2023 Volume 59 Issue 7 Pages 342-352
In a factory production line, if abnormal signals can be automatically detected from time-series fluctuations in power consumption, the defect rate of products can be reduced, and productivity can be improved. Individual period estimation is often needed to detect the abnormals because deviations from the mean period provide valuable information about the failures and the abnormals. This study proposes an online system that estimates individual cycles and detects anomalies from quasi-periodic time series data represented by energy consumption data. The proposed system extracts local patterns by convolution and identifies individual cycles and abnormal signals based on similarity vectors calculated by an attention mechanism. Experimental results show that the proposed system outperforms existing methods on several simulated data sets.
