Applying evolving fuzzy models with adaptive local error bars to on-line fault detection

The main contribution of this paper is a novel fault detection strategy, which is able to cope with changing system states at on-line measurement systems fully automatically. For doing so, an improved fault detection logic is introduced which is based on data-driven evolving fuzzy models. These are sample-wise trained from online measurement data, i.e. the structure and rules of the models evolve over time in order to cope 1.) with high-frequented measurement recordings and 2.) online changing operating conditions. The evolving fuzzy models represent (changing) non-linear dependencies between certain system variables and are used for calculating the deviation between expected model outputs and real measured values on new incoming data samples (-> residuals). The residuals are compared with local confidence regions surrounding the evolving fuzzy models, so-called local error bars, incrementally calculated synchronously to the models. The behavior of the residuals is analyzed over time by an adaptive univariate statistical approach. Evaluation results based on high-dimensional measurement data from engine test benches are demonstrated at the end of the paper, where the novel fault detection approach is compared against static analytical (fault) models.