A method and system for IGBT aging prediction

Abstract

The application is suitable for the technical field of aging prediction, and provides an IGBT aging prediction method, which comprises the following steps: preprocessing and aging feature extraction are performed on an IGBT aging data set, and a model training set and a model test set are divided; an IGBT aging prediction model is constructed, an octopus optimization algorithm is used to perform parameter optimization on the IGBT aging prediction model, and a global optimal parameter combination is obtained; based on the global optimal parameter combination, the IGBT aging prediction model is trained in combination with the model training set; and based on the model test set, the prediction accuracy of the IGBT aging prediction model after training is evaluated. In view of the problems of low prediction accuracy and large prediction uncertainty of the existing IGBT aging prediction model, an IGBT aging prediction model based on an optimized OOA-CNN-BILSTM-Attention is proposed, and the accuracy of IGBT aging prediction is effectively improved.

Description

Technical Field

[0001] This application belongs to the field of aging prediction technology, and in particular relates to an IGBT aging prediction method and system. Background Technology

[0002] Insulated-gate bipolar transistors (IGBTs) are important power electronic components widely used in smart grids, electric vehicles, and renewable energy. Specifically, IGBTs combine the high input impedance of MOSFETs with the low on-state voltage drop of power transistors, offering high efficiency, low loss, high voltage resistance, and high current handling capabilities. However, as IGBTs age and fail under various stresses and other factors, they can eventually cause electrical equipment to stop working. Therefore, it is necessary to predict IGBT aging, allowing for pre-failure maintenance, reducing inspection costs, and improving IGBT operational reliability.

[0003] Current technologies for IGBT aging prediction generally fall into two main categories: model-based and data-driven. Model-based prediction methods can be further divided into physical model-based and analytical model-based methods. For physical models, the lifespan is related to the strain generated by the material under stress. For analytical models, the prediction accuracy mainly depends on the number of variables included in the model; as the number of variables increases, the prediction accuracy improves. Data-driven lifespan prediction methods, on the other hand, do not rely on complex physical models and do not require in-depth modeling of IGBT failure mechanisms. They only need to extract information characterizing the IGBT's aging state from historical data, offering greater adaptability and flexibility.

[0004] However, the aforementioned traditional methods lack the ability to adaptively focus on important features in time-series data, making it difficult to accurately predict IGBT aging trends. Summary of the Invention

[0005] This application provides an IGBT aging prediction method and system, which can solve one of the problems of the prior art mentioned above.

[0006] In a first aspect, embodiments of this application provide an IGBT aging prediction method, including: The IGBT aging dataset is preprocessed and aging features are extracted, and the dataset is divided into a model training set and a model test set. An IGBT aging prediction model was constructed, and the Octopus Optimization Algorithm was used to optimize the parameters of the IGBT aging prediction model to obtain the globally optimal parameter combination. Based on the globally optimal parameter combination and the model training set, the IGBT aging prediction model is trained. Based on the model test set, the prediction accuracy of the trained IGBT aging prediction model is evaluated.

[0007] Furthermore, the IGBT aging dataset is either a Vce-peak dataset or a Vce-on dataset; For the Vce-peak dataset, aging features are obtained by decomposition and reconstruction based on the VMD algorithm. For the Vce-on dataset, the data is smoothed using a simple moving average method, and then simplified using a segmented aggregation approximation method to obtain aging characteristics.

[0008] Furthermore, for the Vce-peak dataset, decomposition is performed using VMD decomposition to obtain aging features, including: The Vce-peak monitoring signal is decomposed into multiple bandpass mode components using the VMD algorithm; Extract the physical parameter sequence, and calculate the correlation coefficient between the physical parameter sequence and each of the bandpass mode components; Based on the correlation coefficient, each bandpass modal component is divided into different modal types, and modal types are screened based on a preset reconstruction threshold to determine the effective modal set and obtain aging characteristics.

[0009] Furthermore, the physical parameter sequence is a time series reflecting the evolution of physical damage within the IGBT; The extraction of the physical parameter sequence includes: Extract the junction temperature curve of the IGBT as the first physical parameter sequence; Rainflow counting is performed on the junction temperature curve to extract the temperature difference amplitude and average junction temperature for each temperature cycle. For each temperature cycle, based on the temperature difference amplitude and combined with the material mechanics formulas, the stress parameters generated in the solder layer are calculated; Based on the stress parameters, the crack propagation rate is obtained by analytically solving the Paris law. Record the crack propagation rate obtained from each temperature cycle to generate a second sequence of physical parameters.

[0010] Furthermore, the mode types include physically degenerate dominant modes, thermo-coupled wave modes, and high-frequency noise-independent modes; The bandpass mode component in the dominant physical degradation mode is highly positively correlated with the crack propagation rate; The bandpass mode component in the thermo-coupled wave mode is related to the junction temperature curve of the IGBT; The bandpass mode component in the high-frequency noise-independent mode is represented as random high-frequency oscillation, which is independent of the evolution law of physical damage inside the IGBT.

[0011] Furthermore, the IGBT aging prediction model includes an input layer, a CNN layer, a BiLSTM layer, an attention mechanism layer, a fully connected layer, and an output layer; wherein, The CNN layer is used for local feature parameter extraction and feature parameter mapping; The BiLSTM layer is used to mine temporal information from the feature parameters; The attention mechanism layer is used to calculate a weight vector for the feature parameter values ​​output by the BiLSTM layer, representing the importance of the current time step to different input parts.

[0012] Furthermore, the step of using the octopus optimization algorithm to optimize the parameters of the IGBT aging prediction model to obtain the globally optimal parameter combination includes: Initialize the octopus population by dividing it into predator individuals and scout individuals, and set the optimization parameter range and fitness function; During the exploration phase, the location of the individual scouts is updated and their location information is fed back to the predator group, and the location of the individual predators is updated. During the development phase, the location of the predator individuals is updated by narrowing the search range; After each location update, check whether the individual's new location exceeds the search space boundary, and change the location if the fitness of the new location is better than the original location; Iterate through all individuals in the octopus population, update the global optimal solution, and obtain the global optimal parameter combination.

[0013] Furthermore, the range of optimization parameters includes the search range for the number of BiLSTM neurons; For the search range of the number of BiLSTM neurons, an autocorrelation function is constructed, the effective memory length of the time series data used for training is calculated, and the search range is adjusted based on the effective memory length. If the IGBT aging dataset is the Vce-peak dataset, then based on the effective memory length and the number of bandpass modal components in the effective modality set, the upper limit of the number of BiLSTM neurons is dynamically calculated.

[0014] Furthermore, the construction of the autocorrelation function and the calculation of the effective memory length of the time series data used for training include: Calculate the autocorrelation coefficient of time series data at different lag orders and generate autocorrelation curves; Set a cutoff threshold, and find the minimum delay order from the autocorrelation curve that first decays and stabilizes within the cutoff threshold range. The minimum delay order is the effective memory length, indicating that after the effective memory length is exceeded, the time series data loses its predictive value for the current IGBT degradation state.

[0015] Secondly, embodiments of this application provide an IGBT aging prediction system, including: The first processing module is used to preprocess the IGBT aging dataset and extract aging features, and to divide the dataset into a model training set and a model test set. The second processing module is used to construct an IGBT aging prediction model and use the Octopus optimization algorithm to optimize the parameters of the IGBT aging prediction model to obtain the globally optimal parameter combination. The third processing module is used to train the IGBT aging prediction model based on the globally optimal parameter combination and the model training set. The fourth processing module is used to evaluate the prediction accuracy of the trained IGBT aging prediction model based on the model test set.

