In order to improve the reliability of industrial robot operation, the abnormal state of the robot needs to be detected in advance, rather than waiting until the fault is serious and causing downtime. In this paper, a current based method for joint state anomaly detection of industrial robot is proposed. Normal joint current data were used to extract motion cycle through the autocorrelation function, and feature extraction of current data was achieved by combining the time-domain feature index and principal component analysis algorithm, and the feature vector was constructed, which was then input into the support vector data description (SVDD) model to complete training and optimization. Finally, abnormal samples were used for testing and verification. The results show that the SVDD model can distinguish the difference between samples well, detect abnormal samples, and has good robustness and generalization.