A Fault Diagnosis Method Based on Frequency Domain Convolution and Temporal Information Decoupling

In recent years, transformer family models have demonstrated excellent performance in time series data research, gradually becoming mainstream in time series analysis. However, due to their unique attention mechanisms, these models suffer from low computational efficiency and are significantly affected by noise. Conversely, convolutional methods have lost momentum in time series research due to their narrow receptive fields. This paper proposes a novel network for fault diagnosis in time series data, based on Fourier analysis and Temporal Convolutional Networks (TCN). This network combines the global receptive field of Fourier analysis with the local receptive field advantages of convolutional methods. It filters out noise during the Fourier analysis phase to achieve better performance in fault diagnosis tasks. Additionally, this method decouples information from time series data, extracting information independently from the time, feature, and variable dimensions. This allows the network to learn complex dependencies among variables, building on previous research. Our study shows that our network can filter noise while capturing both long-term and short-term interactive information. Compared to previous models solely based on convolution, the network offers a broader receptive field and enhanced information decoupling. Experimental results demonstrate that this method achieves excellent performance in fault diagnosis tasks.