Power module testing methods, apparatus, equipment, media and products

Abstract

This application relates to a power module testing method, apparatus, equipment, medium, and product. The method includes: acquiring a set of original waveforms generated by the power module and test condition data of the power module; obtaining an aligned waveform set based on the test condition data and the original waveform set; determining pulse sequence labels and quality control masks associated with the aligned waveform set, whereby the pulse sequence labels record key characteristic parameters of the active pulse excitation applied during the power module test, and the quality control mask characterizes the quality level of each waveform in the aligned waveform set; constructing a platform fingerprint based on the aligned waveform set, test condition data, and quality control mask to obtain fingerprint data; and evaluating the thermal path based on the aligned waveform set, test condition data, and pulse sequence labels to obtain thermal characteristic data. This method can improve the accuracy of identifying latent defects such as thermal paths and electrical micromorphology under short test cycles.

Description

Technical Field

[0001] This application relates to the field of power module testing technology, and in particular to a power module testing method, apparatus, equipment, medium and product. Background Technology

[0002] With the rapid development of power electronic equipment towards higher power density and higher reliability, the quality and long-term stability of power modules, as core components, have become crucial in determining the lifespan of systems.

[0003] In traditional technologies, the testing process heavily relies on direct comparison with specifications. For electrical testing, automated testing equipment applies prescribed current and voltage conditions, acquires the module's response waveform, extracts key parameter points, and compares them with preset absolute thresholds; anything outside these ranges is considered unqualified. For thermal testing, a heating pulse is applied, the change in junction voltage is measured, and the temperature rise is calculated using a known temperature sensitivity coefficient. Then, a single thermal resistance value is calculated based on the steady-state temperature rise and the heating power.

[0004] However, traditional technologies suffer from low accuracy in identifying latent defects such as thermal paths and electrical micromorphology under short testing cycles. Summary of the Invention

[0005] Therefore, it is necessary to provide a power module testing method, apparatus, equipment, medium, and product that can improve the accuracy of identifying latent defects such as thermal paths and electrical micromorphology under short test cycles to address the above-mentioned technical problems.

[0006] In a first aspect, this application provides a power module testing method, comprising: acquiring an original waveform set generated by the power module and test condition data of the power module; obtaining an aligned waveform set based on the test condition data and the original waveform set; determining a pulse sequence label and a quality control mask associated with the aligned waveform set, wherein the pulse sequence label is used to record key characteristic parameters of the active pulse excitation applied during the power module testing process, and the quality control mask is used to characterize the quality level of each waveform in the aligned waveform set; constructing a platform fingerprint based on the aligned waveform set, test condition data, and quality control mask to obtain fingerprint data; performing thermal path evaluation based on the aligned waveform set, test condition data, and pulse sequence label to obtain thermal characteristic data; and identifying latent defects of the power module based on the fingerprint data, thermal characteristic data, test condition data, and quality control mask to obtain the target test result.

[0007] Secondly, this application also provides a power module testing apparatus, comprising: an acquisition module for acquiring a set of original waveforms generated by the power module and test condition data of the power module, and obtaining an aligned waveform set based on the test condition data and the original waveform set; a processing module for determining the pulse sequence label and quality control mask associated with the aligned waveform set, wherein the pulse sequence label is used to record key characteristic parameters of the active pulse excitation applied during the power module test, and the quality control mask is used to characterize the quality level of each waveform in the aligned waveform set; a construction module for constructing a platform fingerprint based on the aligned waveform set, test condition data, and quality control mask, and obtaining fingerprint data; the processing module is further used for performing thermal path evaluation based on the aligned waveform set, test condition data, and pulse sequence label, and obtaining thermal characteristic data; and an identification module for identifying latent defects of the power module based on the fingerprint data, thermal characteristic data, test condition data, and quality control mask, and obtaining target test results.

[0008] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the power module testing method described above.

[0009] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the power module testing method described above.

[0010] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps in the power module testing method described above.

[0011] The aforementioned power module testing methods, devices, equipment, media, and products, through waveform alignment associated with operating conditions, platform fingerprint construction, pulse excitation-assisted thermal path evaluation, and multi-dimensional data fusion decision-making, combined with effective screening of quality control masks, can accurately identify microstructural and thermal path latent defects of power modules within short test cycles, thereby improving the accuracy of identifying latent defects such as thermal paths and electrical micromorphology under short test cycles. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1This is a diagram illustrating the application environment of a power module testing method in one embodiment;

[0014] Figure 2 This is a flowchart illustrating a power module testing method in one embodiment;

[0015] Figure 3 This is a flowchart illustrating the power module testing method in another embodiment;

[0016] Figure 4 This is a structural block diagram of a power module testing device in one embodiment;

[0017] Figure 5 This is an internal structural diagram of a computer device in one embodiment;

[0018] Figure 6 This is a diagram of the internal structure of a computer device in another embodiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0020] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0021] The power module testing method provided in this application embodiment can be applied to, for example, Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed on a cloud or other network server. Terminal 102 responds to the power module's test request and sends the collected raw waveform set and test condition data to server 104. Server 104 obtains the raw waveform set generated by the power module and the power module's test condition data, and obtains an aligned waveform set based on the test condition data and the raw waveform set. It then determines the pulse sequence label and quality control mask associated with the aligned waveform set. The pulse sequence label records the key characteristic parameters of the active pulse excitation applied during the power module test, and the quality control mask characterizes the quality level of each waveform in the aligned waveform set. Based on the aligned waveform set, test condition data, and quality control mask, it constructs a platform fingerprint to obtain fingerprint data. It then performs a thermal path evaluation based on the aligned waveform set, test condition data, and pulse sequence label to obtain thermal characteristic data. Finally, based on the fingerprint data, thermal characteristic data, test condition data, and quality control mask, it identifies the power module's latent defects and obtains the target test result. Server 104 sends the target test results to terminal 102. Terminal 102 can be, but is not limited to, various personal computers, laptops, and tablets. Server 104 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server providing cloud computing services.

[0022] In one exemplary embodiment, such as Figure 2 As shown, a power module testing method is provided, which is applied to... Figure 1 Taking the server in the example of this, the explanation includes:

[0023] Step 201: Obtain the original waveform set generated by the test power module and the test condition data of the power module, and obtain the aligned waveform set based on the test condition data and the original waveform set.

[0024] The original waveform set can be the original electrical signal time series of the power module during the test process, such as voltage and current, which can be directly acquired through a data acquisition system (such as an oscilloscope); the test condition data can be a vector containing DC bus voltage Vdc, gate resistance Rg, and test temperature T; the aligned waveform set can be a standardized waveform dataset obtained by time-domain calibration and elimination of time reference differences from the original waveform set.

[0025] Optionally, an aligned waveform set can be obtained by performing time-domain alignment on the test condition data and the original waveform set. Time-domain alignment aims to eliminate inconsistencies in time bases between different test samples caused by trigger delays, clock drift, and other factors. Specifically, this can be achieved through techniques such as global trigger positioning, cross-correlation fine alignment, and unified time-base resampling.

[0026] Step 202: Determine the pulse sequence label and quality control mask associated with the aligned waveform set. The pulse sequence label is used to record the key characteristic parameters of the active pulse excitation applied during the power module test. The quality control mask is used to characterize the quality level of each waveform in the aligned waveform set.

[0027] Optionally, pulse sequence labels and quality control masks can be obtained by quality gating the aligned waveform set. The quality gating is used to evaluate the signal quality of the aligned waveform, for example, by calculating the signal-to-noise ratio, detecting saturation or clipping distortion, and generating a quality control mask (QC). _mask Among them, QC _mask It can be a Boolean vector, where elements with true values ​​correspond to waveform segments with acceptable quality, and elements with false values ​​correspond to waveform segments with unacceptable quality; or, QC_ mask It can be a weight vector in the range [0,1], and the magnitude of the value reflects the quality of the signal.

[0028] For example, in the time-domain alignment and quality gating steps, more refined sub-steps can be used to obtain high-quality aligned waveform packets. For instance, in the raw data acquisition and preprocessing stage, the probes and fixtures of the acquisition system can first be calibrated for amplitude and phase, and a calibration matrix can be applied to the acquired waveform to correct amplitude and phase distortion while removing DC bias. Subsequently, a bandpass anti-aliasing filter and a notch filter for a specific power frequency or switching frequency can be applied to shape the signal, obtaining a preprocessed waveform packet. In the time-base synchronization and trigger alignment stage, coarse alignment can be performed first using a global trigger signal, followed by precise point-by-point synchronization between samples using a cross-correlation peak fine alignment method. To ensure data consistency across different batches and acquisition devices, the finely aligned waveform can be resampled using a unified time base, and optionally, an estimation and compensation stage for cross-batch clock drift can be added. In the signal quality assessment and repair stage, in addition to scoring the signal-to-noise ratio, saturation, and clipping, some electromagnetic interference (EMI) related indicators can be calculated. For localized spikes or missing data, interpolation filling can be performed using local spline repair. All quality assessment results are then quantified and mapped to the final quality control mask, QC_. mask .

