the first surface). The heat dissipation side of the SiC MOSFET device 4 is firmly attached to the surface of the heating device to simulate different junction temperature environments of the device.

[0075] A waveguide rod made of polyimide is selected as the heat transfer barrier 3. This material has a thermal conductivity of approximately 0.2 W / (m·K), effectively blocking heat conduction from the high-temperature device to the acoustic sensor. The first surface (lower surface) of the heat transfer barrier 3 is contacted with the upper surface (first surface) of the SiC MOSFET device 4 in a dry contact manner (i.e., direct physical contact without coupling agent) to avoid the impact of acoustic coupling agent 2 failure on signal transmission at high temperatures. The acoustic sensor 1 is contacted with the second surface (upper surface) of the heat transfer barrier 3, and acoustic coupling agent 2 is placed between the acoustic sensor 1 and the second surface of the heat transfer barrier 3 to ensure efficient transmission of sound waves from the heat transfer barrier to the sensor.

[0076] Considering the strong electromagnetic interference generated by SiC MOSFET devices during high-speed switching, a differential acoustic emission sensor is preferably used for acoustic sensor 1. Differential acoustic emission sensors utilize the differential amplification principle to effectively cancel common-mode electromagnetic interference noise on the signal line, improving the signal-to-noise ratio of the original signal from the source. In this embodiment, the acoustic coupling agent 2 is a silicon-based coupling agent, forming a good acoustic matching layer between the acoustic sensor 1 and the waveguide rod. This ensures efficient transmission of mechanical stress waves to the acoustic sensor while preventing coupling agent failure or performance degradation due to high temperatures, further guaranteeing the stability and fidelity of signal acquisition.

[0077] During actual testing, the SiC MOSFET device 4 was brought to multiple different preset junction temperatures by adjusting the preset temperature of the heating stage. To simulate a typical operating range, multiple temperature points ranging from 50°C to 150°C at 20°C intervals were set. At each preset temperature, isothermal conditions were maintained for at least 30 minutes to ensure that the SiC MOSFET device 4 reached complete thermodynamic equilibrium, making the device junction temperature consistent with the set temperature of the heating stage.

[0078] At each preset junction temperature, the SiC MOSFET device 4 is adjusted to obtain multiple different turn-off current values ​​during multiple switching transients. This embodiment uses a single-pulse test circuit as the adjustment circuit. By changing the pulse width or bus voltage, the device generates multiple different turn-off current values ​​(e.g., 10A, 20A, 30A, 40A, 50A) at the turn-off moment, with each switching transient corresponding to one of these turn-off current values. In this way, the modulation effect of turn-off current and junction temperature on the mechanical stress wave characteristics of the SiC MOSFET can be systematically quantified.

[0079] The mechanical stress wave signal generated during each switching transient is acquired by acoustic sensor 1, and the corresponding turn-off current value is simultaneously acquired by a current acquisition unit (such as a current probe). This yields a sample set of mechanical stress wave signals covering different preset junction temperatures and different combinations of turn-off currents, which is used for subsequent modeling.

[0080] Step S2: Filter the mechanical stress wave signals under different preset junction temperatures and different turn-off currents.

[0081] To further eliminate the impact of mechanical background noise on acoustic feature extraction, preprocessing of the acquired mechanical stress wave signals at the software level is required. Mechanical background noise is introduced by factors such as structural resonance and environmental vibration of the acquisition platform. This noise overlaps with the effective signal excited by the transient switching of the device in the frequency domain and needs to be suppressed through filtering.

[0082] To accurately determine the frequency band distribution of mechanical background noise and the spectral range of the effective signal, this embodiment performs lead-breaking and single-pulse tests on the acquisition platform. The lead-breaking test generates a simulated acoustic emission signal by breaking a pencil lead, used to identify the structural resonant frequency of the acquisition platform and the frequency band distribution of the mechanical background noise. The single-pulse test acquires the mechanical stress wave signal excited by the actual switching transient of the device (i.e., the signal acquired in step S1). Spectral analysis is performed on the signals acquired by the two tests, and the results are as follows: Figure 3 As shown. Through comparative analysis, it was found that the mechanical background noise introduced by factors such as the resonance of the acquisition platform structure is mainly concentrated in the low frequency range of 0~50kHz, while the effective signal excited by the internal thermal stress of the device still retains abundant energy in the high frequency range above 50kHz.

