Research On Infrared Spectrum Recovery Algorithm Based On The Improved Transformer Model

Nowadays,infrared spectroscopy analysis technology has been widely used in various fields of real life,through which many fields can solve the problems faced by many fields,such as crop variety identification,alloy purity detection and organic structure analysis.However,in the process of infrared spectrum acquisition,due to the aging of infrared spectrometer equipment,the existence of light source and random noise,and the point expansion effect,the collected infrared spectrum tends to degrade,and the characteristics are quite different from the original infrared spectrum,which seriously affects the application of infrared spectrum analysis technology,so it is necessary to restore the degraded infrared spectrum.This thesis proposes two infrared spectrum recovery algorithms,different from the traditional recovery algorithm mainly from the degradation of mathematical model solution analysis or from the infrared spectrum noise removal perspective for recovery,the two algorithms proposed in this thesis based on the infrared spectrum prediction task to achieve the infrared spectrum recovery function,in which the related prediction model is designed to predict the original infrared spectrum from the degraded infrared spectrum,and the predicted result is used as the recovery infrared spectrum.In this thesis,a mask layer is first designed to mask the degraded infrared spectrum,aiming to suppress the noise in the degraded infrared spectrum,and secondly,the eigenvector embedding realized by one-dimensional convolution and the position coding of the learnable parameters are carried out on the infrared spectral sequence.Finally,the multi-head attention mechanism in the Transformer encoder was used to extract features and predict the recovery of infrared spectral sequences.This model training is divided into two phases,the first is pre-training for unsupervised learning,and the second is supervised learning.Experiments show that the algorithm has better recovery performance than traditional methods,both for the qualitative analysis of spectral peaks and flat regions and the comparison of quantitative indicators of recovery performance.On this basis,aiming at the problem that the algorithm of improved Transformer model based on mask is not sufficient in grasping the overall features,this thesis proposes an infrared spectral recovery algorithm based on frequency improvement Transformer model.In this model,two schemes are designed and implemented for frequency preprocessing and frequency attention calculation,namely segmentation selection and block integration.Among them,the frequency preprocessing module is mainly used for noise suppression in the frequency domain,and the frequency attention calculation changes the implementation of the attention mechanism to the frequency domain,which makes better use of the overall characteristics of the frequency domain and optimizes the calculation amount of the model.The simulation shows that the recovery performance of this algorithm is further improved on the basis of the algorithm based on the improved mask Transformer model,and the problem of insufficient grasp of the overall feature is solved.In order to verify the practical application value of this algorithm,this thesis applies it to the infrared spectral classification task and carries out practical application verification experiments.Experiments show that the classification accuracy of the classifier applying this algorithm is up to25 percentage points higher than that of the classifier without this algorithm,which fully verifies the important application value of this algorithm in the field of infrared spectroscopy.