[0029] It's worth noting that the generation of pulse sequence annotations can be achieved through the following steps: Event Timeline Parsing: The test script is read, and event sequences such as StartPulse, EndPulse, and StartMicroHeat are parsed and matched with the global timeline of the aligned waveform packet to obtain the precise timestamp of each event in the actual waveform, forming an event_timeline data structure. Window Index Table Generation: Based on the paired start and end events in the event_timeline, the continuous waveform is divided into multiple independent analysis windows with type labels (such as 'Plateau', 'HeatingSequence'), and a window index table containing window ID, type, and start and end sampling point indices is generated. Pulse Sequence Annotation Mapping: The window index table is mapped to information such as device tag number and batch number contained in the operating conditions, ultimately generating pulse sequence annotations containing complete context information.

[0030] For example, the event timeline parsing can be performed as follows: The system reads the aligned waveform packet, the test script, and the operating condition information. The test script predefines the sequence and nominal duration of each excitation event in the entire test process. The system matches the event sequence in the test script with the global timeline of the aligned waveform packet, parsing out the precise timestamp of each key event (such as the start and end of a power pulse, the injection of a micro-heating sequence, etc.) in the actual waveform data. The output of this process is a data structure called the event timeline (event_timeline). The event_timeline can be a list or a table, where each row represents an event and contains at least three columns of information: event type (e.g., a string such as 'StartPulse', 'EndPulse', 'StartMicroHeat'), event timestamp (e.g., a floating-point number representing the sampling point index or actual time in the waveform data), and parameters related to the event (e.g., the expected current, voltage, etc. of the pulse).

[0031] The method for dividing the analysis window and generating the index table is as follows: After obtaining the event_timeline, it needs to be converted into a direct index of the waveform data to facilitate subsequent algorithm calls. The aim is to meaningfully segment continuous waveform data according to the event timeline and generate a window index table. This algorithm defines different analysis windows by reading pairs of start and end events in the event_timeline. For example, the time interval between a 'StartPulse' event and an 'EndPulse' event constitutes a complete pulse window. The data structure of the window index table can be a table, where each row represents an analysis window and contains the following information: window ID, window type (e.g., 'RiseTime', 'FallTime', 'Plateau', 'HeatingSequence'), start sampling point index, and end sampling point index. In this way, the originally one-dimensional, continuous waveform data is structured into a series of independently addressable analysis windows with type labels.

[0032] The pulse sequence annotation can be obtained as follows: A window index table provides a temporal division of the waveform, but these windows still need to be associated with specific test objects and test content. This step aims to generate the final pulse sequence annotation. Specifically, the system reads the window index table and operating condition information (which typically includes device tag numbers, batch numbers, etc.) and maps this identification information to each analysis window. For example, a pulse sequence annotation entry might look like this: {'Tag Number':'A01','Batch Number':'202508A','Window ID':101,'Window Type':'MicroHeatingPulse','Start Index':50000,'End Index':75000}. For instance, the thermal analysis algorithm accurately extracts the micro-heating segment waveform and corresponding power excitation sequence for analysis based on the entry with the window type "MicroHeatingPulse" in this annotation.

[0033] This embodiment details how to quantify the positioning quality of the platform's initial anchor point and transform it into a sample weight that can be explicitly used in subsequent steps. Specifically, this embodiment includes the following steps: First, the quantitative calculation of the anchor point reliability score (anchor_score). After locating the platform's initial anchor point, its positioning quality needs to be objectively quantitatively evaluated. The anchor_score is a comprehensive score ranging from [0,1], and its calculation can be based on the verification of the distortion and repeatability of signal characteristics near the anchor point. A specific calculation formula can be: anchor_score = w_rep × S_rep + w_dist × S_dist. Where: Srep is the repeatability score, used to measure the stability of the anchor point position in multiple consecutive pulse tests. For example, the standard deviation σt of the anchor point position timestamps in N consecutive pulses can be calculated, then S... rep It can be defined as S rep =exp(σ t / σ0), where σ0 is a normalization constant. σ t The smaller the value, the closer the score is to 1, indicating better repeatability. Sdist is the distortion score, used to measure the similarity between the waveform near the anchor point and the ideal waveform. For example, the cross-correlation coefficient between the actual waveform segment near the anchor point and a standard noise-free platform starting template can be calculated; this coefficient itself can serve as Sdist. A higher correlation coefficient indicates less distortion. wrep and wdist are preset weighting coefficients that satisfy wrep + wdist = 1, used to balance the importance of repeatability and distortion.

[0034] The second step is the mapping from anchor reliability scores to sample weights (KLTR_weight). For ease of use in subsequent statistical learning and scoring, the anchor_score needs to be mapped to a weight, KLTR_weight, that is more suitable as a multiplicative factor. This embodiment uses a sigmoid function for non-linear mapping: KLTR_weight = 1 / (1 + exp(-k(anchor_score - x0))). Here, anchor_score is the result calculated in the previous step. k and x0 are the adjustment parameters of the sigmoid function. For example, k = 10 and x0 = 0.7 can be set. This means that when anchor_score equals 0.7, the weight is 0.5; when anchor_score is much greater than 0.7, the weight rapidly approaches 1; and when anchor_score is much less than 0.7, the weight rapidly approaches 0. Compared to a simple linear mapping, this non-linear mapping can more effectively amplify the weights of high-quality samples while more thoroughly suppressing the influence of low-quality samples.

[0035] The third step is the transfer and application of weights. This step is crucial for establishing a seamless technological effect chain. The calculated KLTR_weight, as metadata strongly bound to the samples, is actively transferred to the physically guided conditional learning and joint decision-making process. Specifically, in the covariance estimation step, each sample is multiplied by its corresponding KLTR_weight when participating in the calculation. Similarly, in the joint anomaly score calculation step, the final fusion score formula will also include KLTR_weight as a global weighting factor. In this way, the impact of a sample deemed unreliable during the alignment stage on the final model and decision will be systematically weakened from beginning to end.

[0036] In this embodiment, the acquisition of the original waveform follows the criterion of fs ≥ 4 × B_analog, where fs is the system sampling frequency and B_analog is the effective analog bandwidth of the measured signal. This criterion aims to oversample far beyond the Nyquist sampling theorem (fs ≥ 2 × B_analog) to accurately capture the fast transient details of the voltage and current waveforms during the power module switching process, which is crucial for subsequent platform micromorphological analysis. The method for determining B_analog is as follows: B_analog is mainly determined by the switching speed of the power module (especially the rise time t). rise and descent time t fall The value of B_analog is determined by the empirical formula B_analog ≈ 0.35 / tr, where tr is the shorter switching time. For example, for an IGBT module with a switching rise time of 50 nanoseconds (ns), the corresponding analog bandwidth is approximately 0.35 / (50×10-9s) = 7MHz. Considering that higher frequency ringing may exist in the test system and actual signals, to retain sufficient design margin, the value of B_analog can be set to 10MHz. Therefore, according to the criterion of fs ≥ 4×B_analog, the required minimum sampling frequency fs should be 4×10MHz = 40MS / s (40 megasamples per second). In practical applications, it is preferable to use acquisition equipment with a higher sampling rate, such as 100MS / s or higher, to ensure the highest signal fidelity.

[0037] Step 203: Construct the platform fingerprint based on the aligned waveform set, test condition data, and quality control mask to obtain fingerprint data.

[0038] The fingerprint data includes, but is not limited to, a conditionally invariant platform micromorphological fingerprint (f*). plt ), a physical consistency residual (r phys ) and a physical consistency weight (w phys )wait.

[0039] Optionally, refined features characterizing the intrinsic physical properties of the device can be extracted from the Miller plateau region of the aligned gate voltage waveform (VGS). The microscopic morphology of this plateau, such as minute bumps, ringing, or slope variations, is closely related to the device's parasitic capacitance and internal structure, and is a key source of information for identifying latent defects. The resulting fingerprint data is a structured set of features. Among them, f* plt It is a multi-dimensional vector used to describe the pure morphological features after removing the influence of operating conditions (such as Vdc, Rg, T), ensuring the comparability of features between different batches and under different test conditions. phys This quantifies the degree of agreement between the extracted features and known semiconductor physics (e.g., the relationship between gate charge Qg and plateau time t_plt). The greater the deviation of the value from zero, the more likely there is an anomaly at the physical mechanism level. phys Is with r phys The accompanying weights are used to represent the reliability of the physical consistency residual.

[0040] Step 204: Evaluate the thermal path based on the aligned waveform set, test condition data, and pulse sequence annotation to obtain thermal characteristic data.

[0041] Thermal characteristic data includes, but is not limited to, the thermal slope and its confidence level.

[0042] Optionally, based on pulse sequence labeling, a specially designed micro-heating power sequence (e.g., multitone or PRBS sequence) capable of eliciting a broadband thermal response is injected into the device. Then, the device's electrical response (e.g., VL) is identified with high precision. CE The change in voltage-temperature rise (Δ*(t)) is used to deduce the internal temperature rise of the device. Finally, within a specific linear operating region, the key thermistor parameter, namely the thermistor slope α, is calculated based on the voltage-temperature rise relationship under two different temperature rise states. T (Units are usually mV / °C), and at the same time assess its confidence level α. Tconf An anomalous alpha T Or a very low α Tconf These could all indicate a problem with the thermal path.

