PEQ-enhanced selective wavelet network for lightweight acoustic welding penetration monitoring

Real-time penetration monitoring is critical for ensuring the structural integrity of laser-arc hybrid welding. While microphone-based acoustic sensing provides a non-intrusive solution, its industrial deployment is severely hindered by non-stationary environmental noise and the prohibitive computational latency of conventional deep learning architectures. To overcome these limitations, this study proposes a physics-augmented sparse wavelet network, designated as SPEQW-Net. The framework replaces black-box preprocessing by utilizing a differentiable cascaded Parametric Equalizer (PEQ) array as a physical front-end for task-adaptive frequency reweighting of acoustic waveforms under industrial interference. Subsequently, a Convolutional Discrete Wavelet Transform (C-DWT) maps the signal into a multiscale representation. Within this module, a targeted sparsity gating mechanism explicitly prunes 50% of the channels with the lowest learned contribution scores, eliminating stochastic environmental artifacts. Finally, a customized MobileNetV3-Small backbone executes lightweight classification. Experimental results on a noise-augmented dataset (5–20 dB SNR) demonstrate that SPEQW-Net achieves an accuracy of 94.8%. With a parameter footprint of only 1.24 million and a computational inference latency of 5.8 ms, the proposed model reduces computational overhead by 89% and accelerates inference by over 7 times compared to ResNet-18. Statistical spectral analysis reveals penetration-state-dependent variations consistent with prior welding-acoustics findings.