Online abnormal interval detection and classification of industrial time series data based on multi-scale deep learning

Abstract

BackgroundThe prosperity of industrial big data promotes the development of anomaly detection in the production process. Abnormal production process is a crucial factor affecting product quality. Current online anomaly detection focuses on individual outliers, obtained via rules based on statistical process control. It is insufficient for the detection of multi-scale anomalies in non-stationary processes. MethodsAn online detection method for abnormal interval based on multi-scale deep learning is proposed. A scale conversion layer containing identity mapping, time domain down-sampling and low-pass filtering is fused with a multi-channel convolution structure to extract multi-scale features of abnormal intervals. Besides, a long short-term memory network is presented to improve the response to long-term features. The proposed network realizes online monitoring of production process via sliding window. Significant findingsThe proposed method has been verified on the UCR benchmark dataset to be superior to comparison methods. It is further applied to the online anomaly detection of the casting speed in continuous casting process. Multi-scale abnormal intervals are accurately detected and classified, and the obtained information including location, magnitude and category can be further employed for process optimization and quality evaluation. The proposed automatic detection method contributes to replace costly manual visual inspections, realizing intelligent production.