CAMMA : A Physics‐Informed Adaptive Mamba Framework for Remaining Useful Life Prediction in Intelligent Predictive Maintenance

ABSTRACT Rolling bearings are one of the most failure‐prone components in rotating machinery, and intelligent predictive maintenance relies on accurate estimation of their remaining useful life (RUL) from vibration measurements. However, the measured vibration is masked by strong noise and is nonmonotonic. Once a defect occurs, its distribution will change sharply, and the monitoring record of a full life cycle is long, so it is difficult to obtain a stable degradation trend and a reliable maintenance cycle. We present CAMMA, a physically adaptive Mamba framework with multihead attention that provides accurate and robust RUL predictions across operating conditions. First, the exponential weighted moving average was used to smooth the high‐frequency noise, and the cumulative energy index (CEI) accumulated the smoothed root mean square response into a monotone nondecreasing alternative term, which represented the dissipated energy at the degraded contact. Then, the adaptive gating unit based on the local signal statistical characteristics and the adaptive Mamba block normalized in the adaptive layer were used to calibrate each channel, which attenuated the running‐in fluctuations while preserving the channels related to degradation. To further capture global dependencies beyond causal recursion, we fuse multihead attention, applied to a fixed‐length observation window, with a Mamba branch whose cost grows linearly in the length of the monitored record. Subsequently, the network regresses a clearly defined health indicator. Experiments on the XJTU‐SY dataset show that CAMMA outperforms LSTM, Transformer, TCN and ILDENet, with average RMSE of 9.55, 8.46 and 7.80 percentage points of health under the three operating conditions, respectively, while ablation and explainability studies verify the contribution of each module.