An intelligent power module dynamic and static characteristic test method based on AI

Abstract

The application relates to an AI-based intelligent power module dynamic and static characteristic test method, and belongs to the field of power semiconductor device testing. The method comprises the following steps: a test system integrating a plurality of AI models and a plurality of station interface boards is built; static characteristic data at discrete temperature points are collected, a first AI model is used to predict characteristic curves of continuous temperature intervals and compared with adaptive threshold values; a plurality of modules of the same batch are simultaneously installed, dynamic tests are sequentially performed, and transient waveforms of active modules and coupling responses of passive modules are synchronously collected; a second AI model is used for defect identification of the dynamic waveforms; a third AI model is used for online reconstruction of dynamic junction temperatures of the modules; starting thermal states of subsequent test modules are adaptively calibrated according to the reconstructed junction temperatures to reproduce thermal interlocking working conditions; a fourth AI model is used for reasoning on the actually measured coupling data, and system adaptability evaluation results of the modules are output; and finally, a test report containing comprehensive judgment conclusions is generated.

Description

Technical Field

[0001] This invention belongs to the field of power semiconductor device testing technology, specifically relating to an AI-based intelligent power module dynamic and static characteristic testing method. Background Technology

[0002] Intelligent power modules (IPMs) are core components of power electronic systems, highly integrating power switching devices, drive circuits, and various protection functions. IPM failure can lead to system shutdown, equipment damage, and even safety accidents; therefore, comprehensive and accurate testing of their dynamic and static characteristics is crucial. Currently, IPM testing is generally performed using automated test equipment (ATE), covering static parameter testing, dynamic parameter testing, protection function testing, and safety and reliability testing.

[0003] However, the existing testing system has the following limitations:

[0004] First, static parameter testing is usually performed only at a few discrete temperature points. When the actual junction temperature of the module in real operating conditions deviates from these discrete test points, the evaluation results will be biased, and the continuous evolution law of the characteristics over the entire temperature range cannot be obtained, resulting in insufficient detection capability for early degradation of temperature-sensitive components.

[0005] Second, in dynamic parameter testing, the interpretation of switching waveforms relies on human experience or fixed thresholds, making it difficult to identify subtle waveform distortions that characterize early latent defects.

[0006] Third, existing tests are always conducted independently on a single module at a constant case temperature, completely failing to detect the dynamic electrical and thermal coupling effects between modules in a real system due to shared DC buses and heat sinks. A high-speed switching action of one module can generate transient common-mode interference and thermal shock to adjacent modules. This coupling effect is completely missing in the single-module testing paradigm, resulting in a large number of latent defects that only surface under system-level coupling stress being systematically missed during the manufacturing process.

[0007] It is evident that existing technologies lack a testing method that can utilize AI technology to deeply explore dynamic and static characteristics and intelligently evaluate module status under simulated real system coupling conditions. Summary of the Invention

[0008] The purpose of this invention is to overcome the problems existing in the prior art and provide an AI-based method for testing the dynamic and static characteristics of intelligent power modules.

[0009] The technical solution of the present invention is as follows:

[0010] A method for testing the dynamic and static characteristics of an AI-based intelligent power module includes the following steps:

[0011] S1. Build an automated testing system with AI processing capabilities. The testing system includes a main control computer, a source measurement unit, a pulse generation unit, a high-speed data acquisition unit, and a multi-station interface board adapted to the IPM module under test. The main control computer is loaded with at least one pre-trained AI model.

[0012] S2. Install the IPM module under test on the multi-station interface board, collect its static characteristic data at multiple temperature points through the source measurement unit, and input the static characteristic data into the first AI model to obtain the static characteristic prediction result of the module in a continuous temperature range.

[0013] S3. Install multiple IPM modules from the same batch simultaneously on the multi-station interface board, perform dynamic testing on each module sequentially through the pulse generation unit, and synchronously acquire the transient waveform of the active module and the coupling response waveform of the passive module using the high-speed data acquisition unit.

[0014] S4. Input the transient waveform into the second AI model to obtain the dynamic defect identification result;

[0015] S5. Calculate the transient power loss based on the transient waveform and input it into the third AI model to obtain the dynamic junction temperature of each module during the test process;

[0016] S6. Based on the dynamic junction temperature, after the current module test is completed, the initial thermal state before the next module test is adaptively adjusted to reproduce the thermal interlocking conditions between modules.

