Anomaly detection of industrial multi-sensor signals based on enhanced spatiotemporal features

摘要

To improve the validity of industrial multi-sensor signals, anomaly detection has become a significant part of industrial signal processing. In practical measurement, industrial multi-sensor signals are mostly fluctuant, and the correlation between the front and back signals is uncertain. These increase the difficulty of the abnormal signal detection. For these problems, this paper proposes an anomaly detection method for industrial multi-sensor signals based on enhanced spatiotemporal features. Firstly, a series of signal preprocessing is performed relying on the characteristics of data set, so that the signals will be processed in a stable state. Then, a stack spatial-temporal autoencoder, which relies on the improved deep stack long short-term memory and autoencoder feature extractors, is proposed to extract features and to reconstruct signals. Next, a high-dimensional unsupervised clusterer is proposed to detect the abnormal signals. Finally, two case studies in magnetic flux leakage (MFL) signals and Tennessee Eastman (TE) benchmark are conducted. MFL signals are the actual signals collected from the experimental platform, and TE benchmark is a public data set. State-of-the-art comparison experiments on feature extraction and abnormal signal detection are performed, and the results show that the proposed method is effective.