A Multi-Channel Attention-Based Variational Autoencoder Framework for Incipient Fault Detection of Rolling Bearings

Abstract Early fault detection of rotating machinery bearings is critical for ensuring the long-term operational safety of industrial equipment. Existing methods typically rely on single-axis vibration signals or simply concatenate multi-sensor data, which limits their ability to capture cooperative degradation mechanisms across spatially distributed measurement points and results in low sensitivity to subtle early stage faults under non-stationary operating conditions. To address these limitations, this paper proposes a Multi-Channel Attention-enhanced Variational Autoencoder (MCA-VAE) framework for early bearing fault detection, which explicitly models the temporal evolution and cross-channel spatial correlations of horizontal and vertical vibration signals in a unified latent space. By incorporating a multi-head selfattention mechanism, the method adaptively enhances the representation of critical degradation patterns. Trained in an unsupervised manner using only fully healthy samples, MCA-VAE constructs a high-confidence latent health distribution, enabling fine-grained degradation stratification and early-stage fault warning through latent statistical distances. Experiments on the IEEE PHM Challenge 2012 bearing dataset demonstrate that MCA-VAE identifies the onset of degradation significantly earlier than existing methods. Ablation studies further confirm the crucial role of multi-channel spatiotemporal modeling and attention mechanisms in enhancing early fault detection performance. This study provides an engineer-friendly and effective solution for high-sensitivity early health monitoring of rotating machinery in complex industrial scenarios.