[0017] S7. Construct the static characteristic data, the coupling response waveform, and the dynamic junction temperature into structured data, input them into the fourth AI model, and obtain the system-level adaptability evaluation results of each module.

[0018] S8. Summarize the above test and evaluation results and generate a comprehensive test report.

[0019] Furthermore, in step S2, the first AI model is a backpropagation neural network with embedded physical constraints.

[0020] Furthermore, the collector-emitter saturation voltage drop predicted by the physical constraints decreases monotonically with increasing temperature, and the collector leakage current predicted by the physical constraints increases monotonically with increasing temperature.

[0021] Furthermore, in step S3, when the passive module is subjected to dynamic testing, its gate is subjected to a nominal forward bias voltage, and its collector is connected to the same DC bus voltage as the active module.

[0022] Furthermore, in step S4, the second AI model is a convolutional neural network model.

[0023] Furthermore, the third AI model mentioned in step S5 is a recurrent neural network model with embedded physical constraints; the physical constraints are based on the heat conduction equation.

[0024] Furthermore, the adaptive adjustment in step S6 includes: taking the maximum junction temperature rise experienced by the next module as a passive module in the previous test as the target preheating heat, and adjusting the initial junction temperature of the next module to the corresponding target value through heating and / or pre-pulse methods.

[0025] Furthermore, the structured data mentioned in step S7 is graph structured data, in which each module serves as a graph node, and the measured electrical coupling strength and thermal coupling effect between modules serve as edge features.

[0026] Furthermore, the fourth AI model is a graph attention network model.

[0027] Compared with the prior art, the beneficial effects of the present invention are:

[0028] 1. This invention utilizes an AI model to deduce characteristic curves within a continuous temperature range from test values ​​at a limited number of discrete temperature points, and compares them with an adaptive threshold. This enables sensitive detection of early degradation in temperature-sensitive systems, overcoming the limitations of traditional discrete point determination.

[0029] 2. This invention uses an AI model to automatically classify defects in standardized dynamic waveforms, replacing manual interpretation, significantly improving detection consistency and efficiency, and can capture subtle waveform distortions that traditional fixed threshold methods cannot detect, thus achieving automatic identification of hidden defects.

[0030] 3. This invention, through multi-module synchronous testing, is the first to simultaneously acquire cross-module common-mode coupling response waveforms and real-time junction temperature dynamics during ATE testing. It utilizes an AI model embedded with physical constraints of thermal conduction to reconstruct the junction temperature online, ensuring that junction temperature changes conform to physical laws. Based on the reconstructed junction temperature, an adaptive thermal calibration mechanism can accurately reproduce the thermal cascading effects in real inverters and stimulate failure modes that only manifest under high-temperature-high-stress coupling conditions.

[0031] 4. This invention utilizes an AI model to integrate electrical coupling strength and thermal coupling effect for cross-module reasoning, identifies hazard source modules and vulnerable modules from the multi-module interaction relationship, and automatically generates pairing taboo suggestions, upgrading factory testing from a single individual qualification judgment to a system-level adaptability pre-assessment, effectively reducing the early failure risk of downstream OEMs. Detailed Implementation

[0032] The technical solutions in the embodiments of the present invention will be clearly and completely described below. 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.

[0033] This embodiment provides an AI-based intelligent power module dynamic and static characteristic testing method. Using a 1700V / 300A rated IGBT-IPM module as the test object, it integrates the functions of a dedicated IPM2.0 test board on the STS8200 general-purpose test platform and deploys multiple AI models to work collaboratively. The entire test system is installed in a temperature-controlled shielded room. The source measurement unit and high-speed acquisition front end are equipped with electromagnetic interference filters and isolation transformers to suppress conducted and radiated interference common in industrial environments, ensuring the accuracy of weak signal (such as Ices leakage current) measurements.

[0034] S1. Build an automated testing system with AI processing capabilities.

[0035] The specific configuration and collaborative working logic of the system hardware architecture are as follows:

[0036] The testing system includes a main control computer, a source measurement unit, a pulse generation unit, a high-speed data acquisition unit, and a multi-station interface board adapted to the IPM module under test. The main control computer is responsible for scheduling the inference tasks of four AI models. The collaborative logic between models is as follows: the backpropagation neural network model and the convolutional neural network model execute in parallel, with their input data coming from the static test of S2 and the dynamic test of S3 / S4, respectively; the recurrent neural network model depends on the dynamic test data of S3 and executes in parallel with the former; the graph attention network model serves as a post-processing step, strictly executing after the above three models have output results, with its input being the structured data from the first three. The main control computer centrally controls and reads data from all lower-level instruments through a PCIe bus interface card and a high-speed USB bus.