[0043] In one embodiment, thermal path evaluation is performed based on an aligned waveform set, test condition data, and pulse sequence annotations to obtain thermal characteristic data. This includes: generating a target power sequence based on annotation parameters associated with the pulse sequence annotations; synchronizing the target power sequence with the electrical response data in the aligned waveform set to obtain target associated data, and determining temperature rise estimation data based on the target power sequence and the electrical response data corresponding to the target power sequence; and estimating the slope of the electrical response data and the temperature rise estimation data within a linear operating range determined based on the test condition data to generate thermal characteristic data.

[0044] The target power sequence can be a micro-heating power sequence covering a wide frequency band; the electrical response data can be point data (such as junction voltage) responding to the target power sequence; the target associated data can be a pairwise dataset composed of the target power sequence and the corresponding electrical response data after time synchronization, where each set of data satisfies the causal correspondence between excitation and response; the temperature rise estimation data can be curve data extracted from the target associated data that characterizes the change of junction temperature inside the power module over time.

[0045] Optionally, a wide-bandwidth micro-heating power sequence can be constructed by designing a gate power excitation consisting of a band-limited multi-tone signal or a pseudo-random binary sequence (PRBS), and then verifying this excitation sequence under electromagnetic interference (EMI) constraints. For example, a multi-tone signal can be designed with its power spectrum energy concentrated at multiple discrete frequency points, ranging from tens of hertz to kilohertz, corresponding to the time constants of different thermal path stages from the chip to the heatsink. This excitation sequence is achieved by controlling the gate drive signal of the power module, thereby generating a total power consumption sequence Ptotal(t) within the device. When designing this sequence, its spectral characteristics also need to be verified to ensure that its high-frequency components do not cause excessive EMI to the test system, meeting the test environment requirements. In this way, even within a millisecond-level test window, the multimodal thermal response of the device can be fully excited, providing a high-quality signal foundation for subsequent accurate identification.

[0046] After obtaining the broadband power consumption excitation P total (t) and synchronously acquired electrical responses (such as junction voltage V) CE After processing *(t), the temperature rise curve ΔT*(t) of the device needs to be derived from it. This is essentially a system identification problem, i.e., solving the convolution equation ΔT*(t)=h(t)*P. total The thermal impulse response h(t) in (t) is used in this embodiment. The sparse deconvolution algorithm based on multiple interleaved pulses (i.e., the ICI algorithm) is an effective means to solve this problem. This method divides the long excitation sequence into multiple interleaved short pulse segments, and utilizes the sparsity of the pulse response to transform the deconvolution problem into an easily solvable sparse optimization problem. In addition, prior constraints based on the physical geometry of the device's thermal resistance-thermal capacity are applied during the solution process. Specifically, the thermal model of the device can be equivalent to a Foster or Cauer network composed of multiple RC links connected in series, whose thermal resistance R th and heat capacity C thAll values ​​must be positive, and their cumulative heat capacity increases monotonically with the extension of the heat path. Incorporating these physical priors as constraints into the optimization problem effectively avoids the solution process from getting trapped in non-physical local optima, significantly improving the accuracy and robustness of the identification results. Finally, by comparing the identified thermal impulse response h(t) with the power consumption sequence P... total By convolving (t), a high-precision temperature rise estimation curve ΔT*(t) can be obtained. Finally, using the obtained temperature rise estimate, the final thermistor slope and its confidence level are calculated under specific electrical conditions.

[0047] For example, obtaining a pure internal junction voltage through precise parasitic parameter de-embedding is a crucial prerequisite for ensuring the accuracy of thermal identification. In this embodiment, the step of analyzing the electrical response in the aligned waveform packet, before employing the interleaved convolution inversion algorithm, further includes: performing phase consistency correction on the voltage and current waveforms in the aligned waveform packet to compensate for timing deviations introduced by the measurement channel. Specifically, due to differences in the physical characteristics and cable lengths of the voltage and current probes, the acquired voltage waveform V... CE There is often a small timing deviation (i.e., phase difference) between the current waveform I(t) and the current waveform I(t), but this deviation can have a significant impact on dynamic analysis. In this embodiment, the strong correlation between the rapid changes in voltage and current (dV / dt and dI / dt) during device switching is utilized. This timing deviation is estimated by calculating the peak position of the cross-correlation function of the two waveforms, and one of the waveforms is shifted along the time axis to achieve precise alignment between the two waveforms in the time domain.

[0048] Based on the phase-corrected waveform, the parasitic inductance and resistance parameters of the circuit are estimated using a controlled regression method, and their parameter uncertainties are evaluated simultaneously. This step is the core of the wire harness / weld wire voltage drop de-embedding (SPLD) technology. In the power module test circuit, the total measured voltage V CE (t) is actually the internal true junction voltage V. CE *(t) is the sum of the dynamic voltage drop generated by external parasitic components (including package leads and bonding wires). This relationship can be modeled as: V CE (t)=V CE *(t)+(R lead +R wire )*I(t)+L lead *(dI(t) / dt). Where R lead L is the parasitic resistance of the package leads. lead For parasitic inductance, R wire This represents the parasitic resistance of the bond wire. To isolate and estimate these parameters, this embodiment employs a controlled regression method. Specifically, waveform segments of the device in specific operating states (e.g., during turn-off or in the linear region of conduction but not saturation) are selected, in which the internal junction voltage V0 is...CE The variation pattern of *(t) is known or can be reasonably approximated (e.g., V at turn-off). CE *(t) is approximately equal to the bus voltage Vdc). In this way, the above equation can be transformed into a equation concerning the unknown parameter R. lead ,L lead ,R wire For linear regression problems, the estimated values ​​R of these parameters can be obtained using the least squares method or more robust regression techniques (such as Huber regression). lead_hat Llead _hat Rwire _hat At the same time, the regression process also outputs the covariance matrix of the parameter estimates, which can be used to assess the uncertainty of these parameters.

[0049] Using the estimated parasitic parameters, the dynamic voltage drop is removed from the original voltage waveform to obtain a purified junction voltage. After obtaining the estimated values ​​of the parasitic parameters, the voltage waveform throughout the testing process can be corrected. The purified junction voltage V0 CE *(t) is calculated using the following formula: V CE *(t)=V CE (t)-(R lead_hat +R wire_hat )*I(t)-L lead_hat *({dI(t) / dt)。 The parameter uncertainty is propagated to generate the purified junction voltage uncertainty, which serves as input for subsequent identification and estimation. Since the estimated values ​​of parasitic parameters themselves contain uncertainty, this uncertainty propagates to V. CE The calculation result of *(t) is used to calculate V using standard error propagation theory. CE *(t) Uncertainty at each time point VCE This uncertainty information is crucial; it will be incorporated into subsequent thermal slope estimation to ultimately determine the confidence level of the slope.

[0050] It is worth noting that, in order to actively excite and expose the multimodal thermal response of the device at different time scales (fast and slow), a micro-heating sequence (MSH) based on band-limited multitone signals was employed. The design of the frequency set of this multitone signal directly affects the breadth and accuracy of thermal response identification. Its design criteria are as follows:

[0051] Logarithmic interval distribution: The heat transfer path within a device, from the chip junction to the module substrate and then to the heat sink, has a wide range of corresponding thermal time constants, typically spanning multiple orders of magnitude. Therefore, the frequency points of multi-tone signals should not be linearly selected, but rather approximately uniformly distributed on a logarithmic coordinate axis. For example, a typical frequency set could be: {10Hz, 22Hz, 46Hz, 100Hz, 215Hz, 464Hz, 1000Hz}.

[0052] Frequency range coverage: The lower limit of the frequency range should be low enough (e.g., 10Hz or lower) to effectively stimulate slow thermal modes related to the package and heat dissipation path. The upper limit of the frequency range should be high enough (e.g., 1kHz or higher) to stimulate fast thermal modes related to the chip itself and its vicinity.

[0053] Avoid harmonic interference: The selected frequency should avoid overlapping or being too close to the system's main switching frequency (such as 10kHz, 20kHz) and its lower harmonics, as well as the power grid frequency (50 / 60Hz) and its harmonics, in order to prevent spectral aliasing and interference.

[0054] Optimizing the Crest Factor: Multitone signals are composed of multiple superimposed sine waves, and their amplitudes may exhibit large peak values. To inject as much average power as possible without exceeding the device's safe operating area (SOA), the initial phase of each frequency component needs to be optimized to minimize the signal's crest factor. Mature techniques such as the Schroeder-phase algorithm can be used to achieve this optimization.

[0055] Optionally, if a multi-tone signal is used for a wide-band micro-heating power sequence, the design criteria for its frequency set are: each frequency point should be approximately uniformly distributed on the logarithmic coordinate axis, the frequency range should cover from a few hertz to several kilohertz, and the system switching frequency and its harmonics should be avoided.

[0056] For example, after constructing a micro-heating power sequence covering a wide frequency band, the method further includes: selecting suitable test segments for micro-heating based on pulse sequence labeling, performing thermal response observability analysis, and applying prior constraints based on the physical geometric relationship between device thermal resistance and thermal capacity. Specifically, this can be implemented by: in the sparse deconvolution optimization solver, applying prior constraints to the thermal resistance R of each level of the thermal network model. th_i and heat capacity C th_i Add nonnegativity constraints (i.e., R) th_i ≥0,C th_i ≥0).

