Optimization Modeling of Near-Infrared Spectral Detection Based on the CKAN Network

This paper proposes a feature extraction model based on the Kolmogorov-Arnold network (KAN network), termed AE-CKAN. The model integrates the KAN network with the reconstruction error from an autoencoder (AE) and, through parameter configuration and the design of regularization techniques, establishes a CKAN network specifically adapted for near-infrared spectroscopy analysis. This approach further mitigates the overfitting and computational demands associated with the learnable activation functions in the KAN network. Experimental results indicate that the prediction set achieves an RMSEP (Root Mean Square Error of Prediction) of 0.0483 and an R2 (Coefficient of Determination) of 0.9114, representing improvements of 1.00% and 5.67% over the CNN and CKAN models, respectively. Notably, the performance in terms of the RPD (Residual Predictive Deviation) metric is particularly outstanding, reaching 3.3603, which is significantly better than both the CNN and CKAN models.