Gas identification with drift counteraction for electronic noses using augmented convolutional neural network

Sensor drift is the most challenging problem in the chemical sensing of an electronic nose system. We have proposed a new pattern recognition approach, namely augmented convolutional neural network (ACNN) to solve a gas discrimination problem over an extended period with high accuracy rates. The ACNN model is a continuously updated machine learning framework that automatically transforms the time-varying gas signals into a multi-dimensional feature matrix, then takes the incremental data to extend the existing model's knowledge with internal parameter tuning, and further compensate the model deviation in virtue of an external adjustment module over the base CNN classifier. Three different cases originating from self-collected short-term data (4-month) to the very long-term public dataset (3-year) are tested to verify the proposed approach compared with the existing e-nose pattern recognition algorithms including support vector machine (SVM), backpropagation neural network (BP), ensemble classifiers (En_wts), and ensemble classifiers with uniform weights (Uni_wts) as well as the normal CNN model. Experiments indicate the presence of e-nose drift over extended periods. Results show that the proposed method can cope well with sensor drift and gets superior performance than the traditional classifiers, with an accuracy increase of over 30% for the worst case.