Power electronic device life prediction method with uncertainty perception

Abstract

The invention relates to the technical field of power electronic device reliability and life prediction, in particular to a power electronic device life prediction method with uncertainty perception, which comprises the following steps: S1, acquiring time sequence observation data of a power electronic device, enhancing the data utilization rate through feature modeling after data preprocessing, and inputting a Transform model; s2, in the training and prediction process of the Transform model, a data uncertainty model and a model uncertainty model are established respectively, and the data uncertainty Aleoric and the model uncertainty Epiric are distinguished; s3, synchronously outputting a residual service life prediction mean value of the power electronic device and a calibrated confidence interval by using the trained Transform model; according to the method, the problem of insufficient reliability of life prediction under sparse data can be effectively solved, and the model uncertainty and the data uncertainty can be distinguished and calibrated, so that a more reliable and explainable decision basis is provided for health management of power electronic devices.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power electronic device reliability and life prediction, in particular to a power electronic device life prediction method with uncertainty perception. BACKGROUND

[0002] In the technical field of power electronic device reliability and life prediction, the residual useful life (RUL) prediction of power semiconductor devices such as SiC MOSFET, IGBT, etc. is of great significance for system health management, state monitoring and operation and maintenance decision-making. However, existing data-driven life prediction methods still have many problems in actual engineering application, especially under the condition of sparse data.

[0003] At present, most data-driven life prediction models tend to output deterministic point estimation results, lacking quantification of prediction uncertainty, leading to "overconfident" predictions in engineering applications, making it difficult to support risk-sensitive decisions. Existing methods generally fail to effectively distinguish between model uncertainty Epistemic caused by insufficient model cognition and data uncertainty Aleatoric caused by data noise, making the prediction interval often too wide or too narrow in practice, lacking a reasonable calibration mechanism, limiting its application in safety margin design.

[0004] Under the condition of sparse data, due to limited sample size, significant measurement noise, and frequent operating condition changes, traditional models are more likely to overfit, leading to prediction results deviating from the true degradation law. Although existing research attempts to quantify uncertainty by introducing Bayesian methods or ensemble learning, these methods generally do not incorporate the physical mechanism of device degradation, lacking prior constraints on the degradation process, making the model lack generalization ability when data is sparse, and reducing the credibility of the prediction results. SUMMARY

[0005] The purpose of the present application is to provide a power electronic device life prediction method with uncertainty perception, which can effectively solve the problem of insufficient life prediction reliability under sparse data, and can distinguish and calibrate model uncertainty and data uncertainty, thereby providing more reliable and interpretable decision-making basis for health management of power electronic devices.

[0006] The purpose of the present application is achieved by the following technical solutions:

[0007] A power electronic device life prediction method with uncertainty perception, the method comprising the following steps:

[0008] Step S1: Obtain time series observation data of the power electronic device, and after data preprocessing, enhance data utilization by feature modeling, and input the Transformer model;

[0009] Step S2: In the training and prediction process of the Transformer model, data uncertainty model and model uncertainty model are established respectively to distinguish between data uncertainty Aleatoric and model uncertainty Epistemic;

[0010] Step S3: Using the trained Transformer model, the remaining useful life prediction mean value and the calibrated confidence interval of the power electronic device are output synchronously;

[0011] In the step S1, the acquisition process of the time series observation data of the power electronic device is:

[0012] The key degradation parameters of the power electronic device are obtained from the power cycle experiment, the on-resistance and threshold voltage of the power electronic device are extracted, the multi-source time series data of the power electronic device are standardized collected, cleaned, synchronized, constructed and labeled; the observation sequence of a single power electronic device is:

[0013]

[0014] Where t k is the time stamp, is the measured degradation amount, is the external stress and working condition, and K represents the total number of sampling time steps;

[0015] The feature data collected is standardized, considering that different sample points have different importance in the sparse data scenario, a weight factor w k is introduced based on the definition of degradation rate or signal-to-noise ratio.

[0016]

[0017] In the formula, α represents the adjustment coefficient, which is used to control the amplification strength of the degradation rate on the weight, which represents the time derivative of the degradation feature, thus obtaining the sparse data standardization formula:

[0018]

[0019] In the formula, x(t k ) represents the original observation value at time t k ; μ x : the mean value of the corresponding feature in the training set, which is used to eliminate the influence of different dimensions; σ x represents the standard deviation of the corresponding feature in the training set, which is used to scale the data to the same scale, and ∈ is a small number to prevent division by zero.

[0020] Subsequently, the data is smoothed, and an adaptive kernel smoothing device is proposed for sparse data. The weight of the observation point with large noise in the window is automatically reduced, thereby improving the smoothing accuracy and robustness of sparse data. The smoothing calculation formula is as follows:

[0021]

[0022] In the formula, x(t k+i ) represents the observation value at time t k+i , Δ represents the sampling interval, h represents the smoothing bandwidth, σ 2 (t k+i ) represents the observation noise variance at this time, j represents the index variable in the smoothing window, and i represents the index of the specific point, which is used to calculate the weight of the current point.

[0023] In step S1, the feature modeling introduces physical prior constraint features, degradation uncertainty perception features, sparse data enhancement features, and cross-modal graph structure features.