[0016] The beneficial effects of the embodiments in this application compared with the prior art are: This application discloses an IGBT aging prediction method. Addressing the issues of low prediction accuracy and high prediction uncertainty in existing IGBT aging prediction models, this application proposes an optimized IGBT aging prediction model based on OOA-CNN-BILSTM-Attention. During training, for the Vce-peak dataset, the model uses the VMD algorithm to decompose the IGBT aging data Vce-peak monitoring into multiple modes and reconstructs the useful modes, thereby extracting and enhancing IGBT degradation features. For the Vce-on dataset, the Vce-on signal is first smoothed using a simple moving average method, and then simplified using a piecewise aggregation approximation method. This simplifies the data while preserving the characteristic trends of the Vce-on data, reducing computational complexity. Furthermore, unlike traditional IGBT aging prediction models that rely on empirical trial-and-error methods for hyperparameter selection, this model employs the Octopus Optimization algorithm to globally optimize parameters such as the learning rate, regularization coefficient, number of BiLSTM neurons, and attention mechanism key values ​​in the IGBT aging model to improve the model's predictive performance. Simultaneously, regarding the search range for the number of BiLSTM neurons, a boundary is constructed based on the features of the input data, transforming the blind "global static grid" into a "local dynamic adaptive domain." This effectively avoids the risk of overfitting and significantly reduces the exploration process of the invalid solution space, thereby significantly improving the convergence speed and final prediction accuracy of the model's hyperparameter optimization. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating an IGBT aging prediction method according to an embodiment of the present invention; Figure 2 It is the degradation curve of the collector-emitter transient peak voltage Vce-peak of the IGBT; Figure 3 This is the IMF component diagram after decomposing the collector-emitter transient peak voltage Vce-peak of the IGBT; Figure 4 This is a graph showing the collector-emitter voltage (Vce-on) data of an IGBT. Figure 5 This is the schematic diagram of LSTM; Figure 6 This is a schematic diagram of a BiLSTM. Figure 7 yes Figure 1 The implementation process of the IGBT aging prediction model in the illustrated embodiment; Figure 8 yes Figure 1 The example shown is a prediction result graph of the Vce-peak model with a training set of 0.8. Figure 9 yes Figure 1 The example shown illustrates the prediction results of the Vce-peak model with a training set of 0.3. Figure 10 yes Figure 1 The example shown illustrates the prediction results of the Vce-on model with a training set of 0.8. Figure 11 yes Figure 1 The example shown illustrates the prediction results of the Vce-on model with a training set of 0.3. Figure 12 yes Figure 1 The Vce-on model prediction results of CRRC IGBT1 with a training set of 0.8 in the embodiment shown are illustrated. Figure 13 yes Figure 1 The Vce-on model prediction results of CRRC IGBT2 with a training set of 0.8 in the embodiment shown are illustrated. Figure 14This is a schematic diagram of the structure of an IGBT aging prediction system provided in an embodiment of the present invention; Figure 15 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0019] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0020] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0021] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0022] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0023] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0024] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0025] Please see Figure 1 As shown, this invention is an IGBT aging prediction method. The failure mechanism of IGBT is the basis for its life prediction. By studying the failure mechanism, a reasonable explanation can be given for the failure behavior. At the same time, the electrical and thermal performance characteristic parameters closely related to the degree of degradation can be identified to make a reasonable life prediction. It is worth noting that IGBT failure behavior is generally divided into chip-related failures and package-related failures. Chip-related failures are mostly instantaneous and mainly include electrostatic discharge, latch-up effect, cosmic ray radiation, overvoltage breakdown, and electrical stress failure caused by overcurrent short circuits. When using power devices, protection circuits are generally configured, so chip-related failures are not likely to occur. Package-related failures are fatigue damage caused by the accumulation of thermal stress. This is mainly due to the heat conduction in the material layers of power devices such as IGBTs during long-term operation, and the temperature gradient caused by the difference in the coefficients of thermal expansion of different parts. This causes changes in the dislocation substructure during cyclic deformation of the material, forming stationary slip bands, which leads to the accumulation of thermal stress differences and eventually cracks. On the solder layer, this usually manifests as solder layer fatigue cracks that gradually propagate and eventually lead to solder layer delamination. On the bond wires, it manifests as bond wire peeling, warping, melting, or even detachment. Under normal circumstances, about 60% of device failures are caused by thermal failure.

[0026] Based on this, this application specifically includes the following steps: S100. Preprocess the IGBT aging dataset and extract aging features, and divide it into a model training set and a model test set. Under normal operating conditions, the lifespan of IGBTs in applications such as photovoltaics is 10 to 15 years. Therefore, accelerated aging tests are used to analyze the reliability of IGBTs. Specifically, accelerated aging tests increase stress to achieve rapid stress accumulation without changing the failure mechanism, so as to accelerate the failure of the device under test. For example, the IGBT accelerated aging test dataset published by NASA's PCoE Research Center uses IRG4BC30KD IGBTs with a rated voltage of 600V, a rated current of 15A, and a TO220 package. A 0.2Ω resistor is connected in series with the IGBT collector as a load, and a 4V regulated power supply is used for power supply. A drive signal with a voltage of 10V, a duty cycle of 40%, and a switching frequency of 10 kHz is applied to the IGBT gate. The package temperature is controlled between 533.15 and 543.15K. During the aging test, parameters such as gate-emitter voltage (Vge), collector-emitter voltage (Vce), and collector-emitter current (Ice) are recorded to form different IGBT aging datasets with corresponding characteristic parameters.

[0027] In some embodiments, the IGBT aging dataset is a Vce-peak dataset or a Vce-on dataset; For the Vce-peak dataset, aging features are obtained by decomposition and reconstruction based on the VMD algorithm. For the Vce-on dataset, the data is smoothed using a simple moving average method, and then simplified using a segmented aggregation approximation method to obtain aging characteristics.

[0028] Specifically, during the degradation process of IGBT, its characteristic parameters change with its operating state. Selecting appropriate degradation characteristic parameters can improve the prediction accuracy of the model. There are many characteristic parameters that can characterize IGBT degradation, including the collector-emitter voltage (Vce-on) signal, which is sensitive to latch-up effect and bond wire detachment; the gate leakage current (Ig), which can characterize gate oxide layer degradation; and the collector current (Ic) signal, gate-emitter voltage (Vge), turn-on / turn-off time (Ton / Toff), etc., which characterize solder layer failure.

[0029] It is worth noting that during IGBT turn-off, a transient voltage (Vce-peak) is generated due to the presence of parasitic components such as stray inductance. This transient voltage will decrease as the IGBT's performance degrades until latch-up occurs. This characteristic can well reflect the performance degradation pattern of the IGBT, such as... Figure 2As shown, in one embodiment of this application, an IGBT aging dataset is constructed based on the transient voltage Vce-peak for model training. Furthermore, the aforementioned transient voltage Vce-peak is the collector-emitter transient peak voltage, which is extracted from the collector-emitter voltage Vce-on. Although the collector-emitter transient peak voltage Vce-peak can characterize the degradation features of IGBTs, relevant literature indicates that the Vce-on signal has stronger performance compared to other aging parameters when considering six aspects: linearity, sensitivity, accuracy, versatility, capability, and online measurement. Therefore, in another embodiment of this application, an IGBT aging dataset is also constructed based on the Vce-on signal for model training.

[0030] In addition, for the raw measurements in the IGBT aging dataset, preprocessing is required during feature extraction to remove outliers that significantly deviate from the normal range in order to ensure data quality. Furthermore, different aging feature extraction methods are used for different types of aging datasets.

[0031] Specifically, the Vce-peak monitoring signal is the peak voltage value of the Vce voltage signal during the turn-off period. That is, the Vce-peak monitoring signal is a feature parameter extracted from the Vce voltage signal. It is difficult to obtain directly under actual working conditions, so its sample data is small and has the problems of high nonlinearity and strong fluctuation. Therefore, for the Vce-peak monitoring signal in the Vce-peak dataset, this application adopts the variational mode decomposition method, namely the VMD algorithm, to decompose it into several intrinsic mode functions and extract key features that can characterize the aging state of IGBT.

[0032] In some embodiments, the step of decomposing the Vce-peak dataset using VMD decomposition to obtain aging features includes: The Vce-peak monitoring signal is decomposed into multiple bandpass mode components using the VMD algorithm; Extract the physical parameter sequence, and calculate the correlation coefficient between the physical parameter sequence and each of the bandpass mode components; Based on the correlation coefficient, each bandpass modal component is divided into different modal types, and modal types are screened based on a preset reconstruction threshold to determine the effective modal set and obtain aging characteristics.

[0033] Specifically, the VMD algorithm, proposed by Konstantin Dragomiretskiy in 2014, can decompose complex signals into several amplitude-modulated and frequency-modulated components and obtain the optimal solution. It is suitable for processing nonlinear and non-stationary signals. Compared with other common EMD decomposition and ensemble empirical mode decomposition methods, the VMD algorithm can effectively solve the mode aliasing problem and improve the stability of the algorithm by introducing constraints and regularization terms. Furthermore, VMD can accelerate the decomposition speed and reduce computational complexity by introducing auxiliary variables and fast Fourier transform. Therefore, this application adopts the VMD algorithm to decompose the failure characteristic parameter Vce-peak monitoring signal, which characterizes the aging state of IGBTs, into multiple modal components, thereby enhancing the aging characteristics and solving the problems of small sample data, high nonlinearity, and strong fluctuations in the failure characteristic parameter Vce-peak monitoring signal.

