Detecting Anomalous Multivariate Time-Series via Hybrid Machine Learning

Abstract

This article investigates the use of hybrid machine learning (HML) for the detection of anomalous multivariate time-series (MVTS). Focusing on a specific industrial use-case from geotechnical engineering, where hundreds of MVTS need to be analyzed and classified, has permitted extensive testing of the proposed methods with real measurement data. The novel hybrid anomaly detector combines two means for detection, creating redundancy and reducing the risk of missing defective elements in a safety relevant application. The two parts are: 1) anomaly detection based on approximately 50 physics-motivated key performance indicators (KPIs) and 2) an unsupervised variational autoencoder (VAE) with long short-term memory layers. The KPI captures expert knowledge on the properties of the data that infer the quality of produced elements; these are used as a type of auto-labeling. The goal of the extension using machine learning (ML) is to detect anomalies that the experts may not have foreseen. In contrast to anomaly detection in streaming data, where the goal is to locate an anomaly, each MVTS is complete in itself at the time of evaluation and is categorized as anomalous or nonanomalous. The article compares the performance of different VAE architectures [e.g., long short-term memory (LSTM-VAE) and bidirectional LSTM (BiLSTM-VAE)]. The results of using a genetic algorithm to optimize the hyperparameters of the different architectures are also presented. It is shown that modeling the industrial process as an assemblage of subprocesses yields a better discriminating power and permits the identification of interdependencies between the subprocesses. Interestingly, different autoencoder architectures may be optimal for different subprocesses; here two different architectures are combined to achieve superior performance. Extensive results are presented based on a very large set of real-time measurement data.