A Noise-Excitation Generative Adversarial Network for Actuator Fault Diagnosis of Multi-legged Robot

This research provides a novel approach for detecting multi-legged robot actuator faults. The most significant concept is to design the Fault Diagnosis Generative Adversarial Network (FD-GAN) to fully adapt to the fault diagnosis problem with insufficient data. We found that it is difficult for methods based on classification and prediction to learn failure patterns without enough data. A straightforward solution is to use massive amounts of normal data to drive the diagnostic model. We introduce frequency-domain information and fuse multi-sensor data to increase the features and expand the difference between normal data and fault data. A GAN-based framework is designed to calculate the probability that the enhanced data belongs to the normal category. It uses a generator network as a feature extractor, and uses a discriminator network as a fault probability evaluator, which creates a new use of GAN in the field of fault diagnosis. Among the many learning strategies of GAN, we find that a key point that can distinguish the two types of data is to use the hidden layer noise with appropriate discrimination as the excitation. We also design a fault location method based on binary search, which greatly improves the search efficiency and engineering value of the entire method. We have conducted a lot of experiments to prove the diagnostic effectiveness of our architecture in various road conditions and working modes. We compared FD-GAN with popular diagnostic methods. The results show that our method has the highest accuracy and recall rate.