E2GKpro: An evidential evolving multi-modeling approach for system behavior prediction with applications

Nonlinear dynamical systems identification and behavior prediction are difficult problems encountered in many areas of industrial applications, such as fault diagnosis and prognosis. In practice, the analytical description of a nonlinear system directly from observed data is a very challenging task because of the too large number of the related parameters to be estimated. As a solution, multi-modeling approaches have lately been applied and consist in dividing the operating range of the system under study into different operating regions easier to describe by simpler functions to be combined. In order to take into consideration the uncertainty related to the available data as well as the uncertainty resulting from the nonlinearity of the system, evidence theory is of particular interest, because it permits the explicit modeling of doubt and ignorance. In the context of multi-modeling, information of doubt may be exploited to properly segment the data and take into account the uncertainty in the transitions between the operating regions. Recently, the Evidential Evolving Gustafson–Kessel algorithm (E2GK) has been proposed to ensure an online partitioning of the data into clusters that correspond to operating regions. Based on E2GK, a multi-modeling approach called E2GKpro is introduced in this paper, which dynamically performs the estimation of the local models by upgrading and modifying their parameters while data arrive. The proposed algorithm is tested on several datasets and compared to existing approaches. The results show that the use of virtual centroids in E2GKpro account for its robustness to noise and generating less operating regions while ensuring precise predictions.

► E2GKpro is an online algorithm for heath detection and prognostics.
► E2GKpro includes clustering and prediction in a single algorithm.
► The complexity of belief functions is circumvented using k-additive measures.
► The theory of belief functions is well suited for PHM applications.