Research On Terahertz Time Domain Spectrum Analysis Based On Machine Learning

Terahertz time-domain spectroscopy has been studied in many fields as a new and effective photoelectric detection technology,which has the characteristics of large bandwidth,high signal-to-noise ratio and strong sensitivity and can quickly obtain the optical parameters of samples,it has great application in terahertz communication,environmental monitoring,food quality and safety,etc.However,the existing food quality detection technology has the disadvantages of low recognition accuracy and long detection time.As an important branch of artificial intelligence,machine learning can simulate human learning behavior,and then complete massive data processing,while maintaining high robustness and efficiency.In order to further explore the application of terahertz time-domain spectroscopy in the field of food safety detection,we obtained the THz spectrum of samples based on the terahertz time domain spectrum.In the meanwhile,the identification and analysis of substances are completed by machine learning algorithms.The main work and contents of this paper are as follows:1.The spectral information of substances is analyzed by clustering algorithm.The spectral data in 0.3-1.6THz band were extracted.Principal component analysis combined with t-distribution dimension reduction algorithm is used to reduce the data dimension to two-dimensional plane,and the clustering effect and model performance of K-means,hierarchical clustering and density clustering are compared.The experimental results show that a group of unlabeled data can be correctly divided into four clusters by using the cluster analysis model,which is in line with the actual situation of the sample.Compared with other models,the K-means clustering effect is improved by 71.2%on average,and there is no noise interference.2.The classification model of supervised learning algorithm is used to classify and identify mixed samples.Evolutionary Support Vector Machine(EA-SVM)algorithm is proposed.The three classification models are SVM,k-nearest neighbor and logic regression as reference,Terahertz time domain spectroscopy technology combined with EA-SVM algorithm improves the classification accuracy and F1 evaluation index by 5.8%and 5.7%respectively,and the average operation efficiency is improved by 55.6%,showing excellent classification performance.The new model with better scalability and higher prediction accuracy is more suitable for qualitative identification of food additives in the future.