AN EDGE-AI PLATFORM FOR ENHANCED DECISION-MAKING IN AGRI-FOOD SUPPLY CHAINS: A PATHWAY TO INDUSTRY 5.0

Abstract

Against the backdrop of the growing link between food waste and climate change, the European Commission (EC) introduced the Industry 5.0 model, aiming to meet the Sustainable Development Goals (SDGs) by promoting sustainable, human-centric and resilient systems. Central to SDGs fulfillment, technology becomes a valuable means for supporting informed decision-making in Agri-Food Supply Chain (AFSC)s. In particular, harnessing real-time data through Edge Intelligence (Edge-AI) with its timely streaming and diminished latency, makes responsive decision-making within immediate reach. From here, this multidisciplinary thesis aims to bridge these modern AFSCs challenges with Edge-AI technology by proposing an innovative decision-making platform. This tool has the potential to enable stakeholders to manage their resources efficiently and intuitively by streamlining visual information on current food demand by identifying two indices: classes and On-Shelf Availability (OSA). This platform is built with bi-objective, lightweight models where 1) agri-food is determined using a custom Convolutional Neural Network (CNN) classifier and 2) OSA is modeled through density maps, where counts are predicted using spatially-aware regression via a Multi-Column Convolutional Neural Network (MCNN). The first attains an accuracy of up to 98.52% in identifying seven agri-food products, while the latter estimates OSA with a Mean Average Error (MAE) of 4.27. To demonstrate the practical value of this proposal, this thesis conducts an empirical validation study by designing a pilot shelf prototype and an interactive decision-making interface through a simulated retail trial. This study provides critical insights into the system's real-life performance, where Jetson Orin Nano prevails with superior inference speed and overall performance. To this end, the synergy of agri-food identification and its associated OSA levels will enable reformulating policies based on real-time demand to reduce food waste and ensure balanced food distribution among retailers. Furthermore, the availability of such data is valuable to be integrated into simulation or digital twins to generate different risk scenarios and asses logistical vulnerabilities under emergencies or uncertainties in anticipation of Industry 5.0 values.