Raman Spectral Identification Using Machine Learning Method

Raman spectroscopy is widely used in scientific and industry research from which scientists can obtain information about the molecular structure of materials.In recent years,the data analysis of Raman spectroscopy has received intensive attention and made great progress,which extended the potential of Raman spectroscopy in real-world applications.At present,the majority of the frontier studies on the classification of Raman spectra focus on binary classification and tri-classification.For more complex Raman spectrum classification problem such as multi-class classification tasks,deep learning methods was reported to outperform other machine learning methods such as SVM and PCA.However,in practical work,experts cannot obtain a large amount of Raman data for the training of deep neural networks,hindering the application of deep learning algorithm.Therefore,it is very important to investigate Raman spectrum’s multi-classification task and relax the deep learning model’s dependency on large-scale datasets.The studies of this thesis are divided into two parts.One is to employ machine learning algorithm to investigate the multi-class classification of RRUFF Raman database and explore the optimal model on this task.The second is to employ transfer learning method to investigate complex classification task of small-scale Raman data.The specific contents and conclusions of the thesis are as follows:1.Six machine learning models(support vector machine,random forest,K-nearest neighbors,fully connected neural network,convolutional neural network and recurrent neural network)were constructed to classify Raman spectra of RRUFF Raman database.The results showed that recurrent neural network(RNN)we built achieved the best performance on this task,it outperformed literature reported best model-convolutional neural network(CNN).Further optimization of the RNN model improved the classification accuracy of the task by 5.8%.Moreover,we found that 1-norm regularization(L1)is very consistent with characteristics of spectroscopic data.The introduction of L1 and data enhancement into the RNN model further improved the classification accuracy by 2.3%.2.Optimized transfer learning method-fine-tuning was employed to classify small-scale Raman spectra.The results showed that transfer learning model performs very well on this task,it improved the classification accuracy by 4.1%compared with non-transfer learning model.