Bayesian igbt residual life prediction method and system fusing physical mechanism

Abstract

This invention proposes a Bayesian method and system for predicting the remaining lifetime of IGBTs based on physical mechanisms, belonging to the field of IGBT remaining lifetime prediction technology. The method includes: collecting time-series data of saturation on-state voltage drop and junction temperature during IGBT operation, and preprocessing this data to obtain a time-series feature sequence; inputting the preprocessed time-series feature sequence into a trained Bayesian long short-term memory network model, injecting physical damage calculated by the Coffin-Manson thermal fatigue model as a dynamic prior into the network weights, and constraining the posterior distribution to approach this prior using KL divergence; extracting deep features from the input sequence using the trained model, performing Monte Carlo sampling, and outputting the predicted mean and variance of the remaining lifetime; inputting the predicted mean into a parameter-free hard-coded layer, and outputting the final corrected predicted remaining lifetime value and confidence level. This invention improves the prediction accuracy, robustness, and reliability of IGBT remaining lifetime prediction.

Description

Technical Field

[0001] This invention belongs to the field of IGBT remaining lifetime prediction technology, and particularly relates to a Bayesian IGBT remaining lifetime prediction method and system that integrates physical mechanisms. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] IGBTs, as core power devices in power electronic converters, are widely used in new energy power generation, electric vehicles, and rail transportation. Operating under complex conditions of high-frequency switching and multi-stress coupling, they are susceptible to aging failures such as solder layer fatigue and bond wire detachment due to factors like temperature fluctuations, sudden load changes, and electromagnetic interference. Statistics show that power semiconductor device failures account for 21% of all power converter failures, with aging failures caused by thermal cycling accounting for as much as 55% of these. Therefore, accurate prediction of IGBT remaining lifespan is crucial for ensuring system safety and reducing operation and maintenance costs.

[0004] Currently, IGBT lifetime prediction methods have mainly gone through four development stages. First, the experience-driven stage, relying on device manuals and fixed lifetime models, resulted in low accuracy and poor generalization ability. Second, the model-driven stage, establishing analytical models based on failure physics theories, such as the Coffin-Manson equation and the Paris formula, while possessing clear physical interpretability, heavily depended on precise material parameters and stress information, making it difficult to adapt to complex and changing field conditions. Third, the purely data-driven stage, relying on machine learning and deep learning algorithms to mine degradation patterns from operational data, avoiding complex physical modeling; however, its prediction results highly depended on the quality and completeness of the training data, and its generalization ability significantly decreased when data distribution was off or failure samples were lacking.

[0005] In recent years, Physical Information Neural Networks (PINNs) have become a research hotspot. These methods incorporate physical rules into the loss function in the form of partial differential equation residuals or regularization terms, guiding the model to learn features that conform to physical laws, thus combining the advantages of physical modeling and data-driven approaches to some extent. However, existing PINN methods still have significant shortcomings in IGBT lifetime prediction: First, model initialization lacks physical guidance, generally using a physically meaningless standard normal distribution as the weight prior, leading to blind parameter optimization in the early training stages; second, physical constraints often employ mathematical methods such as second-order difference smoothing, failing to reflect the microscopic fracture mechanism of solder layer cracks propagating exponentially in the later stages of lifetime; third, fixed constraint weights and a lack of uncertainty quantification make prediction results prone to collapse under harsh conditions such as strong electromagnetic interference; fourth, the lack of monotonicity and hard boundary constraints on prediction results easily leads to unreasonable phenomena such as "lifetime rebound" or "numerical out-of-bounds" phenomena. Summary of the Invention