[0057] In one embodiment, within a linear operating region determined based on test condition data, slope estimation is performed on electrical response data and temperature rise estimation data to generate thermal characteristic data. This includes: selecting multiple operating time points from a set of aligned waveforms operating within the linear operating region determined by the test condition data; extracting at least two data pairs from the electrical response data and temperature rise estimation data based on the multiple operating time points, wherein each data pair includes a voltage value in the electrical response data and a temperature corresponding to the voltage value in the temperature rise estimation data; performing slope estimation calculation based on the at least two data pairs to determine a thermistor slope; evaluating the confidence level of the thermistor slope; and determining the thermal characteristic data based on the confidence level evaluation result and the thermistor slope.

[0058] The operating time point can be a discrete time node selected from the set of aligned waveforms within a specific linear operating region, satisfying preset electrical conditions; the thermistor slope can be the rate at which the junction voltage changes with the junction temperature within a specific linear operating region; and the confidence level can be an index that quantifies the reliability of the thermistor slope estimation result.

[0059] Optionally, in the current waveform of the aligned waveform packet, the positioning device operates at multiple time points within a preset fixed current linear region. This is to accurately measure V, a temperature-sensitive parameter. CE The measurement must be performed under specific and repeatable electrical conditions. In this embodiment, the measurement is performed when the device is turned on and operating at a fixed current I0 far from the saturation region. Specifically, by scanning the current waveform I(t), all values ​​satisfying |I(t)-I0|<ε are identified. I (where ε) I A point in time or a period of time (with a small tolerance) denoted as t I0 At the same time, it is necessary to verify that the device is indeed operating in the linear region at these time points to rule out the saturation effect on V. CE * Interference.

[0060] Using time points, at least two voltage-temperature operating points at different temperatures are determined from the junction voltage estimated by temperature rise and purification. The time point t determined in the previous step is then used... I0 We can see the temperature rise curve ΔT*(t) obtained in Example 2 and the purification junction voltage curve V obtained in Example 3. CE At time (t), the corresponding temperature rise and voltage values ​​are extracted. Due to the design of the micro-heating sequence, at different time points t... I0 The temperature rise of the devices varies. In this embodiment, two time points with significant temperature differences are selected, such as t_1 and t_2, to obtain two voltage-temperature operating points: (V_{CE1}^,T_1) and (V_{CE2}^,T_2), where V_{CE1}^ = V CE^(t_1), T1=Tambient+ΔThat(t1); V_{CE2}^=V CE ^*(t_2), T2=Tambient+ΔThat(t2), T_{ambient} is the ambient temperature.

[0061] Based on at least two voltage-temperature operating points, the thermistor slope of the device is determined through robust regression calculation. The thermistor slope αT is defined as the rate of change of junction voltage with junction temperature at a fixed current I_0. Using the two operating points mentioned above, it can be obtained through simple difference calculation: αT=(V_{CE1}^*-V_{CE2}^*) / (T_1-T_2). If multiple (more than two) operating points are located, robust linear regression (such as the Theil-Sen estimator) can be used to fit these points to obtain the slope αT. This reduces the impact of individual noise points on the results.

[0062] The uncertainty of the purified junction voltage and the range of the temperature rise estimate are then combined to comprehensively assess the confidence level associated with the thermistor slope. This step is crucial to the GCE method, as it determines the confidence level of the thermistor slope α. T A quantitative reliability assessment is provided. Confidence level α T_conf The assessment incorporates two main sources of uncertainty: one is the uncertainty of the purification junction voltage. VCE Secondly, the temperature rise estimate Δ*(t) obtained through inversion also has a confidence interval, i.e., its interval width. Using error propagation theory, the influence of these two sources of uncertainty on the calculation result of αT is synthesized, ultimately yielding a comprehensive confidence score α. Tconf For example, α Tconf It can be defined as 1 minus the normalized standard deviation of αT. A high α Tconf A confidence level (greater than 0.9) indicates that the obtained αT is highly reliable, while a low value indicates that the current αT result is less reliable due to measurement noise or identification instability. This confidence level will play an important role in the final joint decision-making process, used to weight the penalty term for thermal anomalies.

[0063] Step 205: Based on fingerprint data, thermal characteristic data, test condition data, and quality control mask, identify the hidden defects of the power module and obtain the target test results.

[0064] The target test results include, but are not limited to, hierarchical labels and traceability feature reports.

[0065] Optionally, the final decision-making step aims to integrate information from both morphological and thermal dimensions to make a reliable defect judgment. Traditional statistical methods are prone to misjudgment when faced with operating condition drift and complex failure modes. This embodiment introduces a physics-guided learning framework. Specifically, based on historical normal samples, a model is first learned that operates under a specific condition (cond) and is subject to a physical consistency residual (r). phys The guiding and constraining statistical model is specifically embodied in a modified covariance matrix σ. cond Then, for the sample to be tested, its features (f*) are calculated. plt The degree of deviation (i.e., statistical distance) from the physical guidance model, combined with its thermal characteristic parameters, such as Δα. T (t) and α T_conf Physical consistency residual (r) phys Construct a joint anomaly score D star Finally, the comprehensive score is compared with multiple preset thresholds, and graded labels such as normal, suspicious, and abnormal are output. A traceable report containing all key intermediate features and judgment criteria is generated to facilitate subsequent manual review and failure analysis.

[0066] It is worth noting that after performing the physically guided conditional learning and joint decision-making steps, the system can also include a closed-loop management module for report generation and baseline maintenance. For example, after generating hierarchical labels, the system automatically assembles a traceable feature report. This report not only includes the final normal / suspicious / abnormal labels and the joint anomaly score Dstar, but should also list in detail all key intermediate variables that contribute to the decision, such as the conditionally invariant fingerprint f*plt, the thermal slope offset ΔαT, and its confidence level alpha. Tconf Physical consistency residual r phys The report includes snapshots of the covariance matrix σ(cond) used, and provides a complete data traceability chain for subsequent manual review, failure analysis, or process improvement. Furthermore, the system has the capability for adaptive baseline updates. Specifically, for samples judged to be normal and stably released, their feature data can be used to incrementally update the sensitivity matrix B used for conditionally invariant mapping and the covariance matrix σcond used to calculate Mahalanobis distance. This rolling update mechanism allows the discrimination model to automatically adapt to slow, benign production line drift caused by factors such as production materials and equipment aging. Simultaneously, the system monitors the drift rate and data quality of the baseline model. When drastic changes or a continuous decline in data quality are detected, an alarm can be triggered, prompting engineers to investigate or initiate a baseline reset process.

[0067] For example, the baseline adaptive update capability may further include: periodically calculating and recording the change of the covariance matrix σ*(cond) over time (e.g., via the F-norm distance of the matrix), forming a covariance drift log (cov_drift_log). When the drift recorded in this log exceeds a preset threshold, the system can automatically trigger a baseline reset or issue an alert to the engineer.

[0068] In one embodiment, identifying latent defects in the power module based on fingerprint data, thermal characteristic data, test condition data, and quality control mask to obtain the target test result includes: performing statistical estimation based on fingerprint data, test condition data, and quality control mask to obtain an initial conditionalized covariance matrix; establishing structured constraints reflecting physical mechanism consistency based on the physical consistency residuals and weights in the fingerprint data; guiding the contraction of the initial conditionalized covariance matrix towards the structured constraints to generate a physically guided conditionalized covariance matrix; and identifying latent defects in the power module based on the conditionalized covariance matrix and thermal characteristic data to obtain the target test result.

[0069] Optionally, after obtaining the conditionally invariant platform micromorphological fingerprint f*plt of a large number of normal samples, a baseline covariance matrix needs to be estimated for a specific operating condition cond. Since undiscovered outliers may be mixed into the samples, directly calculating the sample covariance matrix is ​​not robust. Therefore, this embodiment applies a robust statistical estimation method, such as the minimum covariance determinant (MCD) algorithm. This algorithm iteratively finds a core, most compact subset of samples and calculates the covariance matrix based on this subset, thus effectively resisting the interference of outliers. Optionally, during the calculation process, QC_ can be used. mask or KLTR weight Weighting the samples further improves the robustness of the estimation. The final result is the initial conditional covariance matrix σ(cond).

[0070] For example, this embodiment introduces the obtained physical consistency information (r phys and w phys We can construct a structured constraint. The goal of this constraint is: if a certain morphological feature dimension is highly correlated with a certain physical law, then the variance of that feature dimension (i.e., the diagonal terms of the covariance matrix) should not be too large; if two morphological feature dimensions are physically independent, then the covariance between them (i.e., the cross terms of the covariance matrix) should tend to zero.