[0037] During the test preparation phase, a square wave calibration signal with a rise time of less than 1 ns is emitted by the pulse generation unit and simultaneously injected into all high-speed synchronous data acquisition channels. The main control computer calculates the fixed time delay difference between each channel using a cross-correlation algorithm and stores the calculated time delay difference as an "inter-channel time delay compensation parameter" in the system. This parameter is automatically compensated in subsequent waveform data alignment to ensure that the sampling point time deviation of all channels is less than 50 ps.

[0038] The high-speed data acquisition unit is a high-speed synchronous data acquisition card; the test system also includes a temperature-sensitive electrical parameter synchronous sampling channel, which is used to acquire the time series of electrical parameters of the built-in temperature-sensitive elements of each module at a rate synchronized with the high-speed synchronous data acquisition card.

[0039] The temperature-sensitive electrical parameter synchronous sampling channel is a key hardware addition to this system. It employs a multi-channel synchronous sampling analog input module configured with 16-bit resolution and a maximum sampling rate of 250 kS / s per channel. The channel inputs are connected to the NTC thermistor voltage divider circuits at each tested module's workstation, achieving synchronization with the high-speed data acquisition card using the same trigger signal. The voltage values ​​across the NTC resistors of each module are collected in a time-division multiplexing manner. The NTC voltage divider circuit uses a 10kΩ pull-up resistor connected in series with the module's built-in NTC resistor (nominal value 10kΩ at 25℃). The voltage divider node is buffered by an instrumentation amplifier before being output to the temperature-sensitive electrical parameter synchronous sampling channel.

[0040] S2. Collect static characteristic data and predict the characteristics of the entire temperature range based on the AI ​​model.

[0041] After four modules under test are simultaneously installed on the multi-station interface adapter board, without establishing electrical connections between modules, the main control computer controls the relay matrix to independently connect the station of the currently tested module to the source measurement unit channel, ensuring that other stations are completely disconnected and that static testing is not affected by adjacent modules. For each module, the source measurement unit measures the collector-emitter saturation voltage drop Vce(sat) and collector leakage current Ices at case temperatures of 25℃ and 150℃, respectively. The measurement conditions for Vce(sat) are 300A rated current, using Kelvin four-wire connection to eliminate the influence of contact resistance, and the measurement conditions for Ices are 1700V rated voltage, with the gate short-circuited. Each measurement is repeated five times and the average value is taken. The data, along with the module serial number, batch number, test temperature, and timestamp, are stored in a database. This database is used to store all historical test data, model parameters, and judgment results, and serves as the data source for model iterative learning and adaptive threshold updates.

[0042] The Vce(sat) and Ice measurements at discrete temperature points are formatted and input into a pre-trained backpropagation neural network. The network's training samples are derived from a large amount of measured data from modules of the same specifications at different temperature points and their corresponding full-temperature-range characteristic curves, which were obtained through calibration using a high-precision temperature control platform. The data is randomly divided into training and validation sets in an 8:2 ratio to ensure the model's generalization ability.

[0043] The core innovation of this network lies in its loss function design. Instead of simply fitting the data, it embeds a physical constraint penalty term based on the Shockley equation. The specific formula is as follows:

[0044] ;

[0045] Parameter description:

[0046] Mean squared error loss measures the deviation between the model's predicted values ​​and the measured values. It is used to calculate gradients and update network weights through backpropagation.

[0047] The physical constraint balance coefficient is set to 0.1. K-fold cross-validation is used to select the coefficient on the validation set that minimizes the physical plausibility (minimum number of violations of the monotonically decreasing trend) and prediction accuracy (R²). 2 The value that achieves the optimal balance is no less than 0.98.

[0048] The physical penalty term based on the Shockley equation is defined as follows.

[0049] ;

[0050] Parameter description:

[0051] N: Number of sampling points within the predicted temperature range (e.g., -40℃ to 175℃, with a step size of 1℃, totaling 216 points).