[0006] To overcome the shortcomings of the existing technologies, this invention provides a Bayesian IGBT remaining lifetime prediction method and system that integrates physical mechanisms. This method enables physical mechanisms to guide deep learning at the gene level, aligns with the fact of accelerated aging in the late stages, eliminates tedious parameter tuning, significantly reduces the incidence of unreasonable predictions, and comprehensively improves the scientific nature and safety of preventive maintenance of equipment.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a Bayesian method for predicting the remaining lifetime of IGBTs that integrates physical mechanisms; Bayesian methods for predicting the remaining lifetime of IGBTs that integrate physical mechanisms include: The timing data of saturation on-state voltage drop and junction temperature during IGBT operation are collected and preprocessed to obtain timing feature sequences. The preprocessed temporal feature sequence is input into the trained Bayesian long short-term memory network model; wherein, the construction process of the Bayesian long short-term memory network model includes: extracting the junction temperature data of the current working condition, calculating the theoretical physical damage by substituting it into the Coffin-Manson thermal fatigue model, constructing the dynamic prior distribution of the model weights, and using KL divergence to constrain the posterior distribution of the weights to approach the dynamic prior distribution. The trained model extracts deep features from the input sequence and performs Monte Carlo sampling, outputting the mean of the remaining lifetime prediction and the variance representing the prediction uncertainty. The predicted mean output by the Bayesian Long Short-Term Memory Network model is input into a parameter-free hard-coded layer, and the final corrected remaining lifetime prediction value and the confidence level represented by the variance are output.

[0008] As a further technical solution, the saturated on-state voltage drop time series data is preprocessed, including: Using the period of the IGBT gate square wave signal as the unit, the average value of the saturation on-state voltage drop timing data in each period is calculated to achieve average downsampling; The saturated on-state voltage drop time series data after downsampling are processed to have zero mean and unit variance. The standardized data are smoothed using an exponential moving average algorithm.

[0009] As a further technical solution, the junction temperature data under the current operating conditions is extracted, including the average junction temperature and junction temperature fluctuation, and then the theoretical physical damage is calculated by substituting it into the Coffin-Manson thermal fatigue model.

[0010] in, Where A is the theoretical fatigue life cycle number, and A is a material constant. Stress index For activation energy, Boltzmann's constant; This refers to the junction temperature fluctuation range; This represents the average junction temperature. Calculate the theoretical physical damage expectation caused by a single thermal cycle:

[0011] in, This represents the expected value of physical damage. This is a scaling function used to map extremely small damage values ​​to a reasonable range of values ​​for the neural network weights. Using the expected value of the physical damage as the mean of the prior distribution, and combining it with a preset empirical constant, a dynamic prior distribution of the Bayesian LSTM network weights is constructed:

[0012] in, It is a dynamic prior distribution; These are preset empirical constants.

[0013] As a further technical solution, deep features are extracted from the input sequence using a trained model, and Monte Carlo sampling is performed to output the mean of the remaining lifetime prediction and the variance representing the prediction uncertainty, including: A Bayesian LSTM structure is used to extract aging features from the preprocessed saturated on-state voltage drop time series data, and the full-cycle hidden state sequence is output. A Bayesian fully connected layer is used to map the hidden state sequence extracted by LSTM to the original remaining lifetime prediction value; Perform Monte Carlo sampling on the same input sample to obtain the original predicted values, and calculate the predicted mean and variance of the original predicted values:

[0014]

[0015] in, To predict the mean, To characterize the variance of prediction uncertainty, The larger the value, the lower the confidence level of the model in the current prediction result, and vice versa.

[0016] As a further technical solution, it also includes constructing a dynamic regulator based on the variance, adaptively adjusting the physical constraint weights, and constructing a composite loss function for model training; the dynamic regulator is:

[0017] in, Based on the weighting coefficient, This is the exponential growth coefficient.

[0018] As a further technical solution, the composite loss function is:

[0019] in, It is a composite loss function; Let be the data fitting term, where Mean square error; Labels represent the actual remaining lifespan. These are the physical constraint terms based on the Paris formula; This is the KL divergence term.