[0034] Therefore, through the VMD algorithm, the original non-stationary signal Vce-pea monitoring signal is adaptively decomposed into K bandpass mode components imf with center frequencies. K Specifically, the Vce-peak monitoring signal is subjected to Hilbert transform, and the signal spectrum is modulated to the corresponding frequency band: ,in, Let be the impact function; The k-th modal component after decomposition; t is time; For corresponding bandpass mode components The center frequency; It is a complex exponential function used to perform frequency shifting operations on signals, thereby shifting the modal components... And its related signals were shifted from the original frequency to a frequency that is Calculate the squared gradient of the demodulated signal for the frequency band of the center frequency. The norm, and its constrained variational model are as follows: ; in, This represents the set of all bandpass mode components obtained from the decomposition. This represents the set of center frequencies corresponding to all bandpass mode components. This represents a constraint condition, specifically that the sum of all bandpass mode components must equal the original input signal. .

[0035] Furthermore, by introducing the Lagrangian function, the constrained problem is transformed into an unconstrained problem, as follows: ; in: For Lagrange daily numbers; As a penalty factor, Represents Lagrange multipliers With constraints The inner product is then calculated using an alternating direction multiplier algorithm, continuously updating until the K imf components are found. ,Right now ,in, It is a time constant. This represents the original input signal in the frequency domain. This represents the frequency domain representation of the k-th bandpass mode component during the (n+1)-th iteration, while also setting the decision precision. When satisfied When the iteration stops, K IMF components with center frequencies are obtained, such as... Figure 3 As shown, in one embodiment, after VDM decomposition of the Vce-peak monitoring signal, seven IMF components are obtained.

[0036] Furthermore, to overcome the "black box" nature of signal processing, this application introduces Paris's law for solder layer fatigue crack propagation rate and a junction temperature fluctuation model to physically calibrate each component. Specifically, during the long-term service of IGBTs, the turn-off peak voltage Vce-peak serves as an important precursor parameter characterizing the device's health status. Its evolution is coupled with slow physical degradation within the device (such as solder layer fatigue and bond wire detachment), transient thermal fluctuations caused by operating conditions, and external measurement noise. Among these, fatigue crack propagation in the IGBT solder layer leads to increased device thermal resistance ( The significant increase in Rth leads to an increase in junction temperature (Tj). According to the physical characteristics of semiconductors, the rise in junction temperature and the degradation of internal interconnect inductance will have a combined effect on the turn-off transient, causing Vce-peak to show a non-stationary downward trend. This trend is an irreversible monotonic trend and is a low-frequency macroscopic degradation scale. The transient thermal fluctuations caused by operating conditions are periodic or quasi-periodic changes in junction temperature Tj caused by load fluctuations, which belong to the medium-low frequency operating condition thermal fluctuation scale. External measurement noise is random high-frequency white noise introduced by electromagnetic interference and sensor accuracy, which belongs to the high-frequency environmental and measurement noise scale.

[0037] When extracting aging features from Vce-peak monitoring signals, high-frequency environmental and measurement noise can severely interfere with the accuracy of aging feature extraction. Therefore, in order to accurately extract the degradation trajectory that can characterize the true health state of IGBTs, physical semantic mapping needs to be performed when extracting aging features from each IMF component of the Vce-peak monitoring signal. Furthermore, noise reduction and reconstruction of the signal based on correlation analysis are required to improve the accuracy of aging feature extraction.

[0038] In some embodiments, the physical parameter sequence is a time series reflecting the evolution of physical damage within the IGBT; The extraction of the physical parameter sequence includes: Extract the junction temperature curve of the IGBT as the first physical parameter sequence; Rainflow counting is performed on the junction temperature curve to extract the temperature difference amplitude and average junction temperature for each temperature cycle. For each temperature cycle, based on the temperature difference amplitude and combined with the material mechanics formulas, the stress parameters generated in the solder layer are calculated; Based on the stress parameters, the crack propagation rate is obtained by analytically solving the Paris law. Record the crack propagation rate obtained from each temperature cycle to generate a second sequence of physical parameters.

[0039] Specifically, junction temperature fluctuations are the most direct driving force for thermomechanical fatigue of the solder layer in power devices such as IGBTs. Therefore, by extracting the junction temperature change sequence of IGBTs, i.e., the junction temperature curve, as the first physical parameter sequence, we can represent the dynamic change process of the temperature of the transistor node inside the IGBT with time or switching cycle, which corresponds to the thermal fluctuation scale of the medium and low frequency operating conditions.

[0040] In some embodiments, the extraction of the junction temperature curve can employ an electro-thermal equivalent network model. Specifically, in simulation software such as MATLAB / Simulink, a Foster or Cauer transistor thermal network model is built, and the real-time power loss Ploss (calculated from the measured collector current Ic and collector-emitter voltage Vce) and the heat sink base plate temperature Tc are input. Through dynamic calculation of the thermal resistance (Rth) and thermal capacity (Cth) network, a continuous virtual junction temperature curve Tj(t) is output. In other embodiments, multiphysics finite element simulation (FEM) can also be used. Junction temperature curve extraction can be achieved by using ANSYS or COMSOL to build a three-dimensional solid model of the device, inputting the actual current and voltage conditions, and solving the temperature change sequence at the center of the chip through the heat conduction equation. This method has the highest accuracy, but the computational load is large, and it is usually used to extract a small number of samples as a reference. In other embodiments, temperature-sensitive electrical parameter (TSEP) calibration and conversion are also used. That is, in a laboratory environment, by measuring temperature-sensitive parameters such as saturation voltage drop or threshold voltage under small current, and combining them with a pre-calibrated "parameter-temperature" relationship curve, the measured electrical signal is converted into a temperature sequence.

[0041] Furthermore, under power cycling or long-term operating conditions, the mismatch in the coefficient of thermal expansion (CTE) between the IGBT, solder layer, and base plate will generate thermal stress, causing cracks to initiate at the edge of the solder layer and propagate towards the center. Therefore, the crack propagation rate (da / dN) reflects the transient accumulation rate of this physical damage, so as to extract a time series reflecting the evolution law of physical damage inside the IGBT, as a second physical parameter series, corresponding to the low-frequency macroscopic degradation scale.

[0042] Specifically, for crack propagation rate, the long-term junction temperature curve Tj(t) of the extracted IBGT is processed by rainflow counting method to extract the temperature difference amplitude (ΔTj) and average junction temperature of each temperature cycle. Rainflow counting method is a cycle counting method used for fatigue analysis. By simulating the path of raindrops flowing along the wavy roof, the complex temperature time history is decomposed into a series of independent cycles, and then the amplitude and mean of each cycle are extracted.

[0043] Furthermore, based on the temperature difference amplitude, the stress parameter ΔK generated in the solder layer during each thermal cycle is calculated using material mechanics formulas. This parameter represents the plastic strain amplitude or stress intensity factor amplitude. The calculated stress parameter is then substituted into the Paris-Erdogan law formula. To obtain the crack propagation rate, where, It is the crack length. da / dN is the number of cycles, and da / dN is the crack propagation rate under the current cycle. C and m are empirical constants related to the properties of the solder material, which are usually obtained by consulting the material handbook or by fitting through accelerated aging tests.

[0044] More specifically, according to the time node or the number of cycles, the calculated da / dN is recorded each time to form a time series P(t) that reflects the evolution law of internal physical damage, that is, the physical parameter sequence.

[0045] Furthermore, the correlation between the physical parameter sequence and each bandpass mode component is calculated, as follows: , where N represents the total number of discrete data points for each bandpass mode component, which depends on the sampling frequency and sampling time of the Vce-peak monitoring signal; P(i) represents the value of the physical parameter sequence at the i-th time point; Represents the arithmetic mean of the entire sequence of physical parameters; This represents the value of the k-th bandpass mode component at time point i; This represents the arithmetic mean of k bandpass mode components.

[0046] Therefore, the correlation coefficients between the physical parameter sequence and each bandpass mode component are calculated and used to measure the bandpass mode components decomposed by VMD. The degree of linear correlation between the physical parameter sequence P(t), representing the actual physical damage signal, and the trend of its change. The larger the value, the more representative the bandpass mode component is of the physical degradation process. Understandably, both the first and second physical parameter sequences are correlated with each bandpass mode component.

[0047] In some embodiments, the mode types include physically degenerate dominant modes, thermo-coupled wave modes, and high-frequency noise-independent modes; The bandpass mode component in the dominant physical degradation mode is highly positively correlated with the crack propagation rate; The bandpass mode component in the thermo-coupled wave mode is related to the junction temperature curve of the IGBT; The bandpass mode component in the high-frequency noise-independent mode is represented as random high-frequency oscillation, which is independent of the evolution law of physical damage inside the IGBT.

