Generative adversarial learning for intelligent fault diagnosis: Matrix kernel-enhanced convolutional networks in robotic systems

In industrial robotics, joint bearing faults account for 30%–35% of failures, yet diagnosing them under variable operating conditions remains challenging due to limited labeled data and complex vibration patterns. This paper proposes a novel IoT-enabled intelligent fault diagnosis framework centered around an enhanced hybrid CGAN-1D-MCNN model with Squeeze-and-Excitation (SE) attention for robust fault diagnosis. First, a conditional GAN (CGAN) synthesizes high-quality fault samples conditioned on fault types, addressing data scarcity by learning the distribution of real vibration signals from a 6-DOF industrial robot (UR-16e) under diverse loads. Next, a matrix Kernel-based 1D-CNN (MCNN) processes multi-channel signals using dynamic matrix kernels, enabling simultaneous spatial feature extraction and cross-channel fusion. SE attention blocks recalibrate channel-wise features via global average pooling and learned excitations. Evaluated on both real-time lab dataset and the public CWRU and Paderborn datasets, CGAN-1D-MCNN achieves 98.90% accuracy, outperforming FMRGAN (94.98%), C-DGAN (96.45%), and GPSC-GAN (95.53%), while maintaining robustness under noise (±2% accuracy drop at 30 dB SNR) and variable loads (0.50-3.0 hp).