A method and system for fatigue identification of IGBT solder layers based on voltage texture

Abstract

This application provides a voltage texture-based method and system for IGBT solder layer fatigue identification, relating to the field of power electronic device health monitoring technology. The method includes: acquiring voltage texture signals of IGBT switching transients over multiple switching cycles using the thermal time constant of the IGBT solder layer as the acquisition benchmark, forming a voltage texture sequence; extracting multiple candidate features from the voltage texture sequence, filtering and combining the candidate features based on the monotonicity index of each candidate feature in the cumulative power cycle count, and forming a solder layer degradation fingerprint vector; inputting the solder layer degradation fingerprint vector into a solder layer fatigue identification model to obtain preliminary identification results; using a deviation assessment model to evaluate the degree of deviation of the solder layer degradation fingerprint vector from the health state distribution, and outputting a confidence score; and correcting the preliminary identification results based on the confidence score to obtain corrected solder layer fatigue identification results. This improves the accuracy, reliability, and physical consistency of IGBT solder layer fatigue identification.

Description

Technical Field

[0001] This invention relates to the field of power electronic device health monitoring technology, and in particular to a method and system for identifying IGBT solder layer fatigue based on voltage texture. Background Technology

[0002] During the power cycling aging process of IGBT modules, solder fatigue is one of the most common failure modes, leading to increased thermal resistance and decreased heat dissipation performance. Traditional solder fatigue detection methods mostly rely on offline physical testing, such as metallographic section analysis or thermal resistance testing. These methods have disadvantages such as being destructive, costly, and unable to be implemented online.

[0003] In recent years, online monitoring methods based on electrical parameters have gradually attracted attention. However, existing methods often only extract simple time-domain features, ignoring the parasitic parameter oscillation information contained in the waveform, and do not consider the monotonic degradation characteristics of features as the power cycle number increases. This results in strong subjectivity in feature selection and inconsistency with actual physical degradation laws. Summary of the Invention

[0004] This invention addresses the technical problems in existing technologies where weld fatigue identification methods only extract simple time-domain features and ignore the monotonic degradation law of features with the number of cumulative power cycles. At the same time, the identification model lacks output monotonicity constraints and confidence evaluation, resulting in low accuracy, poor reliability, and non-compliance with physical degradation consistency. The invention provides an IGBT weld fatigue identification method and system based on voltage texture.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a method for fatigue identification of IGBT solder layers based on voltage texture, comprising: Using the thermal time constant of the IGBT solder layer as the acquisition reference, voltage texture signals of IGBT switching transients within multiple switching cycles are obtained to form a voltage texture sequence. Multiple candidate features are extracted from the voltage texture sequence, and the candidate features are screened according to the monotonicity index of each candidate feature on the cumulative power cycle number. The screened features are combined into a solder layer degradation fingerprint vector. The degradation fingerprint vector of the weld layer is input into the pre-trained weld layer fatigue recognition model to obtain preliminary recognition results. The training loss function of the weld layer fatigue recognition model includes a monotonicity constraint term. A pre-trained deviation assessment model is used to evaluate the degree of deviation of the solder layer degradation fingerprint vector from the health state distribution and output a confidence score. The preliminary identification result is corrected based on the confidence score to obtain the corrected weld fatigue identification result.

[0006] Secondly, the present invention provides an IGBT solder layer fatigue identification system based on voltage texture, comprising: The voltage texture acquisition module is used to acquire voltage texture signals of IGBT switching transients within multiple switching cycles, based on the thermal time constant of the IGBT solder layer, to form a voltage texture sequence. The fingerprint vector construction module is used to extract multiple candidate features from the voltage texture sequence, and to filter the candidate features according to the monotonicity index of each candidate feature on the cumulative power cycle number, and to combine the filtered features into a solder layer degradation fingerprint vector. The fatigue recognition module is used to input the weld layer degradation fingerprint vector into a pre-trained weld layer fatigue recognition model to obtain preliminary recognition results. The training loss function of the weld layer fatigue recognition model includes a monotonicity constraint term. The deviation assessment module is used to evaluate the degree of deviation of the solder layer degradation fingerprint vector from the health state distribution using a pre-trained deviation assessment model, and output a confidence score. The fusion correction module is used to correct the preliminary identification result based on the confidence score to obtain the corrected weld fatigue identification result.

[0007] The beneficial effects of this invention are: Compared with existing technologies, this application first uses the thermal time constant of the IGBT solder layer as the acquisition benchmark to obtain voltage texture signals and construct voltage texture sequences. After extracting candidate features from the voltage texture sequences, it selects feature combinations that monotonically change with the cumulative power cycle number based on the monotonicity index to form the solder layer degradation fingerprint vector. This vector is then input into a solder layer fatigue recognition model with a monotonicity constraint term in the training loss function to obtain preliminary recognition results. At the same time, a deviation evaluation model is used to evaluate the degree of deviation of the solder layer degradation fingerprint vector from the healthy state distribution and output a confidence score. Finally, the preliminary recognition results are corrected based on the confidence score to obtain the corrected solder layer fatigue recognition results.

[0008] Through the above technical solutions, this application constructs a complete technical chain from signal acquisition, monotonic degradation feature screening, model recognition with physical constraints to confidence assessment and result correction. It effectively solves the problems of subjective feature selection, neglect of monotonic degradation law, model output violation of physical consistency and lack of reliable quantification in existing methods, and improves the accuracy, reliability and physical consistency of IGBT solder layer fatigue identification. Attached Figure Description

[0009] Figure 1 A schematic flowchart of the voltage texture-based IGBT solder layer fatigue identification method provided by the present invention; Figure 2This is a schematic diagram of the IGBT solder layer fatigue identification system based on voltage texture provided by the present invention.

[0010] In the attached diagram, the components represented by each number are as follows: Voltage texture acquisition module 11, fingerprint vector construction module 12, fatigue recognition module 13, deviation evaluation module 14, and fusion correction module 15. Detailed Implementation

[0011] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0012] Example 1, as Figure 1 As shown, this embodiment of the invention provides a method for fatigue identification of IGBT solder layers based on voltage texture, including: S10: Using the thermal time constant of the IGBT solder layer as the acquisition reference, the voltage texture signal of the IGBT switching transient within multiple switching cycles is acquired to form a voltage texture sequence.

[0013] During power cycling, the IGBT solder layer undergoes fatigue degradation, leading to increased thermal resistance and affecting the collector-emitter voltage overshoot waveform during turn-off transients. This waveform is sensitive to changes in the solder layer's microstructure, but variations in the device's thermal state across different acquisition cycles can cause waveform differences. Therefore, it is necessary to use the solder layer's own thermal time constant as the acquisition reference to ensure consistent thermal equilibrium of the device in each acquisition, thereby eliminating the influence of temperature drift on the voltage waveform and ensuring that the acquired signal accurately reflects the solder layer's degradation state.

[0014] Specifically, step S10 in the method includes: Based on the material thermal parameters and geometric dimensions of the IGBT solder layer, the initial thermal time constant of the IGBT solder layer is obtained through finite element simulation calculation. Within each acquisition cycle, using the initial thermal time constant as the duration reference, the collector-emitter voltage overshoot waveform of the IGBT module at the moment of turn-off in multiple switching cycles is acquired, and the collector-emitter voltage overshoot waveform is used as a voltage texture signal. All voltage texture signals acquired in multiple consecutive acquisition cycles are arranged in chronological order to form a voltage texture sequence.

[0015] In this embodiment, the initial thermal time constant of the IGBT solder layer is first calculated using finite element simulation based on the material thermal parameters and geometric dimensions of the IGBT solder layer. The material thermal parameters include the thermal conductivity, specific heat capacity, and density of the material used in the IGBT solder layer (e.g., solder, silver solder), which directly affect the heat conduction capability of the solder layer. The geometric dimensions refer to the thickness and area of ​​the solder layer. Finite element simulation is a numerical calculation method that simulates the heat conduction process of the solder layer using a computer. It can accurately calculate the temperature response curve of the solder layer under a step heat load, and the time required for the temperature to rise to 63.2% of the steady-state value is taken as the initial thermal time constant.

[0016] For example, in specific implementation, the material thermal parameters and geometric dimensions of the IGBT solder layer are first input into finite element simulation software (optionally, ANSYS Icepak or COMSOL Multiphysics) to simulate the response process of the solder layer under thermal action, and the initial thermal time constant under healthy conditions is calculated. The initial thermal time constant is used as the reference for subsequent signal acquisition to ensure that the thermal balance state of the device is consistent in different acquisition cycles.

[0017] For example, for a certain model of 1200V / 300A IGBT module, the solder layer thickness is 0.1mm, and the material is Sn96.5 and Ag3.5. The initial thermal time constant is 10ms obtained by finite element simulation calculation.

[0018] Secondly, within each acquisition cycle, using the initial thermal time constant as the duration reference, the collector-emitter voltage overshoot waveform at the moment of turn-off of the IGBT module in multiple switching cycles is acquired, and the collector-emitter voltage overshoot waveform is used as a voltage texture signal. The acquisition cycle refers to the time interval between two consecutive signal acquisitions, which can be dynamically set according to the IGBT's operating frequency and solder layer degradation rate. The collector-emitter voltage is a core electrical parameter of the IGBT. The voltage overshoot waveform at the moment of turn-off refers to the waveform in which the collector-emitter voltage briefly exceeds the steady-state bus voltage when the IGBT is turned off, then falls back and is accompanied by damped oscillations. This waveform is affected by the thermal characteristics and parasitic parameters of the solder layer; solder layer fatigue degradation will cause significant changes in the peak value, oscillation frequency, and decay time constant of the overshoot waveform. In practice, within each acquisition cycle, the initial thermal time constant is used as the acquisition duration reference. For example, the acquisition duration is set as an integer multiple of the thermal time constant, such as 2 times or 5 times. At the turn-off moment of multiple IGBT switching cycles, the collector-emitter voltage overshoot waveform is synchronously acquired through a high-frequency voltage probe and a high sampling rate oscilloscope, and this waveform is used as the voltage texture signal of that cycle.

[0019] For example, if the initial thermal time constant is 10ms, the collector-emitter voltage overshoot waveform at the moment of turn-off is acquired for 5 switching cycles in each acquisition cycle. The acquisition time of each collector-emitter voltage overshoot waveform is 10ms. The average is taken as the voltage texture signal of that cycle to suppress random noise.

