A Vibration-Based Hybrid Deep Learning Approach for Anomaly Detection of Multijoint Industrial Robots

Condition monitoring is essential to ensuring precision and reliability of multi-joint industrial robots. Fault diagnosis is an effective method to identify different types of faults, whereas the tremendous obstacle to applying this method to multi-joint industrial robots is the difficulty of collecting and labeling a substantial quantity of abnormal samples. In this paper, a vibration-based hybrid deep learning approach for anomaly detection of multi-joint industrial robots is proposed to address this problem. A condition monitoring system is established using fiber Bragg grating (FBG) sensors to monitor the vibration of a multi-joint industrial robot. Considering the insufficient vibration information in single-domain data, multi-domain data, including the time domain waveform, frequency spectrum, and time-frequency spectrum of the vibration signal, is used to provide more comprehensive vibration information. Furthermore, the one-dimensional convolutional neural network (CNN) and two-dimensional CNN architectures are designed to automatically extract the representative features from multi-domain data, and these representative features are fused and input into K-nearest neighbors (KNN) to detect the anomalies in robot joints. Only normal samples are used for the training process of the hybridization of CNN and KNN. The results in three experiments illustrate that the proposed anomaly detection method has the satisfactory capability to handle the anomaly detection problem of multi-joint industrial robots.