[0024] Physical prior constraint features:

[0025] The thermal fatigue feature based on the Coffin-Manson model is introduced, and the device physical life model is directly embedded into the deep learning input:

[0026] F CM (t k )=N cyc (t k )·(ΔT j (t k )) -β

[0027] In the formula, N cyc is the cumulative cycle number, ΔT j is the junction temperature cycle amplitude, and β is the material constant.

[0028] Through the equivalent life consumption factor based on the Miner cumulative damage criterion, the remaining life consumption proportion of the device is directly quantified. The cumulative damage value of the device at time t is:

[0029]

[0030] In the formula, N i is the number of cycles applied, N fi is the failure cycle number under the corresponding working condition, i represents the i-th stress working condition, and f(·) represents the life model function. According to the specific failure mechanism, the Arrhenius model, current impact model, radiation dose model, or multi-stress coupling model is selected.

[0031] Degradation uncertainty perception features:

[0032] The prediction distribution parameters are taken as features, and the measurement noise and the working condition disturbance size are reflected through the data features:

[0033]

[0034] In the formula, logσ 2 (x(t k )) represents the log variance output by the network;

[0035] The model cognitive uncertainty is reflected through the model features:

[0036]

[0037] In the formula, M represents the number of samples, represents the prediction result obtained by the mth sampling;

[0038] The uncertainty ratio feature is used to determine whether the prediction error is mainly caused by data noise or model defects:

[0039]

[0040] Sparse data enhancement feature:

[0041] A pseudo-continuous signal is constructed using limited samples, and a neighboring difference residual feature is introduced to represent the information gain of the sparse sampling points:

[0042]

[0043] In the formula, represents the interpolation estimated value;

[0044] Cross-modal graph structure feature:

[0045] The multi-source stress-degradation parameters are modeled as a graph structure, and the adjacency matrix feature is extracted;

[0046] A feature graph G=(V,E) is constructed, where the node set V={x1,x2,…,x d} corresponds to different degradation quantities and stress factors, and the edge weight A ij =corr(x i ,x j ), and the spatial / physical coupling relationship between the multivariate is extracted through the graph Laplacian:

[0047] f graph =Lx,L=D-A

[0048] In the formula, L is the graph Laplacian matrix, x is the node observation value, D represents the degree matrix of the graph, and A is the adjacency matrix of the feature graph;

[0049] The fusion of degradation amount, stress and uncertainty forms an extended health index HI + (t k ), which reflects the device degradation degree, stress level and prediction credibility, and is used as a unified health state input, and the formula is:

[0050]

[0051] In the formula, λ1, λ2, λ3 represent the weighting coefficients, R dson represents the on-resistance, V th represents the threshold voltage,

[0052] Organize these features into a time series tensor, and finally construct a comprehensive feature vector as the model input:

[0053]

[0054] In the step S2, the life degradation monotonicity constraint, the life degradation rate regularization, the energy conservation type constraint and the deterministic shrinkage constraint are introduced in the training of the Transformer model;

[0055] Life degradation monotonicity constraint:

[0056] For the predicted degradation life RUL(tk), it is required to change monotonically with time, if there is a "rebound" or "shock" that violates the monotonicity, the penalty term > 0, otherwise 0, and the constraint loss is defined as:

[0057]

[0058] In the formula, s∈{+1,-1} represents the degradation direction, and the degradation direction s is dynamically used as a parameter, which can adapt to different degradation amounts;

[0059] Life degradation rate regularization:

[0060] The device degradation rate should not appear abnormal shock, and the second-order difference constraint is introduced to constrain the smoothness of the predicted curve, and to avoid non-physical oscillation caused by model overfitting under sparse data;

[0061]

[0062] Energy conservation type constraint:

[0063] Degradation is often related to stress consumption accumulation, and the predicted life consumption Consistency constraint with Miner cumulative damage D(t):

[0064]

[0065] D(t) is the life consumption calculated based on the stress model, consumption ratio converted from RUL predicted by network

[0066] uncertainty shrinkage constraint

[0067] for data variance ask it to match the true residual square, (y(t k )-μ(t k )) 2 ; match

[0068]

[0069] where y(t k ) represents the true observation, μ(t k ) represents the mean predicted by the model

[0070] the life degradation monotonicity constraint, life degradation rate regularization, energy conservation type constraint and certainty shrinkage constraint are combined to comprehensively constrain the physical constraint loss:

[0071] L phy = δ1L mono + δ2L rate + δ3L energy + δ4L unc

[0072] where δ1, δ2, δ3, δ4 are weight coefficients, and the above achieves not only to constrain monotonicity, but also to constrain rate smoothing, energy conservation, uncertainty calibration.

