SFS-PSO: An Improved Data Preprocessing Approach in Fault Diagnosis under Variable Working Conditions

This paper presents an approach for bearing fault diagnosis that leverages sensitive feature selection to tackle the challenge of high intra-class distances and low inter-class distances in data pre-processing of conventional deep transfer learning methods under variable working conditions. Initially, bearing fault signals are transformed from time to frequency domain using the Fast Fourier Transform. Then, these frequency-domain features are refined through a selection process optimized by particle swarm optimization, focusing on those with higher sensitivity weights to reduce intra-class distance and enhance inter-class distance. These selected features are input into domain adversarial neural networks for accurate bearing fault diagnosis under different conditions. Experimental results demonstrate the method's effectiveness and increased diagnostic accuracy.