Event Detection and Predictive Maintenance using Component Echo State Networks

Sammanfattning: With a growing number of sensors collecting information about systems in indus- try and infrastructure, one wants to extract useful information from this data. The goal of this project is to investigate the applicability of Echo State Net- work techniques to time-varying classification of multivariate time series from primarily mechanical and electrical systems. Two relevant technical problems are predicting impending failure of systems (predictive maintenance), and clas- sifying a common event related to the system (event detection). In this project, they are formulated as a supervised machine learning problem on a multivariate time series. For this problem, Echo State Networks (ESN) have proven effective. However, applying these algorithms to new data sets involves a lot of guesswork as to how the algorithm should be configured to model the data effectively. In this work, a modification of the Echo State Network (ESN) model is presented, that helps to remove some of this guesswork. The new algorithm uses specifically structured components in order to facilitate the generation of relevant features by the ESN. The algorithm is tested on two easy event detection data sets, and one hard predictive maintenance data set. The results are compared to Support Vector Machine and Multilayer Perceptron classifiers, as well as to a basic ESN, which is also implemented as a reference. The component ESN successfully generates promising features, and outperforms the minimum complexity ESN as well as the standard classifiers.