LHSNet: large kernel hierarchical shrinkage network with deep adaptive denoising and modern wide-conv architecture for robust fault diagnosis

Traditional intelligent fault diagnosis techniques often struggle with signal distortion in environments characterized by strong multi-source noise, where fault signals are easily submerged and feature extraction becomes extremely unreliable. Furthermore, intelligent diagnosis is difficult to achieve using traditional denoising techniques that require extensive prior expert experience. This paper proposes a novel one-dimensional modern convolutional neural network architecture with soft thresholding denoising modules, tailored for robust mechanical fault detection under complex noise interference. Firstly, adaptive noise reduction modules are developed to assign dynamic filtering parameters to individual vibration signals, which effectively eliminates multi-source noise interference and prevents fault signal distortion. Secondly, an adaptive feature normalization mechanism is integrated to enhance signal discrimination capability while maintaining computational efficiency. Lastly, comparative experiments and ablation studies are conducted under an extended signal-to-noise ratio and diverse noise type backgrounds including Gaussian, pink, and salt-and-pepper noise to verify the superiority of the proposed architecture over existing diagnostic methods. Results have well demonstrated that the proposed approach is able to reliably identify different fault types with over 98.00% median accuracy and 96.70% peak accuracy on standard benchmark datasets. Moreover, the proposed technique maintains high diagnostic performance across diverse operating conditions with less dependency on prior denoising expertise, which shows it is well suitable for practical industrial applications with complex noise interference.