[0020] As a further technical solution, the predicted mean output by the Bayesian Long Short-Term Memory network model is input into a parameter-free hard-coded layer, and the final corrected remaining lifetime prediction value and the confidence level represented by the variance are output, including: At each prediction time, the theoretical minimum physical damage is dynamically calculated based on real-time junction temperature data and the Coffin-Manson-Arrhenius model. The upper envelope of the remaining lifetime at the current moment is dynamically updated by subtracting the theoretical minimum physical damage increment from the corrected remaining lifetime prediction value output in the previous cycle. The predicted mean output by the Bayesian Long Short-Term Memory Network model is projected into the dynamic feasible region formed by the upper envelope and the zero lower bound, and the final corrected remaining lifetime prediction value and confidence level are output.

[0021] A second aspect of the present invention provides a Bayesian IGBT remaining lifetime prediction system that integrates physical mechanisms.

[0022] A Bayesian IGBT remaining lifetime prediction system incorporating physical mechanisms includes: The data acquisition and preprocessing module is configured to: acquire the saturation on-state voltage drop timing data and junction temperature data during IGBT operation, and perform preprocessing to obtain the timing feature sequence; The model construction and feature extraction module is configured to: input the preprocessed temporal feature sequence into the trained Bayesian long short-term memory network model; wherein, the construction process of the Bayesian long short-term memory network model includes: extracting the junction temperature data of the current working condition, calculating the theoretical physical damage by substituting it into the Coffin-Manson thermal fatigue model, thereby constructing the dynamic prior distribution of the model weights, and using KL divergence to constrain the posterior distribution of the weights to approach the dynamic prior distribution; The prediction output module is configured to: extract deep features from the input sequence using a trained model, perform Monte Carlo sampling, and output the mean of the remaining lifetime prediction and the variance representing the prediction uncertainty; The hard-coded correction module is configured to: input the predicted mean output by the Bayesian long short-term memory network model into the parameterless hard-coded layer, and output the final corrected remaining lifetime prediction value and the confidence level represented by the variance.

[0023] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of a Bayesian IGBT remaining lifetime prediction method incorporating physical mechanisms as described in the first aspect of the present invention.

[0024] A fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in a Bayesian IGBT remaining lifetime prediction method incorporating physical mechanisms as described in the first aspect of the present invention.

[0025] The above one or more technical solutions have the following beneficial effects: (1) This invention transforms the analytical model of IGBT thermal fatigue (Coffin-Manson equation) into a dynamic prior distribution at the bottom layer of a Bayesian network. By extracting the junction temperature fluctuation amplitude in real time, calculating the expected value of theoretical physical damage and mapping it to the mean of the prior distribution, the weight distribution of the model already conforms to the real thermodynamic evolution law before it comes into contact with any training data. This mechanism effectively reduces the ineffective optimization space of parameters, fundamentally breaks the black box limitation of pure data-driven models, and significantly improves the physical fidelity and convergence efficiency of the model. In addition, this invention introduces the Paris formula of fracture mechanics into the loss function to construct the physical residual term of the partial differential equation. This mechanism forces the neural network to learn the true rate of microcrack propagation, enabling the model to accurately capture the late-stage accelerated aging mutation trend that is difficult to fit by pure data-driven methods.

[0026] (2) This invention uses Bayesian theory and Monte Carlo sampling output prediction variance as an uncertainty index to provide an intuitive confidence evaluation for operation and maintenance decisions. Simultaneously, a dynamic adjustment mechanism for physical constraint weights based on this uncertainty index is established: when data noise is high and model confidence is low, the physical constraint strength is automatically amplified exponentially to achieve anti-interference self-healing; when data confidence is high, the constraint weights are automatically reduced to prioritize data fitting accuracy. This mechanism eliminates tedious manual parameter tuning and significantly enhances the robustness of the model under complex operating conditions. Furthermore, a parameter-free hard-coded layer based on thermoelectric coupling envelope is set at the prediction output end, overcoming the limitation of existing physical information neural networks where "soft constraints" cannot eliminate non-physical predictions.

[0027] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0028] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0029] Figure 1 This is a flowchart of the method in the first embodiment.

[0030] Figure 2 This is a schematic diagram of the overall method structure of the first embodiment.

