CN-122021326-A - Method and system for predicting residual life of electronic power component by fusing BiLSTM and GPR

CN122021326ACN 122021326 ACN122021326 ACN 122021326ACN-122021326-A

Abstract

The invention provides a method and a system for predicting the residual life of an electronic power component by fusing BiLSTM and GPR, and belongs to the technical field of life prediction of electronic devices. The method comprises the steps of carrying out an accelerated degradation experiment according to a degradation mechanism of an electronic power component to obtain degradation data, constructing and training BiLSTM-GPR mixed models, wherein the models comprise BiLSTM networks and GPR models, extracting nonlinear time sequence features in the degradation data through BiLSTM, outputting point prediction results, inputting the point prediction results into the GPR models to realize prediction uncertainty quantization, outputting probability prediction results comprising confidence intervals, inputting the degradation time sequence data of the electronic power component into the mixed models to predict, obtaining residual life point prediction values, biLSTM effectively capturing the nonlinear time sequence features in the degradation process, providing high-precision point prediction input for GPR, realizing uncertainty quantization on the basis of the GPR, and providing a feasible technical path for realizing accurate prediction of the residual life of the electronic power component by synergetically complementing the two.

Inventors

CHU XIAOXU
CHENG JINJUN
ZHU HAIZHEN
HU BIN
LI CHANGJUN

Assignees

中国人民解放军空军工程大学

Dates

Publication Date: 20260512
Application Date: 20260204

Claims (10)

1. The method for predicting the residual life of the electronic power component by fusing BiLSTM with the GPR is characterized by comprising the following steps of: carrying out an accelerated degradation experiment according to the degradation mechanism of the electronic power component to obtain degradation data of the electronic power component; modeling electronic power component degradation data based on BiLSTM-GPR hybrid model; Constructing and training BiLSTM-GPR mixed models, wherein the mixed models comprise BiLSTM networks and GPR models, extracting nonlinear time sequence characteristics in degradation data through BiLSTM networks, outputting point prediction results, inputting the point prediction results into the GPR models to realize prediction uncertainty quantification, and outputting probability prediction results containing confidence intervals; and (3) inputting the degradation time sequence data of the electronic power component into a BiLSTM-GPR mixed model for prediction to obtain a predicted value of the residual life point of the electronic power component.
2. The method of claim 1, wherein the BiLSTM network comprises two layers of stacked bi-directional LSTM cells with hidden cell numbers of 32 and 16, respectively, and the output layer of the bilstm network generates the final feature representation through an 8-dimensional full connection layer, and the input layer, the output layer and the current information state of the BiLSTM network are as follows: for the output layer in BiLSTM a, For the input of the current information state in BiLSTM, In order to conceal the layer weight matrix, For the output of the hidden layer of the previous period, In order to input the weight matrix, For the input layer in BiLSTM, As a result of the bias term, The state of the previous cycle and the current cycle And The method comprises the following steps: The output of the hidden layer is: wherein the output of the previous period hidden layer Hiding the input of the layer for the current period; Predictive value of current period Expressed as: And The activation functions of BiLSTM layers Sigmond and tanh functions, respectively.
3. The method for predicting the residual life of the electronic power component by fusing BiLSTM and GPR according to claim 2 is characterized in that in the BiLSTM network training process, an Adam optimizer is adopted, an initial learning rate is set to be 0.001, a ReduceLROnPlateau learning rate scheduling strategy is matched to accelerate convergence and avoid local optimization, dropout regularization is introduced between two LSTM layers, the discarding rate is 0.2, 3-fold cross validation is adopted, all super parameters of the GPR network are subjected to end-to-end learning through Exact Marginal Log-Likelihod, the Adam algorithm is adopted in the optimization process, and the initial learning rate is set to be 0.1.
4. The method for predicting remaining life of an electronic power component by fusing BiLSTM and GPR according to claim 1, wherein the GPR model uses a radial basis function as a covariance function, i.e., a kernel function: Establishing a regression model containing noise in a GPR network: Wherein, the In order to observe the value of the value, In order to predict the object of the present invention, For noise, calculating a predicted value according to the prior distribution of the observed value obtained by the noise and the combined prior distribution of the observed value and the predicted target Posterior distribution of (c): Wherein, the Is that Is used for the estimation of the (c), Covariance matrices for the test set are: Is a point predictor of the GPR network.
5. The method for predicting the remaining life of an electronic power component by fusing BiLSTM and GPR according to claim 1, wherein the prediction accuracy of the BiLSTM-GPR hybrid model obtained by training is evaluated by using point predictors including average absolute error, mean square error, root mean square error, average absolute error and a decision coefficient.
6. The method for predicting remaining life of an electronic power component by fusing BiLSTM and GPR as defined in claim 1, wherein the interval prediction accuracy of the BiLSTM-GPR hybrid model is evaluated using a prediction interval coverage, a prediction interval width, and a coverage width criterion.
7. The method for predicting remaining life of an electronic power component by fusing BiLSTM and GPR as defined in claim 1, wherein reliability of the model evaluation is evaluated by a probability integration transformation: Wherein, the In order to be able to predict the value, As a function of probability density.
8. The method for predicting the residual life of an electronic power component by fusing BiLSTM and GPR according to claim 1, wherein the step of obtaining the degradation data of the electronic power component by performing an accelerated degradation test according to the degradation mechanism of the electronic power component comprises the step of performing a reliability test on the electronic power component for a set period of time, wherein the reliability test comprises a high-temperature high-humidity test and a three-dimensional test, the high-temperature high-humidity test is set to 80 ℃ and the relative humidity is set to 80%, the three-dimensional test is set to high-temperature and low-temperature stress circulation, vibration stress and electric stress, the low-temperature condition is-20 ℃, the high-temperature condition is 80 ℃, the high-temperature condition and the low-temperature condition are respectively maintained for 1h, and the electric stress condition is that rated working voltage is applied to the electronic power component to be tested.
9. The electronic power component residual life prediction system integrating BiLSTM and GPR is characterized by comprising an electronic power component degradation data acquisition module, a modeling module and a prediction module; carrying out an accelerated degradation experiment according to the degradation mechanism of the electronic power component to obtain degradation data of the electronic power component; modeling electronic power component degradation data based on BiLSTM-GPR hybrid model; constructing and training BiLSTM-GPR mixed models, wherein the mixed models comprise BiLSTM networks and GPR models, extracting nonlinear time sequence characteristics in degradation data through BiLSTM networks, outputting point prediction results, inputting the point prediction results into the GPR models to realize prediction uncertainty quantification, and outputting probability prediction results containing confidence intervals; and (3) inputting the degradation time sequence data of the electronic power component into a BiLSTM-GPR mixed model for prediction to obtain a predicted value of the residual life point of the electronic power component.
10. The electronic power component remaining life prediction system of fusion BiLSTM with GPR as claimed in claim 9, wherein the modeling module includes a model building unit and a model training unit; The model building unit is based on a characteristic adaptive alignment and probabilistic mapping mechanism of BiLSTM network and GPR network, and is used for an input sequence with given length of T BiLSTM encodes it as a feature vector of fixed dimensions: Wherein, the GPR receives feature vectors for BiLSTM network parameters As input, and model from feature space to remaining life Is a hidden function of (2) Setting the function to obey the prior of the Gaussian process, and calculating an observation value by the hidden function and Gaussian noise; The model training unit is used for training BiLSTM networks and GPR networks, in the BiLSTM network training process, an Adam optimizer is adopted, the initial learning rate is set to be 0.001, a ReduceLROnPlateau learning rate scheduling strategy is matched to accelerate convergence and avoid local optimization, dropout regularization is introduced between two LSTM layers, the discarding rate is 0.2, 3-fold cross validation is adopted, all super parameters of the GPR networks are subjected to end-to-end learning through Exact Marginal Log-Likelihood, an Adam algorithm is adopted in the optimization process, and the initial learning rate is set to be 0.1.