[0052] : Absolute temperature (in K) of the i-th sampling point. The Celsius temperature needs to be converted to Kelvin for differential calculation.

[0053] The partial derivative of the saturated pressure drop with respect to temperature is obtained by predicting the curve at temperature T. i The numerical difference (central difference) at that point is obtained. dT represents the differential component of the saturation pressure drop, and dT represents the differential component of the temperature. According to the Shockley equation, this derivative should be negative, so a penalty is only incurred when the derivative is positive.

[0054] The partial derivative of the leakage current with respect to temperature is also obtained by numerical difference of the predicted curve. This represents a small component of the leakage current. According to semiconductor physics, this derivative should be positive, and a penalty is only incurred when the derivative is negative.

[0055] The penalty imposed on the derivative for being positive or negative forces Vce(sat) to decrease monotonically with increasing temperature and Ice to increase monotonically with increasing temperature throughout the entire predicted temperature range.

[0056] Parameter acquisition method: The derivative is obtained by calculating the central difference after performing a second interpolation on the discrete points output by the already trained network. The penalty term participates directly in gradient backpropagation as part of the loss during the training phase.

[0057] After the model predicts the continuous characteristic curve of the module under test across the entire temperature range of -40℃ to 175℃, the system calls the adaptive threshold envelope generated based on historical good product data for comparison. The adaptive threshold envelope is generated based on kernel density estimation, which performs probability density modeling on the parameter distribution of a large number of historical good products at various temperature points, and determines the upper and lower boundaries with a confidence level of 99.99% (corresponding to a false alarm rate of about a few parts per million).

[0058] If any consecutive segment of the predicted curve of the module under test exceeds 5°C outside the envelope, even if the values ​​at the actual measurement points (25°C, 150°C) are compliant, it will still be judged as a "static characteristic temperature-sensitive anomaly" and marked. This 5°C threshold is selected based on the heat capacity time constant to avoid misjudgment due to single-point noise.

[0059] S3. Perform multi-module coupling dynamic testing and collect cross-module dynamic coupling response and temperature-sensitive data.

[0060] After completing the static characteristic tests of all modules, the relay matrix reconfigures the four stations according to the multi-module coupling test requirements. Taking the first round of testing with M0 as the active module as an example: station A switches to the pulse generation unit, while stations B, C, and D switch to passive monitoring mode. Applying a +15V positive bias voltage to the gate of the passive modules (M1-M3) ensures that their internal IGBTs are in the nominal "waiting to turn on" state. In this state, parameters such as gate input impedance and Miller capacitance are consistent with the normal operation of the lower bridge arm switch of a real inverter. Simultaneously, connecting their collectors to the 1200V DC bus high voltage is to reproduce the voltage stress level of the common-mode coupling path in the real system, ensuring that the coupling current generated by dv / dt through the Miller capacitance can be accurately reproduced.

[0061] A standard double-pulse sequence is applied to M0: the first pulse is 20µs, causing the load inductor current to linearly increase to 300A (simulating rated operating conditions); a dead time of 2µs is applied to prevent bridge arm shoot-through; the second pulse is 2µs. At the instant the second pulse turns off, the high-speed synchronous data acquisition card synchronously records the transient waveforms of Ic and Vce of M0, as well as the Vge coupling response waveforms of M1, M2, and M3, at a sampling rate of 2GSa / s, based on the pre-calibrated and stored inter-channel delay compensation parameters described in S1. The acquisition window covers 200ns before and after the turn-off trigger signal, for a total duration of 400ns, in order to fully capture the approximately 100ns turn-off process of a typical IGBT module at 1200V / 300A and the resulting nanosecond-level oscillation phenomenon.

[0062] Simultaneously, the temperature-sensitive electrical parameter synchronous sampling channel synchronously collects the NTC voltage time series of each module at a rate of 250 kS / s, capturing the junction temperature changes caused by switching transient power losses. M1, M2, and M3 are sequentially switched as active modules, and the above process is repeated, obtaining a total of 4×3 sets of cross-module dynamic coupling response data.