[0071] The process involves guiding the contraction of the initial conditional covariance matrix towards structured constraints. This is achieved by adjusting the diagonal and cross terms to integrate prior physical knowledge, thereby generating a physically guided conditional covariance matrix. This step is the core of physical guidance. Specifically, the guiding contraction of the initial conditional covariance matrix towards structured constraints involves calculating the platform's micromorphological fingerprint (f*plt) and the physical consistency residual (r). phys The correlation matrix between the two is used, and based on this correlation matrix and the physical consistency weight (w) phys This involves performing diagonal loading and cross-term decay on the initially conditional covariance matrix. For example, the i-th dimension of f*plt and r can be calculated. phys Corr(f*) of the j-th dimension plt,i ,r phys,j If the absolute value of this correlation coefficient is large, and the corresponding physical consistency weight w_{phys,j} is also high, then it indicates that feature f* plt,i The changes in δcond are closely related to a certain important physical law. Therefore, when shrinking δcond, the i-th diagonal term should be loaded (i.e., its value should be increased to reflect that its uncertainty should be subject to stronger constraints from physical laws), and the cross term containing the i-th dimension should be decayed more significantly (making it closer to zero to reduce spurious correlations). The shrinkage process can be achieved using a linear shrinkage framework of the Ledoit-Wolf type, but its shrinkage target is no longer a simple diagonal matrix, but a structured target matrix derived from physical consistency information. After this step, the finally generated physical-guided conditional covariance matrix σ*cond retains the main variation patterns of the data and incorporates physical priors, making it more robust and interpretable.

[0072] It is worth noting that the guided contraction of the initial conditional covariance matrix towards structured constraints can be implemented as follows: First, calculate the morphological fingerprint and the physical residual r. phys The Pearson correlation matrix 'Corr' between them, and then based on Corr and weights wp hys Design a structured target matrix T, and finally calculate the covariance matrix using the linear shrinkage formula Σ*=(1-α)Σ+αT, where the shrinkage strength α can be determined through cross-validation.

[0073] For example, the final joint anomaly score D_{star} needs to be compared with a preset threshold Th suspect and Th fail The data is compared to derive a tiered label. The setting of these thresholds has a decisive impact on the system's detection rate and false positive rate. This embodiment provides an adaptive determination method based on historical data statistics:

[0074] Baseline Data Collection: First, during stable production line operation, a large number (e.g., thousands to tens of thousands) of confirmed defect-free normal samples are collected, and their respective joint anomaly scores (Dstar) are calculated. Probability Distribution Fitting: Statistical analysis is performed on this large batch of Dstar scores, their histograms are plotted, and a suitable probability density function (PDF) is fitted, such as a Weiber or gamma distribution, which are commonly used to describe statistics related to failures or anomalies. Percentile-Based Threshold Setting: Based on the fitted probability distribution, the threshold is determined using the inverse function of its cumulative distribution function (CDF). suspect This can be set to the 95th or 98th percentile of the distribution. For example, setting it to the 95th percentile means that in normal samples, 5% of the samples will be marked as suspicious and require secondary inspection, which is a relatively sensitive strategy; Th fail It can be set to a very high quantile, such as 99.9D_{star} score, which is extremely abnormal and far exceeds almost all normal samples, before it will be directly judged as abnormal, thereby controlling the probability of normal samples being misjudged as abnormal (i.e., false positive rate) at an extremely low level (such as 0.1% or 0.01%).

[0075] Among them, the hierarchical threshold Th suspect and Th fail The method for determining it can be: based on the statistical distribution of D_{star} scores of a large number of normal samples, take the 95th percentile and 99.9th percentile as thresholds respectively.

[0076] Optionally, after obtaining the platform micromorphological fingerprint under unchanged conditions, the method further includes: performing a self-check on the correlation between the correction residual of C-INV (i.e., BΔcond) and the operating condition drift Δcond; if the correlation is too high, generating a numerical stability flag num_stability_flag to indicate the potential risk of the current C-INV result. Furthermore, after determining a set of corresponding physical consistency weights for the physical consistency residuals, the method further includes: comparing the weighted physical consistency residuals with a preset threshold; if the threshold is exceeded, generating a phys_alert (physical alarm) signal to trigger a specific retest or alarm process.

[0077] It is worth noting that, assuming a power module sample under test, the following set of input data is obtained through the aforementioned steps: Conditionally invariant platform micromorphological fingerprint (assumed to be 3D): f*plt=[0.2,-0.1,0.3]T. Mean vector of the normal sample group: μplt=[0.0,0.0,0.0]T. Physically guided conditional covariance matrix: Σ (cond)=[[1.0,0.2,0.1],[0.2,1.2,0.3],[0.1,0.3,1.5]]. Thermal characteristic parameters: thermistor slope offset ΔαT=-0.08, with confidence level α. Tconf =0.95. Physical consistency residuals and weights (assuming 2D): rphys=[0.15,-0.2]T, wphys=[0.9,0.8]T. Anchor reliability weights: KLTRweight=0.9. Fusion hyperparameters: w1=1.0, w2=20.0, w3=10.0. The calculation process of the joint anomaly score D_{star} is as follows: 1. Calculate the statistical Mahalanobis distance (MD). First, calculate the inverse matrix σcond of the covariance matrix. -1 ≈[[1.04,-0.18,-0.03],[-0.18,0.95,-0.20],[-0.03,-0.20,0.77]]. Then, calculate the square of the Mahalanobis distance: (MD )2=(f*plt-μplt)T(Σ (cond))-1(f~plt-μplt)(MD ) 2 =0.369; 2. Construct physical penalty terms (Pthermal, Pphysical); Thermal anomaly penalty term: Pthermal=αT_conf∣ΔαT∣=0.95∣-0.08∣=0.076; Physical consistency penalty term: Pphysical=∣∣wphys⊙rphys∣∣1=∣∣[0.9×0.15,0.8×(-0.2)]∣∣1=∣∣[0.135,-0.16]∣∣1=∣0.135∣+∣-0.16∣=0.295; 3. Fusion calculation of joint anomaly score (Dstar), Dstar=KLTRweight×(w1×MD +w2×Pthermal+w3×Pphysical)Dstar = 0.9×(1.0×0.369 + 20.0×0.076 + 10.0×0.295)Dstar = 0.9×(0.369 + 1.52 + 2.95) = 0.9×4.839 ≈ 4.355; 4. Hierarchical decision-making. Assume the preset hierarchical thresholds are: Thsuspect = 4.0 and Thfail = 8.0. Since the calculated combined anomaly score Dstar ≈ 4.355, this value satisfies the condition Thsuspect ≤ Dstar < Thfail (i.e., 4.0 ≤ 4.355 < 8.0). Therefore, the sample of the power module to be tested is finally determined to be at the suspect level, and it is recommended to conduct further retesting or manual analysis. Through this specific numerical case, it clearly demonstrates how information from multiple dimensions such as platform micro-morphology, thermal path characteristics, and physical consistency can be finally fused into a single, interpretable, and quantitative index for defect identification through a structured framework that takes into account various uncertainties and reliabilities.

[0078] In the above power module testing method, through waveform alignment related to working conditions, platform fingerprint construction, pulse excitation assisted thermal path evaluation, and multi-dimensional data fusion decision-making, combined with the effective screening of the quality control mask, it is possible to accurately identify the hidden defects of the micro-structure and thermal path of the power module within a short cycle, thereby improving the accuracy of identifying hidden defects such as thermal paths and electrical micro-morphologies under short test cycles.

[0079] In an exemplary embodiment, as Figure 3 shown, construct a platform fingerprint based on the aligned waveform set, test working condition data, and quality control mask to obtain fingerprint data, including:

[0080] Step 301, perform causal signal processing on the aligned waveform set to determine the starting anchor point of the gate voltage platform;

[0081] Step 302, based on the starting anchor point, segment the platform segment from the aligned waveform set;

[0082] Step 303, according to the test working condition data, apply physical conditional time transmission to the platform segment to generate a time reparameterization mapping;

[0083] Step 304, generate fingerprint data based on the platform segment and the time reparameterization mapping.

[0084] The starting anchor point can be a precise time point at which the gate voltage plateau begins, located through causal signal processing; the plateau segment can be a waveform segment that contains only the gate voltage plateau region, segmented from the aligned waveform set based on the starting anchor point; and the time reparameterization mapping can be a monotonic function generated by PTT that maps the original time axis of the plateau segment to the normalized time axis.

[0085] Optionally, this embodiment illustrates how to achieve precise alignment of the gate voltage plateau waveform across operating conditions using causal time geometry (C-KLTR+PTT) technology. Specifically, it includes applying causal signal processing to the aligned waveform packet to locate the starting anchor point of the gate voltage plateau online. For online, real-time defect detection, the time alignment algorithm must be causal, meaning that the calculation at any time point t can only depend on data at and before time t. Traditional alignment methods based on complete rising or falling edge knees violate causality and cannot be applied online. This embodiment employs a causal knee location and time reparameterization (C-KLTR) technique. Specifically, a causal infinite impulse response (IIR) differentiator is applied to the gate voltage V in the aligned waveform packet. GS (t) is processed to locate the starting anchor point t_ of the gate voltage plateau online without delay. kcausal The IIR differentiator is designed to operate at V... GS A significant peak is generated at the inflection point where the waveform slope changes significantly (i.e., entering the Miller plateau). Combined with a simple plateau voltage threshold criterion, the starting anchor point of the plateau can be located in real time without delay.