[0020] Finally, all voltage texture signals acquired over multiple consecutive acquisition cycles are arranged chronologically to form a voltage texture sequence. Specifically, voltage texture signals acquired over multiple consecutive acquisition cycles (e.g., 10 or 20 acquisition cycles) are arranged sequentially according to their acquisition time to form a voltage texture sequence. Each signal in the voltage texture sequence corresponds to one acquisition cycle, clearly showing the trend of voltage texture changes over time (i.e., with increasing cumulative power cycle count), providing continuous raw data for subsequent extraction of features that monotonically change with solder layer degradation.

[0021] In summary, compared to existing technologies, this application uses the thermal time constant of the IGBT solder layer as the acquisition benchmark to obtain voltage texture signals of IGBT switching transients over multiple switching cycles, thus constructing a voltage texture sequence. By using the physical thermal time constant as the acquisition benchmark, waveform drift caused by differences in thermal states across different acquisition cycles is eliminated, ensuring the consistency and comparability of the voltage texture signals. This provides a stable and reliable raw data foundation for subsequent extraction of candidate features monotonically related to solder layer degradation.

[0022] S20: Extract multiple candidate features from the voltage texture sequence, and filter the candidate features according to the monotonicity index of each candidate feature on the cumulative power cycle number, and combine the filtered features into a solder layer degradation fingerprint vector.

[0023] Voltage texture sequences contain a large amount of raw waveform information, but not all waveform features change monotonically with solder layer fatigue degradation. Some features exhibit random fluctuations due to measurement noise and environmental disturbances, failing to stably indicate the solder layer degradation state. Therefore, it is necessary to extract multi-dimensional candidate features, such as time-domain and parasitic parameter sensitivity, and use monotonicity indices to screen features that change monotonically (increasing or decreasing) with the cumulative power cycle count, eliminating random fluctuation features, and constructing a stable, sensitive, and low-dimensional solder layer degradation fingerprint vector.

[0024] Specifically, step S20 in the method includes: From the voltage texture signal corresponding to each acquisition cycle of the voltage texture sequence, time-domain features and parasitic parameter sensitivity features are extracted. The time-domain features include voltage overshoot peak value, voltage overshoot amplitude to steady-state voltage ratio, and overshoot waveform area. The parasitic parameter sensitivity features include overshoot waveform oscillation frequency offset and overshoot waveform oscillation decay time constant change rate. The extracted time-domain features and the parasitic parameter sensitivity features are used together as candidate features; The feature values ​​of candidate features of the weld layer health status are obtained, the fluctuation standard deviation of the feature values ​​is calculated, and the monotonicity threshold is determined based on the fluctuation standard deviation. Obtain the feature value sequence corresponding to each candidate feature in multiple consecutive acquisition cycles, as well as the cumulative power cycle count sequence corresponding to each acquisition cycle; Based on the feature value sequence and the cumulative power cycle number sequence, calculate the monotonicity index value of the candidate feature as a function of the cumulative power cycle number; Candidate features whose monotonicity index values ​​exceed the monotonicity threshold are selected and combined into a solder layer degradation fingerprint vector.

[0025] In this embodiment, time-domain features and parasitic parameter sensitivity features are first extracted from the voltage texture signal corresponding to each acquisition cycle of the voltage texture sequence. The time-domain features include the voltage overshoot peak value, the ratio of the voltage overshoot amplitude to the steady-state voltage, and the overshoot waveform area. The parasitic parameter sensitivity features include the overshoot waveform oscillation frequency offset and the rate of change of the overshoot waveform oscillation decay time constant. The time-domain features are parameters extracted from the time dimension of the voltage texture signal, directly reflecting the temporal variation of the waveform: the voltage overshoot peak value is the maximum value of the voltage overshoot at the moment of turn-off; solder fatigue can lead to an increase in the voltage overshoot peak value. The ratio of the voltage overshoot amplitude to the steady-state voltage is the ratio of the overshoot amplitude (i.e., the difference between the voltage overshoot peak value and the steady-state voltage) to the steady-state voltage, which can quantify the degree of overshoot. The overshoot waveform area is the area enclosed by the collector-emitter voltage overshoot waveform and the steady-state voltage line, reflecting the duration and intensity of the overshoot.

[0026] Among them, the parasitic parameter sensitivity characteristics reflect the degree to which the voltage waveform is affected by parameters such as parasitic inductance and capacitance of IGBT. Solder layer degradation will lead to changes in parasitic parameters, which in turn will affect the waveform oscillation characteristics: the overshoot waveform oscillation frequency offset is the difference between the actual oscillation frequency and the oscillation frequency under healthy conditions; the overshoot waveform oscillation decay time constant change rate is the ratio of the difference between the current decay time constant and the decay time constant under healthy conditions to the healthy value.

[0027] Specifically, for the voltage texture signal of each acquisition cycle in the voltage texture sequence, the above five features are extracted respectively to ensure that the candidate features can capture the voltage waveform changes caused by solder layer degradation from different dimensions.

[0028] For example, suppose that within one acquisition cycle, a set of key values ​​of the voltage texture signal are obtained from the oscilloscope: steady-state voltage = 1.5V, voltage overshoot peak value = 650V, and the overshoot waveform sampling point sequence is as follows: starting from turn-off, the voltage rises from 1.5V to 650V within 0.2ms, and then decays and oscillates to the steady state within 2ms. The extraction process is as follows: 1. Voltage overshoot peak value: directly read the maximum value of the waveform, i.e., 650V. 2. Ratio of voltage overshoot amplitude to steady-state voltage: overshoot amplitude = 650V - 1.5V = 648.5V, the ratio of voltage overshoot amplitude to steady-state voltage = 648.5 / 1.5 ≈ 432.3. 3. Overshoot waveform area: approximate the region between the collector-emitter voltage overshoot waveform and the steady-state voltage line by dividing it into multiple small trapezoids for numerical integration. For example, a voltage value is sampled every 0.1ms, and the difference between each sampled point and the steady-state voltage is multiplied by the time interval and summed to obtain an integral value of 12V·ms. 4. Overshoot waveform oscillation frequency offset: The decaying oscillation portion after overshoot is taken, and the time interval between two adjacent peaks is measured to calculate the oscillation frequency. Assuming the oscillation frequency is 5MHz in the healthy state and the current measured oscillation frequency is 4.8MHz, then the overshoot waveform oscillation frequency offset = -0.2MHz (negative indicates a frequency decrease). 5. Overshoot waveform oscillation decay time constant change rate: The decay time constant is obtained by exponentially fitting the oscillation envelope. Assuming the decay time constant is 0.5ms in the healthy state and currently 0.6ms, then the overshoot waveform oscillation decay time constant change rate = (0.6-0.5) / 0.5 = 0.2.

[0029] Secondly, the extracted time-domain features and parasitic parameter sensitivity features are used together as candidate features. Specifically, all three extracted time-domain features and two parasitic parameter sensitivity features are used as candidate features. These features reflect the weld layer state from two dimensions: time-domain waveform and parasitic parameter response. They cover the main impacts of weld layer fatigue degradation on voltage texture signals, providing sufficient candidate objects for subsequent screening of strongly correlated features.

[0030] Next, the eigenvalues ​​of candidate features for the weld layer health status are obtained, the standard deviation of the eigenvalue fluctuations is calculated, and the monotonicity threshold is determined based on the standard deviation of the fluctuations. Here, the candidate eigenvalues ​​for the weld layer health status refer to the values ​​corresponding to each candidate feature when the IGBT weld layer is in a healthy state; the standard deviation of the fluctuations is a parameter that measures the dispersion of eigenvalues ​​in a healthy state, reflecting the stable range of the features under healthy conditions; the monotonicity threshold is a critical value for judging the strength of the correlation between candidate features and weld layer fatigue, determined based on the standard deviation of the fluctuations, ensuring that the changing trends of the selected features are significantly different from the fluctuations of the healthy state.

[0031] Specifically, multiple feature values ​​for each candidate feature under healthy conditions are first obtained (e.g., 100 consecutive cycles of data collection without fatigue). The standard deviation σ of the feature value sequence for each candidate feature is then calculated. Since feature changes under healthy conditions mainly originate from measurement noise and minor environmental disturbances, their fluctuations typically follow a normal distribution. According to the statistical 3σ criterion, the true trend change caused by degradation should significantly exceed this normal fluctuation range. Therefore, the monotonicity threshold is set to k times the standard deviation σ, where k is the confidence coefficient. This coefficient can be determined based on the required significance level and the engineering trade-off between sensitivity and false alarm rate; for example, k can be 3 to 5. This ensures that only features with a clear trend exceeding the healthy fluctuation range are selected.

[0032] Furthermore, the feature value sequence corresponding to each candidate feature in multiple consecutive acquisition cycles, and the cumulative power cycle count sequence corresponding to each acquisition cycle are obtained. Specifically, for each candidate feature, its corresponding feature values ​​in multiple consecutive acquisition cycles are arranged in chronological order to form the feature value sequence for that feature; simultaneously, the cumulative power cycle count of the IGBT corresponding to each acquisition cycle (i.e., the total number of power cycles experienced from the start of use to the current cycle) is recorded and arranged in chronological order to form the cumulative power cycle count sequence, ensuring that each value in the feature value sequence matches the corresponding value in the cumulative power cycle count sequence, providing corresponding data pairs for subsequent calculation of monotonicity indicators.

[0033] Furthermore, based on the eigenvalue sequence and the cumulative power cycle count sequence, the monotonicity index value of the candidate feature as a function of the cumulative power cycle count is calculated. The monotonicity index value quantifies the consistency of the trend of the candidate feature as the weld layer fatigue degradation (i.e., the increase in the cumulative power cycle count). A higher monotonicity index value indicates a stronger correlation between the candidate feature and weld layer fatigue. Specifically, by combining the eigenvalue sequence and the cumulative power cycle count sequence, the direction (increasing or decreasing) of each candidate feature as a function of the cumulative power cycle count is analyzed, and the consistency of its trend is calculated to obtain the monotonicity index value.

[0034] For example, if the eigenvalue of a candidate feature increases continuously with the cumulative power cycle number and the direction of change is consistent between two adjacent cycles, then the monotonicity index value of the feature is high; conversely, if the direction of change of the eigenvalue is chaotic, then the monotonicity index value is low.