[0083] Based on the above spectral analysis results, this embodiment preprocesses the mechanical stress wave signal acquired in step S1 using a bandpass filter ranging from 50kHz to 150kHz at the software level. The upper and lower limits of this filtering frequency band are set according to the noise distribution range determined by the lead-breaking test, which can effectively filter out low-frequency mechanical background noise from 0 to 50kHz, while completely preserving the effective signal components within the 50kHz to 150kHz frequency band, ensuring the spectral integrity and signal-to-noise ratio of subsequent acoustic feature extraction. After bandpass filtering, a high signal-to-noise ratio effective mechanical stress wave signal x[n] is obtained, which is used in subsequent feature extraction steps.

[0084] Step S3: Extract an acoustic feature vector from the effective signals of mechanical stress waves under different preset junction temperatures and different turn-off currents, including at least the logarithmic root mean square, peak-to-peak value, and kurtosis, and combine the acoustic feature vector with the corresponding turn-off current to form an input feature vector.

[0085] After bandpass filtering preprocessing of the mechanical stress wave signals, effective mechanical stress wave signal sequences corresponding one-to-one with the acquired operating conditions are obtained. Specifically, each combination of preset junction temperature and each turn-off current value corresponds to one effective mechanical stress wave signal sequence. This step aims to extract multidimensional acoustic features from each effective mechanical stress wave signal sequence that can comprehensively characterize the junction temperature state of the device, providing high-quality input feature vectors for subsequent modeling.

[0086] To balance computational speed and feature robustness, this embodiment constructs a three-dimensional acoustic feature vector F=[F1,F2,F3], which specifically includes three features: root mean square logarithm F1, peak-to-peak value F2, and kurtosis F3.

[0087] The root mean square logarithm is calculated using the following formula:

[0088] (1)

[0089] Where RMS represents the root mean square; N represents the total number of sampling points in the effective signal sequence of the mechanical stress wave; This represents the effective mechanical stress wave signal at the nth sampling point.

[0090] The root mean square (RMS) is used to characterize the nonlinear growth trend of macroscopic acoustic wave energy with temperature. Since the energy of mechanical stress waves increases nonlinearly with increasing junction temperature, directly using the RMS value for modeling may lead to poor model fitting. By taking the logarithm of the RMS, the originally nonlinear growth trend can be linearized, facilitating subsequent machine learning algorithms to learn the mapping relationship between energy and junction temperature.

[0091] Peak-to-peak value is calculated using the following formula:

[0092] (2)

[0093] in, This represents the maximum value of the effective signal sequence of mechanical stress waves. This represents the minimum value of the effective signal sequence of mechanical stress waves.

[0094] Peak-to-peak value is used to characterize the maximum instantaneous thermal shock stress intensity generated during device switching transients. At the moment of device turn-off, the extremely high instantaneous power loss density creates a huge temperature gradient within the chip. The mechanical stress wave excited by the thermoelastic effect manifests as a burst signal in the time domain, and its amplitude range directly reflects the severity of the thermal shock. Peak-to-peak value can effectively capture this transient shock intensity and is an important characteristic sensitive to junction temperature changes.

[0095] Kurtosis is calculated using the following formula:

[0096] (3)

[0097] in, This represents the mean value of the (effective) mechanical stress wave signal; This represents the standard deviation of the (effective) mechanical stress wave signal.

[0098] Kurtosis is used to quantify the steepness and tail thickness of a signal's probability density distribution, and can sensitively capture minute waveform distortions caused by material softening at high temperatures. As junction temperature increases, the Young's modulus and damping coefficient of the SiC MOSFET package material change, causing waveform distortion of mechanical stress waves during propagation, manifested as a change in the probability density distribution of the signal. As a fourth-order statistic, kurtosis is highly sensitive to the impulse components in the signal and the steepness of the waveform, effectively quantifying this temperature-induced waveform change and providing supplementary information for junction temperature monitoring.

[0099] After extracting the three-dimensional acoustic feature vector, the acoustic feature vector F is combined with the turn-off current I corresponding to the current signal sequence to form a multi-dimensional input feature vector X=[F,I]. This input feature vector integrates the temperature-sensitive information contained in the mechanical stress wave signal and the electrical operating condition information of the device. Each input feature vector uniquely corresponds to a preset junction temperature T, providing a well-structured and accurately corresponding training dataset for subsequent modeling.

