In the field of machine olfactory,gas recognition is one of the most important research directions,and the focus of gas recognition is to accurately classify gas information while ensuring efficiency.Although machine learning and deep learning-based classification methods have achieved promising results in the field of gas recognition,accurately and quickly identifying mixed gases remains a challenging problem.Due to the significant potential of deep learning methods in the field of image recognition,this thesis transforms the gas recognition problem into an image recognition problem and leverages the powerful learning capability of deep learning to achieve accurate gas recognition tasks.In addition,considering the catastrophic forgetting issue that general deep learning models may face when applied to gas classification,this thesis creatively proposes a gas recognition method based on Supervised Contrastive Replay(SCR)incremental learning.Therefore,the work of this thesis is as follows:(1)In order to address the issue of low accuracy in gas recognition tasks for some traditional machine learning and deep learning methods,this thesis proposes to use the Markov Transition Field(MTF)method to transform one-dimensional gas data into MTF two-dimensional images.It combines 38 layer deep convolutional neural networks(DCNN)and residual networks(Res Net),which have shown excellent performance in image recognition,for gas recognition.Experimental results on the gas dataset show that the classification accuracy of these two deep learning-based gas recognition methods reaches 93.21% and 93.82% respectively,which is significantly higher than the accuracy of traditional machine learning and deep learning methods.(2)To address the issues of catastrophic forgetting and low efficiency in gas classification for deep learning,this thesis introduces an incremental learning approach based on SCR for classifying mixed gases.The data set is preprocessed using a method that converts the data into grayscale images,and the incremental learning network is used for classification.Experimental results demonstrate that the SCR gas recognition method achieves classification accuracies of 95.27% and 94.63% for 5-class and10-class mixed gases,respectively,and its training time is significantly shorter than typical deep learning models.(3)In the self-built indoor experimental scene,mixed gas information is collected using a sensor array.Based on the above algorithm,the mixed gas information collected in the actual physical experiment is classified.Through analysis of the experimental results,the effectiveness of the above algorithm is verified. |