An anomalous behavior detection of a robot system by using a hierarchical Siamese neural network

Abstract

A robot system anomalous behavior detection is studied in the paper. It is supposed that robots are equipped with many sensors such that a set of environment vectors can be obtained at every time moment. A Siamese neural network (SNN) as a special type of neural networks is proposed to be used in order to detect anomalous behavior. It can be regarded as an alternative to the well-known Mahalanobis distance (MD) often used in anomaly detection. Strategies for efficient training the SNN are considered. In order to simplify the SNN training, we also propose to use a hierarchical neural network. By means of the SNN and the proposed network architectures, the anomalous behavior detection of robots by complex data structures received from the set of sensors can be significantly simplified.