[0100] To further enrich the feature space and improve the prediction accuracy and robustness of the model under complex operating conditions, as an extension of this embodiment, the acoustic feature vector can also include frequency domain features or time-frequency domain features. Specifically, frequency domain features can include frequency domain peak values ​​(i.e., the frequency component with the largest amplitude in the signal spectrum and its amplitude) or signal energy in a specific frequency band (e.g., dividing the 50kHz~150kHz frequency band into several sub-bands and calculating the energy proportion of each sub-band). Time-frequency domain features can be obtained through short-time Fourier transform or wavelet transform, which can simultaneously reflect the variation law of the signal in both time and frequency dimensions. These supplementary features complement the aforementioned three-dimensional time domain features (logarithmic root mean square F1, peak-to-peak value F2, and kurtosis F3), and can comprehensively characterize the characteristics of mechanical stress wave signals from multiple perspectives, further improving the performance of the junction temperature prediction model.

[0101] Step S4: Using the input feature vector X as input and the corresponding preset junction temperature T as output, establish the mapping relationship between the input features and the junction temperature to obtain the junction temperature prediction model.

[0102] After completing the feature extraction in step S3, input feature vectors covering different preset junction temperatures and different turn-off currents, along with their corresponding preset junction temperatures, are obtained. Each input feature vector X... i =[F i ,I i [Based on the three-dimensional acoustic feature vector F] i and the corresponding turn-off current I i The combination constitutes the output label, which is the preset junction temperature T of the i-th sample. i The input feature vectors of all samples are used to construct the input matrix X=[X1,X2,…,X…]. M ] T The corresponding output junction temperatures form the output vector y=[T1,T2,…,T M ] T , where M is the total number of samples (equal to the number of effective signal sequences of mechanical stress waves).

[0103] This embodiment employs a Gaussian process regression algorithm to establish a nonlinear mapping relationship between input features and junction temperature. Gaussian process regression is a nonparametric machine learning method based on a Bayesian framework. It can not only establish a mapping relationship between input and output but also provide an estimate of the uncertainty of the prediction results. Its core idea is to assume that there is a functional relationship y=f(X)+ε between the output junction temperature y and the input feature X, where f(X) follows a Gaussian process prior and ε is Gaussian white noise.

[0104] The modeling process of Gaussian process regression includes the following steps:

[0105] S4.1 Define the kernel function.

[0106] The kernel function (also known as the covariance function) is the core of Gaussian process regression, used to measure the variance between any two input feature vectors X. i and X j The similarity between them, that is, their corresponding output values ​​f(X) i ) and f(X j The covariance of the kernel function is used to determine the model's fitting ability and generalization performance.

[0107] This embodiment employs an additive composite kernel function. The basic idea is to separate the contributions of different dimensions or feature groups in the input feature vector to the output, and construct the overall kernel function by adding multiple sub-kernel functions. The mathematical form of the additive composite kernel function is:

[0108] (4)

[0109] in, , ,…, This indicates that the input feature vector X will be used. i The feature vector is divided into p distinct feature groups, each corresponding to a sub-kernel function. In this embodiment, the input feature vector can be divided into an acoustic feature group (including root mean square logarithm, peak-to-peak value, and kurtosis) and a turn-off current feature group, each modeled using different kernel function forms (such as radial basis function, Marton kernel function, etc.). The advantage of this additive structure is:

[0110] Decoupling nonlinear effects: It can capture the contributions of acoustic features and turn-off current to junction temperature, effectively decoupling the complex nonlinear modulation effect of current and temperature on mechanical stress waves.

[0111] Improved interpretability: The hyperparameters of each sub-kernel function can be optimized independently, which facilitates the analysis of the importance of different features to junction temperature prediction;

[0112] Enhance generalization ability: Avoid overfitting problems that may be caused by a single kernel function and improve the prediction performance of the model under unknown conditions.

[0113] S4.2 Initialize the hyperparameters of the kernel function.

[0114] Each sub-kernel function contains several hyperparameters, for example:

[0115] Signal variance: the overall amplitude range of the control function; length scale: the smoothness of the change of the control function in each input dimension. The larger the length scale, the smoother the change of the function in that dimension; noise variance: characterizes the magnitude of observation noise.