[0048] Specifically, based on the above correlation calculation results, each bandpass mode component after VMD decomposition is assigned a relevant physical meaning, that is, classified into different mode types, including physical degradation mode, thermo-coupled fluctuation mode, and high-frequency noise-independent mode. Among them, the physical degradation dominant mode is usually a low-frequency step or monotonically increasing component, and its amplitude is highly positively correlated with the crack propagation rate of the solder layer. It represents the inherent performance degradation caused by material fatigue and increased thermal resistance, and is the core information carrier for predicting the remaining service life. Thermo-coupled fluctuation mode is usually a mid-frequency fluctuation component. This component is highly consistent with the dynamic change law of the IGBT junction temperature curve, reflecting the transient modulation effect of junction temperature fluctuation on the internal carrier mobility and saturation voltage drop. The high-frequency noise-independent mode is usually manifested as random high-frequency oscillation. This type of component has no statistical correlation with the evolution rate of the above physical failure process and mainly originates from the high-frequency resonance of stray inductance or electromagnetic measurement noise.

[0049] In one embodiment, a first correlation threshold and a second correlation threshold are preset for the physically degraded dominant mode. ≥ Second correlation threshold; for thermocoupled wave modes, first correlation threshold ≤ ≤ Second correlation threshold; for high-frequency noise-independent modes, <First correlation threshold. In a preferred embodiment, the first correlation threshold is 0.3, and the second correlation threshold is 0.6.

[0050] Furthermore, based on a preset reconstruction threshold, modal type screening is performed on the correlation coefficients of the first and second physical parameter sequences to determine the effective modal set. Specifically, the effective modal set is defined. ,in, This represents the correlation coefficient between the k-th bandpass mode component and the crack propagation rate, i.e., the correlation coefficient between the second physical parameter sequence and the bandpass mode component; This represents the correlation coefficient between the k-th bandpass mode component and the junction temperature fluctuation, i.e., the correlation coefficient between the first physical parameter sequence and the bandpass mode component; A preset reconstruction threshold is used to determine whether a modal component is a "useful feature" or "useless noise." It is set based on a first correlation threshold and a second correlation threshold, depending on the subsequent model's requirements for data smoothness. For example, if the first correlation threshold is 0.3 and the second correlation threshold is 0.6, then it is typically set to 0.3 (retaining moderate to high correlation) or 0.6 (retaining only strong correlation). Specifically, if... Setting it to 0.6 indicates strict screening; for physically degenerate dominant modes (ρk≥0.6), they successfully enter the effective mode set. Modes with high frequency noise independence (ρk < 0.3) and high frequency noise-independent modes (ρk < 0.3) are considered unqualified and excluded from reconstruction. Consequently, the reconstructed aging characteristics are extremely smooth, retaining only the core monotonically increasing aging trend. If the reconstruction threshold ρth is set to 0.3, it represents standard screening, and the physically dominant mode (ρk ≥ 0.6) successfully enters the effective mode set. Participating in reconstruction; Thermocoupled wave mode (0.3≤ρk≤0.6): Successfully entered the effective mode set. High-frequency noise and irrelevant modes (ρk<0.3) are precisely intercepted and eliminated, while high-frequency noise and irrelevant modes are involved in the reconstruction.

[0051] Understandably, for each bandpass mode component entering the effective mode set, its correlation coefficient with both the first and second physical parameter sequences must be greater than or equal to a preset reconstruction threshold. That is, traversing all K modes will satisfy the above requirement. The bandpass mode components are classified into the effective mode set.

[0052] Furthermore, the aging features extracted after reconstruction, , This indicates that only the indices within the effective mode set S are cyclically accumulated. Thus, through the above physical mechanism screening and reconstruction process, false fluctuation components unrelated to material fatigue aging in the Vce-peak monitoring signal are eliminated, enabling the reconstructed aging characteristics to more purely and smoothly characterize the true physical health evolution trajectory of the IGBT, laying a reliable data foundation for the subsequent construction of a high-precision life prediction model.

[0053] Furthermore, addressing the issue of excessively large sample size and significant fluctuations in the Vce-on dataset, one embodiment of this application first uses a simple moving average method to smooth the data, suppressing short-term random fluctuations; then, a piecewise aggregation approximation method is used to reduce the dimensionality and simplify the smoothed data, preserving key trend information and reducing the data volume. The processed Vce-on data graph is shown below. Figure 4 As shown.

[0054] Specifically, the simple moving average method is a modified arithmetic average method, a smoothing method used to handle sawtooth waves. It refers to calculating the moving average of each set of data formed by continuous movement of moving periods using the arithmetic average method, and using it as the forecast value for the next period.

[0055] More specifically, the piecewise aggregation approximation method is a feature representation method that boasts advantages such as fast computation speed and simple algorithm implementation, making it more suitable for processing large-scale and structurally complex power load data. Its basic idea is to divide the time series into equal-length segments and calculate the average of these segments as the feature representation of the time series. Specifically, for a data set of length m... Using a data of length w This means that w is divisible by m, which indicates that the method is a time series data dimensionality reduction process with a compression ratio of w / m, and satisfies: Therefore, data S of length m is transformed into data of length w. The method uses the mean of data elements to represent the sequence segment. It takes advantage of the characteristic that time series data changes little in the short term. Even if there are a few sudden changes in the data, the efficiency will be slightly degraded, but the overall correctness of the algorithm will not be affected. Different compression ratios determine the dimension of the data after approximate representation, which can effectively reduce the dimensionality of time series to a certain extent.

[0056] Furthermore, the extracted / simplified aging features, namely the feature values ​​of Vce-peak after decomposition and reconstruction, and the piecewise aggregated approximation of Vce-on, are normalized respectively, mapping all feature values ​​to the [0, 1] interval to eliminate the influence of different features on their dimensions, which facilitates subsequent model training. In addition, a sliding window mechanism is adopted to convert the normalized feature sequence into fixed-length time window samples. Each window serves as an input sample for the model, corresponding to the label of the subsequent aging state. The samples generated by the sliding window are divided into training and testing sets according to a certain ratio for subsequent training and testing of the IGBT-based aging prediction model.

[0057] S200. Construct an IGBT aging prediction model and use the Octopus Optimization Algorithm to optimize the parameters of the IGBT aging prediction model to obtain the globally optimal parameter combination. In some embodiments, the IGBT aging prediction model includes an input layer, a CNN layer, a BiLSTM layer, an attention mechanism layer, a fully connected layer, and an output layer; wherein, The CNN layer is used for local feature parameter extraction and feature parameter mapping; The BiLSTM layer is used to mine temporal information from the feature parameters; The attention mechanism layer is used to calculate a weight vector for the feature parameter values ​​output by the BiLSTM layer, representing the importance of the current time step to different input parts.

[0058] Specifically, CNN is a type of neural network that uses several convolutional layers and pooling layers connected alternately, followed by a fully connected layer for output. The convolutional layers extract different feature parameters from the input through convolution operations, the pooling layers obtain spatially invariant feature parameters from the convolutional layers by reducing the dimensionality of the feature parameters, the feature parameters extracted by the convolutional layers are extracted a second time, and the fully connected layers transform the feature parameters before outputting them through the output layer. This achieves the extraction of local feature parameters from the data.

[0059] Furthermore, LSTM, as an improved RNN, aims to solve the gradient vanishing and gradient exploding problems of traditional RNNs when handling long sequences and long-term dependencies. Its core idea is to introduce structural units called "gates" to control the flow of information and the updating of memory through learning. Specifically, an LSTM unit mainly consists of a forget gate, an input gate, and an output gate. The input gate operates on the input feature parameters, the forget gate determines whether to retain or discard data in the candidate memory unit, and the output gate determines whether to output the feature parameters. The memory unit stores intermediate variables. Its schematic diagram is shown below. Figure 5 As shown, specifically, the input to the LSTM is the hidden layer state H at time t-1. t-1 and the input feature parameters X at the current stage t H t-1 The valid information contained therein is from the memory unit C at time t-1. t-1 and output gate O t-1 Together, they determine how to construct long-distance temporal dependencies through the forward propagation of parameters. The various parameters in LSTM can be represented as follows: ; Among them, I t F is the input gate at time t; t Forget gate at time t; O t H is the output gate at time t; tLet W be the hidden state at time t; σ be the sigmoid function; W xi W xf and W xo W represents the weight parameters of the input gate, forget gate, and output gate at time t, respectively. hi W hf and W ho b represents the weight parameters of the input gate, forget gate, and output gate at time t-1; i b f and b o These are the bias parameters at time t; C t Let be the memory unit at time t.