[0035] Finally, candidate features whose monotonicity index values ​​exceed the monotonicity threshold are selected and combined to form a weld layer degradation fingerprint vector. Specifically, the monotonicity index value of each candidate feature is compared with a preset monotonicity threshold, and candidate features whose monotonicity index values ​​exceed the threshold are selected, i.e., features with strong correlation to weld layer fatigue degradation and stable change trends are selected. These selected features are combined in a preset order (e.g., by feature type or by monotonicity index from high to low) to form the weld layer degradation fingerprint vector. The weld layer degradation fingerprint vector can uniquely and accurately represent the current degradation state of the IGBT weld layer, providing high-quality input features for subsequent model recognition. In this way, features with weak correlation are eliminated, and features that can effectively represent weld layer degradation are retained, constructing an accurate weld layer degradation fingerprint vector.

[0036] Specifically, based on the feature value sequence and the cumulative power cycle number sequence, the monotonicity index value of the candidate feature changing with the cumulative power cycle number is calculated, including: The feature values ​​corresponding to each acquisition cycle in the feature value sequence are arranged in ascending order of the cumulative power cycle number to form a monotonicity analysis feature value sequence. Calculate the difference between the eigenvalues ​​of each adjacent position in the monotonicity analysis eigenvalue sequence, and count the number of positive and negative signs in the difference values; Calculate the absolute value of the difference between the number of occurrences of the positive sign and the number of occurrences of the negative sign, and use the absolute value of the difference as the monotonicity index value.

[0037] In this embodiment, the feature values ​​corresponding to each acquisition cycle in the feature value sequence are first arranged in ascending order of cumulative power cycle count to form a monotonicity analysis feature value sequence.

[0038] Specifically, since the eigenvalue sequence is arranged in chronological order of acquisition time, and the cumulative power cycle count increases with time, each eigenvalue in the sequence needs to be rearranged in ascending order of its corresponding cumulative power cycle count to obtain the monotonicity analysis eigenvalue sequence. In the monotonicity analysis eigenvalue sequence, the order of the eigenvalues ​​corresponds to the increasing order of the weld layer fatigue degradation degree, which can intuitively reflect the changing pattern of the eigenvalues ​​with the degree of degradation.

[0039] Secondly, the differences between eigenvalues ​​at adjacent positions in the monotonicity analysis eigenvalue sequence are calculated, and the frequency of positive and negative differences is counted. The differences reflect the direction of eigenvalue change between two adjacent degradation stages (increasing cumulative power cycle count). Specifically, the difference between every two adjacent eigenvalues ​​in the monotonicity analysis eigenvalue sequence (i.e., the latter minus the former) is calculated, resulting in a series of differences. The frequency of all positive differences (i.e., the number of positive signs) and negative differences (i.e., the number of negative signs) is counted. Positive differences indicate that the eigenvalue increases with increasing degradation, while negative differences indicate that the eigenvalue decreases with increasing degradation.

[0040] For example, if the monotonicity analysis eigenvalue sequence is [2,4,6,8,10], and all adjacent difference values ​​are positive, then the positive sign appears 4 times and the negative sign appears 0 times.

[0041] Finally, the absolute value of the difference between the number of occurrences of positive signs and the number of occurrences of negative signs is calculated, and this absolute value is used as the monotonicity index. Specifically, this calculation method can intuitively quantify the consistency of the feature's changing trend: the larger the absolute value of the difference, the more uniform the trend of the feature's change with the degree of degradation (either continuously increasing or continuously decreasing), and the stronger the monotonicity. The specific formula is: Monotonicity index value = |Number of occurrences of positive signs - Number of occurrences of negative signs|.

[0042] For example, if the positive sign appears 8 times and the negative sign appears 2 times, the monotonicity index is 6, indicating that the feature mainly shows an increasing trend and has strong monotonicity; if the positive sign appears 5 times and the negative sign appears 4 times, the monotonicity index is 1, indicating that the feature changes randomly and has weak monotonicity.

[0043] In summary, compared to existing technologies, this application extracts multiple candidate features from the voltage texture sequence and filters these features based on their monotonicity indices over the cumulative power cycle count. The filtered features are then combined to form a weld layer degradation fingerprint vector. Thus, by extracting multi-dimensional time-domain and parasitic parameter features and using monotonicity indices to eliminate random fluctuation features, a stable, monotonic, and low-dimensional weld layer degradation fingerprint vector is constructed, providing high-quality and highly reliable input features for subsequent fatigue identification models.

[0044] S30: Input the weld layer degradation fingerprint vector into the pre-trained weld layer fatigue recognition model to obtain preliminary recognition results, wherein the training loss function of the weld layer fatigue recognition model includes a monotonicity constraint term.

[0045] After constructing the weld layer degradation fingerprint vector, it needs to be mapped to the weld layer fatigue level. Since weld layer fatigue degradation has a monotonically increasing physical characteristic (i.e., the fatigue level does not decrease with increasing power cycle count), ordinary classification models may produce predictions that violate monotonicity due to noise. Therefore, a monotonicity constraint term needs to be introduced into the training loss function to ensure that the model output conforms to the physical degradation law. Simultaneously, the high-precision classification capability of deep learning models is utilized to output preliminary identification results.

[0046] Specifically, step S30 in the method includes: Construct an initial weld layer fatigue identification model; Multiple sets of training samples are obtained, each set of training samples contains a solder layer degradation fingerprint vector and the corresponding solder layer fatigue level label; Construct a total loss function, which is obtained by adding a classification loss term and a monotonicity constraint term; The initial weld layer fatigue identification model is trained in a supervised manner using the multiple sets of training samples and the total loss function until the preset convergence condition is met, thus obtaining the trained weld layer fatigue identification model. The weld layer degradation fingerprint vector is input into the trained weld layer fatigue recognition model, and the weld layer fatigue level and the classification probability value corresponding to the weld layer fatigue level are output. The weld layer fatigue level and the classification probability value are used together as the preliminary recognition result.

[0047] In this embodiment, an initial weld layer fatigue identification model is first constructed. For example, the initial weld layer fatigue identification model can be constructed using a multi-layer fully connected neural network, mainly composed of an input layer, hidden layers, and an output layer. The specific structure and parameter configuration are as follows: 1. Input layer: The number of nodes in the input layer is equal to the feature dimension of the weld layer degradation fingerprint vector. For example, after screening in step S20, the weld layer degradation fingerprint vector may contain 5 candidate features with monotonicity indices exceeding the threshold. Therefore, the number of nodes in the input layer is set to 5. This dimension can be dynamically determined based on the actual screening results to ensure that the input features match the model structure.

[0048] 2. Hidden Layers: Two hidden layers are set up, with the first layer containing 64 neurons and the second layer containing 32 neurons. The activation function for both layers is ReLU (Rectified Linear Unit). The number of neurons is chosen based on several factors: the input feature dimension is low (5-dimensional), but there may be a non-linear relationship between weld layer degradation and fatigue level, requiring sufficient non-linear fitting capability; experimental experience shows that configurations of 64 and 32 neurons can avoid overfitting while maintaining expressive power in most industrial classification tasks. The number of neurons can also be adjusted according to the number of training samples; the number can be increased when the number of samples is large (e.g., 128-64), and decreased when the number of samples is small (e.g., 32-16). A Dropout layer (dropout rate 0.2) can be added after each layer to enhance generalization ability.

[0049] 3. Output layer: The number of nodes in the output layer is equal to the number of fatigue levels of the weld layer. For example, if the fatigue levels are divided into four levels: healthy, mild fatigue, moderate fatigue, and severe fatigue, then the number of nodes in the output layer is set to 4. The activation function is Softmax, and the classification probability of each level is output. The level corresponding to the maximum output probability is the predicted fatigue level of the weld layer.

[0050] Secondly, multiple sets of training samples are obtained, each containing a weld layer degradation fingerprint vector and a corresponding weld layer fatigue level label. Specifically, the training samples need to cover weld layer states with different fatigue levels to ensure the model has good generalization ability. The weld layer degradation fingerprint vector can be obtained by referring to the feature extraction and screening process in the previous steps, corresponding to the weld layer state under different cumulative power cycle counts; the weld layer fatigue level label is the label of the actual fatigue state of the weld layer, which can be obtained through metallographic section analysis or thermal resistance testing. In practice, a large number of IGBT weld layer samples under different cumulative power cycle counts can be collected, and a weld layer degradation fingerprint vector can be constructed for each sample. The corresponding fatigue level can be labeled using professional detection methods, for example: level 0 - healthy, level 1 - mild, level 2 - moderate, level 3 - severe, forming multiple sets of training samples.

[0051] Next, a total loss function is constructed, which is obtained by adding a classification loss term and a monotonicity constraint term. The total loss function guides model training, ensuring that the model not only accurately classifies fatigue levels but also guarantees that the output conforms to the actual degradation patterns of the weld layer. Specifically, the classification loss term measures the deviation between the model's predicted fatigue level and the true label; the monotonicity constraint term constrains the model's output fatigue level to increase monotonically with the number of cumulative power cycles (i.e., the degree of weld layer fatigue intensifies), preventing the model output from deviating from the actual degradation patterns.

[0052] Furthermore, the initial weld fatigue recognition model is trained in a supervised manner using multiple sets of training samples and a total loss function until a preset convergence condition is met, resulting in a trained weld fatigue recognition model. The preset convergence condition can be set as follows: the total loss value of the model on the validation set no longer decreases for 10 consecutive rounds, and the classification accuracy on the validation set reaches a preset threshold (e.g., 95%).

[0053] For example, the training process of the weld fatigue identification model can refer to the following steps: 1. Data preparation: Randomly divide multiple training samples into training set, validation set and test set in a ratio of 7:1.5:1.5. The training set is used for model parameter update, the validation set is used for hyperparameter adjustment and convergence judgment, and the test set is used for final evaluation of model performance.

