MSFA-Trans: Robust Machinery Fault Diagnosis Through Multiscale Frequency Attention Transformer in Noisy Environments

With the development of industrial production, there is an increasing demand for monitoring the health status of mechanical equipment. However, traditional fault detection methods are susceptible to noise interference during signal extraction, consequently affecting accuracy. Current deep learning techniques also face limitations in modeling frequency features and encoding long-time series data, thereby constraining fault detection capabilities. To address these challenges, we propose the Multi-Scale Frequency Attention Transformer (MSFA-Trans), which introduces time-frequency features with multi-scale frequency attention (MSFA) mechanisms and a Global Frequency Encoding Layer (GFE-Layer) to enhance network performance. Initially, time dynamic convolutions are employed to capture temporal features of sequential data. The GFE-Layer extends global information interaction from the time domain to the frequency domain, synthesizing frequency distribution features and temporal dynamics to comprehensively explore potential fault characteristics in the frequency domain. The MSFA mechanism, based on the Fast Fourier Transform (FFT), models interdependencies among signal frequencies and introduces attention mechanisms in the frequency domain to enhance network representation capabilities. Experimental evaluation on HSA bearing and motor bearing datasets demonstrates the superiority of the MSFA-Trans framework over conventional deep learning models, even in noisy environments, exhibiting robust fault diagnosis performance.