An Online System of Detecting Anomalies and Estimating Cycle Times for Production Lines

Energy consumption data of production machines often exhibit quasi-periodicity, and anomalies are observed when deviations from the quasi-periodicity are detected. For such data, it is crucial to quickly estimate the individual cycles at each time point and detect abnormalities. In this study, we propose a system that satisfies these requirements. The proposed system trains a neural network with an attention mechanism and applies the weight vectors in the mechanism to the two tasks. Experimental results demonstrate that the proposed method outperforms benchmark methods for sensor data that mimic power consumption data of production lines.