Combining reservoir computing and variational inference for efficient one-class learning on dynamical systems

Abstract

Usually, time series acquired from some measure-ment in a dynamical system are the main source of infor-mation about its internal structure and complex behavior. Inthis situation, trying to predict a future state or to classifyinternal features in the system becomes a challenging task thatrequires adequate conceptual and computational tools as well asappropriate datasets. A specially difficult case can be found inthe problems framed under one-class learning. In an attempt tosidestep this issue, we present a machine learning methodologybased in Reservoir Computing and Variational Inference. Inour setting, the dynamical system generating the time series ismodeled by an Echo State Network (ESN), and the parametersof the ESN are defined by an expressive probability distributionwhich is represented as a Variational Autoencoder. As a proof ofits applicability, we show some results obtained in the contextof condition-based maintenance in rotating machinery, wherevibration signals can be measured from the system, our goalis fault detection in helical gearboxes under realistic operatingconditions. The results show that our model is able, after trainedonly with healthy conditions, to discriminate successfully betweenhealthy and faulty conditions.