[0031] Figure 3 This is a system structure diagram of the second embodiment. Detailed Implementation

[0032] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0033] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0034] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0035] The overall approach of this invention addresses the pain points of existing IGBT remaining lifetime prediction methods, such as blind model optimization, violation of physical constraints on late-stage accelerated degradation mechanisms, and lack of confidence. This invention first uses saturated on-state voltage drop time-series data and junction temperature data as input, and injects physical damage calculated by the Coffin-Manson thermal fatigue model as a dynamic prior into a Bayesian neural network to guide the model's initial learning direction. A Bayesian Long Short-Term Memory (LSTM) network model is constructed using Bayesian LSTM and fully connected layers, and Monte Carlo sampling is used to simultaneously output the prediction mean and uncertainty index. The loss weights of the physical constraint terms in the Paris formula are adaptively adjusted based on the prediction uncertainty. Finally, a parameter-free hard-coded layer based on the thermoelectric coupling envelope is set at the output to dynamically correct the prediction results through physical boundary projection. This approach, through multi-level physical mechanism guidance and constraints, achieves a deep integration of data-driven and physical laws, ensuring that the prediction results possess both high accuracy and high physical reliability.

[0036] Example 1 This embodiment discloses a Bayesian IGBT remaining lifetime prediction method that integrates physical mechanisms; like Figure 1 and Figure 2 As shown, a Bayesian method for predicting the remaining lifetime of IGBTs that integrates physical mechanisms includes: Step S1: Collect the saturation on-state voltage drop timing data and junction temperature data during IGBT operation, and perform preprocessing to obtain the timing feature sequence.

[0037] Step S11: Collect timing data of saturation on-state voltage drop during IGBT operation. The data comes from IGBT thermal overstress experiments, ensuring coverage of the complete aging process from brand new to near failure. A sharp drop in signal strength serves as a failure indicator, and invalid data after the failure point is truncated.

[0038] Step S12, process the collected saturated on-state voltage drop time series data. Preprocessing is performed first. Average downsampling, because a square wave signal is applied to the IGBT gate in the experiment, The signal is in square wave form. The calculation is performed within each square wave period. The average value is used to downsample the high-frequency original signal into a periodic low-frequency sequence, reducing data redundancy and computational complexity.

[0039] Subsequently, after downsampling The data is processed to have zero mean and unit variance, and the calculation formula is as follows:

[0040] in, After zero-mean, unit variance processing Standardized data values; After downsampling Original data values; After downsampling The mean of the overall data series, After downsampling The standard deviation of the overall data series.

[0041] Next, the standardized data is smoothed using the exponential moving average algorithm, with a decay factor. The width of the sliding window The core formula is:

[0042] in, The output after exponential moving average smoothing for the current t-th period. Eigenvalues; to They represent all historical data traced back sequentially from the previous cycle to the initial 0th cycle. Standardized values; For the current t-th cycle input Standardized values; the above processing can effectively filter noise and glitches in timing signals, highlighting... The core trend of degradation.

[0043] Finally, the above pretreatment resulted in a stable solution. Temporal feature sequences ,in For the t-th period The processed value, T, represents the total number of periods. A 10-step sliding window of samples is constructed, meaning each input sample consists of 10 consecutive periods. The data is then processed, and the preprocessed feature sequences are output. These 10 output feature sequences are then fed into the subsequent prediction model, corresponding to the remaining lifetime value of the output cycle. The sample format is as follows: .

[0044] Step S2: Input the preprocessed temporal feature sequence into the trained Bayesian Long Short-Term Memory Network model; wherein, the construction process of the Bayesian Long Short-Term Memory Network model includes: extracting the junction temperature data of the current working condition, calculating the theoretical physical damage by substituting it into the Coffin-Manson thermal fatigue model, thereby constructing the dynamic prior distribution of the model weights, and using KL divergence to constrain the posterior distribution of the weights to approach the dynamic prior distribution.

[0045] To address the shortcomings of traditional Bayesian neural networks, which default to using a standard normal distribution with no physical meaning as the weight prior and thus cause the network to blindly seek optimization in the early stages of training, the model constructed in this embodiment injects the laws of thermodynamic evolution into the underlying probability space of deep learning.

