Channel-adaptive generative reconstruction and fusion for multi-sensor graph features in few-shot fault diagnosis

Abstract

Recently, multi-sensor feature fusion has been proven to be an effective strategy for improving the accuracy of few-shot fault diagnosis. However, existing fault diagnosis models based on multi-sensor feature fusion often overlook significant inter-channel discrepancies and struggle to mitigate noise pollution inherent in multi-source signals. To address these limitations, this paper proposes a channel-adaptive generative reconstruction and fusion framework that integrates a contrastive variational graph autoencoder feature fusion (CogFusion) module for robust few-shot fault representation learning. The CogFusion module leverages the generative capability of a contrastive variational graph autoencoder (CGE) to reconstruct noise-suppressed node features while explicitly modeling latent distributions of multi-sensor signals. By incorporating a multi-channel parallel graph contrastive learning strategy, CogFusion enhances discriminative feature separation by contrasting topological structures of positive and negative sample pairs, effectively isolating fault-related patterns from noisy embeddings. To adaptively fuse multi-channel information, a channel discrepancy-guided weighting mechanism dynamically prioritizes high-credibility sensor features, mitigating the impact of low-quality data. To further enhance feature learning in few-shot diagnosis, a dual-scale topological Transformer (DSTT) model is introduced to deeply mine the reconstructed multi-channel topological graph, enabling high-precision few-shot fault diagnosis. Experimental results on the axial flow pump and HUSTgearbox datasets demonstrate that the proposed method outperforms both single-channel and existing multi-sensor feature fusion methods, highlighting its superiority in feature fusion and cross-channel information integration.