[0060] Furthermore, BiLSTM, as an improved algorithm of LSTM, uses LSTM twice, improving the long-term dependencies learned by LSTM and thus increasing the accuracy of the model. The principle diagram of BiLSTM is shown below. Figure 6 As shown, it consists of a forward LSTM and a backward LSTM. The two LSTM layers are based on the same principle, but the data input order is reversed. The feature parameters are input into the forward LSTM in the forward order and into the backward LSTM in the reverse order. The output is obtained by integrating the outputs of the forward and backward LSTMs.

[0061] Therefore, the IGBT aging prediction model of this application is a CNN-BiLSTM-Attention model consisting of an input layer, a CNN layer, a BiLSTM layer, an attention mechanism layer, a fully connected layer, and an output layer. The CNN layer performs local feature parameter extraction and feature parameter mapping to extract useful information from the input feature parameters. The BiLSTM layer is used to mine the temporal information in the feature parameters. At the same time, the attention mechanism is used to calculate the weight vector of the BiLSTM output feature parameter values, which represents the importance of the current time step to different input parts. The use of the attention mechanism can enhance the ability of BiLSTM to process temporal information.

[0062] In addition, to further improve the model's prediction accuracy, such as Figure 7 As shown, this application uses the OOA algorithm to optimize the key value of the attention mechanism, the number of neurons in the BiLSTM, and the regularization coefficient of the model in the above CNN-BiLSTM-Attention model. Therefore, the IGBT aging prediction model of this application is the optimized OOA-CNN-BiLSTM-Attention model.

[0063] In some embodiments, the step of using the octopus optimization algorithm to optimize the parameters of the IGBT aging prediction model and obtain the globally optimal parameter combination includes: Initialize the octopus population by dividing it into predator individuals and scout individuals, and set the optimization parameter range and fitness function; During the exploration phase, the location of the individual scouts is updated and their location information is fed back to the predator group, and the location of the individual predators is updated. During the development phase, the location of the predator individuals is updated by narrowing the search range; After each location update, check whether the individual's new location exceeds the search space boundary, and change the location if the fitness of the new location is better than the original location; Iterate through all individuals in the octopus population, update the global optimal solution, and obtain the global optimal parameter combination.

[0064] Specifically, the Octopus Optimization Algorithm (OOA) is a biomimetic swarm intelligence optimization algorithm inspired by the complex behaviors exhibited by octopuses in their natural environment, including rapid swimming to hunt, fine-grained searching with tentacles, and an iterative optimization mechanism based on visual feedback. The octopus's hunting behavior includes an exploration phase and a development phase. The exploration phase simulates the octopus swimming rapidly by spraying water through its sac to search for prey in a wide area of ​​the sea, while the development phase simulates the octopus using its eight tentacles to crawl along the seabed for fine-grained searching and hunting.

[0065] Furthermore, the specific calculation process of the octopus optimization algorithm is as follows: The Octopus Optimization Algorithm employs a hybrid population structure, dividing the entire group into two classes: predators (Hunters) and scouts (Scouts). This division of labor and cooperation achieves a balance between global exploration and local exploitation. Let the search space dimension be D, and the number of predators be N. h The predator and its eight tentacles together constitute nine search locations, each representing a candidate solution to the optimization problem. The number of scouts is N. s The total population size is then N = 9 × N h +N s .

[0066] Initialize the octopus population by randomly generating it within the search space: ; In the above formula (1): Let i be the component of the i-th search position in the j-th dimension; and are the lower and upper bounds of the j-th dimension variable, respectively; rand is a random number that follows a uniform distribution within the interval [0,1].

[0067] In the exploration phase, the behavior of octopuses using jet propulsion to generate recoil force for rapid swimming is simulated. During this phase, the octopus uses its instantaneous acceleration to search for prey over a wide area to discover potentially high-quality regions. The octopus optimization algorithm achieves global exploration through individual scouts. Scouts can make large-step jumps to cover a wider search space. The scouts' positions are updated using a selection mechanism: several positions are randomly selected from the predator group to generate new positions for the scouts. Specifically, the new positions of the scouts are updated according to equation (2): ; In the above formula (2): For the new position of the reconnaissance soldiers; A location is a random selection from a set of locations chosen from a group of predators. This is the current globally optimal position; α represents the current worst position globally; α is the scaling factor, controlling the search step size; randn is a random number that follows a standard normal distribution.

[0068] After the scouts complete their search, their location information is converted and fed back to the predator group, enabling information sharing among different types of individuals. Simultaneously, individual predators also participate in the global search during the exploration phase, and their location updates are controlled by an adaptive parameter W, as follows: ; In equation (3) above: t is the current iteration number; is the maximum number of iterations; rand is a random number in the interval [0,1]. The adaptive parameter W takes a larger value in the early stage of the algorithm to enhance the global exploration capability; it gradually decreases as the iteration progresses, so that the algorithm can smoothly transition to the local development stage.

[0069] When a predator is in the exploration phase, its position update formula is: ; In the formula: To determine the j-th new position of the i-th predator during the exploration phase; Let j be the current position of the i-th predator in the j-th dimension; and Let be the j-th dimension position of two different individuals randomly selected from the population; W is the adaptive parameter defined by equation (3).

[0070] During the development phase, the behavior of an octopus using its eight tentacles to crawl along the seabed and conduct a detailed search is simulated. In this phase, the octopus conducts a detailed search within the local area where it has discovered prey. Through the flexible movement of its tentacles and real-time perception, it gradually approaches the location of the prey. The octopus optimization algorithm achieves local development through the search unit composed of the predator and its eight tentacles, and conducts detailed exploration near the high-quality solution.

[0071] The position update during the development phase comprehensively considers both the current global optimum and the individual's own position, achieving refined optimization by narrowing the search range. The predator's position update formula during the development phase is shown below: ; In the above formula (5): This represents the j-th new position of the i-th predator during the development phase. This represents the j-th dimension component of the current globally optimal solution; Let be the current position of the i-th predator in the i-th dimension; W is the adaptive parameter defined by equation (3).

[0072] During the development phase, to further enhance local search capabilities, the octopus optimization algorithm introduces a tentacle-based fine-grained search mechanism. Each predator's eight tentacles represent different search directions, forming a local neighborhood around the predator's head. ; In the above formula (6): Let the position of the k-th tentacle be in the j-th dimension; denoted as the position of the predator's head in the j-th dimension; β is the contraction factor, controlling the search range of the tentacles; γ is the decay coefficient, controlling the contraction rate of the search range; rand is a random number in the interval [0,1]; e is a natural constant. This mechanism simulates the process of octopus tentacles gradually contracting and accurately capturing prey.

[0073] Furthermore, to prevent the algorithm from getting trapped in local optima, the Octopus Optimization Algorithm introduces a recoil motion random feedback mechanism, inspired by the recoil and instantaneous acceleration phenomena of an octopus spouting water. This mechanism generates random perturbations through a recoil operator, causing the individual to jump out of the current search area and enter a new space for exploration. Specifically, the trigger probability of the recoil mechanism is... When triggered, the individual's position is updated as follows: ; In the above formula (7): is the j-th dimension new position of the i-th individual after the recoil motion; sign(.) is the sign function, which determines the direction of the perturbation; the meanings of the other signs are the same as those in the previous equations, and will not be repeated here. This mechanism is executed with a certain probability during the algorithm's operation, which enhances the diversity of the population.

[0074] It is worth noting that after each position update, it is necessary to check whether the new position exceeds the search space boundary. If it does, it should be pulled back within the boundary according to equation (8): ; The Octopus Optimization Algorithm employs a greedy selection mechanism, replacing a position only when the fitness of the new position is better than that of the original position. For the minimization problem, the selection mechanism is described by equation (9): ; In equation (9) above: f(.) is the objective function; Let i be the current position of the i-th individual; Let be the candidate new position for the i-th individual.

[0075] Therefore, the above-mentioned octopus optimization algorithm is applied to the IGBT aging prediction model of this application to optimize the parameters of the IGBT aging prediction model, so as to improve the training accuracy and prediction accuracy of the model. Specifically, in the initial state, the octopus population is initialized, that is, the initial population size of the octopus is set to 20 and the maximum number of evolution iterations is 10. At the same time, the corresponding parameter ranges need to be set for the parameters to be optimized. For example, in one embodiment, the learning rate search range is [0.001, 0.01], the regularization parameter search range is [0.0001, 0.01], the BiLSTM neuron number search range is [10, 50], and the attention mechanism key value search range is [2, 50].

[0076] Furthermore, a corresponding fitness function is set as the objective function for determining candidate new positions during the octopus optimization process, and its expression is shown in equation (10): ; In the above equation (10), n is the number of samples; For the actual value of the i-th sample, Let be the predicted value for the i-th sample.

