Assessing the competency of a semi-parametric expert system in the realms of response characterization uncertainty in premixed methanol dual fuel diesel combustion strategies: In critique to RSM

摘要

Engine response characterization endeavours of the day are faced with the contradictory obligations of developing an accurate engine response map in face of an ever-expanding parametric space and decreasing test-bench based prototype development time cycles. Under such imperatives of sparse experimental regimens, the present study focuses on the pivotal notions of reliability in a data-driven multi-variate response modelling endeavour in the complex realms of a premixed methanol dual fuel diesel operation. In the study, the Gene Expression Programming (GEP) technique with its distinguishing features of adaptivity in model evolution has been critically examined for its credibility as an expert system in such emerging combustion strategies. The relevance and competency of the GEP method has been rationalized to this end by comparing its performance with the traditional Response Surface Methodology which has been the mainstay in such engine response characterization objectives. Model integrity has been analysed and compared through a multitude of conventional and improvised statistical goodness-of-fit measures encompassing absolute and relative error metrics. Model reliability was adjudged through an exhaustive uncertainty analysis. Information theoretic measures of Kullback-Leibler divergence and improvised Thiel uncertainty coefficients were employed to compare the quality of system knowledge gained by the competing techniques. Generalization capability was reckoned through all the respective metrics applied across a test data set held blind to the meta-model training sequence. Comparable regression scores across conventional and standardized correlation measures notwithstanding, all GEP evolved meta-models registered substantially lower footprints in all statistical error metrics together with lower uncertainty bandwidths of estimation and information loss across both training and test data simultaneously for all responses explored. The outcome of the study arguably evokes considerable reconsideration on the choice of conventional RSM based characterization strategies in the paradigms of system response uncertainty prevalent in the emergent engine combustion strategies involving significant yet unknown degrees of non-linearity in its parametric operational space.