Research on data augmentation method of ion mobility spectrometry and its application in classification of the simulants of chemical warfare agents

Ion mobility spectrometry is a rapid, highly sensitive analytical method for the gaseous samples with low detection limit. In public security field, it is widely used to detect chemical warfare agents, illegal drugs and explosives. However, due to the difficulty in obtaining samples of chemical warfare agents and hazardous chemicals, there are not enough samples to train the chemometrics models, which resulting in poor accuracy and robustness of the chemometrics models. In order to improve the accuracy and robustness of the chemometrics models, it is necessary to study the data augmentation method of ion mobility spectrometry. In recent years, generative adversarial nets(GAN) has significant progress in the field of image data augmentation field. GAN makes the generated samples obey the real data distribution through adversarial training. However, the generated model will fail without specific constraints. According to the detection principle of ion mobility spectrometry, when the reduced mobility of chemical substance is different, the full width at half-maximum and standard deviation of characteristic ion peak will change accordingly. In this paper, standard deviation-conditional generative adversarial nets(SD-CGAN) data augmentation method was proposed based on the constraint of standard deviation of characteristic ion peak. In order to verify the performance of proposed data augmentation method, dimethyl methylphosphonate and acetone samples were collected to construct the real data sets, which were frequently used as simulants of chemical warfare agents. Pearson correlation coefficient and FID were used to evaluate the basic GAN and proposed SD-CGAN methods. The experimental results indicated that SD-CGAN method could generate high-quality samples. Based on the data augmentation method, deep learning method was used to the classification of multitype samples. It demonstrated that improved one-dimensional CNN method has better performance than traditional(K-NN, SVM) classification methods on the data augmentation set. It has significant meaning to improve the accuracy and robustness for the detection of chemical warfare agents under small real sample size.