[0063] Based on this, the following electrical coupling dimension parameters are extracted from each set of collected coupling response waveforms for subsequent step S7 to construct the graph structure data: common-mode coupling voltage peak amplitude Vcm peak(m,n) dVcm / dt is defined as the voltage value at the sampling point with the largest absolute value in the waveform; the maximum rise slope of the common-mode pulse is dVcm / dt. max(m,n) The oscillation decay rate α is calculated from the maximum value of the numerical differential at the rising edge of the waveform. decay(m,n) The decay time constant and overshoot duration Δt are obtained by fitting the peak value of the waveform oscillation with an exponential envelope. over(m,n) The cumulative duration exceeding the rated withstand voltage of the gate of the module under test in the coupled waveform is defined as the total duration of the voltage. These parameters, along with the corresponding active module m and passive module n identifiers, are stored in the database for use in step S7.

[0064] S4. Preprocess the dynamic waveform and use a convolutional neural network to identify the defect category.

[0065] This step involves extracting 200 sampling points before and after the trigger signal from the transient waveforms of the active module's on / off states recorded by the high-speed synchronous data acquisition card in S3, and performing wavelet denoising. The denoising algorithm uses the Daubechies 4 (db4) wavelet basis and performs a 3-level decomposition. db4 was chosen because its waveform is tightly bound and has good similarity to the transient characteristics of the switching edges; a 3-level decomposition is chosen to separate high-frequency noise from mid- and low-frequency switching characteristics. After denoising, a two-dimensional color waveform image is plotted according to a unified coordinate mapping rule: the horizontal axis represents time (400ns), the left vertical axis represents Vce (0 - 1.2 times the rated voltage), and the right vertical axis represents Ic (0 - 1.5 times the rated current). The Vce curve is plotted in blue (R=0, G=0, B=255), and the Ic curve is plotted in red (255, 0, 0), with a fixed image size of 800×600 pixels. This standardization is a prerequisite for ensuring the consistency of CNN input. Images are input into a pre-trained ResNet-18 convolutional neural network for defect classification, and the judgment threshold of 95% is determined by the optimal point of the ROC curve of the training set.

[0066] The ResNet-18 convolutional neural network was pre-trained before deployment as follows: Training was performed using independently collected historical datasets. The training set contained no fewer than 50,000 standardized waveform images, where defective samples were confirmed through failure physics analysis (e.g., bond wire breakage detected by ultrasonic scanning, gate oxide layer damage detected by microscopy). Data augmentation techniques (including Gaussian noise addition, random pruning, and time-axis perturbation) were used to expand the dataset to a scale balanced with normal samples. Training employed the cross-entropy loss function and the Adam optimizer. Once a classification accuracy exceeding 98% was achieved on the validation set, the model weights were saved, thus completing the pre-training.

[0067] S5. Based on the dynamic junction temperature online reconstruction model, obtain the real-time junction temperature curves of each module.

[0068] The transient waveform data of Vce and Ic recorded by the high-speed synchronous data acquisition card in step S3 above are directly used in this step as time-domain functions Vce(t) and Ic(t) for power loss calculation. The transient power loss is calculated using Vce(t) and Ic(t) of the active module M0. .Will Time series and NTC reference temperatures of the four tested modules The time series data is resampled to a step size of 100µs, where k is the module index variable, k∈{0,1,2,3}. This step size is chosen because the chip-substrate thermal time constant of a typical IPM module is in the tens of milliseconds range; 100µs sampling can fully capture the details of the instantaneous temperature rise during switching, while avoiding excessive computational load. The data input physical information is embedded into a recurrent neural network. This network uses a Long Short-Term Memory (LSTM) structure, and its loss function is designed as follows:

[0069] ;

[0070] Parameter description:

[0071] Mean square error between the estimated junction temperature and the calibration reference value.

[0072] The physical constraint weighting coefficient is set to 0.2. A grid search is used to select a value within the range of 0.05 to 0.5 that minimizes the temperature reconstruction error (mean absolute error) on the validation set and results in a smooth junction temperature change curve (minimum sum of squared second derivatives).

[0073] The physical constraint term, based on the discrete format of the heat conduction equation, is used to force the estimated junction temperature output by the network to strictly satisfy the heat conduction law in the time dimension. That is, it penalizes the deviation between the junction temperature change rate and (self-heating power - heat transfer to the substrate - heat transfer to adjacent modes) / heat capacity at each time step. The specific definition is as follows.

[0074] ;

[0075] Parameter description and acquisition method:

[0076] k, j: Module index variables, k, j∈{0,1,2,3}.

[0077] t: Discrete time step, step size Δt = 100 μs.

