A Novel Lightweight Group Convolutional Neural Network Architecture For Inter-Turn Short Circuit Fault Diagnosis of Induction Motors

The high memory and computational requirements of convolutional neural networks (CNNs) limit their application in intelligent fault diagnosis systems for motors in embedded and online platforms. To tackle this issue, the present paper proposes a novel lightweight group convolutional neural network (LGCNN) architecture. Firstly, a dual-branch multi-scale group convolution block (DM-GroupConvblock) is proposed to facilitate the extraction of multi-scale fault features while optimizing the computational redundancy of the multi-branch structure. Secondly, a residual group convolution module (Res-GroupConvblock) with equal channels is introduced to enhance the learning of high-level discriminative features and accelerate model convergence. Finally, an experimental platform for induction motors is constructed and the effectiveness of the proposed method for inter-turn short circuit fault diagnosis is verified under different loads and different fault severities. The proposed LGCNN architecture demonstrates excellent diagnostic performance in experiments, achieving 99.92% fault diagnosis accuracy with only 1.2M model parameters and 0.6M floating point operations (FLOPs).