[0086] Using the initial anchor point as a reference, offline refinement is performed in conjunction with the complete waveform to accurately segment the platform segment from the aligned waveform packet. The online anchor point t_ kcausal While ensuring real-time performance, its accuracy may be affected by factors such as noise. Therefore, in subsequent offline processing stages (e.g., when performing unified analysis on a batch of data), more comprehensive waveform information can be utilized for refinement, demonstrating the advantages of online-offline dual-mode collaboration. Specifically, taking the t_ obtained online... kcausal As a high-precision search starting point or initial value, within a small time window, it combines complete waveform information (e.g., waveform information after the platform ends can be used non-causally, or collector-emitter voltage V). CE The derivative dV CE The system performs global knee point playback refinement (using the knee point information of / dt) to obtain a more accurate platform start and end point, and segments the platform from the original waveform.

[0087] Based on the operating conditions (i.e., test condition data), a physically conditional time transfer (PTT) is applied to the platform segment to generate a time reparameterization mapping to align with the platform's time-domain scaling caused by different operating conditions. Different operating conditions (such as different bus voltages Vdc or gate resistances Rg) cause variations in the duration of the Miller platform, making direct comparison of the platform morphology under different conditions difficult. To address this issue, this embodiment applies a physically conditional time transfer (PTT) to the platform segment. Essentially, based on the operating condition parameter cond, a monotonic spline function is constructed to nonlinearly map the platform's own time axis (τ) to a normalized time axis (τ). A time-reparameterized mapping is generated. The shape of this spline function is guided by the physical model and adjusted according to the current operating condition (cond) to ensure that the microscopic morphological features inherent in the waveform are not damaged or distorted while stretching or compressing the time axis to align the macroscopic duration. After PTT processing, all platform segments under different operating conditions are mapped to a unified, standardized time domain τ^*, achieving time-domain alignment across operating conditions. Finally, the platform segments and the time-reparameterized mapping are used together for subsequent fingerprint data generation. The precisely segmented and time-stretched aligned platform segments, along with the mapping function (τ→τ*) describing how time stretching is performed, are used together for subsequent fingerprint data generation.

[0088] It is worth noting that the application of causal signal processing and the online location of the starting anchor point of the gate voltage plateau are specifically implemented by using a causal IIR differentiator to calculate V. GS The derivative of (t) is used to determine the initial anchor point t through a joint criterion of the derivative peak value and the voltage amplitude threshold. k_causal Offline refinement is performed by combining the complete waveform, and the specific implementation method is as follows: using t k_causal To find the center, within a small time window, the knee position information of the dVCE / dt waveform is used to perform higher precision correction on the platform's start and end points.

[0089] In one embodiment, generating fingerprint data based on platform segments and time-reparameterized mapping includes: extracting local parameters from the platform segments to determine local normalization factors; generating second-order curvature based on the platform segments and a quality control mask; correcting the second-order curvature according to the time-reparameterized mapping to obtain corrected second-order curvature; normalizing the corrected second-order curvature according to the local normalization factors to obtain morphological features, and constructing a platform fingerprint based on the morphological features to obtain fingerprint data.

[0090] Among them, the local normalization factor can be a normalization factor that reflects the scale of the waveform itself; the second curvature can be a physical quantity obtained by performing second derivative operations on the voltage signal of the platform segment.

[0091] Optionally, this embodiment illustrates how, after obtaining the aligned platform segment, robust and comparable micromorphological features can be extracted using a combined operator integrating VDE2, NAR, and BLC, combined with local normalization. Specifically, this includes: extracting local parameters from the platform segment to determine a local normalization factor for scale normalization. Before extracting morphological features, a normalization factor reflecting the waveform's own scale needs to be determined to eliminate the overall offset of the platform voltage amplitude or slope caused by individual device differences or minor fluctuations in test parameters. Specifically, the step of determining a local normalization factor for scale normalization includes: performing window-constrained gate charge integration on the platform segment to obtain Qg_win; and calculating its average platform slope to obtain dV. GS / dt _win The local normalization factor, denoted as S. local It is based on Qg _win With dV GS / dt _win Determined jointly. For example, S local It can be a vector containing these two values, or a scalar composed of these two values ​​through a predefined function. local This provides a benchmark for subsequent feature normalization.

[0092] Based on the platform segment and quality control mask, a robust second-order curvature segment is generated using a combined operator integrating variational second-order curvature estimation, noise adaptive regularization (NAR), and bandwidth-locked constraint (BLC). The second derivative (or curvature) is extremely sensitive to microscopic changes in the platform's topography, but direct numerical differentiation calculations greatly amplify noise, leading to unstable results. To address this issue, this embodiment employs a combined operator. Specifically, variational second-order curvature estimation (VDE2) is applied by constructing an energy functional and minimizing it, effectively avoiding the instability issues caused by noise-sensitive direct numerical differentiation methods when estimating second-order curvature. This energy functional typically includes a data fidelity term and a regularization term (smoothing term). For example, for the voltage signal V of the platform segment... GS (τ), whose second curvature V GS ''(τ) can be minimized by the following functional E(V) GS We get: E(V) GS '')=∫(G*V GS ''-V GS ) 2 dτ+λ∫(V GS '') 2dτ, where G is the quadratic integral operator and λ is the regularization coefficient. Noise adaptive regularization (NAR) is also integrated into the variational framework described above. Specifically, the value of the regularization coefficient λ is not fixed but adaptively adjusted based on the signal-to-noise ratio of the input signal (which can be estimated from the QC_mask or the local noise level). For samples with high noise, a larger λ is chosen to enhance smoothing, while λ is decreased to preserve more detail.

[0093] Based on the time-reparameterized mapping, bandwidth-locked constraints are applied to the robust second-order curvature segment to correct spectral distortion, followed by normalization using a local normalization factor, thus obtaining a set of corrected topographic features. Physically conditional time transfer (PTT), while aligning platform durations, may unintentionally compress or stretch its spectrum, leading to topographic feature distortion. Bandwidth-locked constraints (BLC) are precisely designed to correct this distortion. Specifically, based on the PTT time mapping function (τ→τ... The corrected coefficients for the spectral energy are calculated and applied to the robust second-order curvature V obtained in the previous step. GS On ''(τ), ensure energy conservation within the critical frequency band. Divide the BLC-corrected second-order curvature by the local normalization factor S obtained in the first sub-step. local Scale normalization is then performed. This yields a set of corrected morphological features f. plt_raw_corrected This morphological feature is a component of fingerprint data. This set of features is robust and eliminates the effects of temporal scaling and individual scale differences, thus exhibiting good comparability.

[0094] For example, the generation of fingerprint data also includes a conditionally invariant mapping step, which is performed after obtaining the corrected morphology features. Specifically, based on the corrected morphology features of a set of reference samples and their corresponding operating conditions, a sensitivity matrix is ​​learned and established to quantify the degree of response of the morphology features to changes in operating conditions. In batch production testing, test operating parameters (such as bus voltage Vdc, gate resistance Rg, etc.) will inevitably have slight inter-batch or intra-batch fluctuations, which will systematically affect the morphology features f. plt_raw_corrected To eliminate this effect, it is first necessary to quantify its extent. Specifically, a subset of devices, referred to as the gold sample or reference sample, can be selected and tested under multiple different, precisely controlled operating conditions to obtain their respective morphological characteristics f_{plt\_raw\_corrected} and corresponding operating condition vector cond. Based on this data, a sensitivity matrix B can be learned through linear regression or a more complex machine learning model. Each element B of this matrix B... ij This represents the sensitivity of the i-th morphological feature to changes in the j-th operating condition parameter.

[0095] For each test sample, the drift of its current operating condition relative to the baseline is measured. Using the sensitivity matrix and this drift, the systematic bias introduced by the operating condition drift is subtracted from the corrected morphological features of the test sample. For any test sample, the actual operating condition *condcurrent* during the test is first recorded. Then, it is compared with a preset baseline operating condition *cond_{baseline}* to obtain the drift vector *Deltacond=cond_{current}-cond_{baseline}*. Using the sensitivity matrix *B* learned in the previous step, the systematic feature bias introduced by the operating condition drift can be estimated as *BΔcond*. Subtracting this estimated bias from the original morphological features yields the corrected features.

[0096] This yields the platform micromorphological index (f*) under unchanged conditions. plt This is then treated as another component of the fingerprint data. The final feature obtained after the above subtraction operation is the platform micromorphological fingerprint f* with unchanged conditions. plt The calculation formula is: f*plt=f plt_raw_corrected -B*Δcond. Theoretically, f*plt is independent of the minute drift of the operating conditions and only reflects the physical properties of the device itself. Therefore, it has extremely high comparability and stability among samples from different batches and at different test times, providing an ideal input for establishing a unified and robust discrimination model.

[0097] In applying Variational Second-Order Curvature Estimation (VDE2), a regularization parameter λ is needed to balance data fidelity and result smoothness. A fixed λ cannot adapt to varying noise levels across different samples. Therefore, this embodiment employs an adaptive adjustment method based on Signal-to-Noise Ratio (SNR). The adjustment formula can be specifically set as: λ = λ0(1 + 1 / SNR)γ. Where: λ is the regularization parameter ultimately used in the VDE2 energy functional; λ0 is a basic regularization coefficient, the value of which can be determined through experimental tuning on a small number of representative samples; SNR is the signal-to-noise ratio of the current platform segment to be processed. SNR can be estimated in various ways, for example, by first performing a preliminary smoothing filter on the signal, and then calculating the ratio of the original signal power to the residual (original signal - smoothed signal) power; γ is a control exponent, typically between 1 and 2, used to adjust the sensitivity of λ to changes in SNR. The adjustment effect of this formula is as follows: when the signal's SNR is high (i.e., the noise is very low), the (1+1 / SNR) term is close to 1, and λ is also close to its fundamental value λ_0, thus preserving more waveform details. When the signal's SNR is low (i.e., the noise is very high), the (1+1 / SNR) term will be significantly greater than 1, causing the value of λ to increase, thereby imposing a stronger smoothing constraint to effectively suppress the interference of noise on the second-order curvature estimation. It is also possible to construct and minimize an energy functional E(VGS′′)=∫(G... The specific algorithm for applying bandwidth locking constraints can be as follows: calculate the Jacobian determinant based on the PTT mapping function, obtain the spectral energy correction factor, perform multiplicative correction on the spectrum of second-order curvature in the Fourier domain, and then return to the time domain through inverse Fourier transform.