[0077] Therefore, during the optimization process, the exploration phase begins, and the scouts update their positions according to equation (2). After completing the search, the scouts' position information is converted and fed back to the predator group, enabling information sharing among different types of individuals. Predator individuals also participate in the global search during the exploration phase, and their positions are updated according to equation (4). Entering the development phase, the position update during the development phase comprehensively considers the current global optimal solution and the individual's own position, achieving refined optimization by narrowing the search range. Predators update their positions during the development phase according to equation (5).

[0078] To further enhance local search capabilities, the Octopus Optimization Algorithm introduces a tentacles fine search mechanism to form a local neighborhood around the predator's head position as shown in Equation (6). To prevent the algorithm from getting stuck in local optima, the Octopus Optimization Algorithm also introduces a recoil motion random feedback mechanism. This mechanism generates random perturbations through recoil operators, causing individuals to jump out of the current search area and enter a new space for exploration. Their positions are updated according to Equation (7). After each position update, it is necessary to check whether the new position exceeds the search space boundary. If it does, it is pulled back into the boundary according to Equation (8).

[0079] Furthermore, the octopus optimization algorithm employs a greedy selection mechanism, replacing only when the fitness of the new position is better than that of the original position. Replacement is performed according to equation (9), traversing all individuals and updating the global optimal solution. and the global optimal fitness value The global optimal solution corresponds to the global optimal parameter combination. The IGBT aging prediction model is then trained based on the global optimal parameter combination. After training, the model is input into the test set for prediction to obtain the predicted value.

[0080] In this application, during the process of deep learning hyperparameter optimization based on the Octopus optimization algorithm, the search range of the number of neurons in the hidden layer of BiLSTM is usually set to a broad static empirical range, such as [10, 50] set in the above embodiment. However, such a static boundary setting that is detached from the inherent properties of the data to be processed has significant limitations: when the search upper limit is too high, it will not only exponentially increase the optimization dimension and computational cost of the algorithm, but also easily lead to model parameter redundancy and thus overfitting; while blindly reducing the search space may miss the global optimal solution.

[0081] To address this issue, this application proposes an adaptive mechanism that dynamically correlates the search domain with data characteristics. Specifically, the essence of BiLSTM neurons is to represent and remember nonlinear dynamic features and long-term dependencies in time series. Therefore, the network capacity of a reasonable neural constant should be strictly matched with the temporal complexity and information memory depth of the input sequence.

[0082] Therefore, this application establishes an explicit mapping relationship between the upper limit Nmax of the number of BiLSTM neurons searched by the OOA algorithm and the "effective memory length" of the input data and its corresponding "number of main modes".

[0083] In some embodiments, the range of optimization parameters includes a search range for the number of BiLSTM neurons; For the search range of the number of BiLSTM neurons, an autocorrelation function is constructed, the effective memory length of the time series data used for training is calculated, and the search range is adjusted based on the effective memory length. If the IGBT aging dataset is the Vce-peak dataset, then based on the effective memory length and the number of bandpass modal components in the effective modality set, the upper limit of the number of BiLSTM neurons is dynamically calculated.

[0084] In this embodiment, the influence of historical states on the current state of the time series data used for training the IGBT aging prediction model is quantified by autocorrelation analysis. Specifically, an autocorrelation function is introduced to calculate the effective memory length Lmemory of the time series data. When the effective memory length Lmemory of the time series data input for training the IGBT aging prediction model is large, it indicates that the signal contains longer-spanning temporal dependence features. In this case, it is necessary to increase the size of the memory units in BiLSTM to broaden the model's receptive field. Conversely, if the sequence approximates a Markov process, it indicates that the signal is short-term dependent, and fewer neurons are needed to meet the fitting requirements.

[0085] In some embodiments, constructing the autocorrelation function and calculating the effective memory length of the time series data used for training includes: Calculate the autocorrelation coefficient of time series data at different lag orders and generate autocorrelation curves; Set a cutoff threshold, and find the minimum delay order from the autocorrelation curve that first decays and stabilizes within the cutoff threshold range. The minimum delay order is the effective memory length, indicating that after the effective memory length is exceeded, the time series data loses its predictive value for the current IGBT degradation state.

[0086] Specifically, for a stationary or quasi-stationary time series X = {x1, x2, ..., xN} of length N, calculate its autocorrelation coefficient r at different lag orders k. k Generate ACF curves, specifically... ,in, Let r0 be the mean of the sequence, and k represent the lag order, k=0,1,2,......; when k=0, r0=1.

[0087] Furthermore, as the delay order k increases, the influence of the historical state on the current state gradually weakens, r k The absolute value of the decay will show a decay trend, therefore a cutoff threshold is set to determine the decay delay time. In some embodiments, a 95% confidence level is set as the cutoff threshold, and its confidence boundary is approximately... . , indicating that if |r k If the correlation falls within and remains within this boundary, it can be considered statistically insignificant, i.e., tending towards white noise; in other embodiments, the truncation threshold is set based on empirical threshold criteria, such as setting a specific percentage of decay to the initial value, for example, decaying to e.-1 (Approximately 0.368) or set directly to 0.05.

[0088] Furthermore, we find the smallest delay order k that causes the ACF curve to first decay and stabilize within the aforementioned cutoff threshold. After spanning Lmemory time steps, the historical data of the sequence has lost its significant predictive value for the current degradation state. Therefore, the upper limit of the number of BiLSTM neurons is determined based on this Lmemory.

[0089] Furthermore, if the time series data approximates a first-order Markov process, it means that the state at the next moment depends only on the state at the current moment and is independent of earlier historical trajectories. Therefore, it is necessary to verify whether the time series data belongs to this "short-term dependence." Specifically, it is necessary to combine the ACF and the partial autocorrelation function (PACF) for double verification. For the ACF, if the sequence is a Markov process, it is equivalent to a first-order autoregressive process in time series analysis. Its ACF curve will not suddenly drop to 0 at a certain point, but will show rapid exponential decay or decay accompanied by damped oscillations. Therefore, if the ACF curve shows a significant exponential decline and the corresponding Lmemory value is very small, for example, Lmemory < If the correlation coefficient is positive, it is initially determined to be a short-term dependency. Furthermore, since the ACF includes the sum of "direct correlation" and "indirect correlation," a partial autocorrelation function (PACF) is needed for rigorous confirmation to eliminate the transitive influence of intermediate variables. Specifically, the PACF measures x... t With x t-k The pure direct correlation between them means that for sequences approximating Markov processes, the PACF will be truncated instantaneously after a lag of 1 order (or a very low order p), meaning the PACF value drops directly into the ±1.96N confidence interval and remains essentially zero. Time series data determined by the rapid decay of ACF and the 1st-order truncation of PACF exhibit obvious "Markov short-term dependence characteristics," thus requiring a reduction in the number of BiLSTM neurons.

[0090] Furthermore, for non-stationary degradation signals, which often involve multiple timescales of evolutionary features, each independently evolving physical mode (such as trend terms or periodic fluctuation terms) requires a certain number of neurons in the feature space for mapping. Therefore, for Vce-peak monitoring signals in the Vce-peak dataset, it is necessary to consider the number of effective modes in the effective mode set obtained during the aforementioned decomposition and reconstruction process of the Vce-peak monitoring signals. To quantify the frequency domain complexity of the data, specifically... The more features there are, the higher the dimension of the feature space required to reconstruct the original signal, and the upper limit of the number of neurons required should also be expanded accordingly.

[0091] Therefore, the upper limit of the number of neurons in a BiLSTM is... , where [.] is the floor function; ω1 and ω2 are the influence weight coefficients of effective memory length and effective mode number, respectively. It can be understood that the effective mode number is the number of bandpass mode components in the effective mode set; This is the basic redundancy bias term; while the lower limit of the number of BiLSTM neurons. A constant to ensure the network's fundamental nonlinear mapping capability.

[0092] In summary, by introducing the aforementioned data feature-driven boundary construction strategy, the search range for the number of BiLSTM neurons in the OOA algorithm is transformed from a blind "global static grid" to a "local dynamic adaptive domain." This not only gives the optimization algorithm clear physical and data science guidance and effectively avoids the risk of overfitting, but also significantly reduces the exploration process of the invalid solution space, thereby significantly improving the convergence speed and final prediction accuracy of the model's hyperparameter optimization.

[0093] S300. Based on the globally optimal parameter combination and the model training set, train the IGBT aging prediction model. In this embodiment, the IGBT aging dataset is divided into a model training set and a model test set according to a preset ratio. During the training process, based on the globally optimal parameter combination obtained in step S300 above, the model training set is combined with multiple iterations until the preset iteration conditions are met to obtain the IGBT aging prediction model, which is used to predict the aging of IGBTs.