[0078] The estimated junction temperature (°C) of the module with module index k at time t is output by the LSTM network.

[0079] Equivalent thermal capacity (J / K) of the module chip with module index k. Method of obtaining: From the transient thermal impedance curve provided by the manufacturer. Extracted by Δt / ΔZ th The limit is determined at t→0, or by the thermal time constant. The results are obtained by reverse engineering, and finite element simulation is used for verification when necessary.

[0080] Equivalent thermal resistance (K / W) from the module chip with module index k to the NTC temperature measurement point on the substrate. Method of obtaining: Measured offline using the "dual-interface" method according to JESD51-14 standard, or... Steady-state value of the curve.

[0081] The coupling thermal resistance (K / W) between module k and module j. Method of obtaining the coefficient: Calculate the thermal coupling coefficient by establishing a multi-module model using multiphysics finite element simulation (such as ANSYS Icepak), and then correct it using an infrared thermal imager for actual measurement.

[0082] The self-heating power (W) of the module with module index k at time step t. This value is 0 for passive modules; for active modules, it is calculated from transient Vce × Ic.

[0083] : The substrate reference temperature (°C) measured by the built-in NTC thermistor of module k. It is read by the voltage divider circuit and converted by the calibration table.

[0084] After reconstruction, the following thermal coupling dimension parameters are extracted from the junction temperature curves of each module for subsequent step S7 to construct the graph structure data: For each pair of active module m and passive module n, extract ΔTj. couple(n;m) dTj / dt is defined as the peak value of the junction temperature change of passive module n within the transient switching state of m and the subsequent thermal response time window (300µs); max(n;m) , defined as the maximum value of the junction temperature change rate of n within the aforementioned time window. Simultaneously, the coupling response difference D is calculated. response(m,n) Extract gate voltage waveform segments from passive module n within 50ns windows before and after the switching transient of active module m, and calculate the dynamic time warping (DTW) distance between the two segments to quantify the degree of disturbance caused by the switching action of module m to the gate behavior of module n. These parameters, along with the corresponding identifiers of active module m and passive module n, are stored in the database for use in step S7.

[0085] S6. Based on the reconstructed dynamic junction temperature, adaptively calibrate the initial thermal state of subsequent test modules.

[0086] Before M1 is activated as the active module after the M0 test is completed, the target preheating junction temperature is calculated. The formula for calculating the target junction temperature is:

[0087] ;

[0088] : The current shell temperature (°C) set by the heating unit below the M1 workstation.

[0089] The maximum transient junction temperature rise (°C) of module M1 during the previous M0 switching process was reconstructed. This value was extracted and stored in step S5.

[0090] If calculated If the junction temperature exceeds 80% of the maximum allowable junction temperature in the module datasheet (this 80% is the engineering standard for derating, leaving a safety margin), it will be forcibly truncated to that threshold.

[0091] S7. Construct graph-structured data and use graph attention networks for multi-module coupled reasoning.

[0092] This step organizes the various test and reconstruction data obtained from S2 to S6 into a graph structure. The specific construction method is as follows: four tested modules M0 to M3 are used as four graph nodes. The feature vector of each node is composed of two parts: the first part is the static characteristic data obtained in step S2, and the second part is the dynamic defect identification result. Two directed edges are constructed between any two different modules, for a total of 12 directed edges. The feature vector of each directed edge is composed of the following measured physical quantities: electrical coupling dimension parameters and thermal coupling dimension parameters.

[0093] The electrical coupling dimension parameter is the peak amplitude Vcm of the common-mode coupling voltage collected by S3. peak(m,n) Maximum rising slope dVcm / dt max(m,n) Oscillation decay rate α decay(m,n) and the overshoot duration Δt exceeding the gate rated withstand voltage. over(m,n) ;

[0094] The thermal coupling dimension parameter is the transient change in passive module junction temperature ΔTj caused by the switching of the active module extracted in S5. couple(n;m) The time derivative of the temperature rise, dTj / dt max(n;m) The dynamic time warping distance D of the passive module gate voltage waveform before and after coupling. response(n;m) .

[0095] Cross-module causal reasoning is performed using graph attention networks. When aggregating information between nodes, the graph attention network employs a physically modified attention weight formula. This formula modifies the calculation of attention coefficients in standard graph attention networks by adding a physical correction term δj→i to the attention logit value. This correction term is mapped from the measured electrical and thermal coupling strengths of S3 and S5, ensuring that modules with stronger coupling have greater weights during information aggregation, thus providing physically interpretable reasoning results.