[0116] Before training begins, these hyperparameters need to be initialized. This embodiment can use empirical values ​​for initialization, such as setting the signal variance to 1, the length scale of each dimension to 1 / 10 of the input feature range, and the noise variance to 1 / 100 of the output junction temperature variance. Alternatively, random initialization can be used, selecting the optimal initial values ​​through multiple trials.

[0117] S4.3 Based on the defined kernel function and input feature vectors, construct the covariance matrix between the input feature vectors.

[0118] Based on the defined kernel function and the input feature vectors of all modeling samples, a covariance matrix K is constructed among the training samples. The covariance matrix K is an M×M symmetric positive definite matrix whose elements... Let K represent the covariance between the i-th sample and the j-th sample. After considering the effect of noise, the complete covariance matrix K is... y This involves superimposing noise variance onto the original covariance matrix K.

[0119] S4.4. Calculate the logarithmic marginal likelihood based on the covariance matrix, and with the goal of maximizing the logarithmic marginal likelihood, solve for the optimal hyperparameters of the kernel function through an optimization algorithm to obtain the junction temperature prediction model.

[0120] Hyperparameter optimization for Gaussian process regression employs the method of maximizing the log-marginal likelihood. The expression for the log-marginal likelihood function is:

[0121] (5)

[0122] Where P represents the probability density function; the first term on the right-hand side of the expression is the data fitting term, which encourages the model to better fit the training data; the second term is the complexity penalty term, which prevents the model from becoming too complex and causing overfitting; and the third term is a constant term.

[0123] With the goal of maximizing the aforementioned logarithmic marginal likelihood, the optimal hyperparameters of the kernel function are solved using an optimization algorithm. This embodiment can employ the conjugate gradient method or a quasi-Newton method (such as the L-BFGS algorithm) for optimization. During optimization, the partial derivatives of the logarithmic marginal likelihood with respect to each hyperparameter need to be calculated, and the hyperparameter values ​​are iteratively updated until convergence. After optimization, the optimal set of hyperparameters that maximizes the marginal likelihood is obtained, at which point the model achieves the best balance between fitting performance and generalization ability on the training data.

[0124] Through the above steps, the defined kernel function, optimal hyperparameters, and training sample data are obtained. These three elements together constitute the trained Gaussian process regression model, which is the junction temperature prediction model described in this embodiment. This model can not only predict the junction temperature of new input feature vectors but also simultaneously output the prediction variance to characterize the uncertainty of the prediction results.

[0125] The modeling method of this invention is not limited to Gaussian process regression; other machine learning algorithms, such as artificial neural networks, support vector regression, and decision tree ensembles, can also be used. These algorithms can also establish a nonlinear mapping relationship between input features and junction temperature, achieving high-precision, non-invasive junction temperature monitoring. Those skilled in the art can select appropriate machine learning algorithms for modeling based on actual application scenarios and data characteristics, all of which fall within the scope of protection of this invention.

[0126] Step S5: Real-time acquisition of the current mechanical stress wave signal and current turn-off current of the SiC MOSFET device under test, extraction of the acoustic feature vector of the current mechanical stress wave signal, and input of the combination with the current turn-off current into the junction temperature prediction model. The junction temperature prediction model outputs the current junction temperature prediction value of the SiC MOSFET device under test.

[0127] After completing the modeling in step S4, a trained junction temperature prediction model was obtained. This model is built based on the Gaussian process regression algorithm and includes a predefined additive composite kernel function, optimized hyperparameters, and training sample data collected during the modeling phase. This step aims to apply the model to actual operating conditions to achieve online, non-invasive real-time monitoring of the junction temperature of SiC MOSFET devices.

[0128] Under actual operating conditions, the SiC MOSFET device under test is in normal working condition, and its switching transients are triggered normally by the system's main controller or drive circuit. In this embodiment, the signal acquisition process is started synchronously each time a switching transient occurs.

[0129] Using the acoustic sensor set in step S1, the mechanical stress wave signal (i.e., the current mechanical stress wave signal) excited by the SiC MOSFET device under test during the current switching transient is acquired in real time. Simultaneously, the turn-off current value (i.e., the current turn-off current) corresponding to this switching transient is acquired synchronously through a current acquisition unit (such as a current probe). It should be noted that the turn-off current value acquired at this time is the real-time current value of the device under the current actual operating conditions, not the preset current value obtained through single-pulse test circuit adjustment during the modeling stage.

