Rolling Bearing Fault Prediction via Conv-LSTM with Spatiotemporal Graph Representation

Traditional approaches for bearing vibration analysis predominantly emphasize temporal feature extraction, often neglecting critical spatial correlations. This study proposes a novel fault prediction framework integrating Convolutional and Long Short-Term Memory networks (Conv-LSTM) with spatiotemporal graph representations. Our method transforms one-dimensional vibration signals into two-dimensional spatiotemporal graphs through systematic data fusion, enabling simultaneous capture of temporal dynamics and spatial dependencies. The hybrid Conv-LSTM architecture synergizes convolutional operations for spatial pattern recognition with LSTM’s temporal modeling capabilities. Experimental validation on full-lifecycle bearing data demonstrates superior performance over conventional models, achieving 97% prediction accuracy with 0.32 MSE. This advancement enables proactive maintenance decision-making in industrial applications.