DCT-ISTA: A Network Block to Denoise Raw Vibration Signal with Interpretability for Mechanical Fault Diagnosis

Mechanical fault diagnosis is a crucial task in the field of high-value equipment prognostic and health management. Despite the numerous applications of deep neural networks in this domain, their lack of interpretability hinders their further adoption. In this paper, we propose a novel interpretable network block named DCT-ISTA for the diagnosis of rotating machinery faults, which is obtained by unrolling the classical Iterative Soft Thresholding Algorithm from the compressive sensing field and integrating prior knowledge from the fault diagnosis domain. DCT-ISTA combines sparse coding with Discrete Cosine Transform (DCT), formulated as a network through algorithmic unrolling, and integrates it as a data denoising method with backbone networks to enhance their performance. We validate the proposed method using simulated data with three different backbone networks. Through accuracy evaluation and analysis of signal reconstruction results, we demonstrate the effectiveness of our approach.