MCVAE-GAN: A Novel Diagnosis Method for Solving Harmonic Reducer Fault Signal Scarcity and Sample Diversity Problem

Abstract

Harmonic reducers are vital for robotics and aerospace due to their ability to deliver high torque and reduction, but they face challenges in fault diagnosis because of their complex design and specific operating conditions. The scarcity of real-world fault data and signal sample diversity for these reducers exacerbates the problem, leading to a significant data imbalance that hinders the effectiveness of deep learning-based fault diagnosis methods. To address this, we propose a novel approach, the multiconditional variational autoencoder generative adversarial network (MCVAE-GAN), featuring a multihead attention mechanism (MHAM) fusion to enhance sample diversity and tackle fault data scarcity. The MCVAE-GAN employs MHAM in its encoder, decoder, and discriminator to improve data processing and analysis. The encoder uses MHAM to focus on crucial data attributes, the decoder enhances sample characteristics by managing conditional and latent variables interaction, and the discriminator distinguishes between real and generated samples by concentrating on key features. The expanded dataset is used to train a classification network to complete fault diagnosis. Our method significantly outperforms contrasting data generation techniques in fault diagnosis, demonstrating superior sample imbalance management and showcasing its applicability in real-world engineering scenarios.