A novel wavelet convolutional network for monitoring interfacial bonding quality in LDED using AE signals

Abstract

Laser cladding is widely used for surface modification and repair of metallic materials, where the interfacial bonding quality is critical, especially in dissimilar material cladding. However, existing evaluation methods rely on experience or destructive testing, lacking real-time and quantitative capabilities, which limits broader application. The strong interactions between laser, powder and substrate result in complex multimodal mixing within the acoustic emission (AE) signals, making it challenging for traditional purely data-driven deep learning methods. To address this, this paper proposes a wavelet deep learning framework with physical interpretability. First, a comprehensive descriptive label for the cladding layer was constructed based on two key forming indicators: flatness ratio and dilution rate. Then, a wavelet packet decomposition (WPD) layer adaptively decomposed the AE signals, reducing frequency aliasing. Based on this, a parallel wavelet convolution layer was designed to extract physical features from the decomposed sub-band information using wavelet convolution kernels. To further enhance the feature representation ability, an attention module was introduced to optimizing feature expression. Finally, multi-layer neural networks were used to map the features to the comprehensive forming labels, achieving accurate monitoring of interfacial quality. Experimental results demonstrate a 96.18 % accuracy in interfacial bonding quality identification. Moreover, feature visualization results confirm the significant role of the parallel wavelet convolution layer in improving the distinguishability, while the attention module can effectively perceive energy fluctuations within the frequency bands, promoting the aggregation of similar samples and alleviating the boundary overlap between easily confused categories.