[0073] the data uncertainty model is modeled

[0074] the observation noise variance is decomposed into three parts of sensor, interpolation / sparse sampling, process disturbance

[0075]

[0076] sensor noise:

[0077]

[0078] where θ T is a temperature sensitive coefficient. θ0 is a constant term, representing the baseline noise level, T j (t k ) represents the junction temperature, θ I represents a current sensitive coefficient, I(t k ) represents the device working current, θ V represents a voltage sensitive coefficient, V(t k ) represents the device working voltage; logσ 2 : log modeling is used to ensure that the variance is non-negative and changes monotonically with stress;

[0079] The interpolation / sparse term calculation formula related to local sampling sparsity and curvature is:

[0080]

[0081] In the formula, η0: constant term, that is, minimum interpolation noise, η1: sampling interval sensitivity coefficient, Δt k : adjacent observation time interval, the larger the interval, the larger the interpolation error, η2: curvature sensitivity coefficient, The second derivative of the degradation curve, that is, curvature, the larger the curvature, the larger the error in sparse sampling;

[0082] The process noise calculation formula is:

[0083]

[0084] In the formula, ζ0: constant term, representing the reference process noise; ζ1: comprehensive stress sensitivity coefficient;

[0085] S total (t k ): comprehensive stress factor;

[0086] ζ2: degradation rate sensitivity coefficient;

[0087] The first derivative of the degradation curve, that is, rate, the faster the degradation, the larger the process noise;

[0088] The data likelihood loss function is constructed as:

[0089]

[0090] The model uncertainty model is modeled as:

[0091] Considering that different sampling results have different reliabilities, the variance comes not only from the output dispersion, but also from the difference between the model uncertainty model and the internal representation. The weight ω m is defined, and the uncertainty is:

[0092]

[0093] In the formula, Based on the similarity calculation of the intermediate hidden layer representation h m and the mean , τ represents the temperature parameter, which controls the weight distribution.

[0094] The model uncertainty model variance is decomposed into intra-layer uncertainty and inter-layer uncertainty. The innovative introduction of hierarchical decomposition can analyze whether the uncertainty comes from the lack of shallow feature learning or the fluctuation of deep prediction head:

[0095]

[0096] In the formula, l represents a network layer index, β l represents an allocation coefficient, controls the influence of different layers;

[0097] In addition to the variance, the KL divergence between the prediction distributions is introduced to comprehensively define the Epistemic uncertainty:

[0098]

[0099] In the formula, p m (y|x) represents the prediction distribution obtained by the mth sampling, KL(·||·) represents the Kullback-Leibler divergence, and λ KL represents a weight factor;

[0100] The uncertainty modeling formula is:

[0101]

[0102] In the formula, represents a weighted sampling variance, represents hierarchical variance decomposition, represents a KL divergence correction term, and α, β and γ represent adjustment coefficients;

[0103] The model uncertainty loss function is constructed as:

[0104] The final uncertainty optimization target is L unc-total = L ale + L epi , L ale represents a data likelihood loss; and L epi is a model loss.

[0105] The beneficial effects of the present application are:

[0106] By introducing physical prior constraints and feature modeling, the robustness of the model under small sample and sparse data conditions is improved, the mean absolute error MAE can be reduced by 20%, the mean square error MSE is reduced by 15%, and the distortion of the prediction result due to insufficient training samples is avoided;

[0107] A prediction framework is established to represent the model uncertainty Epistemic and the data uncertainty Aleatoric at the same time, the prediction mean and confidence interval are output, the deviation between the prediction interval coverage rate PICP and the nominal confidence is less than ±5%, the interval width PINAW is narrowed by 10% on average while ensuring the coverage rate, and a basis is provided for engineering safety margin design and risk assessment;

[0108] The model parameter quantity is reduced by 30% compared with a traditional deep integration method, and the single inference time is shortened by more than 20%. BRIEF DESCRIPTION OF DRAWINGS

[0109] The application will be described in further detail below with reference to the drawings and specific implementation methods.

[0110] Figure 1 is a power electronic device life prediction method flowchart with uncertainty of the application;

[0111] Figure 2 is a comparative schematic diagram of the remaining useful life prediction and uncertainty analysis of the silicon carbide MOSFET #10 by the deterministic Transformer only point estimation;

[0112] Figure 3 is a comparative schematic diagram of the remaining useful life prediction and uncertainty analysis of the silicon carbide MOSFET #10 by the Transformer using heteroscedastic Gaussian likelihood and MC-Dropout;

[0113] Figure 4 is a comparative schematic diagram of the remaining useful life prediction and uncertainty analysis of the silicon carbide MOSFET #10 of the application. DETAILED DESCRIPTION

[0114] The application will be described in further detail below with reference to the drawings and specific implementation methods.

[0115] As Figures 1 to 4 shown below, the steps and functions of a power electronic device life prediction method with uncertainty perception are described in detail;

[0116] A power electronic device life prediction method with uncertainty perception, the method comprising the following steps:

[0117] Step S1: Obtain the time series observation data of the power electronic device, and after data preprocessing, enhance the data utilization rate through feature modeling, and input the Transformer model;

[0118] Step S2: In the training and prediction process of the Transformer model, respectively establish a data uncertainty model and a model uncertainty model, and distinguish between data uncertainty Aleatoric and model uncertainty Epistemic;

[0119] Step S3: Use the trained Transformer model to simultaneously output the remaining useful life prediction mean value and the calibrated confidence interval of the power electronic device;

[0120] In the step S1, the acquisition process of the time series observation data of the power electronic device is:

[0121] The key degradation parameters of power electronic devices are obtained from power cycle experiments, and the on-resistance and threshold voltage of power electronic devices are extracted. The multi-source time series data of power electronic devices are standardized, cleaned, synchronized, constructed and labeled. The observation sequence of a single power electronic device is as follows:

[0122]

[0123] Where t k is the timestamp, is the measured degradation quantity, such as R DS(on) , V TH , junction temperature estimation gate charge, etc. is the external stress and working condition, such as current I and duty cycle, and K represents the total number of time steps obtained by sampling;

[0124] The collected feature data is standardized. Considering that different sample points have different importance in the sparse data scenario, a weight factor w k is introduced based on the definition of degradation rate or signal-to-noise ratio.