[0046] Specifically, the theoretical physical damage is first calculated by substituting the average junction temperature and junction temperature fluctuation under the current operating conditions from the collected junction temperature data into the Coffin-Manson thermal fatigue model:

[0047] in, Where A is the theoretical fatigue life cycle number, and A is a material constant. Stress index For activation energy, Boltzmann's constant; This refers to the junction temperature fluctuation range; This represents the average junction temperature. Subsequently, the expected theoretical physical damage caused by a single thermal cycle was calculated:

[0048] in, This represents the expected value of physical damage. This is a scaling function used to map extremely small damage values ​​to a reasonable range of values ​​for the neural network weights. Next, the expected value of the physical damage is used as the mean of the prior distribution, and combined with a preset empirical constant, a dynamic prior distribution of the Bayesian LSTM network weights is constructed:

[0049] in, It is a dynamic prior distribution; These are preset empirical constants.

[0050] This setup both anchors the network's initial subconsciousness to the vicinity of real physical laws through the mean and preserves a fault-tolerant buffer space for subsequent feature fine-tuning through the empirical constant variance. Finally, the KL divergence between the posterior distribution of the network weights and this dynamic prior distribution is calculated. It is used as a regularization term in the total loss function to guide the model's parameter updates during training to stay true to the underlying physical laws.

[0051] Step S3: Extract deep features from the input sequence using the trained model, perform Monte Carlo sampling, and output the mean of the remaining lifetime prediction and the variance representing the prediction uncertainty.

[0052] The preprocessed periodic low-frequency time series feature sequence The input is fed into a Bayesian LSTM for deep feature extraction. Unlike traditional deterministic neural networks, the Bayesian LSTM network models its internal weight matrix and bias parameters as probability distributions controlled by mean and variance, enabling it not only to capture degradation features in time-series data but also to perceive the uncertainty of model parameters.

[0053] First, a two-layer Bayesian LSTM structure is used for extraction. The aging characteristics in time-series data are represented by an LSTM layer with 80 units per layer. All weights and biases in this layer are modeled as normally distributed, rather than as fixed values. .in, The mean and standard deviation of the weighted distribution are given. The mean and standard deviation of the bias distribution are trainable parameters. The final output is the full-cycle hidden state sequence. ,Include Aging information.

[0054] A three-layer Bayesian fully connected layer is used to map the full-cycle hidden state sequence extracted by LSTM to the original remaining lifetime prediction value. The weights and biases of each layer are also modeled as a normal distribution, ultimately outputting unconstrained original prediction values. .

[0055] Finally, perform M=10 Monte Carlo samplings on the same input sample (i.e., independently sample 10 times from the weight distribution of the Bayesian layer) to obtain 10 sets of original predicted values. Calculate the forecast mean and uncertainty:

[0056]

[0057] in, To predict the mean, The variance, which characterizes the uncertainty of prediction, is the uncertainty index. The larger the value, the lower the confidence level of the model in the current prediction result, and vice versa.

[0058] Furthermore, the obtained uncertainty metrics are used in turn to guide model training, specifically: According to uncertainty Adjusting the weight of physical constraint terms in the total loss in real time The core formula is:

[0059] in, Based on the weighting coefficient, This is an exponential growth coefficient, which can be fine-tuned according to actual working conditions. When When the noise level increases (due to high data noise and low model confidence), Exponential scaling strengthens physical constraints, forces the model to follow aging patterns, and filters out interference. When When reducing (high data confidence, high model confidence), Reduce or weaken the constraints, and prioritize ensuring the accuracy of data fitting.

[0060] The total loss consists of the data fitting term, the physical constraint term, and the KL divergence term, as shown in the formula:

[0061] in, It is a composite loss function; For data fitting terms; These are the physical constraint terms based on the Paris formula; This is the KL divergence term.

[0062] Among them, the data fitting term Used to predict the mean Compared with the actual remaining life label The mean squared error ensures that the model fits the aging patterns in the data.

