Detecção de pontos de mudança em séries temporais utilizando uma formulação neural/fuzzy/Bayesiana: aplicação na detecção das falhas

Abstract

In this work, the problem of detecting till two change points in time series is handled by using a new neural/fuzzy/Bayesian technique. This proposed technique is split into a three-step formulation, namely: the first step is performed by a Kohonen neural network classification algorithm that defines the model to be used in the case of one change point or two change points in the time series. The second step consists of a fuzzy clustering to transform the initial data in the time series, with arbitrary distribution, into a new one that can be approximated by a beta distribution. Also, the fuzzy cluster centers are determined by using the Kohonen neural network classification algorithm used in the first step. The last step consists in using the Metropolis-Hastings algorithm to appropriately perform the detection of the change points in the transformed time series generated by the second step, with beta distribution. The main contribution of the proposed approach in this work, related to previous one in the Literature, is to allow to detect till two change points in time series with the correct model selection. Simulation results are presented in this work to illustrate the effectiveness of the proposed approach.