[0125]

[0126] In the formula, α represents an adjustment coefficient for controlling the amplification strength of the degradation rate on the weight, which represents the time derivative of the degradation feature. Thus, the sparse data standardization formula is obtained as follows:

[0127]

[0128] In the formula, x(t k ) represents the original observation value at time t k ; μ x represents the mean value of the corresponding feature in the training set, which is used to eliminate the influence of different dimensions; σ x represents the standard deviation of the corresponding feature in the training set, which is used to scale the data to the same scale, and ∈ is a small number to prevent division by zero.

[0129] Subsequently, the data is smoothed, and an adaptive kernel smoothing processor is proposed for device sparse data. The observation points with larger noise in the window automatically reduce the weight, thereby improving the smoothing accuracy and robustness of sparse data. The smoothing calculation formula is as follows:

[0130]

[0131] In the formula, x(t k+i ) represents the observation value at time t k+i , Δ represents the sampling interval, h represents the smoothing bandwidth, and σ 2 (tk+i ) represents the observation noise variance at this moment, j represents the index variable within the smoothing window, i represents the index of the specific point, and is used to calculate the weight of the current point;

[0132] Based on noise perception adaptive kernel smoothing, automatically reduce the weight of high noise points, improve the effective signal-to-noise ratio under sparse sampling, and reduce the noise variance estimation error by 18%; The rate weight w k Strengthen the degradation acceleration zone, the late prediction deviation of RUL is reduced by 10-15%.

[0133] In order to improve the prediction accuracy and robustness under sparse data conditions, the original observation of power electronic devices is modeled by features, not limited to conventional statistical features and time-frequency features. Feature modeling introduces physical prior constraint features, degradation uncertainty perception features, sparse data enhancement features and cross-modal graph structure features;

[0134] Physical prior constraint features:

[0135] Introduce thermal fatigue features based on Coffin-Manson model, and directly embed device physical life model into deep learning input:

[0136] F CM (t k )=N cyc (t k )·(ΔT j (t k )) -β

[0137] In the formula, N cyc is the cumulative cycle number, ΔT j is the junction temperature cycle amplitude, and β is the material constant;

[0138] Through the equivalent life consumption factor based on Miner's cumulative damage criterion, the remaining life consumption proportion of the device is directly quantified. The cumulative damage value of the device at time t:

[0139]

[0140] In the formula, N i is the number of cycles applied, N fi is the failure cycle number under the corresponding working condition, i represents the i-th stress condition, and f(·) represents the life model function. According to the specific failure mechanism, select Arrhenius model, current impact model, radiation dose model or multi-stress coupling model;

[0141] Degradation uncertainty perception features:

[0142] The application not only uses the degradation amount, but also predicts the distribution parameter as a feature, and reflects the measurement noise and the working condition disturbance size through the data features:

[0143]

[0144] In the formula, logσ 2 (x(t k )) represents the log variance output by the network;

[0145] Reflect the model cognitive uncertainty through the model features:

[0146]

[0147] In the formula, M represents the sampling number, represents the prediction result obtained by the mth sampling;

[0148] Use the uncertainty ratio feature to determine whether the prediction error is mainly caused by data noise or model defects:

[0149]

[0150] Sparse data enhancement feature:

[0151] Pseudo-continuous signals are constructed using limited samples, and a neighboring difference residual feature is introduced to represent the information gain of sparse sampling points:

[0152]

[0153] In the formula, represents the interpolation estimated value;

[0154] Cross-modal graph structure feature:

[0155] Model the multi-source stress-degradation parameter as a graph structure, and extract the adjacency matrix feature;

[0156] Construct a feature graph G=(V,E), wherein the node set V={x1,x2,…,x d} corresponds to different degradation amounts and stress factors, and the edge weight A ij =corr(x i ,x j ), and the spatial / physical coupling relationship between the multiple variables is extracted through the graph Laplacian feature:

[0157] f graph =Lx,L=D-A

[0158] In the formula, L is the graph Laplacian matrix, x is the node observation value, D represents the degree matrix of the graph, and A is the adjacency matrix of the feature graph;

[0159] The fusion of degradation amount, stress and uncertainty forms an extended health index HI + (t k ), which reflects the device degradation degree, stress level and prediction credibility, and is used as a unified health state input, and the formula is:

[0160]

[0161] In the formula, λ1, λ2, λ3 represent weighting coefficients, R dson represents the on-resistance, V th represents the threshold voltage,

[0162] Organize these features into a time series tensor, and finally construct a comprehensive feature vector as the model input:

[0163]

[0164] The fusion of physical prior features makes the model fit the real degradation trajectory more closely, and the overall mean absolute error is reduced by 20-35%;