[0130] The acquired current mechanical stress wave signal is processed using the same preprocessing procedure as in step S2. Specifically, a bandpass filter of 50kHz to 150kHz is applied to the current mechanical stress wave signal. This filter frequency band is set based on the noise distribution range determined by the previous lead-breaking test, which can effectively filter out low-frequency mechanical background noise while retaining the high-frequency components of the effective signal. After bandpass filtering, the effective signal sequence of the current mechanical stress wave is obtained.

[0131] For the preprocessed current effective mechanical stress wave signal sequence, the same feature extraction method as in step S3 is applied to extract the three-dimensional acoustic feature vector of the current signal (i.e., the current acoustic feature vector), which specifically includes: root mean square logarithm, peak-to-peak value, and kurtosis.

[0132] The extracted current acoustic feature vector is combined with the synchronously acquired current off-current to form the current input feature vector. This vector has the exact same dimensions and structure as the input feature vector used in the modeling phase.

[0133] The current input feature vector is then fed into the junction temperature prediction model trained in step S4. This model is based on the Gaussian process regression algorithm, and its prediction process is as follows:

[0134] First, calculate the covariance vector between the current input feature vector and the input feature vectors of all modeling samples;

[0135] Then, combine the inverse of the covariance matrix that has been calculated and stored during the modeling phase. Given the output junction vector y of the modeling sample, calculate the predicted mean corresponding to the current input. The predicted mean is the estimated junction temperature corresponding to the current input feature vector, which is the real-time junction temperature of the SiC MOSFET device under test in the current switching transient.

[0136] In addition, the model can also calculate the prediction variance (σ). * ) 2 Prediction variance can be used to construct prediction confidence intervals, for example, at a 95% confidence level, the actual junction temperature is located at... Within the specified range. This uncertainty estimate is crucial for assessing the reliability of the forecast results and guiding subsequent proactive thermal management decisions.

[0137] The calculated current junction temperature estimate and its prediction variance are output to the upper-level monitoring system or active thermal management unit. This junction temperature information can be used in the following application scenarios:

[0138] Active thermal management: When the estimated junction temperature exceeds the preset safety threshold, the system can automatically take measures such as frequency reduction, current limiting or active cooling to prevent the device from failing due to overheating;

[0139] Health status assessment: By monitoring the junction temperature trend over a long period of time and combining it with the device's operating history, the aging degree of the packaging material and the remaining lifespan of the device can be assessed.

[0140] Fault warning: When the junction temperature prediction value fluctuates abnormally or the prediction variance increases significantly, the potential fault risk can be warned in advance, prompting maintenance or replacement.

[0141] Design optimization: The accumulated junction temperature data can be used to optimize heat dissipation design, improve control strategies, and enhance the overall reliability and power density of the system.

[0142] For each switching transient of the device, steps S1 to S5 can be repeated to achieve continuous, online, and real-time monitoring of the junction temperature of the SiCMOSFET device. Since the entire monitoring process only requires the acquisition of acoustic and current signals and does not require invasive electrical measurements, it will not affect the normal operation of the device and naturally meets the electrical isolation requirements.

[0143] The real-time monitoring method of this embodiment has proven its effectiveness in verification tests. Junction temperature inversion tests were conducted using the trained junction temperature prediction model on verification data independent of the modeling set. The results showed a root mean square error (RMSE) as low as 1.26℃, a mean absolute error (MAE) of 0.92℃, and the absolute error of prediction within the 95% confidence interval was controlled within 2.8℃. This fully demonstrates the high accuracy and robustness of the method of this invention under complex operating conditions.

[0144] Example 2

[0145] The difference between this embodiment and Embodiment 1 is that the input feature vector of the junction temperature prediction model consists only of acoustic feature vectors and does not include turn-off current, thereby realizing independent junction temperature monitoring without electrical measurement assistance.

[0146] Steps S1 and S2 in this embodiment are the same as in Embodiment 1, and will not be repeated here.

