Deep Complex Wavelet Denoising Network for Interpretable Fault Diagnosis of Industrial Robots With Noise Interference and Imbalanced Data

Abstract

Fault diagnosis (FD) of industrial robots (IRs) plays an increasingly indispensable role in modern manufacturing. Fault-related component obscurity by strong noise, feature exploitation insufficiency with scarce fault samples, and limited physical interpretation hinder existing diagnostic models’ application to IRs. A deep, complex wavelet denoising network (DCWDN) is, thus, proposed to achieve high-performance and interpretable FD with robustness against noise and class-imbalanced data. Hereinto, a dual-tree cascade autoencoder with trainable convolutional filters is constructed. Significantly, complex wavelet conditions such as orthogonality, approximate analyticity, and sparsity are imposed on the filters to structure their optimization. Meanwhile, shrinkage-based denoising with learnable thresholds is integrated to suppress noise-related components. The proposed DCWDN organically combines the data adaptivity of deep learning (DL) and wavelets’ time-frequency representation ability. Its interpretability is embodied through the explainable structure, learned scientifically meaningful filters, and extracted coefficients with explicit fault indications. Case studies on real IR datasets and experimental drivetrain benchmarks are conducted to demonstrate the effectiveness and superiority of the proposed method.