Condition monitoring of an asphalt production plant through physical parameter estimation based on nonlinear neural models

This thesis was developed in collaboration with Marini – Fayat Group and the ACTEMA research group of the University of Bologna. The work focuses on the development of diagnostic tools for an asphalt production plant, with particular attention to the filtration subsystem. The objective of the study is the estimation of two physical parameters: the filter fluidic resistance and the drum resistance. Variations in these parameters may indicate changes in operating conditions or the progressive degradation of plant components, making their estimation a valuable tool for condition monitoring. The thesis investigates two different estimation approaches: the first method is based on a linearized representation of the system. The plant model is linearized around an equilibrium point, and the system parameters are identified using the Recursive Least Squares algorithm. The identified transfer function is then used to extract frequency-domain features, specifically the Bode magnitude values, which are mapped to the two parameters of interest through a neural network capable of estimating both resistances simultaneously. The second approach addresses the problem directly in the nonlinear domain. In this case, a feedforward neural network is trained to estimate the filter resistance from NLARX-type regressors built from past samples of the system input and output signals. This estimator is then integrated within a Simulink model to evaluate its performance in a realistic simulation environment. The results show that neural-network-based estimators can effectively infer parameters that are not directly measurable, providing a promising solution for the implementation of condition monitoring strategies in industrial asphalt plants.