Enhancing Particle Filter Performance for High Accuracy State Estimation and RUL Prediction

Accurately estimating the failure state and predicting the remaining useful life (RUL) of a system are important for effective predictive maintenance, and the particle filter (PF) method has been widely applied in estimating the failure state of nonlinear system. However, the PF method has some limitations that lead to performance degradation, such as particle degeneracy and multiple solutions in the posterior probability density function (pdf). To address these limitations, this article proposes an enhancing PF (EPF) algorithm that integrates Gaussian mutation (GM) and density-based spatial clustering of applications with noise (DBSCAN) for robust failure state estimation and reliable RUL prediction. Specifically, the GM method adaptively adjusts low-weight particles into high-likelihood regions, mitigating degeneracy and ensuring sufficient particle diversity. Meanwhile, DBSCAN clustering isolates the influence of subsolutions on the state prediction result, enhancing the accuracy of the estimated final point. Building on EPF improvements, we proposed a new framework combining EPF and the long short-term memory (LSTM) model. Combining the power of EPF fault state estimation with the learning capabilities of LSTM, our method provides a data-driven process for reliable long-term prediction. Experimental results with a well-known nonlinear benchmark system, real lithium-ion batteries, and bearing datasets show that EPF + LSTM can track and predict RUL effectively compared with other methods.