[0054] 2. Model Training: The weld layer degradation fingerprint vectors in the training set are used as input features, and the corresponding weld layer fatigue level labels are used as supervision labels. The total loss function is used to calculate the error between the model output and the true label. The optimizer is Adam, with an initial learning rate of 0.001, a batch size of 32, and a maximum of 200 training epochs. After each training epoch, the total loss and classification accuracy are calculated on the validation set. When the total loss on the validation set no longer decreases for 10 consecutive epochs (optionally, a decrease of less than 0.001 is considered sufficient) and the classification accuracy is not lower than 95%, the model is considered to have converged, and training is stopped.

[0055] 3. Model Saving: Save the model parameters with the minimum total loss on the validation set as the trained weld layer fatigue recognition model. Finally, evaluate the model's classification accuracy on the test set to ensure it meets the requirements of practical applications.

[0056] Finally, the weld layer degradation fingerprint vector is input into the trained weld layer fatigue recognition model, which outputs the weld layer fatigue level and the corresponding classification probability value. The weld layer fatigue level and classification probability value together constitute the preliminary recognition result. Here, the weld layer fatigue level is the classification result of the weld layer fatigue recognition model on the current weld layer state, such as "moderate fatigue"; the classification probability value is the confidence level of the weld layer fatigue recognition model in predicting this fatigue level, ranging from 0 to 1. The closer the probability value is to 1, the higher the confidence of the weld layer fatigue recognition model in the classification result. Specifically, the weld layer degradation fingerprint vector to be identified is input into the trained weld layer fatigue recognition model. By analyzing the vector features, the corresponding weld layer fatigue level and classification probability value are output, which together constitute the preliminary recognition result.

[0057] For example, the weld fatigue identification model outputs a four-dimensional probability vector, such as [0.1, 0.7, 0.2, 0.0], to the input weld degradation fingerprint vector, representing a healthy probability of 0.1, a mild probability of 0.7, a moderate probability of 0.2, and a severe probability of 0.0, respectively. Then, the mild fatigue with the highest probability is determined as the weld fatigue level, and the corresponding classification probability value is 0.7.

[0058] Specifically, the process of constructing the total loss function includes: Based on the cross-entropy loss function, a classification loss term is constructed; For the same IGBT module, the two solder layer degradation fingerprint vectors corresponding to two adjacent acquisition cycles are respectively input into the initial solder layer fatigue recognition model to obtain the first solder layer fatigue level output value and the second solder layer fatigue level output value. Obtain the first cumulative power cycle number corresponding to the fatigue level output value of the first weld layer, and the second cumulative power cycle number corresponding to the fatigue level output value of the second weld layer; If the first cumulative power cycle count is greater than the second cumulative power cycle count, and the first weld fatigue level output value is less than the second weld fatigue level output value, then the difference between the two is taken as the value of the monotonicity constraint term; otherwise, the value of the monotonicity constraint term is zero. The classification loss term is added to the monotonicity constraint term to obtain the total loss function.

[0059] In this embodiment, a classification loss term is first constructed based on the cross-entropy loss function. Specifically, the cross-entropy loss function is a commonly used loss function in classification tasks, which can effectively measure the deviation between the model's predicted probability distribution and the true label distribution. For example, a classification cross-entropy loss function can be selected, with the following formula: Assuming there are K fatigue levels, for a training sample, its one-hot encoding of the true label is represented as y=[ , ,…, ], where the actual fatigue level of the weld layer corresponds to =1 indicates that the sample belongs to class c, and the rest are 0. The predicted probability distribution output by the model is y^=[ , ,…, Then, the formula for calculating the classification cross-entropy loss function L is: .

[0060] The smaller the value of the classification loss term, the closer the probability distribution predicted by the model is to the true label, and the higher the classification accuracy. For example, if the true label is "moderate fatigue", its one-hot encoding is [0,0,1,0], and the model predicts probabilities of [0.1,0.2,0.6,0.1], then the classification cross-entropy loss function L = -(0×log0.1 + 0×log0.2 + 1×log0.6 + 0×log0.1) ≈ 0.51, which is used as the value of the classification loss term.

[0061] Secondly, for the two solder layer degradation fingerprint vectors corresponding to two adjacent acquisition cycles of the same IGBT module, they are input into the initial solder layer fatigue identification model to obtain the first solder layer fatigue level output value and the second solder layer fatigue level output value. Specifically, for the same IGBT module, the increasing cumulative power cycle count in two adjacent acquisition cycles corresponds to a greater degree of solder layer fatigue. The first solder layer fatigue level output value corresponds to the acquisition cycle with a higher cumulative power cycle count (higher fatigue level), and the second solder layer fatigue level output value corresponds to the acquisition cycle with a lower cumulative power cycle count (lower fatigue level). In detail, two adjacent acquisition cycles of the same IGBT module are selected, and their solder layer degradation fingerprint vectors are constructed respectively. These are input into the initial solder layer fatigue identification model to obtain two corresponding fatigue level output values, which are used to subsequently determine whether the output of the initial solder layer fatigue identification model conforms to a monotonicity pattern.

[0062] Next, obtain the first cumulative power cycle number corresponding to the first weld layer fatigue level output value, and the second cumulative power cycle number corresponding to the second weld layer fatigue level output value. Specifically, record the cumulative power cycle number of the acquisition period corresponding to the first weld layer fatigue level output value as the first cumulative power cycle number, and record the cumulative power cycle number of the acquisition period corresponding to the second weld layer fatigue level output value as the second cumulative power cycle number.

[0063] Furthermore, if the first cumulative power cycle count is greater than the second cumulative power cycle count, and the output value of the first weld layer fatigue level is less than the output value of the second weld layer fatigue level, then the difference between the two is used as the value of the monotonicity constraint term; otherwise, the value of the monotonicity constraint term is zero. Specifically, the role of the monotonicity constraint term is to penalize situations where the model output does not conform to the actual degradation law of the weld layer: when the first cumulative power cycle count (higher fatigue level) is greater than the second cumulative power cycle count, but the output value of the first weld layer fatigue level (a larger value indicates more severe degradation) is less than the output value of the second weld layer fatigue level, it indicates that the model output violates the monotonicity law—the fatigue level increases while the predicted level decreases. In this case, the difference between the two is used as the value of the monotonicity constraint term to penalize the model; if the model output conforms to the monotonicity law, that is, the output value of the first weld layer fatigue level is greater than or equal to the output value of the second weld layer fatigue level, then the value of the monotonicity constraint term is zero.

[0064] For example, if the first cumulative power cycle count is 10,000 times, the second cumulative power cycle count is 8,000 times, the first weld layer fatigue level output value is 2 (representing moderate fatigue), and the second weld layer fatigue level output value is 3 (representing severe fatigue), then the value of the monotonicity constraint term is 3-2=1; conversely, if the first weld layer fatigue level output value is 3 and the second weld layer fatigue level output value is 2, then the value of the monotonicity constraint term is 0.

[0065] Finally, the classification loss term and the monotonicity constraint term are added together to obtain the total loss function. Specifically, the formula for the total loss function is: Total Loss Function = Classification Loss Term + Monotonicity Constraint Term. Here, the classification loss term ensures accurate model classification, while the monotonicity constraint term ensures that the model output conforms to the monotonic law of solder layer degradation. The combination of these two ensures that the trained model can more accurately identify the fatigue state of the IGBT solder layer. Thus, the total loss function balances the model's classification accuracy and the monotonicity of the output results.

[0066] In summary, compared to existing technologies, this application inputs the weld layer degradation fingerprint vector into a pre-trained weld layer fatigue recognition model to obtain preliminary recognition results. Thus, by introducing a monotonicity constraint term, the output of the weld layer fatigue recognition model conforms to the physical law that weld layer fatigue increases monotonically with power cycles. Simultaneously, the high classification accuracy of deep learning provides a reliable foundation for subsequent confidence level correction.

[0067] S40: Use a pre-trained deviation assessment model to evaluate the deviation of the solder layer degradation fingerprint vector from the health state distribution and output a confidence score.

[0068] The classification probability value output by the weld fatigue identification model only represents the model's own confidence in the classification result, but it cannot reflect whether the current weld degradation fingerprint vector deviates from the distribution of healthy state samples in the training data. When the weld layer is severely degraded or exhibits abnormal failure modes, the weld degradation fingerprint vector may fall into the low-probability region of the healthy distribution, in which case the confidence of the weld fatigue identification model may be artificially high. Therefore, it is necessary to additionally evaluate the degree of deviation between the weld degradation fingerprint vector and the healthy state distribution, and output a confidence score for subsequent correction.

[0069] Specifically, step S40 in the method includes: Obtain a pre-trained deviation assessment model, wherein the deviation assessment model is constructed based on a generative adversarial network and includes an encoder and a discriminator; The solder layer degradation fingerprint vector is input into the deviation evaluation model, and the encoder maps the solder layer degradation fingerprint vector to the latent space feature representation. The latent space feature representation is input into the discriminator, which outputs the probability value that the solder layer degradation fingerprint vector belongs to the healthy state distribution. The probability value is used as the confidence score.

[0070] In this embodiment, a pre-trained deviation assessment model is first obtained. This model is constructed based on a generative adversarial network (GAN) and includes an encoder and a discriminator. Specifically, GANs have powerful feature extraction and distribution recognition capabilities, making them suitable for assessing the deviation between the solder layer degradation fingerprint vector and the healthy state distribution. The encoder maps the input solder layer degradation fingerprint vector into a low-dimensional latent space feature representation, simplifying the feature dimension while retaining core information. The discriminator determines whether the input latent space feature representation belongs to the healthy state feature distribution and outputs the corresponding probability value.

[0071] Secondly, the solder layer degradation fingerprint vector is input into the deviation assessment model, and the encoder maps the solder layer degradation fingerprint vector to a latent space feature representation. Specifically, the latent space feature representation is a low-dimensional abstract representation of the solder layer degradation fingerprint vector, which can retain the core features of the vector while reducing the data dimensionality, making it easier for the discriminator to perform distribution recognition. In practice, the solder layer degradation fingerprint vector to be evaluated is input into the deviation assessment model, and the encoder performs feature extraction and dimensionality compression on the solder layer degradation fingerprint vector, mapping it to a latent space feature representation.