[0098] Optionally, the step of locating the starting anchor point of the gate voltage platform online further includes: for each located starting anchor point, examining the distortion and repeatability of its signal characteristics to generate an anchor point reliability score that quantifies its positioning quality. The starting anchor point t of the platform was located using methods such as a causal IIR differentiator. k_causal However, due to signal noise, interference, or abnormal behavior of the device itself, signal distortion (such as ringing or overshoot) or significant position drift (poor repeatability) may occur near the anchor point of some samples. This embodiment quantifies and evaluates these situations. Specifically, it calculates indicators such as the signal-to-noise ratio of the waveform segment near the anchor point, the difference from the standard template, or the standard deviation of the anchor point position in continuous pulses. These indicators are then combined into a single anchor reliability score, `anchor_score`, ranging from [0,1], using a preset evaluation function. The closer the score is to 1, the higher the positioning quality and the more reliable the alignment result of the sample.

[0099] The anchor reliability score is converted into an anchor reliability weight (KLTR) using a preset function. weight This weight is passed to subsequent covariance estimation and anomaly scoring processes. After obtaining the anchor_score, it needs to be transformed into a weight that can be directly used for weighted calculation. This preset function can be a simple linear mapping or a non-linear sigmoid function, such as KLTR. weight =1 / (1+exp(-k(anchor score -x0))), where k and x0 are parameters controlling the shape of the function. This function's purpose is to optimize the KLTR of samples with high anchor scores. weight Close to 1; for samples with very low anchor_score, their KLTR is... weight Approaching 0, thus achieving a smooth weight transition. Most importantly, this calculated KLTR... weight It is not merely the end point of this step; it will be passed as sample-bound metadata to all subsequent batch learning and joint decision-making steps. For example, it will be used in the calculation of the weighted covariance matrix and the final aggregation of joint outliers, thus ensuring that samples with poor alignment quality and unreliable from the source have a lower impact on the final statistical model and discrimination results.

[0100] The anchor_score is calculated as follows: the anchor reliability score can be calculated using the formula: anchor_score = w_rep * S_rep + w_dist * S_dist. Where S... rep The repeatability score is calculated based on the standard deviation of the anchor point's position timestamps in a continuous pulse; S_{dist} is the distortion score calculated based on the cross-correlation coefficient between the waveform near the anchor point and the standard template. The specific mapping function for KLTR_weight can be: a function that converts the anchor point reliability score into an anchor point reliability weight via a preset function, specifically an S-shaped function.

[0101] KLTRweight=1 / (1+exp(-10(anchor_score-0.7)))

[0102] It is worth noting that the steps of constructing the platform micromorphological fingerprint data also include the following sub-steps to generate the physical consistency portion of the fingerprint data: In the space composed of multiple interrelated physical features such as gate charge, platform time, and platform slope, a set of preset physical mechanism edges are established, each mechanism edge representing an expected physical constraint relationship. Based on semiconductor device physics, there are inherent and definite physical connections between some macroscopic features of the platform. This embodiment abstracts these connections as physical mechanism edges. For example: Mechanism edge 1 (Qg–t_plt): the platform duration t plt It should be related to the gate charge Q flowing through the platform. _g_plt (This can be obtained by integrating the gate current) It is approximately proportional, because Q g_plt ≈Ig×t plt The gate current Ig is approximately constant during the plateau; Mechanism 2 (Qg – ΔV_plt): The gate charge Q_{g\_plt} of the plateau is also related to the voltage width DeltaV of the plateau. plt (Voltage difference at the start and end of the platform) related; Mechanism side 3 (t_plt–slope): Platform duration t plt With the average slope of the platform plt They typically exhibit an inverse correlation. These mechanistic edges together form a Physical Consistency Graph (PC-Graph), which defines the macroscopic physical laws that normal devices should satisfy.

[0103] For each test sample, the deviation of its various physical characteristics from a set of physical mechanism edges is calculated, thus quantifying it as physical consistency residuals. For each test sample, the aforementioned macroscopic physical characteristics (such as t) are first extracted from its waveform. plt ,Qg _plt ,slope plt (etc.). Then, these feature values ​​are substituted into the preset mechanistic edge relationships to calculate the degree of deviation from the expectation. For example, for mechanistic edge 1, t can be obtained first through regression of a set of normal samples. plt With Q _g_plt The standard linear relationship is established, and then the distance from the current sample point to this line is calculated. This distance is defined as a physical consistency residual. This calculation is performed for each mechanism edge, ultimately resulting in a vector composed of multiple residual values, i.e., the physical consistency residual r. phys r physEach element corresponds to the degree of violation of a physical law. Based on the residual distribution and the quality control mask, a set of corresponding physical consistency weights is determined for this physical consistency residual. Different mechanism edges may have different constraint strengths and reliability; at the same time, if the original waveform quality of a sample is not high (indicated by QC_mask), then the reliability of its calculated residual should also be reduced accordingly. Therefore, it is necessary to define a residual vector r. phys Each element is assigned a corresponding weight. This weight is the physical consistency weight w. phys The value of r can be determined by combining the statistical distribution of the residual in a large number of normal samples (e.g., the more concentrated the distribution, the more stable the pattern, and the higher the weight) with the QC_mask value of the current sample. phys and w phys Vectors will serve as important prior knowledge, guiding the statistical model to better align with physical reality in subsequent covariance learning and joint decision-making steps.

[0104] For example, this embodiment illustrates the final joint anomaly score calculation process. It integrates the outputs of all previous embodiments (including morphological fingerprints, thermal characteristic parameters, physical consistency residuals, physically guided covariance matrices, and sample weights) to obtain a comprehensive evaluation score, and clearly demonstrates α. Tconf and KLTR weight The final application. In this embodiment, the subsequent joint anomaly score calculation specifically includes: using a physically guided conditional covariance matrix to measure the deviation of the platform micromorphological fingerprint in the fingerprint data from the normal sample group, in order to calculate a statistical Mahalanobis distance. Mahalanobis distance (MD) is an effective distance metric that considers the correlation between features. This step uses the obtained, more robust physically guided covariance matrix Sigma^(cond) to calculate the degree of deviation of the morphological fingerprint tilde{f}{plt} of the test sample from the center mu{plt} of the normal sample group. The calculation formula is:

[0105] MD^=\sqrt{(\tilde{f}{plt}-\mu{plt})^T(\Sigma^(cond))^{-1}(\tilde{f}{plt}-\mu{plt})}. The larger the value of MD^, the more the plateau micromorphology of the sample deviates from that of the normal population.

[0106] Meanwhile, based on the thermal characteristic parameters and the physical consistency residuals in the fingerprint data, one or more physical penalty terms for penalizing the inconsistency between the thermal path and the physical model are constructed. Relying solely on the statistical distance MD^* may not be able to effectively detect certain specific defects directly related to the physical mechanism. Therefore, physical penalty terms are additionally constructed in this step. Specifically, based on the thermal characteristic parameters, a thermal anomaly penalty term weighted by its confidence level (α Tconf ), and combined with the physical consistency residuals, together constitute one or more physical penalty terms. For example, the thermal anomaly penalty term can be defined as: Pthermal = αT_conf∣ΔαT∣, where \DeltaαT is the offset of the thermal sensitivity slope relative to the reference. The meaning of this formula is that only when the offset of the thermal sensitivity slope is large and the confidence level of its measurement is also high, a large penalty is given. This effectively avoids misjudgment caused by inaccurate measurement. The physical consistency penalty term can be defined as: P_{physical}=||w phys \odotr phys ||_1, that is, the weighted L1 norm of r phys , where \odot represents element-wise multiplication.

[0107] Fuse the statistical Mahalanobis distance and the physical penalty terms to generate a comprehensive joint anomaly score. This step performs the final fusion of all the above information. Specifically, the generated anchor reliability weight (KLTRweight) is used to weight and aggregate the statistical Mahalanobis distance and the physical penalty terms. An exemplary fusion formula is: Dstar = KLTRweight(w1×MD +w2×Pthermal+w3×Pphysical), where w1, w2, and w3 are preset hyperparameters for balancing the importance of different types of anomalies. By introducing KLTRweight, it is ensured that the influence of samples with poor alignment quality on the final score is suppressed, thereby improving the robustness of the entire decision-making system.