[0165] In the step S2, the life degradation monotonicity constraint, the life degradation rate regularization, the energy conservation type constraint and the deterministic shrinkage constraint are introduced in the training of the Transformer model;

[0166] Life degradation monotonicity constraint:

[0167] For the predicted degradation life RUL(tk), it is required to change monotonically with time, if there is a "rebound" or "shock" that violates the monotonicity, the penalty term > 0, otherwise it is 0, and the constraint loss is defined as:

[0168]

[0169] In the formula, s∈{+1,-1} represents the degradation direction, and the degradation direction s is dynamically used as a parameter, which can adapt to different degradation amounts;

[0170] Life degradation rate regularization:

[0171] The device degradation rate should not appear abnormal shock, and the second-order difference constraint is introduced to constrain the smoothness of the predicted curve, and to avoid non-physical oscillation caused by model overfitting under sparse data;

[0172]

[0173] Energy conservation type constraint:

[0174] Degradation is often related to stress consumption accumulation, and the predicted life consumption Consistency constraint with Miner cumulative damage D(t):

[0175]

[0176] D(t) is the life consumption calculated based on the stress model, The consumption ratio obtained by converting the RUL predicted by the network;

[0177] Uncertainty contraction constraint:

[0178] To ensure that the prediction uncertainty is not arbitrary noise, but "calibrated confidence", for the data variance It is required to match the true residual square, (y(t k )-μ(t k )) 2 ;

[0179]

[0180] In the formula, y(t k ) represents the true observation value, and μ(t k ) represents the mean value predicted by the model;

[0181] The life degradation monotonicity constraint, the life degradation rate regularization, the energy conservation type constraint and the certainty contraction constraint are combined to comprehensively constrain the physical constraint loss:

[0182] L phy =δ1L mono +δ2L rate +δ3L energy +δ4L unc

[0183] In the formula, δ1δ2δ3δ4 are weight coefficients, and the above realizes not only the constraint of monotonicity, but also the constraint of rate smoothing, energy conservation and uncertainty calibration;

[0184] The monotonicity and curvature regularization make the physical consistency rate of the prediction curve≥95%, effectively inhibiting the prediction rebound and shock.

[0185] The present application introduces a double-channel probability modeling framework in life prediction, simultaneously describes the data uncertainty caused by data noise and the model uncertainty caused by insufficient model cognition, and proposes a multi-level fusion and calibration mechanism, to realize high-confidence RUL prediction;

[0186] The data uncertainty model is modeled:

[0187] The observation noise variance is decomposed into three parts of sensor, interpolation / sparse sampling and process disturbance

[0188]

[0189] Sensor noise:

[0190]

[0191] In the formula, θ T This represents the temperature sensitivity coefficient. θ0 is a constant term representing the reference noise level, T j (t k ) represents the junction temperature, θ I I(t) represents the current sensitivity coefficient. k ) represents the device operating current, θ V V(t) represents the voltage sensitivity coefficient. k ) represents the device operating voltage; logσ 2 : Use logarithmic modeling to ensure that the variance is non-negative and changes monotonically with stress;

[0192] The formula for calculating the interpolation / sparse term related to local sampling sparsity and curvature is as follows:

[0193]

[0194] In the formula, η0: constant term, i.e., minimum interpolation noise; η1: sampling interval sensitivity coefficient; Δt k η2: Curvature sensitivity coefficient; the larger the interval between adjacent observations, the greater the interpolation error. The second derivative of the degradation curve is its curvature. The greater the curvature, the greater the error in sparse sampling.

[0195] The formula for calculating process noise is:

[0196]

[0197] In the formula, ζ0: constant term, representing the reference process noise; ζ1: comprehensive stress sensitivity coefficient;

[0198] S total (t k ): Comprehensive stress factor;

[0199] ζ2: Sensitivity coefficient for degradation rate;

[0200] The first derivative of the degradation curve is the rate; the faster the degradation, the greater the process noise.

[0201] The data likelihood loss function is constructed as follows:

[0202]

[0203] The model is modeled based on uncertainty.

[0204] Considering the varying reliability of different sampling results, variance stems not only from output dispersion but also from differences in internal model representations due to model uncertainty. Therefore, weights ω are defined. mThe uncertainty is:

[0205]

[0206] where, Based on the intermediate hidden layer representation h m The similarity calculation with the mean , τ represents the temperature parameter, and controls the weight distribution;

[0207] The model uncertainty is decomposed into intra-layer uncertainty and inter-layer uncertainty, and the hierarchical decomposition is introduced to analyze whether the uncertainty comes from the insufficient learning of shallow features or the fluctuation of deep prediction heads:

[0208]

[0209] where, l represents the network layer index, represents the prediction contribution of the m-th sample at the l-th layer, β l represents the allocation coefficient, which controls the influence of different layers;

[0210] In addition to the variance, the KL divergence between the prediction distributions is also introduced to comprehensively define the Epistemic uncertainty:

[0211]

[0212] where, p m (y|x) represents the prediction distribution obtained by the m-th sample, KL(·||·) represents the Kullback–Leibler divergence, λ KL represents the weight factor;

[0213] The uncertainty modeling formula is:

[0214]

[0215] where, represents the weighted sampling variance, represents the hierarchical variance decomposition, represents the KL divergence correction term, α, β, γ represent the adjustment coefficients;

[0216] The model uncertainty loss function is constructed as:

[0217] The final uncertainty optimization objective is L unc-total = L ale + L epi , L ale represents the data likelihood loss; L epi is the model loss; the uncertainty matches the real error distribution, and the nominal deviation is ≤3–5%, and the interval width is narrowed by 10–25%.

