A lightweight ConvFormer for bearing fault diagnosis using multichannel vibration signals

Abstract

Rolling bearing fault diagnosis based on vibration signals is essential for the reliable operation of rotating machinery. However, many deep learning models still suffer from high computational cost and limited deployment efficiency, especially when multichannel signals are used to capture richer fault information. To address this issue, this paper proposes MDAFConvFormerLite, a lightweight ConvFormer-based model for bearing fault diagnosis using multichannel vibration signals. Specifically, the proposed model combines a multiscale disentangled adaptive fusion (DAF) block for local feature extraction and a lightweight Transformer operating on compressed temporal tokens, enabling joint modeling of fine-grained impulsive responses, medium-scale temporal structures, and global contextual dependencies with limited computational overhead. Experiments conducted on the MAFAULDA seven-class diagnosis task under a strict raw-filelevel data split and unified preprocessing protocol show that MDAFConvFormerLite achieves 96.0% accuracy and 96.1% Macro-F1. In addition, the proposed model remains compact, with only 116.327 K parameters and 174.458 MFLOPs. These results indicate that MDAFConvFormerLite provides an effective and lightweight solution for intelligent bearing fault diagnosis.