[0072] Finally, the latent space feature representation is input into the discriminator, which outputs the probability value that the solder layer degradation fingerprint vector belongs to the healthy state distribution. This probability value is used as the confidence score. Specifically, the discriminator compares the latent space feature representation with the latent space feature distribution under healthy conditions and outputs the probability value that the feature representation belongs to the healthy state distribution. This probability value is the confidence score. The closer the confidence score is to 1, the smaller the deviation of the current solder layer degradation fingerprint vector from the healthy state distribution, the closer the solder layer state is to healthy, and the higher the reliability of the preliminary identification result. Conversely, the closer the confidence score is to 0, the greater the deviation, and the lower the reliability of the preliminary identification result.

[0073] Specifically, the training process of the deviation assessment model includes: Collect multiple solder layer degradation fingerprint vectors corresponding to the IGBT solder layer being in a healthy state to form a healthy state sample set; Multiple sets of training auxiliary samples are collected. Each set of training auxiliary samples contains a solder layer degradation fingerprint vector and a corresponding solder layer physical degradation degree calibration value. An initial deviation assessment model is constructed, which includes an initial encoder and an initial discriminator; Determine the training loss function, which is composed of an adversarial loss term and a physical consistency constraint term; The initial deviation assessment model is trained iteratively in multiple rounds, and the following steps are performed in each round of iteration: The weld layer degradation fingerprint vector in the health status sample set is input into the initial encoder, the initial encoder outputs the latent space feature representation, the latent space feature representation is input into the initial discriminator, the initial discriminator outputs the true / false discrimination result, and the value of the adversarial loss term is calculated based on the true / false discrimination result; The solder layer degradation fingerprint vector in the training auxiliary sample is input into the initial encoder to obtain the auxiliary latent space feature representation. The auxiliary latent space feature representation is input into the initial discriminator to obtain the healthy distribution deviation output value. The monotonicity correlation coefficient between the healthy distribution deviation output value and the solder layer physical degradation degree calibration value in the training auxiliary sample is calculated. The value of subtracting the monotonicity correlation coefficient is used as the value of the physical consistency constraint term. The value of the adversarial loss term is added to the value of the physical consistency constraint term to obtain the value of the training loss function, and the parameters of the initial encoder and the initial discriminator are updated according to the value of the training loss function; The training continues until the preset convergence condition is met, at which point the deviation evaluation model is obtained.

[0074] In this embodiment, multiple solder layer degradation fingerprint vectors corresponding to IGBT solder layers in a healthy state are first collected to form a healthy state sample set. This healthy state sample set is used to enable the deviation assessment model to learn the feature distribution of the solder layer degradation fingerprint vectors in a healthy state. Specifically, multiple IGBT solder layer samples in a healthy state are collected. Through the aforementioned feature extraction and filtering process, a solder layer degradation fingerprint vector for each IGBT solder layer sample is constructed, and these vectors are then organized into a healthy state sample set.

[0075] For example, 5,000 healthy IGBT solder layer samples are collected to construct 5,000 solder layer degradation fingerprint vectors, forming a healthy sample set.

[0076] Secondly, multiple sets of training auxiliary samples are collected. Each set of training auxiliary samples contains a solder layer degradation fingerprint vector and a corresponding solder layer physical degradation degree calibration value. The training auxiliary samples are used to ensure that the output of the deviation assessment model remains consistent with the actual physical degradation degree of the solder layer. The solder layer physical degradation degree calibration value is obtained through professional physical detection methods (e.g., observing the solder layer crack area using a metallographic microscope or through thermal resistance testing), and is used to quantify the actual degradation degree of the solder layer. Its value range can be set from 0 to 1, where 0 represents a healthy state and 1 represents a severely degraded state. In specific implementation, multiple sets of IGBT solder layer samples with different degradation degrees are collected, and a solder layer degradation fingerprint vector is constructed for each sample. Then, the physical degradation degree is calibrated using professional detection methods such as metallographic microscopy to obtain the corresponding solder layer physical degradation degree calibration value. The solder layer degradation fingerprint vector and the solder layer physical degradation degree calibration value are combined to form the training auxiliary samples. This ensures that the training auxiliary samples cover the entire range from healthy to severely degraded.

[0077] It should be noted that the calibration value of the physical degradation degree of the weld layer is dynamically determined through professional physical testing methods. For example, the proportion of crack area or crack length in the weld layer can be observed using a metallographic microscope. Based on the measured crack area ratio, it can be calibrated to the range of 0 to 1 through linear mapping. For example, 0 represents healthy (no cracks) and 1 represents severe degradation.

[0078] Next, an initial deviation assessment model is constructed, which includes an initial encoder and an initial discriminator. For example, the initial deviation assessment model can be constructed using a multi-layer fully connected network: 1. Initial encoder: mainly composed of an input layer, hidden layers, and an output layer. The number of nodes in the input layer is equal to the number of features in the solder layer degenerate fingerprint vector; the hidden layer has two layers, with 64 neurons in the first layer and 32 neurons in the second layer, both using ReLU (Rectified Linear Unit) activation functions to progressively extract high-order features from the solder layer degenerate fingerprint vector; the output layer has a preset latent space dimension (e.g., 3D), using a linear activation function to map the input vector to a low-dimensional latent space feature representation. The weights of each layer are initialized using Xavier, and the biases are initialized to 0.

[0079] 2. Initial Discriminator: Primarily composed of an input layer, hidden layers, and an output layer. The number of nodes in the input layer equals the dimension of the latent space output by the initial encoder (e.g., 3D). Two hidden layers are used: the first layer has 32 neurons, and the second layer has 16 neurons, both employing ReLU activation. The output layer is a single neuron using the Sigmoid activation function, outputting the probability value (range 0-1) that the current input latent space feature representation belongs to the healthy state distribution. Weights in each layer are initialized using Xavier, and biases are initialized to 0.

[0080] Thus, through the above structural configuration, the initial encoder can compress the high-dimensional weld layer degradation fingerprint vector into a low-dimensional latent space feature representation, and the initial discriminator can determine whether it belongs to the healthy state distribution based on the latent space feature representation. Together, they constitute the initial deviation assessment model, providing a foundation for subsequent adversarial training and physical consistency constraints.

[0081] Furthermore, a training loss function is determined, which is the sum of an adversarial loss term and a physical consistency constraint term. The adversarial loss term constrains the adversarial training of the initial encoder and initial discriminator, enabling the initial discriminator to distinguish between the latent space features of healthy samples and those generated, while simultaneously ensuring that the latent space features generated by the initial encoder closely approximate the prior distribution. A smaller value for the adversarial loss term indicates a stronger discriminative ability of the initial discriminator, and a more accurate reflection of the health status of the latent space features generated by the initial encoder. For example, the adversarial loss can be calculated using the loss form of a Wasserstein Generative Adversarial Network (WGAN).

[0082] The physical consistency constraint term is used to ensure that the deviation of the health distribution output from the model is consistent with the actual physical degradation of the weld layer. Specifically, it requires that the deviation of the health distribution monotonically increases with the calibration value of the weld layer's physical degradation. The physical consistency constraint term is calculated by subtracting the correlation coefficient (e.g., the Spearman correlation coefficient) between the output value of the health distribution deviation and the calibration value of the weld layer's physical degradation. A smaller value indicates a closer consistency between the model output and actual physical laws; a larger value indicates a more severe violation of monotonicity, resulting in a greater penalty imposed on the model during training.

[0083] Furthermore, the initial deviation assessment model is trained iteratively through multiple rounds. In each iteration, the following steps are performed: the weld layer degradation fingerprint vector from the healthy state sample set is input into the initial encoder, which outputs a latent space feature representation. This latent space feature representation is then input into the initial discriminator, which outputs a true / false discrimination result. The adversarial loss term is calculated based on this result. A smaller adversarial loss term indicates a stronger discriminative ability of the initial discriminator, and the more accurately the latent space features generated by the initial encoder reflect the healthy state.

[0084] For example, the adversarial loss can be calculated using the loss form of Wasserstein Generative Adversarial Network (WGAN): Assume that the healthy state sample set contains the latent space feature representations z1, z2, z3, z4 of 4 healthy samples, and the initial discriminator outputs D(z1) = 0.9, D(z2) = 0.8, D(z3) = 0.95, and D(z4) = 0.85 respectively; at the same time, 4 noise vectors are randomly sampled. , , , The initial discriminator outputs D( )=0.2,D( )=0.3,D( )=0.25,D( )=0.2. Therefore, the adversarial loss of the Wasserstein Generative Adversarial Network (WGAN) is... =(0.9+0.8+0.95+0.85) / 4-(0.2+0.3+0.25+0.2) / 4=0.6375, which is used as the value of the adversarial loss term. This value represents the distance between healthy samples and prior distribution samples in the discriminator output. During training, the discriminator is optimized by maximizing this distance, while the encoder is optimized by minimizing this distance.

[0085] Furthermore, the weld layer degradation fingerprint vector from the training auxiliary samples is input into the initial encoder to obtain the auxiliary latent space feature representation. This auxiliary latent space feature representation is then input into the initial discriminator to obtain the health distribution deviation output value. The monotonic correlation coefficient between the health distribution deviation output value and the weld layer physical degradation degree calibration value in the training auxiliary samples is calculated. The value minus the monotonic correlation coefficient is used as the value of the physical consistency constraint term. Here, the health distribution deviation output value is the degree of deviation between the auxiliary latent space feature representation output by the initial discriminator and the health state distribution (i.e., 1 minus the confidence score). The monotonic correlation coefficient measures the monotonic correlation between the health distribution deviation output value and the physical degradation degree calibration value, ranging from -1 to 1. The closer the monotonic correlation coefficient is to 1, the stronger the positive correlation between the two, meaning the deviation increases with the degree of degradation, which conforms to physical laws. The smaller the value of 1 minus the monotonic correlation coefficient, the more consistent the model output is with actual physical laws.

[0086] For example, the Spearman correlation coefficient can be used to calculate the monotonic correlation coefficient between the output value of the health distribution deviation and the calibration value of the physical degradation degree of the weld layer in the training auxiliary samples: Assume that the calibration value of the physical degradation degree of the 5 samples in the training auxiliary samples is d=[0,0.5,0.2,0.8,1.0], and the corresponding output value of the health distribution deviation (i.e., 1 minus the discriminator output probability) is p=[0.1,0.3,0.6,0.9,0.95]. First, d and p are arranged in ascending order and their rankings are recorded: the ranking of each value in the original sequence of d is [1,3,2,4,5], and the ranking of each value in the original sequence of p is [1,2,3,4,5]. The ranking difference Δi=rank(di)-rank(pi)=[0,1,-1,0,0] is calculated, and then the Spearman correlation coefficient is obtained. =1- =1-0.1=0.9, where n refers to the number of weld layer degradation fingerprint vectors in the training auxiliary samples. Then the value of the physical consistency constraint term =1-0.1=0.9, which means that the deviation assessment model output is highly consistent with the actual physical law and the penalty is small.

