A Siamese Autoencoder Preserving Distances for Anomaly Detection in Multi-robot Systems

Abstract

A Siamese autoencoder preserving distances for preprocessing sensor data in the multi-robot system anomaly detection is proposed. It can be viewed as two identical autoencoders with shared weights by the encoder parts. The proposed neural network reduces the dimensionality of the input observations in order to simplify the use of the Mahalanobis distance in anomaly detection. Moreover, it reduces the dimensionality preserving the original data structure. The Siamese autoencoder also increases the distance between anomalous observations and centers of sliding windows. The network allows us to detect the anomalous behavior of robots taking into account a complex data structure received from sensors. Numerical experiments illustrate the outperformance of the Siamese autoencoder.