Research On Automatic Feature Extraction And Recognition Of Terahertz Spectrum

Terahertz time-domain spectroscopy is a detection technique with high resolution,coherence,low energy,fingerprint spectrum and high signal-to-noise ratio,which makes the technology quantify substances such as drugs,explosives,drugs,food and agricultural products.And qualitative analysis has been widely used.In the current qualitative and quantitative analysis of materials using terahertz time-domain spectroscopy,it is often encountered that there is no obvious characteristic peak in the terahertz time-domain spectrum.How to extract and analyze these spectra is an important issue.To this end,this paper focuses on the feature extraction and effective recognition classification of terahertz spectroscopy,and achieves the following results:This paper proposes an identification method based on bidirectional long-and short-term memory networks to automatically extract terahertz spectral features.The long-term and short-term memory unit can effectively solve the problem that the original terahertz spectral data has high dimensionality and make the model difficult to train.Combined with the model's bi-directional spectral information,the architecture model can enhance the model to automatically extract effective feature information from similar terahertz spectral data.ability.The experimental results show that the method has the advantage of directly extracting the effective features of the terahertz spectrum,thus achieving the effective recognition and classification of similar terahertz spectral data.Because the above model automatically extracts the terahertz spectral features and recognition methods,it is difficult to adjust the parameters and the model cannot be self-scaling model complexity according to the size of the data set.Therefore,this paper proposes a multi-grain cascaded forest algorithm(multi-Grained).Cascade Forest,gcForest)automatically extracts the identification method of terahertz spectral features.The multi-granularity scanning stage is used to enhance the terahertz spectral samples,and the order relationship of the terahertz spectral sequence data features is mined as much as possible.Then the data structure is processed layer by layer in the cascade forest stage to enhance the characterization and learning ability of thealgorithm.Improve the prediction accuracy of the model.The experimental results show that the proposed method has the ability to directly extract effective features from similar terahertz spectra,and the model complexity can be adaptively scaled,which is suitable for high-precision identification of terahertz spectral data sets of different sizes.The experimental results of the above two methods show that the terahertz spectral data sets with different complexity of the fifteen compounds of the three classes have good recognition results.It can be used for qualitative identification in the fields of drug analysis,quality control,explosives and food testing.