An Attention Conditional Regularized Least Squares Generative Adversarial Network for Gearbox Fault Diagnosis

Gearbox plays a role in mechanical equipment such as power transmission, speed, and torque conversion. However, in large and complex industrial scenarios, the acquisition of gearbox fault data is often expensive, and relying on a small amount of fault data to achieve intelligent fault identification is a challenging task. To address this challenge, we propose an intelligent diagnosis method based on Attention Conditional Regularized Least Squares Generative Adversarial Networks (ACLGAN). First, the diversity of original samples is increased by introducing an overlapping segmentation strategy. Then, based on the least squares loss function, the conditional regularization term is incorporated to alleviate the issues of unstable model training, disappearing gradient, and exploding gradient. At the same time, the Conditional Block Attention Mechanism (CBAM) is adopted to further enhance the quality of the generated samples. Finally, the real samples and the obtained fake samples are fed into the designed classifier based on deep convolutional neural network (DCNN) to realize fault diagnosis. We validated the applicability of ACLGAN using the PHM2009 gearbox dataset, and the results show that the intelligent diagnosis method based on ACLGAN can generate high quality simulation data and better recognize six various fault states of gearboxes.