Multivariate Time Series Anomaly Detection On Improved HTM Model
Hui Zeng, Xiaoyong Zhao, Lei Wang · 2021 IEEE International Conference on Computer Science, Electronic Information Engineering and Intelligent Control Technology (CEI) · 2021
In recent years, industrial big data has attracted much attention as the key technical support of “Intelligent Manufacturing” and “Industrial Internet”. And as the dependence of intelligent manufacturing on digitalization continues to increase, data quality problems caused by device and system failures, harsh environment, improper scheduling and management, duplication or missing of data fields, etc., have more significant impacts on industrial processes. Therefore, the anomaly detection of industrial big data is particularly important. Among the methods onto time series data for anomaly detection, HTM(Hierarchical Temporal Memory) algorithm performs well in the unsupervised univariate time series data anomaly detection, but the capability of original HTM model for detecting multivariate time series anomaly data is insufficient. However, the multivariate data anomaly detection is common in industry and the performance requirements for data anomaly detection are relatively high. Thus, this paper proposes an improved HTM algorithm model - MSP-HTM(Multiple Spatial Poolers HTM) model. The MSP-HTM model respectively encode the value of different dimensions at the same time, and then put the result from encoder into spatial pooler respectively, finally the temporal memory layer merge result from spatial poolers, and predict future data. Experiments show that the MSP-HTM model can improve performance by processing the multivariate time series data in parallel and improve the effect of data anomaly detection.