Investigation of Bonds Between Network Convolution and Time-Frequency Transforms for Ex-Ante Interpretable Machine Health Prognosis

In the era of big data and intelligent sensing, deep neural networks provide new impetus for prognostics and health management (PHM) with their powerful feature extraction capabilities. However, the pursuit of performance through increased network depth and complexity concurrently escalates the number of hyperparameters and model intricacy, thereby exacerbating the inherent opaque nature and restricting their deployment in complex industrial settings. To address this dilemma, this article develops a machine health prognosis framework with ex-ante interpretability based on complex domain time-frequency network (CDTFN). Specifically, this article first investigates the intrinsic bonds between network convolution and time-frequency transforms. Building upon this foundation, four complex observation operators with trainable parameters are designed for extracting fault-related time-frequency information, embedding it into the CDTFN as a preprocessing layer. Simultaneously, by extending the forward and backward propagation mechanisms of real-valued networks to the complex domain, the proposed CDTFN gains the capability to fuse complex-valued time-frequency information and establish end-to-end mapping from feature representation layers to prediction labels. The effectiveness and accuracy of the proposed prognosis framework based on CDTFN are verified by public and self-built run-to-failure rolling bearings datasets. The detailed experimental results further demonstrate its distinct advantages in interpretability and generalization capability.