Adaptive Multiple-Kernel SVM with Joint Kernel-Weight and Hyperparameter Optimization for IR Spectroscopy

Abstract

Infrared spectroscopy analysis technology is widely used in material identification and other fields due to its non-destructive and fast characteristics. However, the high-dimensional, nonlinear, and noisy nature of the data poses challenges to the robustness of feature extraction and classification models; Although SVM performs well in spectral classification, its performance depends on kernel function selection, and traditional single kernel processing of complex infrared spectral data has limited capabilities. To this end, this study focuses on the SVM composite kernel intelligent selection strategy and proposes a framework of integrated intelligent optimization algorithm. With the goal of maximizing classification accuracy and minimizing model complexity, the optimal weighted combination is searched in candidate kernel pools such as polynomial kernel and RBF (Radial Basis Function) kernel. The core innovation is the collaborative automatic optimization of composite kernel weight coefficients and SVM parameters; Based on the ATR-FTIR dataset of 120 organic compounds, experiments were conducted using a modified PSO (Particle Swarm Optimization) algorithm (optimizing inertia weights and learning factors, constructing a fusion fitness function for collaborative optimization) and an outer 5-fold cross validation. The results showed that the optimized composite kernel model achieved a classification accuracy of 94.1% ± 1.2%, which was 11.5 percentage points higher than the traditional RBF kernel (82.6% ± 1.8%). The convergence speed of the optimization iteration was 47% ± 3.5% higher than the standard PSO, and the prediction accuracy on the near-infrared drug/food dataset reached 96.67% ± 0.9%. The generalization and robustness were significant, providing a new paradigm for intelligent feature analysis of complex infrared spectral data and improving the accuracy and automation level of spectral substance identification.