Embedded Artificial Intelligence Approach for Gas Recognition in Smart Agriculture Applications Using Low Cost MOX Gas Sensors

Abstract

Smart agriculture represents one of the greatest potential scenarios in the field of the Internet of Things (IoT). Sensory and connectivity technologies together with big data analysis make the farming more accurate and controlled, overcoming the current model of intensive production. It allows farmers to accurately assess the amounts of water, fertilizers, and pesticides needed. Consequently, it guarantees better and healthier crops as well as costs reduction. The adoption of low cost and low power smart air probes based on gas, humidity, pressure and temperature sensors for monitoring air quality and environment conditions and crop emissions, is arising to become a fundamental instrument to be used to provide decision support to the farmers. Moreover, to embed edge computing in the probe for pre-process measurements enhance the potentiality of sensors increasing efficiency and reducing the amount of information traveling in the cloud. By focusing on the sensory features of the platform, the paper aims to exploit the developing a gas recognition algorithm which embeds a neural network and that is able to recognize different types of gas such as NH3, CH4, N20 on the basis of data coming from a gas sensor. Metal-oxide-semiconductor (MOX) gas sensors are generally low-cost sensors made of a single sensing material and are not characterized by an excellent selectivity with respect to different gases; for this reason, an embedded algorithm able to identify the gas type is the key feature to enable the adoption of the MOX technology in smart agriculture applications. This paper presents an AI method to enhance the selectivity feature of a MOX gas sensor. An artificial neural network (ANN) has been implemented and trained in Python environment, using different machine learning tools, such as Keras and scikit-learn. The trained ANN was able to recognize four types of gas detected by the embedded MOX gas sensor in lab conditions. By using X-CUBE-AI tool, the C-model implementation of the pre-trained ANN was generated and embedded in a low power STM32 microcontroller used in the smart air probe.