Description

Method and system for predicting residual life of electronic power component by fusing BiLSTM and GPR Technical Field The invention relates to the field of reliability engineering, and provides a method and a system for predicting the residual life of an electronic power component by fusing BiLSTM and a GPR. Background The electronic power component is an important component in the field of power electronics, and is widely applied to safety sensitive scenes such as automobile electronics, industrial control, rail transit, aerospace and the like which need efficient electric energy control. However, the degradation phenomenon of the device, which occurs as it is operated for a long time, causes the working performance thereof to gradually decrease. Thereby affecting the reliability and safety of the whole system. Meanwhile, the degradation process of the electronic power component is not an occasional event, but a progressive process with obvious time sequence dependency and accumulation effect, the early stress effect and damage accumulation can continuously influence the subsequent performance change, and certain statistical regularity is shown in the whole life cycle of the device. Reliability problem research on electronic power components has become a key to promote long-term stable operation of power electronic systems. Thus, accurate predictions of the state of health (SOH) and Remaining Useful Life (RUL) of the device are critical to the application of electronic power components. The DC-DC power module is a power electronic device for converting direct current voltage into direct current voltage of different levels, and is an indispensable key component in the electronic system at present. In the long-term operation process, the DC-DC power supply module can be degraded due to factors such as temperature stress, humidity stress, electric stress and vibration, and the degradation mechanism covers the most typical failure mode in the electronic power component, so that the multi-stress accelerated degradation experiment is carried out on the DC-DC power supply module, not only is the RUL of the DC-DC power supply module facilitated to be predicted, but also a reference basis can be provided for reliability modeling, fault Prediction and Health Management (PHM) of other electronic power components, and the method has important significance in improving the robustness of the system, prolonging the service life and realizing intelligent operation and maintenance. The RUL prediction method of the electronic power component can be mainly classified into two types, a model-based method and a data-driven-based method. The model-based method needs to construct an equivalent circuit model and an electrochemical model for the device according to a modeling mechanism, and further deduces a performance evolution rule closely related to an intrinsic degradation mechanism. The method has high prediction precision and strong interpretability, but is complex in modeling and time-consuming in solving, and particularly for devices with complex internal structures and remarkable multi-physical field coupling of electronic power components, the difficulty and the cost for building a physical model are too high, so that the method is limited to be widely applied to actual engineering. The data driving method uses a statistical correlation method to mine the degradation information of the device and the evolution rule of the health state from the historical data, and has become the main stream of life Prediction and Health Management (PHM), and the existing data driving method can be divided into a statistical analysis method and a machine learning method. The statistical analysis method generally relies on assumptions of monotonicity, stationarity and the like of a degradation process, and the introduced principal component analysis method is difficult to characterize nonlinearity and non-stationarity of a device under complex working conditions and is sensitive to noise. Meanwhile, in order to overcome the limitation of the statistical method, researchers gradually turn to a more flexible machine learning method, and the traditional machine learning method is similar to a traditional machine learning method, and although the problem of excessively strong model assumption is solved to a certain extent, the long-term dependency relationship of device degradation is difficult to effectively capture. Therefore, researchers have proposed solving this problem by a Recurrent Neural Network (RNN) by which to construct health indicators for health status assessment of devices, which, while solving the timing dependency problem, is unstable in prediction of long-term degradation trend. The long-short-term memory network (LSTM) effectively relieves the problem of gradient disappearance in the RNN by introducing a gating mechanism, can effectively transfer and express information in a long-time sequence, and can not cause useful inf