[0218] The uncertainty-aware lifetime prediction of SiCMOSFET under sparse data is taken as an example to illustrate the implementation process of the power electronic device lifetime prediction method with uncertainty awareness of the present application in detail.

[0219] The uncertainty-aware power electronic device lifetime prediction method of the present application is verified for its improved engineering applicability and prediction accuracy on the basis of sparse observation data of a batch of 1200V SiCMOSFETs in power cycle tests. The test objects are 20 devices (No. #01-#20) of the same batch.

[0220] The test conditions are set as follows: the on-state current of the device is I on = 30A, the duty cycle is 50%, and the ambient temperature is The original data are sampled at 1Hz to obtain the on-state resistance R DS(on) , the threshold voltage V TH , the estimated junction temperature , the bus voltage V DS , the current I and the switching frequency f sw . The failure criterion is ΔR DS(on) ≥ 20%, and the failure time T f corresponds to the failure.

[0221] The sequence is sparsified by extracting one point every 10 minutes, and about K≈1200 observations are formed for a single device. The training / validation / testing is divided at the device level: #01-#14 for training, #15-#17 for validation, and #18-#20 for testing.

[0222] In the preprocessing stage, the “data preprocessing module” is used to suppress the high-variance noise points caused by sparsity. The window w=3, the bandwidth h=2Δ, and the numerical stability term ∈=10 -6 . At the same time, in order to highlight the key information amount of the degradation acceleration zone, the rate-driven weight (α=0.4) is used. The standardization parameters (μ,σ of Z-score) are only estimated in the training set and are frozen in the validation / test stage to avoid data leakage.

[0223] The feature modeling follows the “physics-data-uncertainty” fusion idea. In order to explicitly introduce the physical mechanism, the physical features driven by thermal fatigue and electrical stress are extracted. The thermal fatigue feature adopts the Coffin-Manson form, and in order to let the model identify the error source, the uncertainty-aware feature is constructed. The data uncertainty noise is modeled by “source decomposition”, and the model uncertainty is obtained by 30 times of MC-Dropout deep sampling. Considering the coupling structure between the degradation amount and the stress amount, the cross-modal graph feature is constructed.

[0224] Further, to integrate degradation magnitude, stress intensity and confidence in a single channel, we define an extended health index HI + (t k ), λ1=0.6, λ2=0.25, λ3=0.15. This index rises with both degradation and noise, and can be directly used as a maintenance threshold.

[0225] The final feature vector contains the original degradation and its derivative physical life features {F CM (t), D(t)}, uncertainty features sparse enhancement residuals δx(t), cross-modal graph features {f graph (t), L 2 x(t)}, and extended health index HI + (t) (each channel is Z-score normalized and masked).

[0226] Aleatoric uncertainty is decomposed into sensor, interpolation / sparse sampling, and process disturbance, with respective initialization ranges: θ . ∈ [0, 0.5], η . ∈ [10 -4 , 10 -2 ], ζ . ∈ [10 -4 , 10 -2 ].

[0227] Epistemic uncertainty adopts MC-Dropout (M=30, dropout=0.1) with temperature parameter τ=5.0 for weighted variance, and hierarchical decomposition covering 4 layers of encoder, β l ={0.2, 0.3, 0.3, 0.2}, and KL term measures the difference between sampled distributions.

[0228] The model backbone adopts a physical information Transformer (4 / 2 layers of encoder / decoder, 8 attention heads, 256 hidden dimensions, 512 feed-forward dimensions), with accumulated damage added in positional encoding. Optimizer adopts Adam with initial learning rate 10 -3 cosine annealing to 10 -5 , batch size 64, and 200 epochs; loss weights are set as α=0.6, β=0.8, γ=0.4.

[0229] The present invention module shows significant advantages compared to Comparative Example 1 (remove Aleatoric decomposition and physical constraints) and Comparative Example 2 (only normal MC-Dropout). The MAE mean of RUL point estimation is 82 (Comparative Example 1: 108; Comparative Example 2: 123), the MSE is 1.34x10^4 (Comparative Example 1: 1.62x10^4; Comparative Example 2: 1.85x10^4), and the negative log-likelihood NLL is 1.72 (Comparative Example 1: 2.06; Comparative Example 2: 2.21). In terms of interval quality, the prediction interval coverage probability is 94.1% (closer to nominal), and the prediction interval normalized average width is 0.142 (Comparative Example 1: 0.189, Comparative Example 2: 0.205), showing the characteristics of "narrow and accurate" interval calibration.

