SKR-ShuffleNet v2: efficient lightweight network rail fastener fault diagnosis method based on an attention mechanism

Rail fastener condition is crucial for rail vehicle safety. However, due to the high physical costs and low inspection accuracies that characterise traditional inspection tasks, intelligent fault diagnosis is becoming an important option among current mechanical system maintenance methods. Deep learning models, such as the convolutional neural network (CNN), have been successfully applied to fault diagnosis tasks with good results. However, considering the large sizes, long detection times, and high requirements for field equipment hardware when using general neural networks, this paper proposes a rail fastener detection method based on improvements to the lightweight network ShuffleNet v2. This method effectively increases the detection accuracy and efficiency for rail fastener. First, the faulty rail fastener sample dataset is expanded using data augmentation and labelled for use as network input. Then, the selective kernel attention mechanism is integrated into the basic unit to improve its channel focus, and a new optimiser, Ranger21, is adopted into the model to solve the problem of AdamW instability in the initial state, thereby increasing the fault classification detection accuracy. Experimental results show that the detection accuracy and precision of the improved model increased by 1.45% and 2.00%, respectively, which verifies the effectiveness of the algorithm.