[0094] S400. Based on the model test set, evaluate the prediction accuracy of the trained IGBT aging prediction model.

[0095] In this embodiment, the prediction accuracy of the IGBT aging prediction model after training needs to be evaluated. For example, the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) can be used as evaluation indicators for the IGBT aging prediction model. The smaller the three evaluation indicators are, the closer the predicted value is to the true value and the more accurate the result is.

[0096] In this application, the Vce-peak and Vce-on datasets are extracted from the IGBT aging data shared by NASA research centers and used as IGBT aging data for model training, testing, and validation.

[0097] In some embodiments, the Vce-peak dataset is used as the IGBT aging dataset. The OOA-CNN-BiLSTM-Attention model obtained after optimization in step S200 is trained and validated. For example, the aging features of the Vce-peak monitoring signal after data preprocessing are imported into the IGBT aging prediction model of this application, i.e., the optimized OOA-CNN-BiLSTM-Attention model, and the conventional CNN-BiLSTM-Attention model under the prior art for comparative training and validation. The corresponding model training set accounts for 0.8 and the model test set accounts for 0.2. The training effect is as follows. Figure 8 As shown. If the training set accounts for 0.3 and the test set accounts for 0.7, the training effect is as follows. Figure 9 As shown.

[0098] from Figure 8 and Figure 9 As can be seen, regardless of the proportion of the training set, the optimized model's prediction results deviate less from the true values ​​compared to the unoptimized model, and the prediction accuracy is higher. Moreover, the optimized model can still maintain high accuracy even when the proportion of the training set is small, while the unoptimized model has a larger error when the proportion of the training set is small. This proves that the IGBT aging prediction model proposed in this application can be applied to situations with a small amount of data.

[0099] In some embodiments, the Vce-on dataset is used as the IGBT aging dataset. The OOA-CNN-BiLSTM-Attention model obtained after optimization in step S200 is trained and validated. For example, the aging features of the Vce-on signal after data preprocessing are imported into the IGBT aging prediction model of this application, i.e., the optimized OOA-CNN-BiLSTM-Attention model, and the conventional CNN-BiLSTM-Attention model under the prior art for comparative training and validation. The corresponding model training set accounts for 0.8, and the model test set accounts for 0.2. The training effect is as follows. Figure 10 As shown. If the training set accounts for 0.3 and the test set accounts for 0.7, the training effect is as follows. Figure 11 As shown.

[0100] from Figure 10 and Figure 11 As can be seen, using the collector-emitter voltage Vce-on as aging data, the optimized model's prediction results deviate less from the true values ​​compared to the unoptimized model, and the prediction accuracy is higher. Moreover, the optimized model can still maintain high accuracy even when the training set is small, while the unoptimized model has a larger error when the training set is small.

[0101] Therefore, whether the collector-emitter transient peak voltage Vce-peak or the collector-emitter voltage Vce-on is used as aging data, the IGBT aging prediction model proposed in this application has higher accuracy and is more suitable for situations with a small amount of data, proving that the IGBT aging prediction model of this application has superiority.

[0102] To further verify the feasibility of the OOA-CNN-BiLSTM-Attention model proposed in this application, we built LSTM model, CNN-LSTM model, BILSTM model and CNN-BILSTM-Attention model respectively for horizontal comparison.

[0103] Specifically, the first 10 data points of the Vce-peak monitoring signal were used as input, and the 11th data point was used as output. A sliding time window was used to construct 409 time series. Then, the first 80% of the time series were selected as the model training set, and the remaining 20% ​​were selected as the model test set. After the model training was completed, the model test set was used for prediction. The average value of the prediction error evaluation index after multiple trials is shown in Table 1.

[0104] Table 1. Lateral Comparison of Vce-peak Monitoring Signals (0.8 Training Set) and Model Prediction Errors

[0105] As shown in Table 1, the OAA-CNN-BILSTM-Attention lifetime prediction model exhibits the best prediction performance among all compared models. Specifically, the OAA-CNN-BILSTM-Attention model has the highest coefficient of determination (R²), indicating that it best fits the IGBT aging data. Furthermore, the prediction errors of the OAA-CNN-BILSTM-Attention model are all lower than the other four models, compared to LSTM, CNN-LSTM, BILSTM, and CNN-BILSTM-Attention. The proposed aging prediction model showed significant improvements over several basic prediction models, with the mean absolute error (MAE) decreasing by 49.69%, 39.04%, 36.14%, and 12.98%, the mean absolute percentage error (MAPE) decreasing by 47.81%, 45.26%, 41.07%, and 16.83%, and the root mean square error (RMSE) decreasing by 42.01%, 39.6%, 36.26%, and 24.17%, respectively. These improvements demonstrate that the proposed aging prediction model can be used for IGBT aging prediction and shows improvement over several basic prediction models. Furthermore, the optimized model demonstrates a certain degree of improvement over the unoptimized model, proving that the optimization of the model in this application has a certain effect.

[0106] Table 2. Lateral Comparison of Vce-peak Monitoring Signals (0.3 Training Set) and Model Prediction Errors

[0107] Table 3. Lateral Comparison of Model Prediction Errors for Vce-on Signals (0.8 Training Set)

[0108] Table 4. Lateral Comparison of Vce-on Signal (0.3 Training Set) and Model Prediction Error

[0109] As shown in Tables 1, 2, 3, and 4, regardless of whether the Vce-peak monitoring signal or the Vce-on signal is used for prediction, and regardless of the proportion of the model training set, the prediction performance of the optimized model is better than that of the unoptimized model, proving that the model proposed in this application has universality and superiority.

[0110] To verify the practicality of the optimized OOA-CNN-BILSTM-Attition model proposed in this application, the collector-emitter voltage Vce-on of IGBTs under actual operating conditions at the CRRC Zhuzhou Group factory was used as aging data. The preprocessed Vce-on values ​​of CRRC IGBT1 were imported into the optimized OOA-CNN-BILSTM-Attition model for training and verification. The training set comprised 0.8 units of the model, and the test set comprised 0.2 units. The training results are as follows: Figure 12 As shown.

[0111] from Figure 12 As can be seen from the data, the model of this application can predict the changing trend of the collector-emitter voltage Vce-on of IGBT1 under actual working conditions. Moreover, before the failure of IGBT, there is a significant sudden drop in the collector-emitter voltage Vce-on. The model of this application can accurately predict the sudden drop trend, proving that the model of this application can effectively predict the aging trend of IGBT under actual working conditions. Furthermore, compared with the unoptimized model, the deviation between the predicted value and the actual value is smaller, and the prediction accuracy is higher, which is sufficient to prove that this application has certain practicality under actual working conditions.

[0112] In some embodiments, the Vce-on values ​​of the CRRC IGBT2 after data preprocessing are also imported into the optimized OOA-CNN-BILSTM-Attention model proposed in this application for training and verification. The training set accounts for 0.8 and the test set accounts for 0.2, and the training effect is as follows. Figure 13 As shown.

[0113] from Figure 13As can be seen, the model in this application can predict the changing trend of the collector-emitter voltage Vce-on of IGBT2 under actual operating conditions. Compared with IGBT1, IGBT2 does not show a significant sudden drop in the later stage. The model in this application accurately predicts its changing trend, and no sudden drop occurs. This proves that the model in this application can effectively predict the aging trend of IGBTs under actual operating conditions and has a certain degree of universality. Moreover, compared with the unoptimized model, the deviation between the predicted value and the actual value is smaller, and the prediction accuracy is higher, which is sufficient to prove that this application has a certain degree of practicality under actual operating conditions.

[0114] Therefore, the model proposed in this application was validated using IGBT aging data shared by NASA research centers and IGBT aging data under actual operating conditions at the CRRC Zhuzhou Group factory, demonstrating the effectiveness and superiority of the model in improving performance. Through comparative analysis of different models, the following conclusions can be drawn: 1) The Octopus optimization algorithm is used to globally optimize parameters such as learning rate, regularization coefficient, number of BiLSTM neurons and attention mechanism key value in the IGBT aging prediction model, thereby improving the accuracy of the IGBT lifetime prediction model. 2) Using the VMD algorithm to extract the degradation features of IGBTs and reconstruct the useful modes can reduce the uncertainty of IGBT aging model prediction and improve accuracy; using a combination of simple moving average method and piecewise aggregation approximation method can preserve the aging features of Vce-on and simplify the amount of calculation, making it better for prediction. 3) The proposed model can accurately predict the aging of IGBTs even with limited monitoring data, and its long-term prediction performance is good, which can provide a reference for the monitoring and maintenance of IGBTs in complex environments.

