A lightweight gas classification and concentration prediction method based on PTQ-CNN by using an electronic nose system

Abstract

As a device for gas detection, electronic nose (E-nose) has been widely used in many fields. In recent years, the emergence of deep learning has greatly improved the gas detection accuracy. However, numerous parameters make it difficult to deploy deep learning algorithms, hindering the real-world application of the E-nose. In this work, we designed a convolutional neural network (CNN) and post-training quantization (PTQ) was used to compress the model size to obtain a lightweight model. First, we designed an E-nose and an efficient gas data acquisition platform to collect sufficient gas data containing four different categories of gas for model training. In the process of designing the CNN, the effects of batch size, data length and number of convolutional kernels on the performance of the model have been discussed to obtain the best performance of the CNN. Then, using the PTQ to compress the CNN model, a lightweight model PTQ-CNN is obtained. The experimental results show that CNN still maintains a high accuracy in performing the tasks of concentration prediction and gas classification after being compressed by 3.75 × . Specifically, the average values of R², MAE and MSE of PTQ-CNN for gas concentration prediction were 0.993, 1.524 and 5.856, respectively, and the classification accuracy of the gas reached 99.89 %.