Autoregressive Semi‐Connected Autoencoder for Remaining Useful Life Prediction of Rolling Bearings With Threshold‐Free First Prediction Time Detection

ABSTRACT Remaining useful life (RUL) prediction of rolling bearings is critical for machinery prognostics. Accurately constructing health indicators (HIs) is a vital prerequisite, yet physical HIs are highly susceptible to noise. To address this, we propose a novel framework based on the Autoregressive Semi‐Connected Autoencoder (ARSCAE). As the core architecture, ARSCAE integrates semi‐connected weight constraints and autoregressive temporal mechanisms, overcoming traditional fully‐connected limitations while reducing parameters and overfitting risks. Specifically, the decoder utilizes fully connected layers, while the encoder introduces sparsified semi‐connected layers. By imposing a strictly lower‐triangular mask on the weight matrix, the encoder restricts each output to depend solely on historical inputs, strictly enforcing temporal causality to enhance HI trendability. Furthermore, an autoregressive smoothing regularization is formulated to suppress training instability. Concurrently, leveraging the unilateral inhibition of the ReLU function, the virtual HI stays at zero during the healthy stage and transitions to positive values upon degradation, enabling threshold‐free, adaptive first prediction time (FPT) detection. Finally, a GRU network maps the HI sequences to the RUL. Experimental results on the XJTU‐SY and a self‐constructed dataset under high‐temperature conditions demonstrate that ARSCAE constructs superior monotonic HIs and achieves state‐of‐the‐art precision in RUL prediction.