Pattern discovery in time series using autoencoder in comparison to nonlearning approaches

In technical systems the analysis of similar situations is a promising technique to gain information about system’s state, its health or wearing. Very often, cannot be defined but need discovered as recurrent patterns within time series data system under consideration. This paper addresses assessment different approaches discover frequent variable-length in series. Because success artificial neural networks (NN) various research fields, special issue this work applicability NNs problem pattern discovery Therefore we applied and adapted Convolutional Autoencoder compared it classical nonlearning based on Dynamic Time Warping, discretization well Matrix Profile. These have also been adapted, fulfill our requirements like potentially scaled from noisy We showed performance (quality, computing time, effort parametrization) those an extensive test with synthetic sets. Additionally transferability other sets tested by using real life vehicle data. demonstrated ability Autoencoders unsupervised way. Furthermore tests showed, that able quality approaches.