DeepConAD: Deep and Confidence Prediction for Unsupervised Anomaly Detection in Time Series

Abstract

The current digital era of Industrial IoT and Automotive Technologies have made it standard for a large number of sensors to be installed on machines or vehicle, capture and exploit time-series data from such sensors for health monitoring tasks such as anomaly detection, fault detection, as well as prognostics. Anomalies or outliers are unexpected observations which deviate significantly from the expected observations and typically correspond to critical events. Current literature demonstrates good performance of Autoencoder for anomaly and novelty detection problems due to its efficient data encoding in an unsupervised manner. Despite the unsupervised nature of autoencoder-based anomaly detection methods, they are limited by the identification of anomalies using a threshold that is defined based on the distribution of reconstruction cost. Often, it is difficult to set a precise threshold when the distribution of reconstruction cost is not known. Motivated by this, we proposed a new unsupervised anomaly detection method (DeepConAD) that combined Autoencoder with forecasting model and used uncertainty estimates or confidence interval from forecasting model to identify anomalies in multivariate time series. We performed an experimental evaluation and comparison of DeepConAD with two other anomaly detection methods using Yahoo benchmark dataset, which contain both real and synthetic time series. Experimental results show DeepConAD outperforms other anomaly detection methods in most of the cases.