Anomaly Detection Using Deep Learning

Abstract

Deep learning has revolutionized many fields that process large amounts of data such as images, video, audio, speech, and text. Anomaly detection, however, is among the areas that still require major advancements. Based on the key traits of deep learning, which are the need for very little hand engineering, and the ability to effectively use large amounts of data, its applications in anomaly detection could be very beneficial. In this thesis, an anomaly detection architecture is proposed that consists of two models: a normal model, which is a time series forecaster and predicts the next expected behavior of the system under healthy conditions, and an anomaly detector that identifies any failure by comparing the expected values with the actual observations. Deep architectures such as convolutional neural networks and recurrent neural networks have been incorporated in the design of the normal model, and conventional machine learning methods, including one-class SVM, isolation forests, multi-layer perceptron, decision trees, and random forests are used for the anomaly detector. The proposed architecture has been applied to two problems: pipeline leak detection and condition monitoring and fault detection of small induction motors. The results of both applications have proved very promising and indicate the capacity for further improvements to come.