Monitoraggio Intelligente dei Macchinari Agricoli per una Maggiore Efficienza e Sostenibilità

Ensuring both the productivity and sustainability of agriculture is one of the defining challenges of our time. Global food demand is rising, while natural resources are under increasing pressure. Within this context, agricultural mechanisation plays a dual role: it is a powerful driver of productivity, but also a significant contributor to fuel consumption, greenhouse gas emissions, and resource inefficiency. Improving the monitoring and management of machinery is therefore a key step towards reducing environmental impact and increasing farm profitability. This thesis explores how smart telemetry systems and data-driven methods can be used to monitor, classify, and evaluate the performance of agricultural machinery. By focusing on data streams already produced by modern tractors, through the Controller Area Network (CAN-bus) and Global Navigation Satellite Systems (GNSS), and applying machine learning algorithms, this work demonstrates new ways to extract actionable knowledge from raw machine data. The research contributes to the development of Digital Agriculture, where digital technologies, robotics, and analytics enhance both efficiency and sustainability. The first research strand investigates the use of CAN-bus data for tractor state classification during ploughing operations. Modern machines record hundreds of parameters in real time, such as engine speed, torque, fuel rate, and wheel speed. A Random Forest classifier was trained on labelled datasets to identify key operational states: idle, moving, turning, and working, with high accuracy. Importantly, the analysis showed that even a small subset of features (engine speed, fuel rate, wheel speed) can provide reliable classification, opening the possibility of developing simplified monitoring systems that remain robust across different brands and models. This approach also enabled detailed analysis of fuel consumption patterns, offering farmers a new way to identify inefficiencies and reduce costs. The second strand focuses on a minimal-input GNSS-only approach. Many farms, especially small and medium-scale ones, lack access to proprietary CAN-bus data or advanced telemetry platforms. To address this gap, the thesis tested a workflow that relies exclusively on GNSS coordinates and derived kinematic features. A Random Forest model enriched with trajectory and heading data achieved around 90% accuracy in classifying operational states. Although accuracy dropped slightly in unseen conditions (74%), the results confirm that even low-cost GNSS loggers or smartphone GPS modules can deliver meaningful insights into machine performance and field efficiency. This makes the method particularly valuable in contexts where affordability and accessibility are critical. The third strand applies telemetry to autonomous agricultural robots, comparing their performance in maize sowing and headland turning with conventional tractors. Results highlighted both the potential efficiency gains of robotic operations and the current limitations, such as manoeuvring complexity at field boundaries. These findings provide an evidence base for assessing the readiness of robotic technologies and guiding their integration into farm operations. Together, these three investigations illustrate the versatility of telemetry for improving decision-making in mechanised agriculture. CAN-bus analysis enables high-resolution monitoring of modern fleets; GNSS-only approaches democratise access to performance insights for smaller farms; and robotic telemetry supports the transition towards autonomous systems. Across all cases, the combination of machine learning and real-world field data proves essential to convert raw signals into usable information.