A Real-Time Health Management Method for Rolling Mills Based on RPC-Former

Abstract

In an aluminum hot rolling process, where support bearings undergo sudden load changes and rapid speed variations several times per minute, these variations lead to significant baseline shifts in the vibration and temperature signals. However, true fault cases only appear occasionally after several months. Therefore, the supervised learning methods usually presented in laboratory experiments do not have much practical application value. In this paper, we propose the Reconstruction-Prediction Coupling Transformer (RPC-Former), an unsupervised model which is trained directly on the normal operating data. A single backbone is used to reconstruct the current time windows and forecast the short-term future states. The continuous sensor records are normalized using Reversible Instance Normalization (RevIN) and divided into temporal segments to deal with the baseline drift problem. A dynamic Channel and Spatial Adaptation (CASA) coordinate attention mechanism is employed to extract cross-channel and time-block features. The reconstruction and prediction errors constitute a comprehensive health index for multi-level warnings. The model was tested on the hot rolling line data and achieved a joint root mean square error (RMSE) of 0.06247 in the ablation tests and detected a reducer bearing fault 4.8 days earlier than the manual detection.