Acoustic Feature-Driven Cross-Domain Fault Diagnosis of Industrial Equipment: An Adaptive Feature Fusion Approach

Fault diagnosis of industrial equipment is crucial to ensure production safety, extend equipment life, and improve maintenance efficiency. Traditional acoustic diagnosis methods are susceptible to noise interference and domain offset in complex industrial scenarios, and their generalisation capability is limited. In this article, a cross-domain fault diagnosis method based on harmonic-percussion source separation (HPSS) and adaptive dual-channel feature fusion (ADFF) is proposed. Firstly, the HPSS algorithm is used to decompose the Log-Mel spectrogram into harmonic and percussion components to effectively separate the fault features from noise; secondly, the ADFF module is designed to dynamically fuse the time-frequency features of the two types of components through the dual-channel convolutional network and the squeeze-excitation (SE) attention mechanism to enhance the cross-domain generalisation capability. Experiments on a multi-condition drone fault dataset show that the diagnostic accuracy of the proposed approach in cross-equipment and cross-load scenarios is improved by up to 7.56% compared with the traditional approach.