[0147] Step S3: Extract an acoustic feature vector from the effective signals of mechanical stress waves under different preset junction temperatures and different turn-off currents. This acoustic feature vector includes at least the root mean square, peak-to-peak value, and kurtosis. Use this acoustic feature vector as the input feature vector.

[0148] From the effective signal sequence of mechanical stress wave corresponding to each preset junction temperature and each turn-off current value, a three-dimensional acoustic feature vector F=[F1,F2,F3] is extracted, which specifically includes three features: root mean square F1, peak-to-peak value F2, and kurtosis F3.

[0149] Unlike Example 1, the three-dimensional acoustic feature vector F extracted in this step is directly used as the input feature vector and does not include the turn-off current. Each input feature vector F i A single preset junction temperature T corresponds to a unique junction temperature. i This constitutes a training dataset based on pure acoustics.

[0150] Step S4: Construct a pure acoustic prediction model based on Gaussian process regression.

[0151] This step is similar to step S4 in Embodiment 1, but the input features are only three-dimensional acoustic feature vectors. Specifically, the pure acoustic feature vector F obtained in step S3 is used. i As input, the corresponding preset junction temperature T i As output, the Gaussian process regression algorithm is used to establish a nonlinear mapping relationship between the input features and the junction temperature, thus obtaining a pure acoustic prediction model.

[0152] Step S5: During real-time monitoring, only the current mechanical stress wave signal needs to be collected using an acoustic sensor. After filtering in step S2, the current acoustic feature vector is extracted and input into the pure acoustic prediction model constructed in step S4. The model outputs the current junction temperature prediction. There is no need to collect the turn-off current.

[0153] The method described in this embodiment has been validated in ablation experiments. Using a purely acoustic prediction model incorporating only three-dimensional acoustic features for junction temperature prediction, the root mean square error (RMSE) was 1.43℃, the mean absolute error (MAE) was controlled within 1.1℃, and the absolute prediction error was limited to within 3.3℃ at a 95% statistical confidence interval. This result demonstrates that the mechanical stress wave signal itself carries an inherent temperature sensitivity characteristic completely independent of current changes. Even in scenarios without electrical measurement assistance or with drastic load fluctuations, the method of this embodiment can still provide high-precision and robust independent junction temperature monitoring.

[0154] This embodiment, together with Embodiment 1, constitutes the complete technical solution of the present invention. Embodiment 1 employs a fusion model of "acoustic characteristics + turn-off current," providing the highest prediction accuracy (RMSE 1.26℃) under normal operating conditions. This embodiment employs a pure acoustic prediction model, which can still independently complete the junction temperature monitoring task (RMSE 1.43℃) even in special scenarios such as current sensor failure, severe load fluctuations, or when no electrical measurement assistance is required. The two embodiments complement each other, covering a complete application range from normal to extreme operating conditions, fully demonstrating the adaptability and robustness of the technical solution of the present invention.

[0155] Example 3

[0156] This embodiment provides a mechanical stress wave monitoring system for SiC MOSFET junction temperature, including a heat transfer barrier, a heating device, an adjustment circuit, an acoustic sensor, a current acquisition unit, a feature extraction unit, a model building unit, and a real-time monitoring unit. The connection method of the heat transfer barrier, heating device, adjustment circuit, acoustic sensor, and current acquisition unit is the same as in Embodiment 1.

[0157] The input terminals of the feature extraction unit are connected to the acoustic sensor and the current acquisition unit, respectively, to extract features from the acquired mechanical stress wave signal and the cut-off current. Specifically, the feature extraction unit first preprocesses the raw mechanical stress wave signal acquired by the acoustic sensor, including applying a bandpass filter from 50kHz to 150kHz. This filtering frequency band is set according to the noise distribution range determined by the lead-breaking test, and is used to filter out low-frequency mechanical background noise and retain the effective signal. After filtering, the effective mechanical stress wave signal sequence is obtained.

[0158] The feature extraction unit extracts a three-dimensional acoustic feature vector from the effective signal sequence of mechanical stress waves corresponding to each preset junction temperature and each turn-off current value. This feature vector includes at least the logarithmic root mean square, peak-to-peak value, and kurtosis. The input feature vector is composed of the three-dimensional acoustic feature vector and the corresponding turn-off current.

