Bayesian Meta Random Graph Implicit Enhanced Network for Lithium-Ion Batteries Remaining Useful Life Prediction With Few-Shot and Uncertainty Interference

Accurately predicting and evaluating the remaining useful life (RUL) of lithium-ion batteries (LIBs) is of great significance to develop a maintenance program for replacing batteries according to the local conditions and effectively guarantee the normal operation of the equipment. Existing LIBs RUL prediction methods mainly utilize sufficient labeled data to achieve the capture of degraded tracks; however, labeled data are usually sparse and susceptible to outliers and missing values in practical applications, resulting in degraded data that are difficult to be effectively distinguished, which reduces the LIBs RUL prediction performance. To address this challenge, a new Bayesian meta random graph implicit enhanced network (BayesMRG) is proposed for LIBs RUL prediction in the case of few-shot and uncertainty interference. The model analyzes the feature representations of the learning nodes in each task in a bilevel optimization framework and summarizes the pairwise dependencies of the feature representations based on the changes of the tasks. Then, a random feature function is introduced to regulate the approximation accuracy of the features, and the random graph features are formed by multilayer update aggregation. In addition, the dual causal variable semantics are designed to mine the causal relationships between degraded tasks and variables in random graph features. Comprehensive experiments are conducted on three public battery datasets. Compared with the existing methods, the BayesMRG model improves the root-mean-square error (RMSE) and mean absolute error (MAE) by more than 43% and 42% on the NASA and CALCE datasets, respectively, and the prediction error on the MIT dataset is within ±0.01.