SMLFormer: A Lightweight and Robust Framework for Intelligent In-Transit Diagnosis
Tao Zhou, Dechen Yao, Jianwei Yang, X. Q. Li · IEEE Transactions on Instrumentation and Measurement · 2024
In real industrial environments, bearing vibration data is collected by sensors, which are often embedded with noise, resulting in traditional models that cannot effectively identify fault types. In addition, with the high computational cost of the traditional models, it is difficult to achieve the deployment of in-transit equipment detection. In order to address these problems, this article proposes a lightweight fault diagnosis model based on a separable multiscale convolution and linear transformer, called “SMLFormer.” First, a separable multiscale convolution block (SMCB) is constructed by the idea of separable convolution, which can effectively reduce the parameters and computational complexity. Second, lightweight linear attention can extract beneficial features in the global scope and improve the capability for fault diagnosis. Third, a novel soft-threshold structure is proposed to reduce the complexity associated with linear dimensionality reduction while effectively mitigating the noise. Experimental results on the two datasets demonstrate that SMLFormer can accomplish higher diagnostic accuracy and lower parameters.