Remaining Useful Life Prediction of Lithium-Ion Batteries Using Neural Networks with Adaptive Bayesian Learning

With smart electronic devices delving deeper into our everyday lives, predictive maintenance solutions are gaining more traction in the electronic manufacturing industry. It is imperative for the manufacturers to identify potential failures and predict the system/device's remaining useful life (RUL). Although data-driven models are commonly used for prognostic applications, they are limited by the necessity of large training datasets and also the optimization algorithms used in such methods run into local minima problems. In order to overcome these drawbacks, we train a Neural Network with Bayesian inference. In this work, we use Neural Networks (NN) as the prediction model and an adaptive Bayesian learning approach to estimate the RUL of electronic devices. The proposed prognostic approach functions in two stages-weight regularization using adaptive Bayesian learning and prognosis using NN. A Bayesian framework (particle filter algorithm) is adopted in the first stage to estimate the network parameters (weights and bias) using the NN prediction model as the state transition function. However, using a higher number of hidden neurons in the NN prediction model leads to particle weight decay in the Bayesian framework. To overcome the weight decay issues, we propose particle roughening as a weight regularization method in the Bayesian framework wherein a small Gaussian jitter is added to the decaying particles. Additionally, weight regularization was also performed by adopting conventional resampling strategies to evaluate the efficiency and robustness of the proposed approach and to reduce optimization problems commonly encountered in NN models. In the second stage, the estimated distributions of network parameters were fed into the NN prediction model to predict the RUL of the device. The lithium-ion battery capacity degradation data (CALCE/NASA) were used to test the proposed method, and RMSE values and execution time were used as metrics to evaluate the performance.