Performance evaluation of lightweight pattern recognition algorithms for portable environmental monitoring electronic noses

Abstract

The bionic electronic nose mimics the olfactory system of living organisms, enabling the detection and differentiation of odors or gas mixtures. With appropriate data processing and pattern recognition algorithms, it can be used in environmental quality analysis and process control. The existing electronic nose algorithms have problems such as data processing difficulties, model complexity and high computational cost. In this study, we investigated the performances of various pattern recognition algorithms for classifying and identifying different gases, especially the performances of the lightweight algorithms, based on publicly available dataset. First, the response signals from the sensor array were converted into response maps. Then, several pattern recognition algorithms, including HOG_SVM, VGG16, SqueezeNet, ShuffleNet v2, MobileNet v3, and GhostNet were used to identify the gas species in unmixed and mixed gas datasets. And the performances of the pattern recognition algorithms were evaluated. The results show that VGG16 achieves the highest accuracy, attaining a perfect 100% on the unmixed gas dataset and 95.14% on the mixed gas dataset. Among lightweight models, SqueezeNet demonstrated superior performance, achieving an accuracy of 94.44% on the mixed gas dataset. Furthermore, SqueezeNet has the fewest parameters, totaling just 0.74 million. Meanwhile, MobileNet v3 is notable for its minimal computational cost, requiring only 0.059 billion Floating Point Operations Per Second (FLOPs). Deployment on an edge computing platform, such as the NVIDIA Jetson Xavier, enabled MobileNet v3 to achieve optimal performance in terms of inference latency and memory consumption.