Abstract: In order to make up for the lack of selectivity of a single gas sensor in the face of a variety of gases and to identify a variety of gases more accurately, an improved convolutional neural network based on structural reparameterization technology and depth-separable convolution technology is proposed. It integrates the multi- branch convolution structure during model training into the single branch simple convolution layer during inference. In addition to simplifying the complexity of the inference model, the feature extraction ability of the model for gas response data is greatly enhanced. When this method is applied to a common data set of gas sensor array containing 10 common VOCs, the recognition accuracy reaches 96. 46%, and the accuracy reaches 97. 44% after adjusting the complexity of the model and adding the convolutional layer.

Key words:  , e-nose, gas classification and identification, structure re-parameterized, convolutional neural networks(CNN)