A Hybrid Prognostic Approach Combining a Stacked Multihead Self-Attention Autoencoder and Multimodal Relevance Sampling Particle Filter

Accurate prediction of the remaining useful life (RUL) of bearings is critical for optimizing maintenance strategies, preventing unexpected failures, and enhancing repair management. Despite the increasing interest in hybrid-driven prognostic methods, several challenges remain: 1) constructing a high-quality health indicator (HI) is essential for accurate prognostics, but existing methods are hindered by feature redundancy and insufficient extraction of deep latent information, resulting in unreliable HIs and 2) the particle filter (PF), a widely used RUL estimation method, can degrade into a prior estimation technique when measurement data are missing during prediction, leading to particle impoverishment and reduced accuracy. To address these challenges, this study introduces an innovative hybrid-driven framework for predicting the RUL, comprising two main components: HI construction and a closed-loop RUL prediction algorithm. For robust HI construction, a heuristic search method based on the minimum redundancy maximum relevance (mRMR) criterion is used to select optimal features, effectively mitigating redundancy. These features are then used as guiding labels to train a stacked multihead self-attention autoencoder (SMSAE), which extracts deep, latent features from raw data, going beyond the shallow representations of the guiding labels. The framework further employs Bayesian transformer (BTransformer) models for one-step state predictions, which serve as virtual measurements in the multimodal relevance sampling particle filter (MRSPF) algorithm, alleviating particle impoverishment and refining posterior distributions. Experimental validation on benchmark datasets demonstrates the superior predictive accuracy and reliability of the proposed framework.