Research On Terahertz Spectrum Absorption Peaks Extraction And Deep Learning Recognition Algorithms

Terahertz spectroscopy can identify substances by detecting the vibration and rotation information of substance molecules,and has broad application prospects in food safety and public safety.Many substances have specific absorption peaks in their terahertz spectrum,and absorption peak characteristics can be used to identify substances.However,the absorption peaks of different components in the mixture spectrum may overlap with each other,making it difficult to identify weak peaks that are masked.On the other hand,due to the limitations of current terahertz spectroscopy hardware technology,spectral measurement time is long,and the sample production process is complicated.It is difficult to obtain a large number of spectral samples for learning features.In response to the above problems,this thesis studies absorption peaks extraction and deep learning algorithms of terahertz spectrum.Traditional absorption peaks extraction algorithms are divided into two independent steps of peak finding and fitting.Results of peak finding may not be suitable for fitting,and the number of absorption peaks cannot be adjusted according to the effect of the fitting,which is prone to false peaks,omit peaks and inaccurate peak positions.To this end,this thesis proposes an interval combination multi-peak fitting algorithm.The spectral band is divided into different intervals using valley points of the curve after the curve is greatly smoothed.Adjacent intervals can be combined into a larger interval for separate multi-peak fitting,and a genetic algorithm is used to solve the optimal interval combination scheme.Measure spectra for each pure substance repeatedly,and use density clustering algorithm to select characteristic absorption peaks.This thesis proposes an absorption peaks matching identification algorithm,which uses characteristic absorption peaks of pure substances as standard data.The algorithm can calculate the similarity of the absorption peak between the absorption spectrum containing any number of absorption peaks and standard substances,so as to realize substance identification and mixture component detection.Deep learning technology can automatically extract characteristics of the spectrum,regardless of whether the substance has absorption peaks or not.Convolutional neuralnetworks are currently the most commonly used deep learning models for spectral recognition problems,but require large amounts of data during training.At this stage,the amount of terahertz spectrum data is small,and convolutional neural networks are prone to overfitting.To this end,this thesis designs a multi-scale strided convolutional neural network,uses a multi-scale convolution structure to extract multiple types of features,and uses strided convolution instead of the pooling layer for feature compression to avoid loss of position information during the pooling process,and improve the spectral recognition rate finally.Algorithmic experiments were performed by actual absorption spectrum data which were obtained using terahertz spectrograph.The experimental data of the absorption peaks extraction show that the interval combination multi-peak fitting algorithm can improve the resolution and positioning accuracy of aliasing peaks based on the second derivative peaks finding method.Using data of pure substances as standard data,the first two and three recognition rates of six mixtures with different degrees of aliasing reached 91.7% and 97.5%,respectively.The deep learning experimental data show that the recognition effect of feature compression using strided convolution is better than pooling operation.The multi-scale strided convolutional neural network proposed in this thesis had a recognition rate of 99.65%for 40 substances,which was higher than the classic convolutional neural network at97.64%.The network can effectively retain the position information in the spectrum and reduce overfitting.