[0108] Compare the joint anomaly score with a preset classification threshold to determine the classification label. Finally, compare the calculated joint anomaly score D_{star} with a set of preset thresholds (such as Th suspect and Th fail ). If Dstar < Thsuspect, it is determined to be normal; if Thsuspect ≤ Dstar < Thfail, it is determined to be suspicious; if Dstar ≥ Thfail, it is determined to be abnormal. Thus, the recognition of the hidden defect mode of a single power module is completed.

[0109] In this embodiment, online real-time accurate positioning of the starting anchor point of the gate voltage platform is achieved through causal signal processing. Pure platform segments are segmented based on the anchor point, and then the time-domain scaling differences across operating conditions are eliminated through physical conditional time transmission. Finally, fingerprint data that only reflects the inherent physical properties of the device and has comparability and high stability across operating conditions is generated. This not only meets the short cycle requirements of batch testing, but also provides reliable core feature support for the identification of latent defects.

[0110] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0111] Based on the same inventive concept, this application also provides a power module testing apparatus for implementing the power module testing method described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more power module testing apparatus embodiments provided below can be found in the limitations of the power module testing method described above, and will not be repeated here.

[0112] In one exemplary embodiment, such as Figure 4 As shown, a power module testing device is provided, including: a processing module 41, a construction module 42, and an identification module 43, wherein:

[0113] Processing module 41 is used to determine the pulse sequence label and quality control mask associated with the aligned waveform set. The pulse sequence label is used to record the key characteristic parameters of the active pulse excitation applied during the power module test. The quality control mask is used to characterize the quality level of each waveform in the aligned waveform set.

[0114] Module 42 is used to construct the platform fingerprint based on the aligned waveform set, test condition data and quality control mask, and obtain fingerprint data.

[0115] The processing module 41 is also used to perform thermal path evaluation based on the aligned waveform set, test condition data and pulse sequence labeling to obtain thermal characteristic data.

[0116] The identification module 43 is used to identify hidden defects in the power module based on fingerprint data, thermal characteristic data, test condition data and quality control mask, and obtain the target test results.

[0117] In one embodiment, the processing module 41 is further configured to: generate a target power sequence based on the annotation parameters associated with the pulse sequence annotation; synchronize the target power sequence with the electrical response data in the aligned waveform set to obtain target associated data, and determine temperature rise estimation data based on the target power sequence and the electrical response data corresponding to the target power sequence; and perform slope estimation on the electrical response data and temperature rise estimation data within a linear working range determined based on the test condition data to generate thermal characteristic data.

[0118] In one embodiment, the processing module 41 is further configured to: select multiple operating time points from a set of aligned waveforms operating in a linear operating region determined by the test condition data; extract at least two data pairs from the electrical response data and temperature rise estimation data based on the multiple operating time points, wherein the data pairs include the voltage value in the electrical response data and the temperature corresponding to the voltage value in the temperature rise estimation data; perform slope estimation calculation based on the at least two data pairs to determine the thermistor slope; evaluate the confidence level of the thermistor slope, and determine the thermal characteristic data based on the confidence level evaluation result and the thermistor slope.

[0119] In one embodiment, the construction module 42 is further configured to: perform causal signal processing on the aligned waveform set to determine the starting anchor point of the gate voltage platform; segment the platform segment from the aligned waveform set based on the starting anchor point; apply physical conditional time transmission to the platform segment according to the test condition data to generate a time reparameterization mapping; and generate fingerprint data based on the platform segment and the time reparameterization mapping.

[0120] In one embodiment, the construction module 42 is further configured to: extract local parameters from the platform segment and determine the local normalization factor; generate a second-order curvature based on the platform segment and the quality control mask; correct the second-order curvature according to the time reparameterization mapping to obtain the corrected second-order curvature; normalize the corrected second-order curvature according to the local normalization factor to obtain morphological features, and construct a platform fingerprint based on the morphological features to obtain fingerprint data.

[0121] In one embodiment, the identification module 43 is further configured to: perform statistical estimation based on fingerprint data, test condition data, and quality control mask to obtain an initial conditional covariance matrix; establish structured constraints reflecting physical mechanism consistency based on the physical consistency residuals and weights in the fingerprint data; perform guided contraction of the initial conditional covariance matrix to the structured constraints to generate a physical-guided conditional covariance matrix; and identify latent defects of the power module based on the conditional covariance matrix and thermal characteristic data to obtain the target test results.

[0122] Each module in the aforementioned power module testing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0123] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores target test results. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a power module testing method.

[0124] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 6As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When the computer program is executed by the processor, it implements a power module testing method. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0125] Those skilled in the art will understand that Figure 5 and Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0126] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0127] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0128] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0129] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0130] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0131] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0132] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A power module testing method, characterized in that, The method includes: Obtain the original waveform set generated by the test power module and the test condition data of the power module, and obtain the aligned waveform set based on the test condition data and the original waveform set; Determine the pulse sequence label and quality control mask associated with the aligned waveform set. The pulse sequence label is used to record the key characteristic parameters of the active pulse excitation applied during the power module test, and the quality control mask is used to characterize the quality level of each waveform in the aligned waveform set. The platform fingerprint is constructed based on the aligned waveform set, the test condition data, and the quality control mask to obtain fingerprint data; Thermal path evaluation is performed based on the aligned waveform set, the test condition data, and the pulse sequence labeling to obtain thermal characteristic data; Based on the fingerprint data, thermal characteristic data, test condition data, and quality control mask, the hidden defects of the power module are identified, and the target test results are obtained.

2. The method according to claim 1, characterized in that, The step of evaluating the thermal path based on the aligned waveform set, the test condition data, and the pulse sequence annotation to obtain thermal characteristic data includes: Generate the target power sequence based on the annotation parameters associated with the pulse sequence annotation; The target power sequence is synchronized with the electrical response data in the aligned waveform set to obtain target associated data, and temperature rise estimation data is determined based on the target power sequence and the electrical response data corresponding to the target power sequence. Within the linear operating range determined based on the test condition data, the slope of the electrical response data and the temperature rise estimation data is estimated to generate thermal characteristic data.

3. The method according to claim 2, characterized in that, The step of estimating the slope of the electrical response data and the temperature rise estimation data within the linear operating range determined based on the test condition data to generate thermal characteristic data includes: Multiple working time points are selected from the set of aligned waveforms operating in the linear working area determined by the test condition data; Based on the multiple working time points, at least two data pairs are extracted from the electrical response data and temperature rise estimation data. The data pairs include the voltage value in the electrical response data and the temperature corresponding to the voltage value in the temperature rise estimation data. Based on the at least two data pairs, slope estimation calculations are performed to determine the thermal slope. Assess the confidence level of the thermodynamic slope and determine the thermal characteristic data based on the confidence level assessment result and the thermodynamic slope.

4. The method according to claim 1, characterized in that, The step of constructing a platform fingerprint based on the aligned waveform set, the test condition data, and the quality control mask to obtain fingerprint data includes: The aligned waveform set is subjected to causal signal processing to determine the starting anchor point of the gate voltage plateau; Using the starting anchor point as a reference, platform segments are segmented from the aligned waveform set; Based on the test condition data, physical conditional time transmission is applied to the platform segment to generate a time reparameterization mapping; The fingerprint data is generated based on the platform segment and the time reparameterization mapping.

5. The method according to claim 4, characterized in that, The step of generating fingerprint data based on the platform segment and the time reparameterization mapping includes: Local parameters are extracted from the platform segment to determine the local normalization factor; Based on the platform segment and the quality control mask, a second-order curvature is generated; The second-order curvature is corrected according to the time reparameterization mapping to obtain the corrected second-order curvature; The corrected second-order curvature is normalized according to the local normalization factor to obtain morphological features, and a platform fingerprint is constructed based on the morphological features to obtain fingerprint data.

6. The method according to any one of claims 1-5, characterized in that, The step of identifying latent defects in the power module based on the fingerprint data, thermal characteristic data, test condition data, and quality control mask to obtain the target test result includes: Based on the fingerprint data, the test condition data, and the quality control mask, statistical estimation is performed to obtain the initial conditionalized covariance matrix. Based on the physical consistency residuals and weights in the fingerprint data, a structured constraint reflecting the consistency of the physical mechanism is established. The initial conditional covariance matrix is ​​guided to shrink towards the structured constraints to generate a physically guided conditional covariance matrix. Based on the conditional covariance matrix and the thermal characteristic data, the hidden defects of the power module are identified, and the target test results are obtained.

7. A power module testing device, characterized in that, The device includes: The acquisition module is used to acquire the original waveform set generated by the test power module and the test condition data of the power module, and to obtain the aligned waveform set based on the test condition data and the original waveform set. The processing module is used to determine the pulse sequence label and quality control mask associated with the aligned waveform set. The pulse sequence label is used to record the key characteristic parameters of the active pulse excitation applied during the power module test, and the quality control mask is used to characterize the quality level of each waveform in the aligned waveform set. The construction module is used to construct a platform fingerprint based on the aligned waveform set, the test condition data, and the quality control mask to obtain fingerprint data; The processing module is further configured to perform thermal path evaluation based on the aligned waveform set, the test condition data, and the pulse sequence labeling to obtain thermal characteristic data; The identification module is used to identify the hidden defects of the power module based on the fingerprint data, the thermal characteristic data, the test condition data, and the quality control mask, and to obtain the target test result.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images