Corrosion Detection of Gearbox Based on Wavelet Time–Frequency Attention Fusion Network

Marine gearboxes operating long-term in high-temperature, high-humidity, and high-salinity mist marine environments are highly susceptible to corrosion faults, posing significant threats to the reliability and safety of shipboard equipment. Ultrasonic testing (UT) technology, with its noncontact and remote capabilities, is well-suited for inspecting complex workpieces in such adverse conditions. To address the limitations of existing quantitative corrosion detection methods for gearboxes, while fully leveraging the nondestructiveness and information integrity advantages of ultrasonic nondestructive testing technology, this article proposes the wavelet time–frequency attention fusion network (WAFN), an ultrasonic signal-based method for gearbox corrosion detection. The method first constructs an optimized four-channel parallel ConvNeXt network (multichannel time–frequency feature extraction (MCFE) module) for deep feature extraction. Subsequently, a Transformer encoder module is introduced to fuse global features and capture cross-channel spatial dependencies. Then, a symmetric multichannel cross-attention feature fusion (CAFF) module realizes adaptive weighted fusion of local and global features. Finally, a supervised collaborative contrast loss (SCCL) training mechanism is designed, combining feature loss and classification loss to pull features of the same corrosion level closer while pushing features of different levels apart. This effectively mitigates interference from intraclass variations and blurred interclass feature boundaries inherent in quantitative data, achieving quantitative nondestructive detection of gearbox corrosion. Experimental results show that the proposed model achieves higher comprehensive accuracy on the two datasets and in the supplementary experiments with actual plates, verifying the effectiveness of this method.