Feature Extraction Using Convolutional Neural Networks for Anomaly Detection

Abstract

Anomaly detection is an import field of study, which has many applications, e.g., fraud and disease detection. It consists of identifying non-conforming patterns regarding an expected behavior. Despite the improvements provided by deep learning techniques in several areas, their use for anomaly detection is not widespread. The main reason is the difficulty to learn discriminative models when all the information available regards one class, or the classes are highly unbalanced. We propose a new deep learning-based solution for the anomaly detection problem. It consists of a hybrid system, composed of a feature extractor and a one-class classifier. The feature extractor is a convolutional neural network, trained as a regressor to learn a predefined distribution. The classifier is the one-class support vector machine, which performs the anomaly detection by using the outputs provided by the feature extractor. We used a gearbox failure diagnosis data set to assess the performance of our proposal. We also compared our anomaly detection system with other deep learning-based techniques commonly found in the literature. Our proposal presented an average accuracy close to 0.95, outperforming techniques based on the reconstruction error and hybrid models.

Using the Kullback-Leibler Divergence and Kolmogorov-Smirnov Test to Select Input Sizes to the Fault Diagnosis Problem Based on a CNN Model

“…They are applied on several domains, e.g. speech and audio recognition [1], fault detection and diagnosis [2,3], feature extraction on images and videos [3,4]. Those techniques learn to represent the input information at different complexity levels along with their intermediate layers.…”