Samir Brahim Belhaouari

Abstract. This paper introduces a novel gas identification method based on the Cluster-k-Nearest Neighbor (C-k-NN) algorithm, integrating k-NN’s accuracy with k-means clustering efficiency. Feature selection leverages new statistical metrics, specifically \(\text{Metric1}_j(l,k) = \frac{| \bar{x}_j(l) - \bar{x}_j(k) |}{\sqrt{ \frac{S_j(l)^2}{n_l} + \frac{S_j(k)^2}{n_k} }}\) and \(\text{Metric2}_j(l,k) = 1 - A_{\text{overlap}}(l,k)\), where \(A_{\text{overlap}}(l,k) = \int_{-\infty}^{+\infty} \min(\phi_l(x), \phi_k(x)) dx\). The proposed Tree C-k-NN achieves 100% accuracy by minimizing classification time and misclassification. The approach is validated on six gases, with a performance of 98.7% for C-k-NN and 100% for Tree C-k-NN, and further confirmed using public datasets.
    

Illustration of the proposed experimental design.