[0159] The model building unit is connected to the feature extraction unit. It is used to establish the mapping relationship between the input features and the junction temperature by taking the input feature vector output by the feature extraction unit as input and the corresponding preset junction temperature as output, so as to obtain the junction temperature prediction model.

[0160] In this embodiment, the model building unit uses a Gaussian process regression algorithm to establish a mapping relationship between input features and junction temperature, thereby obtaining a junction temperature prediction model. The model building unit stores the trained junction temperature prediction model and its parameters (including kernel function form, optimal hyperparameters, training sample data, etc.) in the model memory for the real-time monitoring unit to access.

[0161] The real-time monitoring unit is connected to the model memory of the acoustic sensor, the current acquisition unit, the feature extraction unit, and the model building unit, respectively. It is used to acquire the current mechanical stress wave signal and the current turn-off current of the SiCMOSFET device under test in the current switching transient state through the acoustic sensor and the current acquisition unit during the real-time monitoring phase.

[0162] The real-time monitoring unit calls the feature extraction unit to perform the same preprocessing and feature extraction on the current mechanical stress wave signal as in the modeling stage, obtaining the current acoustic feature vector, which is then combined with the current turn-off current to form the current input feature vector. The real-time monitoring unit then inputs the current input feature vector into the junction temperature prediction model stored in the model memory. The model calculates the predicted mean as the current junction temperature estimate and can simultaneously output the prediction variance to characterize the uncertainty of the prediction result.

[0163] The real-time monitoring unit also includes a result output module, which outputs the current junction temperature estimate to the upper-level monitoring system, active thermal management unit, or human-machine interface for subsequent active thermal management, health status assessment, or fault early warning.

[0164] As a variant embodiment of the system of the present invention, the model building unit can construct a purely acoustic prediction model, that is, the input feature vector consists only of the acoustic feature vector extracted by the feature extraction unit and does not contain turn-off current information. Correspondingly, the real-time monitoring unit only needs to collect the current mechanical stress wave signal during the real-time monitoring phase, without needing to collect the current turn-off current, to independently complete the junction temperature monitoring. This variant embodiment is suitable for special scenarios such as current sensor failure, severe load fluctuations, or when electrical measurement assistance is not required.

[0165] As another variant of the system of the present invention, the feature extraction unit can be extended to extract frequency domain features or time-frequency domain features, including frequency domain peaks or signal energy of a specific frequency band, which together with the three-dimensional time domain features constitute a higher-dimensional input feature vector.

[0166] The above description only discloses specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or modifications that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring the mechanical stress wave of SiC MOSFET junction temperature, characterized in that, The method includes: The SiC MOSFET device is heated to multiple preset junction temperatures. At each preset junction temperature, the SiC MOSFET device is adjusted to obtain multiple different turn-off current values ​​during multiple switching transients, with each switching transient corresponding to one of the turn-off current values. The mechanical stress wave signal excited by each switching transient is collected by an acoustic sensor, and the turn-off current corresponding to that switching transient is collected synchronously by a current acquisition unit to obtain the mechanical stress wave signal under different preset junction temperatures and different turn-off currents. The acoustic feature vectors, including at least the root mean square, peak-to-peak value, and kurtosis, are extracted from the mechanical stress wave signals under different preset junction temperatures and different turn-off currents. The acoustic feature vectors are then combined with the corresponding turn-off currents to form the input feature vectors. Using the input feature vector as input and the corresponding preset junction temperature as output, a mapping relationship between the input features and the junction temperature is established to obtain a junction temperature prediction model; The current mechanical stress wave signal and current turn-off current of the SiC MOSFET device under test are acquired in real time. The acoustic feature vector of the current mechanical stress wave signal is extracted and combined with the current turn-off current and input into the junction temperature prediction model. The junction temperature prediction model outputs the current junction temperature prediction value of the SiC MOSFET device under test.

2. The method for monitoring the mechanical stress wave of SiC MOSFET junction temperature according to claim 1, characterized in that, The acoustic sensor is indirectly connected to the SiC MOSFET device via a heat transfer barrier, specifically configured as follows: The first surface of the heat transfer barrier is dry-contacted to the first surface of the SiC MOSFET device, the acoustic sensor is abutted to the second surface of the heat transfer barrier, and an acoustic coupling agent is coated between the acoustic sensor and the second surface of the heat transfer barrier; the heating device is disposed on the second surface side of the SiC MOSFET device opposite to the first surface.

