Abstract: In view of the shortcomings of traditional remaining useful life (RUL) prediction methods in modeling the complex spatiotemporal dependencies and noise resistance of multi-sensor data, a RUL prediction model is proposed, integrating a spatial-temporal graph convolutional network (STGCN), a soft-threshold residual attention mechanism, and a bidirectional gated recurrent unit (BiGRU). The model first extracts spatial and temporal features from multi-sensor data through STGCN to model the topological relationship and dynamic evolution between equipment components. Secondly, a BiGRU unit is introduced to capture deep temporal dependencies, and combined with a soft-threshold residual attention mechanism, the perception of key degradation features is enhanced and noise interference is suppressed. Finally, accurate prediction of the RUL of the machine equipment is achieved. Experiments on the PHM2010 and NASA datasets show that this method still has excellent prediction performance under various noise interferences, significantly outperforming existing methods.