[0087] For example, conversely, if the ranking order of p is exactly the opposite of d, for instance, if p's ranking is [5,4,3,2,1] and d's ranking is [1,2,3,4,5], then Δi are [-4,-2,0,2,4] respectively, with a sum of squares of 40, and the Spearman correlation coefficient is calculated. =1- =1-240 / 120=-1, then the value of the physical consistency constraint term =1-(-1)=2. It can be seen that when the monotonicity is poor, the value of the physical consistency constraint term is larger, thus penalizing the deviation assessment model.

[0088] Furthermore, the value of the adversarial loss term is added to the value of the physical consistency constraint term to obtain the value of the training loss function; the parameters of the initial encoder and initial discriminator are then updated based on the value of the training loss function. Specifically, the value of the adversarial loss term is added to the value of the physical consistency constraint term to obtain the value of the training loss function, and then the gradient of the parameters of each layer of the initial encoder and initial discriminator is calculated through backpropagation based on this value. The parameters are then updated using an optimizer to reduce the value of the loss function, thereby improving the evaluation accuracy and physical consistency of the model.

[0089] For example, the Adam optimizer can be used for parameter updates. Let the adversarial loss for the current round be 0.6375, and the physical consistency constraint be 0; then the training loss value is 0.6375. The gradient of the loss with respect to all weights and biases in the initial encoder and initial discriminator is calculated through backpropagation. Then, the Adam optimizer is configured with a learning rate of 0.001, a first-order moment decay coefficient β1 = 0.9, and a second-order moment decay coefficient β2 = 0.999. The parameters are updated based on the calculated gradients. After multiple iterations, the loss function value gradually converges.

[0090] It should be noted that if the numerical ranges of the adversarial loss term and the physical consistency constraint term differ significantly, they can be multiplied by preset weighting coefficients before being summed to ensure that their contributions to the total loss are of similar magnitude. The specific values ​​of the weighting coefficients can be optimized through cross-validation or based on the monotonicity correlation coefficients of the model on the validation set, and can be determined by those skilled in the art based on actual training results. This weighting process is common knowledge in the field.

[0091] Finally, the training continues until a preset convergence condition is met, resulting in a fully trained deviation evaluation model. The preset convergence condition can be set as follows: the value of the training loss function no longer decreases over 10 consecutive iterations. When this preset convergence condition is met, training stops, the parameters of the encoder and discriminator are saved, and the fully trained deviation evaluation model is obtained.

[0092] In summary, compared to existing technologies, this application employs a pre-trained deviation assessment model to evaluate the degree of deviation of the weld layer degradation fingerprint vector from the health state distribution, and outputs a confidence score. Thus, by learning the health state distribution and outputting the deviation probability through a generative adversarial network, a quantitative basis for the reliability of the initial identification results is provided, effectively solving the problem of artificially high confidence levels in traditional classification models and providing reliable weight parameters for subsequent corrections.

[0093] S50: Based on the confidence score, the preliminary identification result is corrected to obtain the corrected weld fatigue identification result.

[0094] After initial identification and confidence assessment, preliminary results of the weld fatigue level and corresponding confidence scores were obtained. Since the preliminary identification results may be affected by model errors or data distribution deviations, and the confidence score reflects the degree of deviation between the current weld degradation fingerprint vector and the healthy state distribution, the preliminary identification results were corrected based on the confidence score to obtain corrected weld fatigue identification results, thus achieving a more reliable final identification result.

[0095] Specifically, step S50 in the method includes: The classification probability value is obtained from the preliminary identification result, and the classification probability value is multiplied by the confidence score to obtain the first corrected weight; Retrieve multiple historical samples from the historical case database that have the highest similarity to the weld layer degradation fingerprint vector, and obtain the weld layer fatigue level corresponding to the multiple historical samples; A weighted vote is performed on the fatigue levels of the weld layers corresponding to the multiple historical samples, and the fatigue level corresponding to the weighted vote result is used as the historical reference fatigue level. The fatigue level of the IGBT weld layer obtained from the most recent inspection is used as the lower limit constraint of the fatigue level. Multiply the first correction weight by the weld fatigue level in the preliminary identification result, multiply the value minus the first correction weight by the historical reference fatigue level, add the two products and compare them with the fatigue level lower limit constraint, and take the larger value as the corrected weld fatigue identification result.

[0096] In this embodiment, the classification probability value is first obtained from the preliminary identification result, and then multiplied by the confidence score to obtain the first correction weight. The first correction weight is used to quantify the reliability of the preliminary identification result. The classification probability value is the confidence score of the identification result output by the weld layer fatigue identification model (range 0-1); the confidence score is the probability value (range 0-1) output by the deviation assessment model indicating that the current weld layer degradation fingerprint vector belongs to the healthy state distribution. Multiplying the two yields the first correction weight. A larger first correction weight indicates a more reliable preliminary identification result, and a greater proportion of the preliminary identification result is retained during correction.

[0097] For example, if the classification probability is 0.7 and the confidence score is 0.3, then the first adjustment weight = 0.7 × 0.3 = 0.21.

[0098] Secondly, the system retrieves multiple historical samples with the highest similarity to the solder layer degradation fingerprint vector from the historical case database, and obtains the solder layer fatigue level corresponding to these samples. Specifically, the historical case database stores a large number of past IGBT solder layer detection cases, each containing a solder layer degradation fingerprint vector and the corresponding actual solder layer fatigue level. Cosine similarity or Euclidean distance can be used to calculate the similarity between the current solder layer degradation fingerprint vector to be identified and the fingerprint vector of each historical sample in the historical case database. Multiple historical samples with the highest similarity (e.g., 5) are then selected, and the solder layer fatigue levels corresponding to these samples are extracted as historical references.

[0099] Next, a weighted vote is performed on the fatigue levels of the weld layers corresponding to the multiple historical samples, and the fatigue level corresponding to the weighted voting result is used as the historical reference fatigue level. Specifically, firstly, a weight is assigned based on the similarity between each historical sample and the current weld layer degradation fingerprint vector. The higher the similarity, the closer the historical sample is to the current state, and the greater the weight should be given to its corresponding weld layer fatigue level in the vote. The similarity can be calculated using cosine similarity or the reciprocal of Euclidean distance and normalized to the 0-1 range. Then, for each weld layer fatigue level, the weights of its corresponding historical samples are accumulated to obtain the weighted vote count for that level. Finally, the weld layer fatigue level with the highest weighted vote count is selected as the historical reference fatigue level.

[0100] For example, five similar historical samples were retrieved, three of which showed moderate fatigue (cosine similarity of 0.9, 0.85, and 0.8 respectively), and two showed mild fatigue (cosine similarity of 0.7 and 0.65 respectively). First, the similarity was normalized: the total = 0.9 + 0.85 + 0.8 + 0.7 + 0.65 = 3.9. The normalized values ​​for each similarity were 0.9 / 3.9 = 0.231, 0.85 / 3.9 = 0.218, 0.8 / 3.9 = 0.205, 0.7 / 3.9 = 0.179, and 0.65 / 3.9 = 0.167. Therefore, the total weight for moderate fatigue was 0.231 + 0.218 + 0.205 = 0.654, and the total weight for mild fatigue was 0.179 + 0.167 = 0.346. The weighted votes for moderate fatigue are higher, therefore the historical reference fatigue level is moderate fatigue.

[0101] Furthermore, the fatigue level of the IGBT solder layer obtained from the most recent inspection is used as the lower limit constraint for the fatigue level. Specifically, solder layer fatigue is an irreversible cumulative degradation process. As the cumulative power cycle count increases, the fatigue level will only intensify or remain unchanged, not lessen. Therefore, the fatigue level of the IGBT solder layer obtained from the most recent inspection (numerically represented, e.g., healthy = 0, mild = 1, moderate = 2, severe = 3) is used as the lower limit constraint for this correction, ensuring that the corrected identification result is not lower than this level, thus conforming to the physical degradation law.

[0102] Finally, the first correction weight is multiplied by the weld fatigue level in the preliminary identification result, and the value minus the first correction weight is multiplied by the historical reference fatigue level. The two products are then added together and compared with the lower limit constraint of the fatigue level. The larger value is taken as the corrected weld fatigue identification result. The specific calculation consists of two steps: 1. Weighted fusion: The weld fatigue level in the preliminary identification result is weighted and fused with the historical reference fatigue level value (e.g., moderate = 2). The fusion level = (first correction weight × weld fatigue level value in the preliminary identification result) + ((1 - first correction weight) × historical reference fatigue level value). Among them, the first correction weight reflects the credibility of the preliminary identification result. The larger the weight, the higher the proportion of the preliminary identification result in the fusion.

[0103] 2. Lower limit constraint: Compare the lower limit constraint values ​​of the fusion level and fatigue level, and take the larger value as the corrected weld layer fatigue identification result. If necessary, continuous values ​​can be rounded or mapped back to discrete levels according to the threshold.

[0104] For example, if the initial identification result indicates slight fatigue (level 1, classification probability 0.7) and a confidence score of 0.3, then the first correction weight is 0.21; the historical reference fatigue level is moderate fatigue (level 2); and the most recent detection level is slight fatigue (level 1, i.e., lower limit 1). The fusion level is calculated as 0.21 × 1 + 0.79 × 2 = 1.79. Comparing this with the lower limit 1, the larger value of 1.79 is taken. Based on the set threshold (e.g., ≥1.5 is judged as moderate, <1.5 as slight), since 1.79 ≥ 1.5, the corrected weld fatigue identification result is moderate fatigue. This result integrates the initial identification and historical reference, and also satisfies the lower limit constraint of irreversible fatigue, ensuring physical consistency.

[0105] In summary, compared to existing technologies, this application corrects the preliminary identification result based on the confidence score to obtain a corrected weld layer fatigue identification result. Thus, by integrating the preliminary identification result, confidence score, historical similar cases, and fatigue irreversibility lower limit constraint, the accuracy and reliability of weld layer fatigue identification are effectively improved, avoiding potential biases that may arise from a single model.

