Abstract: To achieve accurate and non-destructive identification of composite filter rods with natural plant-derived porous material particles (Natural Plant-derived Porous particles, NP), terahertz time-domain spectroscopy (THz-TDS) was adopted as the core detection technology to acquire the terahertz absorption coefficient spectra of all prepared filter rod samples. Three classical spectral preprocessing methods, namely Multiplicative Scatter Correction (Multiplicative Scatter Correction, MSC), S-G(Savitzky-Golay, SG) smoothing algorithm, and Standard Normal Variate (Standard Normal Variate, SNV) transformation, were individually implemented to process the raw original spectra. The preprocessing procedures were designed to mitigate adverse interference caused by sample surface unevenness, optical scattering effects, baseline drift and random noise, so as to enhance spectral signal quality and highlight inherent characteristic differences between different sample types. On the basis of the preprocessed spectral datasets, Principal Component Analysis (PCA) was further utilized for dimensionality reduction of the high-dimensional spectral data. By extracting principal components that cover the vast majority of spectral variance information, this procedure can effectively eliminate data redundancy and multicollinearity among spectral variables, reducing the computational load of subsequent classification models while retaining the core feature information of the samples. For the classification recognition task, multiple supervised machine learning classification models were constructed, including Support Vector Machine (Support Vector Machine, SVM) configured with three distinct kernel functions (Linear kernel, Polynomial kernel, and Radial Basis Function (RBF) kernel), Logistic Regression (Logistic Regression, LR), K-Nearest Neighbor (K-Nearest Neighbor, KNN), Decision Tree (Decision Tree, DT), and Random Forest (Random Forest, RF). Finally, the effects of different combinations of preprocessing strategies and classification models on classification performance, recognition accuracy and generalization capability were systematically compared and comprehensively evaluated. The comprehensive results of spectral feature analysis and classification model performance evaluation are elaborated as follows: ①The absorption coefficient in the 0.75 - 2.50 THz frequency band serves as the core spectral feature for characterizing Natural Plant-derived Porous particles filter rods. The 0.82-1.20 THz and 1.64-1.98 THz bands present high sensitivity to the additive content and microscopic spatial morphology of the substance. ②After MSC preprocessing, the cross-validation accuracy of all models on the training set exceeds 87%, and both the SVM (RBF) and SVM (Linear) models can reach 93.21%. ③PCA was applied to perform feature extraction on the absorption spectral data. Among the models constructed with the first three principal components that account for a cumulative variance contribution rate of 92%, the MSC-SVM (RBF) model exhibits superior comprehensive performance. The model delivers an overall accuracy above 95%, with a test set accuracy of 95.83% (Macro-F1 score: 95.83%, Kappa coefficient: 95.24%). It achieves 100% recognition accuracy for samples in categories A2, A3, B2 and C0, while the recognition accuracy for categories including B1 and B3 ranges from 86.67% to 93.33%. ④For the rest of the classification models coupled with MSC preprocessing, the test set classification accuracies are listed respectively: 94.17% for the MSC-LR model, 92.50% for the MSC-KNN model, 87.50% for the MSC-DT model, and 88.33% for the MSC-RF model. The MSC-SVM (RBF) model can be applied to research on non-destructive detection and identification of different types of these NP particles composite filter rods, and also provides a reference method for quality control during the manufacturing process of NP particles and other types of particle-based composite filter rods.