Optimally pruned extreme learning machine with ensemble of regularization techniques and negative correlation penalty applied to automotive engine coldstart hydrocarbon emission identification

In this investigation, the authors test the efficacy and reliability of optimally pruned extreme learning machine with ensemble of regularization techniques to identify the exhaust gas temperature (Texh) and the engine-out hydrocarbon emission (HCraw) during the coldstart operation of automotive engines. These variables have significant impact on the cumulative tailpipe emissions (HCcum) during a coldstart phenomenon, which is the number one emission-related problem for today's spark-ignited (SI) engine vehicles. To do so, the concepts of ensemble computing with negative correlation learning (NCL) and pruning of neurons are used in tandem to cope with difficulties associated with extracting knowledge from collected database. In the proposed framework, the regularization strategies are adopted to help us increasing the numerical stability of identifier while mitigating the redundant complexity of hidden neurons. Moreover, to increase the generalization of identifier and also reduce the effects of uncertainty, an ensemble of independent OP-ELM with NCL selection criterion called OP-ELM-ER-NCL is taken into account. To endorse the valid performance of OP-ELM-ER-NCL for modeling the characteristics of engine over the coldstart phenomenon, its performance is compared to a set of well-known identification systems, i.e. standard extreme learning machine (ELM), back-propagation neural network (BPNN), OP-ELM with different types of regularization, ensemble of regularized OP-ELM without negative correlation (OP-ELM-ER), and an ensemble ELM with a constrained linear system of leave-one-out outputs (E-LL), in terms of both accuracy and computational complexity. The simulation results indicate that the proposed identifier is really capable of capturing the knowledge of collected database. It is observed that its resulted accuracy and robustness are comparable with those obtained by identification methods available in the literature. Besides, using NCL strategy aids the ensemble to select the most effective regularization techniques and remove the redundant (ineffective) ones, which consequently decreases the complexity of final ensemble.