Multi-condition Fault Diagnosis for Rotating Machinery Using Vibration Images and Joint Distribution Adaptation

On basis of vibration images and transfer learning, this work presents a method for fault diagnosis of rotating machinery. Vibration images can be regarded as a combination of several traditional single-point sensors, which provide abundant full-field vibration information. The frequency-domain features of vibration acceleration are extracted according to the motion phase in images. Then proposed method utilizes a transfer learning method, called joint distribution adaptation, to construct a shared feature space. In this space, the marginal probability distribution and conditional distribution of vibration signals in two different conditions are minimized. Afterwards, the k-Nearest Neighbor classifier is used in this space for fault diagnosis. Experimental results manifest the proposed method can effectively diagnose faults under different working conditions and between different machineries, and the diagnosis accuracy is above 90%.