Anomaly detection using Long Short Term Memory Networks and its applications in Supply Chain Management

Abstract

Anomaly detection has been becoming an important problem in several domains. In this paper, we propose a new method to detect anomalies in time series based on Long Short Term Memory (LSTM) networks. After being trained on normal data, the networks are used to predict interested steps in time series. The difference between the predicted values and observed values is calculated as prediction errors. Then we use a kernel estimator of the quantile function to compute a threshold, which is used to determine anomalous observations. The performance of proposed method is illustrated through an example of anomaly detection of consumer demand in supply chain management. The numerical experiment shows that our approach achieve a higher level of detection accuracy and a lower percentage of false alarm rate compared to the previous One-Class Support Vector Machine method.