[0096] ;

[0097] Parameter description:

[0098] ， : Feature vectors of nodes i and j.

[0099] Learnable shared linear transformation matrix.

[0100] Attention vector The transpose of this vector maps the concatenated vector to a scalar logit value.

[0101] ‖: Vector concatenation operation.

[0102] : The set of neighboring nodes of node i.

[0103] : Physical correction term, obtained by mapping edge feature vectors through a miniature fully connected network.

[0104] The network will ultimately output the probability of the hazard source. Vulnerability probability And suitability score.

[0105] S8. Generate a test report containing a comprehensive judgment conclusion.

[0106] The main control computer summarizes the test and reasoning results of all the aforementioned steps and generates a comprehensive test report for each module under test.

[0107] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for testing the dynamic and static characteristics of an AI-based intelligent power module, characterized in that: Includes the following steps: S1. Build an automated testing system with AI processing capabilities. The testing system includes a main control computer, a source measurement unit, a pulse generation unit, a high-speed data acquisition unit, and a multi-station interface board adapted to the IPM module under test. The main control computer is loaded with at least one pre-trained AI model. S2. Install the IPM module under test on the multi-station interface board, collect its static characteristic data at multiple temperature points through the source measurement unit, and input the static characteristic data into the first AI model to obtain the static characteristic prediction result of the module in a continuous temperature range. S3. Install multiple IPM modules from the same batch simultaneously on the multi-station interface board, perform dynamic testing on each module sequentially through the pulse generation unit, and synchronously acquire the transient waveform of the active module and the coupling response waveform of the passive module using the high-speed data acquisition unit. S4. Input the transient waveform into the second AI model to obtain the dynamic defect identification result; S5. Calculate the transient power loss based on the transient waveform and input it into the third AI model to obtain the dynamic junction temperature of each module during the test process; S6. Based on the dynamic junction temperature, after the current module test is completed, the initial thermal state before the next module test is adaptively adjusted to reproduce the thermal interlocking conditions between modules. S7. Construct the static characteristic data, the coupling response waveform, and the dynamic junction temperature into structured data, input them into the fourth AI model, and obtain the system-level adaptability evaluation results of each module. S8. Summarize the above test and evaluation results and generate a comprehensive test report.

2. The method for testing the dynamic and static characteristics of an AI-based intelligent power module according to claim 1, characterized in that: In step S2, the first AI model is a backpropagation neural network with embedded physical constraints.

3. The method for testing the dynamic and static characteristics of an AI-based intelligent power module according to claim 2, characterized in that: The physical constraints force the prediction that the collector-emitter saturation voltage drop decreases monotonically with increasing temperature, and the forced prediction that the collector leakage current increases monotonically with increasing temperature.

4. The method for testing the dynamic and static characteristics of an AI-based intelligent power module according to claim 1, characterized in that: In step S3, during dynamic testing, the gate of the passive module is subjected to a nominal forward bias, and its collector is connected to the same DC bus voltage as the active module.

5. The method for testing the dynamic and static characteristics of an AI-based intelligent power module according to claim 1, characterized in that: In step S4, the second AI model is a convolutional neural network model.

6. The method for testing the dynamic and static characteristics of an AI-based intelligent power module according to claim 1, characterized in that: The third AI model mentioned in step S5 is a recurrent neural network model with embedded physical constraints; the physical constraints are based on the heat conduction equation.

7. The method for testing the dynamic and static characteristics of an AI-based intelligent power module according to claim 1, characterized in that: The adaptive adjustment in step S6 includes: taking the maximum junction temperature rise experienced by the next module as a passive module in the previous test as the target preheating heat, and adjusting the initial junction temperature of the next module to the corresponding target value through heating and / or pre-pulse methods.

8. The method for testing the dynamic and static characteristics of an AI-based intelligent power module according to claim 1, characterized in that: The structured data mentioned in step S7 is graph structured data, in which each module serves as a graph node, and the measured electrical coupling strength and thermal coupling effect between modules serve as edge features.

9. The method for testing the dynamic and static characteristics of an AI-based intelligent power module according to claim 1, characterized in that: The fourth AI model is a graph attention network model.