3. The method for monitoring the mechanical stress wave of SiC MOSFET junction temperature according to claim 2, characterized in that, The heat transfer barrier is a waveguide rod made of polyimide; the acoustic coupling agent is a silicon-based coupling agent.

4. The method for monitoring the mechanical stress wave of SiC MOSFET junction temperature according to claim 1, characterized in that, Before extracting the acoustic feature vector, the method further includes filtering the mechanical stress wave signal; the frequency band of the filtering is set according to the noise distribution range determined by the lead breakage test, so as to filter out low-frequency mechanical background noise and retain the effective signal.

5. The method for monitoring the mechanical stress wave of SiC MOSFET junction temperature according to claim 1, characterized in that, The root mean square logarithm is calculated using the following formula: ； Where F1 represents the logarithmic root mean square; RMS represents the root mean square; and N represents the total number of sampling points. This represents the mechanical stress wave signal at the nth sampling point; The peak value is calculated using the following formula: ； Where F2 represents the peak-to-peak value; The kurtosis is calculated using the following formula: ； Where F3 represents kurtosis; This represents the mean value of the mechanical stress wave signal; This represents the standard deviation of the mechanical stress wave signal.

6. The method for monitoring the mechanical stress wave of SiC MOSFET junction temperature according to claim 1, characterized in that, The acoustic feature vector also includes frequency domain features or time-frequency domain features; the frequency domain features include frequency domain peaks or signal energy in a specific frequency band.

7. The method for monitoring the mechanical stress wave of SiC MOSFET junction temperature according to claim 1, characterized in that, Machine learning algorithms are used to establish a mapping relationship between input features and junction temperature.

8. The method for monitoring the mechanical stress wave of SiC MOSFET junction temperature according to claim 7, characterized in that, The machine learning algorithm is a Gaussian process regression algorithm, which establishes a mapping relationship between input features and junction temperature, including: Define a kernel function; the kernel function is an additive composite kernel function, used to measure the covariance between input feature vectors; Initialize the hyperparameters of the kernel function, including signal variance, length scale, and noise variance; Based on the defined kernel function and the input feature vectors, construct the covariance matrix between the input feature vectors; The logarithmic marginal likelihood is calculated based on the covariance matrix, and the optimal hyperparameters of the kernel function are obtained by optimizing the algorithm to maximize the logarithmic marginal likelihood, thereby obtaining the junction temperature prediction model.

9. The method for monitoring the mechanical stress wave of SiC MOSFET junction temperature according to claim 1, characterized in that, The input feature vector consists only of the acoustic feature vector and does not include the turn-off current; the junction temperature prediction model is a pure acoustic prediction model used for junction temperature monitoring without electrical measurement assistance.

10. A mechanical stress wave monitoring system for SiC MOSFET junction temperature, characterized in that, The system includes: A heating device, located on the surface of the SiC MOSFET device, is used to bring the SiC MOSFET device to multiple different preset junction temperatures; The adjustment circuit is used to adjust parameters at each preset junction temperature so that the SiC MOSFET device obtains multiple different turn-off current values ​​in multiple switching transients, and each switching transient corresponds to one of the turn-off current values. Acoustic sensors are used to acquire mechanical stress wave signals generated during each switching transient. The current acquisition unit is used to synchronously acquire the turn-off current corresponding to each switching transient. The feature extraction unit is used to extract acoustic feature vectors, including at least logarithmic root mean square, peak-to-peak value and kurtosis, from mechanical stress wave signals under different preset junction temperatures and different turn-off currents, and combine the acoustic feature vectors with the corresponding turn-off currents to form an input feature vector. The model building unit is used to establish a mapping relationship between the input features and the junction temperature by taking the input feature vector as input and the corresponding preset junction temperature as output, so as to obtain a junction temperature prediction model. The real-time monitoring unit is used to acquire the current mechanical stress wave signal and current turn-off current of the SiC MOSFET device under test through the acoustic sensor and current acquisition unit during the real-time monitoring phase, and call the feature extraction unit to extract the current acoustic feature vector, which is then combined with the current turn-off current and input into the junction temperature prediction model. The junction temperature prediction model outputs the current junction temperature prediction value of the SiC MOSFET device under test.

Images