A Lightweight Transformer-Based Model for Efficient Industrial Fault Diagnosis

The complexity of industrial systems has triggered an ever-increasing demand for accurate fault diagnosis. However, traditional methods are typically unreliable, particularly in scenarios with limited data. This paper proposes the Separable Multi-Scale Convolutional MLP Network (SMCMN), a lightweight fault diagnosis framework designed to address the challenges of deploying complex Transformer models in resource-constrained environments. The SMCMN framework reduces the complexity associated with the Transformer’s multi-head self-attention mechanisms and traditional CNN cross-channel convolution by redesigning the Multi-Scale Separable Convolution (MSSC) block. This block incorporates large kernel sizes and dropout layers to enhance accuracy. Furthermore, the framework replaces conventional attention mechanisms with a Multi-Layer Perceptron (MLP) layer, thereby reducing model complexity. Extensive testing across two diverse datasets has demonstrated that the SMCMN framework exhibits superior robustness and performance, proving to be an effective solution for fault diagnosis in practical applications, even under varying noise levels.