Development of novel combustion risk index for flammable liquids based on unsupervised clustering algorithms

Abstract

Current liquid flammability classification mainly relies on flash point and its risk is largely dependent on consequence and probability. However, combustions of liquefied marine fuels have their uniqueness, leading to a less consistent with the common classification. This work aims at classifying flammable liquids in compression ignition engines for further safety evaluation. Besides liquid flammability characteristics, flame propagation and aerosol formulation are considered. Two unsupervised machine learning clustering algorithms, k-means and spectral clustering, are applied to the collected liquid compounds database. To consider both cluster cohesion and separation, the global mean silhouette value is used to find the optimal number of clusters and to evaluate the clustering performance. The results show that the spectral clustering outperforms k-means on classifying the risk ratings for all proposed models, while the clustering accuracy of the optimal model has been doubled by employing spectral clustering algorithm. Moreover, principal component analysis and star coordinate diagrams are presented to visualize high dimensional data to 2-D graphs. Finally, the overall liquid safety performance is evaluated by a novel combustion risk index via the weight values determined by the information entropy approach. This index can be used to explore inherently safer fuels in the process industries.