Near‐Infrared Spectroscopy Combined With Multistrategy Fusion Zebra Optimization Algorithm and KAN for Grade Identification of Baijiu Base Liquor

ABSTRACT

To address the strong subjectivity of manual evaluation, the low efficiency of traditional physicochemical detection, as well as the issues of high variable dimensionality, redundant information, and subtle inter‐class differences in near‐infrared (NIR) spectral data during the grade identification of Baijiu base liquor, this paper proposes a grade identification method for Baijiu base liquor based on multistrategy fusion zebra optimization algorithm (MFZOA) feature selection and Kolmogorov–Arnold network (KAN). First, a Fourier transform near‐infrared (FT‐NIR) spectrometer was employed to acquire the spectral information of strong‐flavor Baijiu base liquor samples, and the GA‐SPXY method was used to partition the samples into training and validation sets, thus improving the representativeness of sample division. Subsequently, a binary MFZOA feature selection model oriented to classification tasks was constructed, in which candidate feature subsets were encoded as zebra individuals, and the classification performance was used as the fitness evaluation criterion. MFZOA integrates good point set initialization to improve the uniformity of the initial population distribution, a nonlinear dynamic convergence factor to balance global exploration and local exploitation, and adaptive Gaussian mutation to enhance population diversity, thereby improving the search for discriminative wave number subsets. Finally, the selected feature variables were fed into the KAN model, whose learnable univariate functions on network edges enable flexible nonlinear mapping between spectral features and grade labels. Experimental results demonstrate that MFZOA can effectively screen out the characteristic wave numbers relevant to base liquor grade discrimination, and the obtained feature subset exhibits strong class discriminability. Under the same feature input conditions, the KAN model outperforms the compared models including random forest (RF), support vector machine (SVM), multilayer perceptron (MLP), and one dimensional convolutional neural network (1D‐CNN), achieving an accuracy (ACC) of 95.86%, a recall of 96.94%, and a macro‐F1 of 95.52%. The findings indicate that the proposed MFZOA–KAN method can realize rapid and nondestructive grade identification of Baijiu base liquor, and can provide a methodological reference for intelligent grading and quality control in the Baijiu production process.