Data-driven inference modeling based on an on-line Wang-Mendel fuzzy approach

Abstract To address the modeling of continuous production process with dynamic and nonlinear characteristics, an on-line Wang-Mendel fuzzy inference model is proposed in this paper, which extracts the fuzzy rules from the raw data without prior knowledge. Considering the state of the industrial system changes dynamically online, typical data samples with support degrees are used to describe the sample distribution characteristics in each fuzzy region. With respect to online dynamic learning process, a self-evolutionary strategy with adaptive memory factors is proposed, and the fuzzy rule structures are updated gradually by a stochastic gradient descent method. For on-line updating, a loss function is constructed by considering the inference errors and the previous rule structure, in order to achieve continuous learning without forgetting the knowledge learned before. In the reasoning process, a sparse fuzzy reasoning approach is designed for extrapolating the knowledge of the fuzzy regions without sample data. To verify the effectiveness of the proposed method, chaotic time series data with noises and industrial practical data coming from a steel plant are employed for experimental analyses. The experimental results show that, the proposed method is capable of describing a variety of dynamic features and exhibiting high accuracy for the industrial data.