[0230] Table 1 Index tie results for #18-#20 devices

[0231] Indicator Invention Comparative Example 1 Comparative Example 2 Mean Squared Error 82 108 123 Mean Absolute Error 1.34e4 1.62e4 1.85e4 Prediction Interval Coverage Probability 94.1％ 87.5％ 96.8％ Prediction Interval Normalized Average Width 0.142 0.189 0.205

[0232] Explanation:

[0233] Comparative Example 1 = remove Aleatoric decomposition (with single variance) + no physical constraints;

[0234] Comparative Example 2 = keep Aleatoric decomposition, but no calibration / monotonic / hierarchical Epistemic, only normal MC-Dropout.

Claims

1. A power electronic device lifetime prediction method with uncertainty perception, characterized by: The method comprises the following steps: Step S1: Obtain the time sequence observation data of the power electronic device, and after data preprocessing, enhance the data utilization rate through feature modeling, and input the Transformer model; Step S2: In the training and prediction process of the Transformer model, a data uncertainty model and a model uncertainty model are established respectively, and the data uncertainty Aleatoric and the model uncertainty Epistemic are distinguished; Step S3: Using the trained Transformer model, the mean value of the remaining service life prediction of the power electronic device and the calibrated confidence interval are output synchronously.

2. The power electronic device lifetime prediction method with uncertainty awareness according to claim 1, wherein: In the step S1, the acquisition process of the time sequence observation data of the power electronic device is as follows: Obtain the key degradation parameters of the power electronic device from the power cycle experiment, extract the on-resistance and threshold voltage of the power electronic device, and standardize, clean, synchronize, construct and label the multi-source time sequence data of the power electronic device; the observation sequence of a single power electronic device is as follows: where t k is a timestamp, is a measured degradation amount, is an external stress and working condition, and K represents the total number of sampling time steps. For the collected feature data, standardization processing is performed, and considering that different sample points have different importance in a sparse data scenario, a weight factor w is introduced based on the definition of degradation rate or signal-to-noise ratio k ; In the formula, α represents an adjustment coefficient, used to control the amplification strength of the degradation rate on the weight, The time derivative of the degradation feature is represented, and thus, the sparse data standardization processing formula is obtained: In the formula, x(t k ) represents the original observation value at time t k ; μ x : the mean of the corresponding feature in the training set, used to eliminate the influence of different dimensions; σ x represents the standard deviation of the corresponding feature in the training set, used to scale the data to the same scale, and ∈ is a small number to prevent division by zero; Then, the data is smoothed, and an adaptive kernel smoothing processor is proposed for sparse data of the device, the weight of the observation point with large noise in the window is automatically reduced, so as to improve the smoothing accuracy and robustness of the sparse data, and the smoothing calculation formula is as follows: In the formula, x(t) k+i ) represents time t k+i The observed values, where Δ represents the sampling interval, h represents the smoothing bandwidth, and σ 2 (t k+i ) represents the observation noise variance at that moment, j represents the index variable within the smoothing window, and i represents the index of the specific point, used to calculate the weight of the current point.

3. The power electronic device lifetime prediction method with uncertainty awareness according to claim 1, wherein: In the step S1, the feature modeling introduces physical prior constraint features, degradation uncertainty perception features, sparse data enhancement features and cross-modal graph structure features.

4. The power electronic device life prediction method with uncertainty perception according to claim 3, characterized in that: Physical prior constraint features: Introduce the thermal fatigue features based on the Coffin-Manson model, and directly embed the device physical life model into the deep learning input: F CM (t k )=N cyc (t k )·(ΔT j (t k )) -β In the formula, N cyc is the cumulative cycle number, ΔT j is the junction temperature cycle amplitude, and β is a material constant. Through the equivalent life consumption factor based on the Miner cumulative damage criterion, the residual life consumption proportion of the device is directly quantified, and the cumulative damage value of the device at time t is: N fi = f(·) where N i is the number of cycles applied, N fi is the number of cycles to failure under the corresponding stress condition, i denotes the i-th stress condition, and f(·) denotes a life model function, which is selected according to the specific failure mechanism as the Arrhenius model, the current shock model, the irradiation dose model, or the multi-stress coupling model. Degradation uncertainty perception features: Take the prediction distribution parameters as features, reflect the measurement noise and working condition disturbance size through data features: where logσ 2 (x(t k )) represents the log variance output by the network; Reflect the model cognitive uncertainty through model features: In the formula, M represents the number of sampling times, represents the prediction result obtained by the mth sampling. Use the uncertainty ratio feature to judge whether the prediction error is mainly caused by data noise or model defects: Sparse data enhancement features: Construct a pseudo-continuous signal using a limited sample, and introduce a neighboring difference residual feature to represent the information gain of the sparse sampling point: In the formula, is expressed as an interpolated estimate; Cross-modal graph structure features: Model the multi-source stress-degradation parameters as a graph structure, and extract the adjacency matrix feature; A graph G = (V, E) is constructed, where the node set V = {x1, x2,..., x d} corresponds to different degradation quantities and stress factors, and the edge weight A ij = corr(x i , x j ) extracts the spatial / physical coupling relationship between multiple variables by graph Laplacian. f graph = Lx, L = D - A In the formula, L is the graph Laplacian matrix, x is the node observation value, D represents the degree matrix of the graph, and A is the adjacency matrix of the feature graph; The fusion of degradation amount, stress and uncertainty forms the extended health index HI + (t k ), which reflects the device degradation degree, stress level and prediction confidence, and is used as a unified health state input, and the formula is: where λ1, λ2, λ3 represent weighting coefficients, R dson denotes the on-resistance, V th denotes the threshold voltage, organizing these features into a temporal tensor, and finally constructing a comprehensive feature vector as the model input:

5. The power electronic device lifetime prediction method with uncertainty awareness according to claim 1, wherein: In the step S2, the training of the Transformer model introduces life degradation monotonicity constraint, life degradation rate regularization, energy conservation constraint and deterministic shrinkage constraint.

