Sviluppo di un sensore virtuale basato su machine learning per l'analisi dinamica in tempo reale di un quadrilatero articolato con non linearità

Accurate force estimation is essential for ensuring efficiency, reliability, and safety in mechanical systems. While physical sensors provide direct measurements, their cost, maintenance, and installation challenges often make them impractical, especially in dynamic environments. This thesis introduces an AI-driven virtual sensor for real-time force estimation in a four-bar mechanism, addressing key non-linearities such as joint clearance, backlash, and damping. Traditional analytical and simulation-based methods, while effective, struggle with computational intensity and adaptability, making machine learning a compelling alternative. The research begins with modeling and simulating the four-bar mechanism in MATLAB’s Simscape Multibody, capturing realistic dynamic behaviors under varying conditions. The impact of camera frame rate, the number of captured frames, and maximum velocity on prediction accuracy is thoroughly examined, followed by an in-depth analysis of clearance, backlash, and damping effects, bridging the gap between idealized simulations and real-world mechanical imperfections. A machine learning-based virtual sensor is then developed to predict reaction forces with high precision, trained on simulation-derived datasets that undergo feature engineering and normalization. The study evaluates multiple feature selection strategies, revealing that a hybrid approach combining angular differences (velocity-related features) with absolute angular positions yields the most accurate force predictions. The neural network model, fine-tuned through Bayesian hyperparameter optimization, achieves exceptional accuracy under controlled conditions. However, as crank velocity increases and non-linearities become more pronounced, predictive performance declines, exposing the limitations of traditional neural networks in handling out-of-distribution scenarios. This research highlights the need for advanced data augmentation techniques, physics-informed machine learning, and adaptive architectures to enhance predictive robustness in real-world applications. By integrating AI with physics-based modeling, this thesis presents a scalable and efficient framework for intelligent virtual sensing, paving the way for more adaptive and cost-effective force estimation solutions in engineering.