A lightweight mechanical fault diagnosis framework based on dynamic separable convolution and broadcast self-attention

To address issues such as the large number of parameters, high computational complexity, and inadequate real-time performance in existing CNN-Transformer hybrid fault diagnosis models, we propose a lightweight fault diagnosis framework, LWConvFormer. This framework comprises two core innovative modules: a dynamic separable multi-scale convolutional module that employs a gated network for adaptive feature extraction, thereby reducing computational load while enhancing adaptability to complex fault modes; and a broadcast self-attention module that substitutes traditional matrix multiplication with broadcast operations, thereby decreasing computational complexity from a quadratic to a linear level. Experimental results based on the planetary gearbox at Xi'an Jiaotong University and the QPZZ-II type rotating machinery test bench demonstrate that LWConvFormer maintains excellent diagnostic performance across various noise levels. The number of parameters and computational load are reduced by a factor of 6 to 10 compared to mainstream methods, while the training speed increases by nearly 7 times. This framework effectively balances diagnostic accuracy, model lightweighting, and noise resistance, offering an efficient solution for real-time fault diagnosis in industrial settings.