Identification Of Chenopodium Based On Terahertz Spectroscopy And Machine Learning

Chen Pi is a kind of valuable medicinal food,which is widely used in daily life and has a market value of tens of billions of dollars,but the quality of Chenpi in the market varies,and the illegal businessmen use substandard as good,which not only damages the interests of consumers,but also affects the brand image of Chenpi.Therefore,it is very important to explore a real-time identification method with easy operation and short detection period.Terahertz wave is located between microwave and infrared in the electromagnetic spectrum,and it is in the middle of infrared and microwave in the spectrum,so it has excellent characteristics such as fingerprint spectrum,high security,low energy,etc.It is these characteristics that make terahertz have a wide range of applications in material identification,national defense,biomedical and other fields.In this thesis,we first classify different chenopods using a combination of terahertz spectroscopy and machine learning,and then analyze the chenopod spectra for feature importance to find the feature points in the spectra that play a key role in the classification results.The main work is as follows:1.Terahertz spectroscopy with machine learning approach to classify seven types of Chen Pi.The raw data of different chenopods are preprocessed and four classification models,namely,support vector machine,random forest,back propagation neural network and kernel limit learning machine,are built respectively.From the experimental results,the best accuracy of the four types of models for the data of terahertz timedomain spectrometer is 94.2857%,90%,82.4571% and 91.4286 %.For the data from the terahertz vector network analyzer,the best accuracies of the four classes of models are98.5714%,98.5714%,88.5714%,and95.7143%,respectively,with the support vector machine model having the best performance and moderate training time.The above results show that the identification method based on terahertz spectroscopy and machine learning can achieve the identification of Chenopodium.2.Feature importance analysis of terahertz spectrum.The spectrum was divided into different frequency bands by equal intervals,LASSO algorithm,partial least squares algorithm and mutual information algorithm to analyze the feature importance of the spectrum and build a support vector machine based classification model.The experimental results show that the classification models established by terahertz time-domain spectrometer data in different frequency bands have significant differences in accuracy,among which the accuracy of 0-1THz band is 91.4286%,while the accuracy of the rest of the bands does not exceed 53%.And the accuracy of the data from the terahertz vector network analyzer is basically the same in each frequency band,which is greater than 95%.The accuracy of the classification models built using the top 10 features in importance is no less than87%,with the highest being 95.7143%.This indicates that these 10 features play a major role in the classification process,and these features basically lie within the frequency range of 0.5T-0.85 THz,indicating the existence of feature points in this frequency band for different chenopodium spectra.