[0063] Physical constraints ,in, The internal crack propagation coefficient of the IGBT; The stress intensity factor amplitude is generated by the thermomechanical coupling alternating stress acting on the tip of the microcrack, and its magnitude is driven by real-time junction temperature fluctuations. The crack propagation index of the material is used; the Paris formula from fracture mechanics is introduced as a physical loss term, with weights... It is adaptively controlled by a dynamic regulator.

[0064] The physical essence of IGBT core aging failures (such as solder layer fatigue and bond wire detachment) is the thermomechanical fatigue crack propagation process of the metal material caused by the mismatch of the coefficients of thermal expansion (CTE) between the chip, solder, and copper substrate under frequent switching alternating thermal stress. Since the growth rate of this microcrack completely follows the kinetic law defined by the Paris equation, and crack propagation directly leads to an increase in device thermal resistance and an exponentially accelerated decline in late-stage lifetime, the crack evolution trend characterized by the Paris equation is used as the physical residual loss. This forces the lifetime decay rate predicted by the neural network to conform to the actual microscopic material fracture law, thereby accurately capturing the abrupt change trend of "late-stage accelerated aging" that is difficult for purely data-driven models to fit.

[0065] The KL divergence term is used to constrain the weight distribution of the Bayesian layer. Close to the prior distribution The calculation formula is:

[0066] in, For weight The corresponding posterior distribution variance; For weight The square of the corresponding posterior distribution mean.

[0067] This measure prevents the weight distribution from being too wide or the mean from deviating from a reasonable range, ensuring the authenticity of uncertainty quantification, while avoiding model overfitting.

[0068] Step S4: Input the predicted mean output by the Bayesian Long Short-Term Memory Network model into the parameterless hard-coded layer, and output the final corrected remaining lifetime prediction value and the confidence level represented by the variance.

[0069] At the end of the prediction phase, this embodiment innovatively incorporates a parameterless hard-coded layer to perform dynamic feasible region projection driven by physical mechanisms. This layer has no parameters and does not increase the computational overhead of online inference; instead, it forces the prediction mean to be adjusted through real-time environment awareness. To impose bottom-line physical constraints and prevent "blindly optimistic" predictions due to lagging deep learning data, the specific steps and formulas are as follows: First, at each prediction time t, the thermo-electric coupling mapper synchronously reads the real-time junction temperature fluctuation amplitude for that period. With average junction temperature Substituting this into the classic Coffin-Manson-Arrhenius thermodynamic fatigue analytical model, the theoretically maximum number of cycles the device can withstand under this severe condition is calculated. :

[0070] in, This represents the theoretically maximum number of cycles that the device can withstand.

[0071] Based on this, the theoretical minimum physical damage increment caused by thermal shock in a single cycle can be calculated:

[0072] in, This represents the theoretical minimum physical damage increment.

[0073] Secondly, a dynamic upper bound envelope boundary is constructed. Based on the final effective remaining lifetime of the cache in the previous cycle, the theoretical minimum damage for the current cycle is subtracted to deduce the most conservative upper bound for the lifetime allowed at the current moment:

[0074] in, This represents the most conservative upper limit of the lifespan allowed at the current moment. This represents the final effective remaining lifetime of the cache from the previous cycle. This is the lifetime-scale mapping coefficient. It is used when the system experiences extreme conditions (such as a short circuit causing a sudden increase in temperature). It will magnify instantly. The envelope will experience a significant, step-like downward movement. If there is no historical value for the first period, then set... .

[0075] Finally, dynamic feasible region projection correction is performed. The original predictions output by the Bayesian network are projected onto the projection of the feasible region. Within the dynamic physically feasible domain formed by the baseline 0:

[0076] in, This is the final output value for the predicted remaining lifetime. These are the original predicted values.

[0077] Through this projection operator, if the network's predicted value... Greater than This violates the objective law that physical damage is inevitable, and the body will be forcibly cut off to the level of the envelope without mercy; at the same time, it ensures that the lifespan is not negative.

