Abstract

 In industrial environments, fault signals are often buried by external interference signals and noise, and their intensity weakens due to complex signal transmission paths. Conventional signal processing approaches may misdiagnose the system’s health condition by focusing on interferences rather than fault features. To overcome these challenges, this study proposes a domain knowledge-informed AI to effectively enhance fault features. The proposed neural network incorporates convolutional layers designed to capture fault impulse responses and employs a loss function that considers both the periodicity and sparsity of faults. The proposed method was validated using an accelerated gearbox test dataset. Experimental results demonstrate that, unlike conventional signal processing approaches that amplify interferences, the proposed method effectively enhances fault features and improves diagnostic reliability.