Abstract: To address the low classification accuracy caused by spectral overlap and similar spectral characteristics in detecting aldehydes using surface-enhanced Raman scattering (SERS) spectroscopy technology, a hybrid model integrating convolutional neural network (CNN) and support vector machine (SVM), termed CNN-SVM, is proposed in this work. The model employs CNN to automatically extract both local and global features of SERS spectra and utilizes SVM to enhance classification performance. Experimental results on 1,500 SERS spectral data of various aldehyde substances collected on a self-built platform show that the CNN-SVM model achieves a classification accuracy of 97.33% on the test set, significantly outperforming the least squares support vector machine model (88.67%) or the CNN (92.00%) model separately. Meanwhile, the CNN-SVM model substantially reduces spectral processing time and memory consumption compared to a pure CNN architecture. Due to effectively overcoming the issue of spectral overlap, the proposed CNN-SVM hybrid model delivers high accuracy, high computational efficiency, and low resource requirements, providing a viable strategy for real-time, highly sensitive monitoring of aldehydes in complex environments and broadening the application of SERS spectroscopy technology in food safety and environmental monitoring.

Key words: spectral analysis, classification model, surface-enhanced Raman scattering, aldehyde pollutants, deep learning, construction machinery exhaust