A Transmission Axle Fault Diagnosis System for Massive Rapid Transit System With Enhanced Attention-Based Shuffle Networks

Abstract

The transmission axle faults can cause severe damage to gearbox components and jeopardize motor power transmission. A transmission axle fault (TAF) diagnosis method for train propulsion systems in mass rapid transit (MRT) networks is proposed, utilizing an Enhanced Attention-based Shuffle Network (EASN). Time-domain vibration signals, directly acquired from sensors mounted on the transmission axles, are employed as input to the EASN model. The proposed model is specifically designed for deployment on lightweight AI edge devices, with an emphasis on computational efficiency and real-time performance for onboard diagnostic applications. The EASN architecture is composed of multiple building blocks, referred to as Split-Attention Shuffle Units (SASUs). Each SASU integrates a Shuffling Processing Module (SPM) and a Split Attention Module (SAM) in a cascaded configuration. While the SPM is based on ShuffleNet, which is known for its computational efficiency but relatively lower classification accuracy, the proposed SASU mitigates this limitation through the introduction of an even channel shuffling mechanism combined with a hybrid attention strategy. The hybrid attention scheme leverages both Spatial Excitation (SPE) and Squeeze-and-Excitation (SnE) mechanisms, significantly enhancing the network’s diagnostic accuracy without compromising its lightweight design. Experimental results demonstrate that the proposed EASN achieves a Top-1 classification accuracy of 93.9%, representing a 5.6% improvement over ResNet-50 while reducing the model size by 98.8%. Compared with lightweight models such as MobileNet V2, EASN improves accuracy by 19.4% with only a 28.6% increase in parameter size. These findings indicate that EASN offers an effective balance between diagnostic accuracy and model compactness, making it well-suited for real-time, edge-based fault detection in mass rapid transit systems.