[0115] Please see Figure 14 As shown, the present invention also provides an IGBT aging prediction system, the system comprising: The first processing module 1401 is used to preprocess the IGBT aging dataset and extract aging features, and to divide the dataset into a model training set and a model test set. The second processing module 1402 is used to construct an IGBT aging prediction model and use the Octopus optimization algorithm to optimize the parameters of the IGBT aging prediction model to obtain the globally optimal parameter combination. The third processing module 1403 is used to train the IGBT aging prediction model based on the global optimal parameter combination and the model training set. The fourth processing module 1404 is used to evaluate the prediction accuracy of the trained IGBT aging prediction model based on the model test set.

[0116] It is understandable that, such as Figure 1 The content of the IGBT aging prediction method embodiments shown is applicable to the IGBT aging prediction system embodiments. The specific functions implemented by the IGBT aging prediction system embodiments are the same as those shown in the examples. Figure 1 The IGBT aging prediction method shown in the embodiment is the same, and the beneficial effects achieved are the same as those shown. Figure 1 The beneficial effects achieved by the IGBT aging prediction method embodiment shown are also the same.

[0117] It should be noted that the information interaction and execution process between the above systems are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0118] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0119] Please see Figure 15 As shown, this embodiment of the invention also provides a computer device 15, including: a memory 1502 and a processor 1501, and a computer program 1503 stored in the memory 1502. When the computer program 1503 is executed on the processor 1501, it implements the IGBT aging prediction method as described in any of the above methods.

[0120] The computer device 15 may be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device 15 may include, but is not limited to, a processor 1501 and a memory 1502. Those skilled in the art will understand that... Figure 15 The computer device 15 is merely an example and does not constitute a limitation on the computer device 15. It may include more or fewer components than shown, or combine certain components, or different components, such as input / output devices, network access devices, etc.

[0121] The processor 1501 may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0122] In some embodiments, the memory 1502 may be an internal storage unit of the computer device 15, such as a hard disk or memory of the computer device 15. In other embodiments, the memory 1502 may be an external storage device of the computer device 15, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the computer device 15. Furthermore, the memory 1502 may include both internal and external storage units of the computer device 15. The memory 1502 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory 1502 can also be used to temporarily store data that has been output or will be output.

[0123] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the IGBT aging prediction method as described in any of the above methods.

[0124] In this embodiment, if the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographic device / computer device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0125] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for predicting IGBT aging, characterized in that, include: The IGBT aging dataset is preprocessed and aging features are extracted, and the dataset is divided into a model training set and a model test set. An IGBT aging prediction model was constructed, and the Octopus Optimization Algorithm was used to optimize the parameters of the IGBT aging prediction model to obtain the globally optimal parameter combination. Based on the globally optimal parameter combination and the model training set, the IGBT aging prediction model is trained. Based on the model test set, the prediction accuracy of the trained IGBT aging prediction model is evaluated. The IGBT aging dataset is the Vce-peak dataset; For the Vce-peak dataset, decomposition and reconstruction are performed based on the VMD algorithm to obtain aging features, including: The Vce-peak monitoring signal is decomposed into multiple bandpass mode components using the VMD algorithm; Extract the physical parameter sequence, and calculate the correlation coefficient between the physical parameter sequence and each of the bandpass mode components; Based on the correlation coefficient, each bandpass mode component is divided into different mode types, and based on a preset reconstruction threshold, mode types are screened to determine the effective mode set and obtain aging characteristics. The physical parameter sequence is a time series that reflects the evolution of physical damage inside the IGBT; The extraction of the physical parameter sequence includes: Extract the junction temperature curve of the IGBT as the first physical parameter sequence; Rainflow counting is performed on the junction temperature curve to extract the temperature difference amplitude and average junction temperature for each temperature cycle. For each temperature cycle, based on the temperature difference amplitude and combined with the material mechanics formulas, the stress parameters generated in the solder layer are calculated; Based on the stress parameters, the crack propagation rate is obtained by analytically solving the Paris law. Record the crack propagation rate obtained from each temperature cycle to generate a second sequence of physical parameters; The modal types include physically degenerate dominant modes, thermo-coupled wave modes, and high-frequency noise-independent modes; The bandpass mode component in the dominant physical degradation mode is highly positively correlated with the crack propagation rate; The bandpass mode component in the thermo-coupled wave mode is related to the junction temperature curve of the IGBT; The bandpass mode component in the high-frequency noise-independent mode is represented as random high-frequency oscillation, which is independent of the evolution law of physical damage inside the IGBT.

2. The method as described in claim 1, characterized in that, The IGBT aging prediction model includes an input layer, a CNN layer, a BiLSTM layer, an attention mechanism layer, a fully connected layer, and an output layer; wherein, The CNN layer is used for local feature parameter extraction and feature parameter mapping; The BiLSTM layer is used to mine temporal information from the feature parameters; The attention mechanism layer is used to calculate a weight vector for the feature parameter values ​​output by the BiLSTM layer, representing the importance of the current time step to different input parts.

3. The method as described in claim 2, characterized in that, The step of using the octopus optimization algorithm to optimize the parameters of the IGBT aging prediction model to obtain the globally optimal parameter combination includes: Initialize the octopus population by dividing it into predator individuals and scout individuals, and set the optimization parameter range and fitness function; During the exploration phase, the location of the individual scouts is updated and their location information is fed back to the predator group, and the location of the individual predators is updated. During the development phase, the location of the predator individuals is updated by narrowing the search range; After each location update, check whether the individual's new location exceeds the search space boundary, and change the location if the fitness of the new location is better than the original location; Iterate through all individuals in the octopus population, update the global optimal solution, and obtain the global optimal parameter combination.

4. The method as described in claim 3, characterized in that, The range of optimization parameters includes the search range for the number of BiLSTM neurons; For the search range of the number of BiLSTM neurons, an autocorrelation function is constructed, the effective memory length of the time series data used for training is calculated, and the search range is adjusted based on the effective memory length. If the IGBT aging dataset is the Vce-peak dataset, then based on the effective memory length and the number of bandpass modal components in the effective modality set, the upper limit of the number of BiLSTM neurons is dynamically calculated.

5. The method as described in claim 4, characterized in that, The construction of the autocorrelation function and the calculation of the effective memory length of the time series data used for training include: Calculate the autocorrelation coefficient of time series data at different lag orders and generate autocorrelation curves; Set a cutoff threshold, and find the minimum delay order from the autocorrelation curve that first decays and stabilizes within the cutoff threshold range. The minimum delay order is the effective memory length, indicating that after the effective memory length is exceeded, the time series data loses its predictive value for the current IGBT degradation state.

6. An IGBT aging prediction system, characterized in that, include: The first processing module is used to preprocess the IGBT aging dataset and extract aging features, and to divide the dataset into a model training set and a model test set. The second processing module is used to construct an IGBT aging prediction model and use the Octopus optimization algorithm to optimize the parameters of the IGBT aging prediction model to obtain the globally optimal parameter combination. The third processing module is used to train the IGBT aging prediction model based on the globally optimal parameter combination and the model training set. The fourth processing module is used to evaluate the prediction accuracy of the trained IGBT aging prediction model based on the model test set. The IGBT aging dataset is the Vce-peak dataset; For the Vce-peak dataset, decomposition and reconstruction are performed based on the VMD algorithm to obtain aging features, including: The Vce-peak monitoring signal is decomposed into multiple bandpass mode components using the VMD algorithm; Extract the physical parameter sequence, and calculate the correlation coefficient between the physical parameter sequence and each of the bandpass mode components; Based on the correlation coefficient, each bandpass mode component is divided into different mode types, and based on a preset reconstruction threshold, mode types are screened to determine the effective mode set and obtain aging characteristics. The physical parameter sequence is a time series that reflects the evolution of physical damage inside the IGBT; The extraction of the physical parameter sequence includes: Extract the junction temperature curve of the IGBT as the first physical parameter sequence; Rainflow counting is performed on the junction temperature curve to extract the temperature difference amplitude and average junction temperature for each temperature cycle. For each temperature cycle, based on the temperature difference amplitude and combined with the material mechanics formulas, the stress parameters generated in the solder layer are calculated; Based on the stress parameters, the crack propagation rate is obtained by analytically solving the Paris law. Record the crack propagation rate obtained from each temperature cycle to generate a second sequence of physical parameters; The modal types include physically degenerate dominant modes, thermo-coupled wave modes, and high-frequency noise-independent modes; The bandpass mode component in the dominant physical degradation mode is highly positively correlated with the crack propagation rate; The bandpass mode component in the thermo-coupled wave mode is related to the junction temperature curve of the IGBT; The bandpass mode component in the high-frequency noise-independent mode is represented as random high-frequency oscillation, which is independent of the evolution law of physical damage inside the IGBT.

Images