6. The power electronic device life prediction method with uncertainty perception according to claim 1, characterized in that: Life degradation monotonicity constraint: For the predicted degradation life RUL(tk), it is required to change monotonically with time. If there is a "rebound" or "shock" that violates monotonicity, a penalty > 0 is imposed, otherwise 0. The constraint loss is defined as: In the formula, s∈{+1,-1} represents the degradation direction. The degradation direction s is dynamically taken as a parameter, which can adapt to different degradation amounts; Life degradation rate regularization: The device degradation rate should not have abnormal shock. The second-order difference constraint is introduced to constrain the smoothness of the prediction curve, avoiding non-physical oscillation caused by model overfitting under sparse data; Energy conservation constraint: Degradation is often related to stress consumption accumulation, defining the life consumption prediction Consistency constraint with Miner cumulative damage D(t): D(t) is the life consumption calculated based on the stress model, a consumption ratio converted from the RUL predicted by the network; Uncertainty shrinkage constraint: For data variance It is required that it is matched with the true residual square, (y(t k )-μ(t k )) 2 Match; In the formula, y(t) k ) represents the actual observed value, μ(t) k ) represents the mean predicted by the model.

7. The power electronic device lifetime prediction method with uncertainty awareness according to claim 6, wherein: The life degradation monotonicity constraint, life degradation rate regularization, energy conservation constraint and certainty shrinkage constraint are combined to form a comprehensive physical constraint loss: L phy = δ1L mono + δ2L rate + δ3L energy + δ4L unc In the formula, δ1, δ2, δ3, δ4 are weight coefficients. The above realizes not only the constraint of monotonicity, but also the constraint of rate smoothness, energy conservation and uncertainty calibration.

8. The power electronic device lifetime prediction method with uncertainty awareness according to claim 1, wherein: The data uncertainty model is modeled: The observation noise variance is decomposed into three parts of sensor, interpolation / sparse sampling and process disturbance Sensor noise: where θ T is the temperature sensitivity coefficient. θ0is the constant term, representing the baseline noise level, T j (t k ) is the junction temperature, θ I is the current sensitivity coefficient, I(t k ) is the device operating current, θ V is the voltage sensitivity coefficient, V(t k ) is the device operating voltage; logσ 2 : log-based modeling, ensuring non-negative variance and monotonic change with stress; The interpolation / sparse item calculation formula related to local sampling sparsity and curvature is: wherein η0: constant term, i.e. minimum interpolation noise, η1: sampling interval sensitivity coefficient, Δt k : adjacent observation time interval, the larger the interval, the larger the interpolation error, η2: curvature sensitivity coefficient, derivative of the degeneration curve, i.e. curvature, the larger the curvature, the larger the error when sparsely sampling; The process noise calculation formula is: In the formula, ζ0: constant term, represents the reference process noise; ζ1: comprehensive stress sensitivity coefficient; S total (t k ) : integrated stress factor; ζ2: degradation rate sensitivity coefficient; The first derivative of the degradation curve, i.e. the rate, the faster the degradation, the greater the process noise; The data likelihood loss function is constructed as:

9. The power electronic device lifetime prediction method with uncertainty awareness according to claim 8, wherein: The model uncertainty model is modeled: Considering the credibility of different sampling results is different, the variance not only comes from the dispersion of output, considering the model uncertainty model internal representation difference, define the weight ω m , uncertainty is: wherein Based on the intermediate hidden layer representation h m Similarity to the mean of the similarity computation, τ denotes a temperature parameter, controlling the weight distribution; The model uncertainty model variance is divided into intra-layer uncertainty and inter-layer uncertainty. The innovative hierarchical decomposition can analyze whether the uncertainty is caused by insufficient shallow feature learning or deep prediction head fluctuation: In the formula, l represents a network layer index, βml represents a prediction contribution of the mth sampling at the lth layer, l βml represents a prediction contribution of the mth sampling at the lth layer, In addition to variance, the KL divergence between prediction distributions is also introduced to comprehensively define Epistemic uncertainty: where p m (y | x) denotes the predictive distribution obtained from the m-th sample, KL(· || ·) denotes the Kullback-Leibler divergence, λ KL denotes the weight factor; Uncertainty modeling formula: wherein denotes the weighted sampling variance, denotes the hierarchical variance decomposition, denotes the KL divergence correction term, and a, b, g denote adjustment factors; The model uncertainty loss function is constructed as:

10. The power electronic device lifetime prediction method with uncertainty awareness according to claim 9, wherein: The final uncertainty optimization objective is L unc-total = L ale + L epi , L ale represents the data likelihood loss; L epi is the model loss.

Images