Multimodal Forecasting Methodology Applied to Industrial Process Monitoring

Abstract

Industrial process modeling represents a key factor to allow the future generation of industrial manufacturing plants. In this regard, accurate models of critical signals need to be designed in order to forecast process deviations. In this work, a novel multimodal forecasting methodology based on adaptive dynamics packaging and codification of the process operation is proposed. First, a target signal is decomposed by means of the empirical mode decomposition in order to identify the characteristic intrinsic mode functions. Second, such dynamics are packaged depending on their significance and modeling complexity. Third, the operating condition of the considered process, reflected by available auxiliary signals, is codified by means of a self-organizing map and presented to the modeling structure. The forecasting structure is supported by a set of ensemble adaptive-neurofuzzy-inference-system-based models, each one being focused on a different set of signal dynamics. The performance and effectiveness of the proposed method are validated experimentally with industrial data from a copper rod manufacturing plant and performance comparison with classical approaches. The proposed method shows improved performance and generalization over classical single-model approaches.