[0078] Finally, the output is a remaining lifetime prediction with high physical fidelity. It also outputs uncertainty indicators simultaneously, providing confidence support for operation and maintenance decisions. This enables the system to immediately issue reasonable warnings of a precipitous drop in lifespan when faced with sudden strong thermal stress, reducing the incidence of unreasonable predictions.

[0079] Example 2 This embodiment discloses a Bayesian IGBT remaining lifetime prediction system that integrates physical mechanisms; like Figure 3 As shown, a Bayesian IGBT remaining lifetime prediction system incorporating physical mechanisms includes: The data acquisition and preprocessing module is configured to: acquire the saturation on-state voltage drop timing data and junction temperature data during IGBT operation, and perform preprocessing to obtain the timing feature sequence; The model construction and feature extraction module is configured to: input the preprocessed temporal feature sequence into the trained Bayesian long short-term memory network model; wherein, the construction process of the Bayesian long short-term memory network model includes: extracting the junction temperature data of the current working condition, calculating the theoretical physical damage by substituting it into the Coffin-Manson thermal fatigue model, thereby constructing the dynamic prior distribution of the model weights, and using KL divergence to constrain the posterior distribution of the weights to approach the dynamic prior distribution; The prediction output module is configured to: extract deep features from the input sequence using a trained model, perform Monte Carlo sampling, and output the mean of the remaining lifetime prediction and the variance representing the prediction uncertainty; The hard-coded correction module is configured to: input the predicted mean output by the Bayesian long short-term memory network model into the parameterless hard-coded layer, and output the final corrected remaining lifetime prediction value and the confidence level represented by the variance.

[0080] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0081] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in a Bayesian IGBT remaining lifetime prediction method incorporating physical mechanisms as described in Example 1.

[0082] Example 4 The purpose of this embodiment is to provide an electronic device.

[0083] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in a Bayesian IGBT remaining lifetime prediction method incorporating physical mechanisms as described in Example 1.

[0084] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0085] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0086] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A Bayesian IGBT residual life prediction method fusing physical mechanisms, characterized in that, The application relates to a method for predicting the remaining life of an IGBT, comprising: Collecting and preprocessing saturation on-voltage time series data and junction temperature data during IGBT operation to obtain a time series feature sequence; Inputting the preprocessed time series feature sequence into a trained Bayesian long short-term memory network model; wherein the construction process of the Bayesian long short-term memory network model comprises: extracting the junction temperature data of the current working condition, calculating the theoretical physical damage by substituting the Coffin-Manson thermal fatigue model, constructing the dynamic prior distribution of the model weight, and constraining the posterior distribution of the weight to tend to the dynamic prior distribution through KL divergence; Extracting deep features from the input sequence through the trained model, performing Monte Carlo sampling, and outputting the remaining life prediction mean value and the variance representing the prediction uncertainty; Inputting the prediction mean value output by the Bayesian long short-term memory network model into a parameter-free hard coding layer to output the final corrected remaining life prediction value and the confidence represented by the variance.

2. The physics-of-fusion based Bayesian IGBT residual life prediction method of claim 1, wherein, The preprocessing of the saturation on-voltage time series data comprises: Calculating the average value of the saturation on-voltage time series data in each period in units of IGBT gate square wave signal period to realize average down-sampling; Performing zero-mean unit-variance processing on the down-sampled saturation on-voltage time series data; Using the exponential moving average algorithm to smooth the standardized data.

3. The physics-of-fusion based Bayesian IGBT residual life prediction method of claim 1, wherein, Extracting the junction temperature data of the current working condition, including the average junction temperature and the junction temperature fluctuation, and calculating the theoretical physical damage by substituting the Coffin-Manson thermal fatigue model: wherein, Nf is the theoretical fatigue life cycle number, A is a material constant, n is the stress exponent, Ea is the activation energy, kB is the Boltzmann constant; ΔTj is the junction temperature fluctuation amplitude; Tj is the average junction temperature; Calculating the expected value of the theoretical physical damage caused by a single thermal cycle: wherein, is a physical impairment expectation value; is a scale mapping function for mapping minimal impairment values to a reasonable numerical interval of neural network weights; Taking the physical damage expectation value as the mean value of the prior distribution, combining a pre-set empirical constant, and constructing the dynamic prior distribution of the Bayesian LSTM network weight: wherein, is a dynamic prior distribution; is a preset empirical constant.

