Reconstructive reservoir computing for anomaly detection in time-series signals

We propose reconstructive reservoir computing (RRC) which performs better anomaly detection in time-series signals than forecasting-based methods. In this paper, re-construction means that a neural network generates past input signals. RRC reconstructs a past normal signal for anomaly detection using an echo state network which can learn quickly and stably. We expect that it is easier to restore a past normal signal than to predict an unknown future normal signal. For anomaly detection, we compute an instantaneous recon-struction error. The reconstruction error larger than a threshold is a sign of anomaly. We conduct experiments using sound data obtained from a pump. In the experiments, we pay attention to a time lag between input and output to be reconstructed since we assume that an excessive time lag makes reconstruction difficult due to signal attenuation in the network. Experimental results show that if the time lag is moderate, the reconstruction error of the normal signal is lower than the forecasting error of the same signal. Furthermore, we show that RRC with the appropriate time lag has a better anomaly detection performance index than forecasting-based methods.