CLAD: A Deep Learning Framework for Continually Learning in Anomaly Detection

Abstract

The rapid development and frequent revolutions in information technology (like Edge Computing, Wireless Sensor Network) highlight the significance of Internet of Things. Nowadays, a variety of infrastructures from diverse fields are limited by external and internal environmental factors. In actual operation, these factors may cause serious anomalies and aggravate the burden of facilities’ maintenance. With a rise in the number and running time, the performance of these facilities would be unstable and complex. This paper proposes a deep learning framework (CLAD) to do adaptive anomaly detection. It implements dynamic anomaly thresholding based on the prediction of incremental Long Short-Term Memory. This framework employs a replay buffer to solve the decline of detection accuracy. With this framework, models can keep perfect detection accuracy under a quite high load (time and number). This framework is valuable for further research of adaptive anomaly detection.