[0106] In summary, the embodiments of this application have at least the following technical effects: Compared to existing technologies, this application first uses the thermal time constant of the IGBT solder layer as the acquisition benchmark to obtain voltage texture signals of IGBT switching transients within multiple switching cycles and constructs a voltage texture sequence. After extracting multiple candidate features from the voltage texture sequence, the candidate features are screened based on their monotonicity index on the cumulative power cycle count. Candidate features with monotonicity index exceeding a threshold are combined into a solder layer degradation fingerprint vector, thereby eliminating random fluctuation features. The solder layer degradation fingerprint vector is input into a pre-trained solder layer fatigue recognition model. The training loss function of this model includes a monotonicity constraint term to ensure that the model output conforms to the physical law that solder layer fatigue increases monotonically with the cumulative power cycle count, thus obtaining preliminary recognition results. Simultaneously, a deviation evaluation model based on generative adversarial networks is used to evaluate the deviation of the solder layer degradation fingerprint vector from the healthy state distribution and output a confidence score. Finally, the preliminary recognition results are corrected based on the confidence score. The classification probability value is multiplied by the confidence score to obtain the first correction weight. Combined with the weighted voting results of similar samples in the historical case library and the fatigue level lower limit constraint, the corrected solder layer fatigue recognition results are obtained.

[0107] Through the above technical solutions, this application constructs a complete technical chain, from voltage texture signal acquisition based on thermal time constant, monotonically degrading feature screening, and physical constraint-based solder layer fatigue identification model training, to deviation assessment and confidence correction based on generative adversarial networks. This effectively solves the technical problems of existing methods, such as strong subjectivity in feature selection, neglect of feature monotonically degrading laws, model output violating the physical consistency of solder layer fatigue, and lack of quantitative evaluation of the credibility of identification results. It improves the accuracy, reliability, and physical consistency of IGBT solder layer fatigue identification, and provides a high-confidence monitoring method for the health management of power devices.

[0108] Example 2, as Figure 2 As shown, based on the same inventive concept as the voltage texture-based IGBT solder layer fatigue identification method provided in Embodiment 1, this embodiment of the invention also provides an IGBT solder layer fatigue identification system based on voltage texture, including: The voltage texture acquisition module 11 is used to acquire voltage texture signals of IGBT switching transients within multiple switching cycles, based on the thermal time constant of the IGBT solder layer, to form a voltage texture sequence. The fingerprint vector construction module 12 is used to extract multiple candidate features from the voltage texture sequence, and to filter the candidate features according to the monotonicity index of each candidate feature on the cumulative power cycle number, and to combine the filtered features into a solder layer degradation fingerprint vector. The fatigue recognition module 13 is used to input the weld layer degradation fingerprint vector into a pre-trained weld layer fatigue recognition model to obtain a preliminary recognition result, wherein the training loss function of the weld layer fatigue recognition model includes a monotonicity constraint term. Deviation assessment module 14 is used to assess the degree of deviation of the solder layer degradation fingerprint vector relative to the health state distribution using a pre-trained deviation assessment model, and output a confidence score. The fusion correction module 15 is used to correct the preliminary identification result based on the confidence score to obtain the corrected weld fatigue identification result.

[0109] Specifically, the voltage texture acquisition module 11 is used for: Based on the material thermal parameters and geometric dimensions of the IGBT solder layer, the initial thermal time constant of the IGBT solder layer is obtained through finite element simulation calculation. Within each acquisition cycle, using the initial thermal time constant as the duration reference, the collector-emitter voltage overshoot waveform of the IGBT module at the moment of turn-off in multiple switching cycles is acquired, and the collector-emitter voltage overshoot waveform is used as a voltage texture signal. All voltage texture signals acquired in multiple consecutive acquisition cycles are arranged in chronological order to form a voltage texture sequence.

[0110] Specifically, the fingerprint vector construction module 12 is used for: From the voltage texture signal corresponding to each acquisition cycle of the voltage texture sequence, time-domain features and parasitic parameter sensitivity features are extracted. The time-domain features include voltage overshoot peak value, voltage overshoot amplitude to steady-state voltage ratio, and overshoot waveform area. The parasitic parameter sensitivity features include overshoot waveform oscillation frequency offset and overshoot waveform oscillation decay time constant change rate. The extracted time-domain features and the parasitic parameter sensitivity features are used together as candidate features; The feature values ​​of candidate features of the weld layer health status are obtained, the fluctuation standard deviation of the feature values ​​is calculated, and the monotonicity threshold is determined based on the fluctuation standard deviation. Obtain the feature value sequence corresponding to each candidate feature in multiple consecutive acquisition cycles, as well as the cumulative power cycle count sequence corresponding to each acquisition cycle; Based on the feature value sequence and the cumulative power cycle number sequence, calculate the monotonicity index value of the candidate feature as a function of the cumulative power cycle number; Candidate features whose monotonicity index values ​​exceed the monotonicity threshold are selected and combined into a solder layer degradation fingerprint vector.

[0111] Specifically, based on the feature value sequence and the cumulative power cycle number sequence, the monotonicity index value of the candidate feature changing with the cumulative power cycle number is calculated, including: The feature values ​​corresponding to each acquisition cycle in the feature value sequence are arranged in ascending order of the cumulative power cycle number to form a monotonicity analysis feature value sequence. Calculate the difference between the eigenvalues ​​of each adjacent position in the monotonicity analysis eigenvalue sequence, and count the number of positive and negative signs in the difference values; Calculate the absolute value of the difference between the number of occurrences of the positive sign and the number of occurrences of the negative sign, and use the absolute value of the difference as the monotonicity index value.

[0112] The fatigue recognition module 13 is specifically used for: Construct an initial weld layer fatigue identification model; Multiple sets of training samples are obtained, each set of training samples contains a solder layer degradation fingerprint vector and the corresponding solder layer fatigue level label; Construct a total loss function, which is obtained by adding a classification loss term and a monotonicity constraint term; The initial weld layer fatigue identification model is trained in a supervised manner using the multiple sets of training samples and the total loss function until the preset convergence condition is met, thus obtaining the trained weld layer fatigue identification model. The weld layer degradation fingerprint vector is input into the trained weld layer fatigue recognition model, and the weld layer fatigue level and the classification probability value corresponding to the weld layer fatigue level are output. The weld layer fatigue level and the classification probability value are used together as the preliminary recognition result.

[0113] Specifically, the process of constructing the total loss function includes: Based on the cross-entropy loss function, a classification loss term is constructed; For the same IGBT module, the two solder layer degradation fingerprint vectors corresponding to two adjacent acquisition cycles are respectively input into the initial solder layer fatigue recognition model to obtain the first solder layer fatigue level output value and the second solder layer fatigue level output value. Obtain the first cumulative power cycle number corresponding to the fatigue level output value of the first weld layer, and the second cumulative power cycle number corresponding to the fatigue level output value of the second weld layer; If the first cumulative power cycle count is greater than the second cumulative power cycle count, and the first weld fatigue level output value is less than the second weld fatigue level output value, then the difference between the two is taken as the value of the monotonicity constraint term; otherwise, the value of the monotonicity constraint term is zero. The classification loss term is added to the monotonicity constraint term to obtain the total loss function.

[0114] Specifically, the deviation evaluation module 14 is used for: Obtain a pre-trained deviation assessment model, wherein the deviation assessment model is constructed based on a generative adversarial network and includes an encoder and a discriminator; The solder layer degradation fingerprint vector is input into the deviation evaluation model, and the encoder maps the solder layer degradation fingerprint vector to the latent space feature representation. The latent space feature representation is input into the discriminator, which outputs the probability value that the solder layer degradation fingerprint vector belongs to the healthy state distribution. The probability value is used as the confidence score.

[0115] Specifically, the training process of the deviation assessment model includes: Collect multiple solder layer degradation fingerprint vectors corresponding to the IGBT solder layer being in a healthy state to form a healthy state sample set; Multiple sets of training auxiliary samples are collected. Each set of training auxiliary samples contains a solder layer degradation fingerprint vector and a corresponding solder layer physical degradation degree calibration value. An initial deviation assessment model is constructed, which includes an initial encoder and an initial discriminator; Determine the training loss function, which is composed of an adversarial loss term and a physical consistency constraint term; The initial deviation assessment model is trained iteratively in multiple rounds, and the following steps are performed in each round of iteration: The weld layer degradation fingerprint vector in the health status sample set is input into the initial encoder, the initial encoder outputs the latent space feature representation, the latent space feature representation is input into the initial discriminator, the initial discriminator outputs the true / false discrimination result, and the value of the adversarial loss term is calculated based on the true / false discrimination result; The solder layer degradation fingerprint vector in the training auxiliary sample is input into the initial encoder to obtain the auxiliary latent space feature representation. The auxiliary latent space feature representation is input into the initial discriminator to obtain the healthy distribution deviation output value. The monotonicity correlation coefficient between the healthy distribution deviation output value and the solder layer physical degradation degree calibration value in the training auxiliary sample is calculated. The value of subtracting the monotonicity correlation coefficient is used as the value of the physical consistency constraint term. The value of the adversarial loss term is added to the value of the physical consistency constraint term to obtain the value of the training loss function, and the parameters of the initial encoder and the initial discriminator are updated according to the value of the training loss function; The training continues until the preset convergence condition is met, at which point the deviation evaluation model is obtained.

[0116] The fusion correction module 15 is specifically used for: The classification probability value is obtained from the preliminary identification result, and the classification probability value is multiplied by the confidence score to obtain the first corrected weight; Retrieve multiple historical samples from the historical case database that have the highest similarity to the weld layer degradation fingerprint vector, and obtain the weld layer fatigue level corresponding to the multiple historical samples; A weighted vote is performed on the fatigue levels of the weld layers corresponding to the multiple historical samples, and the fatigue level corresponding to the weighted vote result is used as the historical reference fatigue level. The fatigue level of the IGBT weld layer obtained from the most recent inspection is used as the lower limit constraint of the fatigue level. Multiply the first correction weight by the weld fatigue level in the preliminary identification result, multiply the value minus the first correction weight by the historical reference fatigue level, add the two products and compare them with the fatigue level lower limit constraint, and take the larger value as the corrected weld fatigue identification result.