4. The physics-of-fusion based Bayesian IGBT residual life prediction method of claim 1, wherein, Extracting deep features from the input sequence through the trained model, performing Monte Carlo sampling, and outputting the remaining life prediction mean value and the variance representing the prediction uncertainty, comprising: Using the Bayesian LSTM structure to extract the aging features in the preprocessed saturation on-voltage time series data and output the full-cycle hidden state sequence; Using the Bayesian fully connected layer to map the hidden state sequence extracted by the LSTM into the original remaining life prediction value; Performing Monte Carlo sampling on the same input sample to obtain the original prediction value, and calculating the prediction mean value and the variance of the original prediction value: wherein, is the mean of the prediction, is the variance of the prediction uncertainty, The larger the value, the lower the confidence of the model in the current prediction result, and vice versa.

5. The physics-of-fusion based Bayesian IGBT residual life prediction method of claim 4, wherein, Further comprising a dynamic regulator constructed based on the variance for adaptively adjusting the physical constraint weight, and a composite loss function for model training; the dynamic regulator is: wherein, is a base weight coefficient, is an exponential growth coefficient.

6. The physics-of-fusion based Bayesian IGBT residual life prediction method of claim 5, wherein, The composite loss function is: wherein, is a composite loss function; is a data fitting term, wherein, is a mean squared error; is a true residual life label; is a physical constraint term based on Paris formula; is a KL divergence term.

7. The physics-of-fusion Bayesian IGBT residual life prediction method of claim 1, wherein, Inputting the prediction mean value output by the Bayesian long short-term memory network model into a parameter-free hard coding layer to output the final corrected remaining life prediction value and the confidence represented by the variance, comprising: At each prediction time, dynamically calculating the theoretical minimum physical damage based on the real-time junction temperature data and the Coffin-Manson-Arrhenius model; Dynamically updating the upper envelope line of the remaining life at the current time based on the corrected remaining life prediction value output in the last period minus the theoretical minimum physical damage increment of the current period. The predicted mean output by the Bayesian Long Short-Term Memory Network model is projected into the dynamic feasible region formed by the upper envelope and the zero lower bound, and the final corrected remaining lifetime prediction value and confidence level are output.

8. A Bayesian IGBT residual life prediction system fusing physical mechanisms, characterized by, include: The data acquisition and preprocessing module is configured to: acquire the saturation on-state voltage drop timing data and junction temperature data during IGBT operation, and perform preprocessing to obtain the timing feature sequence; The model building and feature extraction module is configured to: input the preprocessed temporal feature sequence into the trained Bayesian long short-term memory network model; wherein, the construction process of the Bayesian long short-term memory network model includes: extracting the junction temperature data of the current working condition, calculating the theoretical physical damage by substituting it into the Coffin-Manson thermal fatigue model, constructing the dynamic prior distribution of the model weights, and using KL divergence to constrain the posterior distribution of the weights to approach the dynamic prior distribution; The prediction output module is configured to: extract deep features from the input sequence using a trained model, perform Monte Carlo sampling, and output the mean of the remaining lifetime prediction and the variance representing the prediction uncertainty; The hard-coded correction module is configured to: input the predicted mean output by the Bayesian long short-term memory network model into the parameterless hard-coded layer, and output the final corrected remaining lifetime prediction value and the confidence level represented by the variance.

9. A computer-readable storage medium having stored thereon a program, characterized in that, When executed by the processor, the program implements the steps in the Bayesian IGBT remaining lifetime prediction method that incorporates physical mechanisms as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the Bayesian IGBT remaining lifetime prediction method that integrates physical mechanisms as described in any one of claims 1-7.

Images