[0117] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

Claims

1. A fatigue identification method for IGBT solder layers based on voltage texture, characterized in that, The method includes: Using the thermal time constant of the IGBT solder layer as the acquisition reference, voltage texture signals of IGBT switching transients within multiple switching cycles are obtained to form a voltage texture sequence. Multiple candidate features are extracted from the voltage texture sequence, and the candidate features are screened according to the monotonicity index of each candidate feature on the cumulative power cycle number. The screened features are combined into a solder layer degradation fingerprint vector. The degradation fingerprint vector of the weld layer is input into the pre-trained weld layer fatigue recognition model to obtain preliminary recognition results. The training loss function of the weld layer fatigue recognition model includes a monotonicity constraint term. A pre-trained deviation assessment model is used to evaluate the degree of deviation of the solder layer degradation fingerprint vector from the health state distribution and output a confidence score. The preliminary identification result is corrected based on the confidence score to obtain the corrected weld fatigue identification result. Among them, the voltage texture signal of IGBT switching transients within multiple switching cycles is acquired using the thermal time constant of the IGBT solder layer as the acquisition reference, forming a voltage texture sequence, including: Based on the material thermal parameters and geometric dimensions of the IGBT solder layer, the initial thermal time constant of the IGBT solder layer is obtained through finite element simulation calculation. Within each acquisition cycle, using the initial thermal time constant as the duration reference, the collector-emitter voltage overshoot waveform of the IGBT module at the moment of turn-off in multiple switching cycles is acquired, and the collector-emitter voltage overshoot waveform is used as a voltage texture signal. All voltage texture signals acquired in multiple consecutive acquisition cycles are arranged in chronological order to form a voltage texture sequence. Specifically, a pre-trained deviation assessment model is used to evaluate the deviation of the solder layer degradation fingerprint vector from the healthy state distribution, and outputs a confidence score, including: Obtain a pre-trained deviation assessment model, wherein the deviation assessment model is constructed based on a generative adversarial network and includes an encoder and a discriminator; The solder layer degradation fingerprint vector is input into the deviation evaluation model, and the encoder maps the solder layer degradation fingerprint vector to the latent space feature representation. The latent space feature representation is input into the discriminator, which outputs the probability value that the solder layer degradation fingerprint vector belongs to the healthy state distribution. The probability value is used as the confidence score.

2. The IGBT solder layer fatigue identification method based on voltage texture according to claim 1, characterized in that, Multiple candidate features are extracted from the voltage texture sequence, and the candidate features are filtered based on the monotonicity index of each candidate feature on the cumulative power cycle number. The filtered features are combined into a solder layer degradation fingerprint vector, including: From the voltage texture signal corresponding to each acquisition cycle of the voltage texture sequence, time-domain features and parasitic parameter sensitivity features are extracted. The time-domain features include voltage overshoot peak value, voltage overshoot amplitude to steady-state voltage ratio, and overshoot waveform area. The parasitic parameter sensitivity features include overshoot waveform oscillation frequency offset and overshoot waveform oscillation decay time constant change rate. The extracted time-domain features and the parasitic parameter sensitivity features are used together as candidate features; The feature values ​​of candidate features of the weld layer health status are obtained, the fluctuation standard deviation of the feature values ​​is calculated, and the monotonicity threshold is determined based on the fluctuation standard deviation. Obtain the feature value sequence corresponding to each candidate feature in multiple consecutive acquisition cycles, as well as the cumulative power cycle count sequence corresponding to each acquisition cycle; Based on the feature value sequence and the cumulative power cycle number sequence, calculate the monotonicity index value of the candidate feature as a function of the cumulative power cycle number; Candidate features whose monotonicity index values ​​exceed the monotonicity threshold are selected and combined into a solder layer degradation fingerprint vector.

3. The IGBT solder layer fatigue identification method based on voltage texture according to claim 2, characterized in that, Based on the feature value sequence and the cumulative power cycle number sequence, calculate the monotonicity index value of the candidate feature as a function of the cumulative power cycle number, including: The feature values ​​corresponding to each acquisition cycle in the feature value sequence are arranged in ascending order of the cumulative power cycle number to form a monotonicity analysis feature value sequence. Calculate the difference between the eigenvalues ​​of each adjacent position in the monotonicity analysis eigenvalue sequence, and count the number of positive and negative signs in the difference values; Calculate the absolute value of the difference between the number of occurrences of the positive sign and the number of occurrences of the negative sign, and use the absolute value of the difference as the monotonicity index value.

4. The IGBT solder layer fatigue identification method based on voltage texture according to claim 1, characterized in that, The weld layer degradation fingerprint vector is input into a pre-trained weld layer fatigue recognition model to obtain preliminary recognition results. The training loss function of the weld layer fatigue recognition model includes a monotonicity constraint term, including: Construct an initial weld layer fatigue identification model; Multiple sets of training samples are obtained, each set of training samples contains a solder layer degradation fingerprint vector and the corresponding solder layer fatigue level label; Construct a total loss function, which is obtained by adding a classification loss term and a monotonicity constraint term; The initial weld layer fatigue identification model is trained in a supervised manner using the multiple sets of training samples and the total loss function until the preset convergence condition is met, thus obtaining the trained weld layer fatigue identification model. The weld layer degradation fingerprint vector is input into the trained weld layer fatigue recognition model, and the weld layer fatigue level and the classification probability value corresponding to the weld layer fatigue level are output. The weld layer fatigue level and the classification probability value are used together as the preliminary recognition result.

5. The IGBT solder layer fatigue identification method based on voltage texture according to claim 4, characterized in that, The process of constructing the total loss function includes: Based on the cross-entropy loss function, a classification loss term is constructed; For the same IGBT module, the two solder layer degradation fingerprint vectors corresponding to two adjacent acquisition cycles are respectively input into the initial solder layer fatigue recognition model to obtain the first solder layer fatigue level output value and the second solder layer fatigue level output value. Obtain the first cumulative power cycle number corresponding to the fatigue level output value of the first weld layer, and the second cumulative power cycle number corresponding to the fatigue level output value of the second weld layer; If the first cumulative power cycle count is greater than the second cumulative power cycle count, and the first weld fatigue level output value is less than the second weld fatigue level output value, then the difference between the two is taken as the value of the monotonicity constraint term; otherwise, the value of the monotonicity constraint term is zero. The classification loss term is added to the monotonicity constraint term to obtain the total loss function.

6. The IGBT solder layer fatigue identification method based on voltage texture according to claim 1, characterized in that, The training process of the deviation assessment model includes: Collect multiple solder layer degradation fingerprint vectors corresponding to the IGBT solder layer being in a healthy state to form a healthy state sample set; Multiple sets of training auxiliary samples are collected. Each set of training auxiliary samples contains a solder layer degradation fingerprint vector and a corresponding solder layer physical degradation degree calibration value. An initial deviation assessment model is constructed, which includes an initial encoder and an initial discriminator; Determine the training loss function, which is composed of an adversarial loss term and a physical consistency constraint term; The initial deviation assessment model is trained iteratively in multiple rounds, and the following steps are performed in each round of iteration: The weld layer degradation fingerprint vector in the health status sample set is input into the initial encoder, the initial encoder outputs the latent space feature representation, the latent space feature representation is input into the initial discriminator, the initial discriminator outputs the true / false discrimination result, and the value of the adversarial loss term is calculated based on the true / false discrimination result; The solder layer degradation fingerprint vector in the training auxiliary sample is input into the initial encoder to obtain the auxiliary latent space feature representation. The auxiliary latent space feature representation is input into the initial discriminator to obtain the healthy distribution deviation output value. The monotonicity correlation coefficient between the healthy distribution deviation output value and the solder layer physical degradation degree calibration value in the training auxiliary sample is calculated. The value of subtracting the monotonicity correlation coefficient is used as the value of the physical consistency constraint term. The value of the adversarial loss term is added to the value of the physical consistency constraint term to obtain the value of the training loss function, and the parameters of the initial encoder and the initial discriminator are updated according to the value of the training loss function; The training continues until the preset convergence condition is met, at which point the deviation evaluation model is obtained.

7. The IGBT solder layer fatigue identification method based on voltage texture according to claim 1, characterized in that, The preliminary identification result is corrected based on the confidence score to obtain the corrected weld fatigue identification result, including: The classification probability value is obtained from the preliminary identification result, and the classification probability value is multiplied by the confidence score to obtain the first corrected weight; Retrieve multiple historical samples from the historical case database that have the highest similarity to the weld layer degradation fingerprint vector, and obtain the weld layer fatigue level corresponding to the multiple historical samples; The fatigue levels of the weld layers corresponding to the multiple historical samples are weighted and voted on, and the fatigue level corresponding to the weighted voting result is used as the historical reference fatigue level. The fatigue level of the IGBT weld layer obtained from the most recent inspection is used as the lower limit constraint of the fatigue level. Multiply the first correction weight by the weld fatigue level in the preliminary identification result, multiply the value minus the first correction weight by the historical reference fatigue level, add the two products and compare them with the fatigue level lower limit constraint, and take the larger value as the corrected weld fatigue identification result.

8. A voltage texture-based IGBT solder layer fatigue identification system, characterized in that, The method for performing IGBT solder layer fatigue identification based on voltage texture as described in any one of claims 1-7 includes: The voltage texture acquisition module is used to acquire voltage texture signals of IGBT switching transients within multiple switching cycles, based on the thermal time constant of the IGBT solder layer, to form a voltage texture sequence. The fingerprint vector construction module is used to extract multiple candidate features from the voltage texture sequence, and to filter the candidate features according to the monotonicity index of each candidate feature on the cumulative power cycle number, and to combine the filtered features into a solder layer degradation fingerprint vector. The fatigue recognition module is used to input the weld layer degradation fingerprint vector into a pre-trained weld layer fatigue recognition model to obtain preliminary recognition results. The training loss function of the weld layer fatigue recognition model includes a monotonicity constraint term. The deviation assessment module is used to evaluate the degree of deviation of the solder layer degradation fingerprint vector from the health state distribution using a pre-trained deviation assessment model, and output a confidence score. The fusion correction module is used to correct the preliminary identification result based on the confidence score to obtain the corrected weld fatigue identification result.

Images