Data-Driven Bayesian Modeling Using Approximate Bayesian Computation for Heat Exchanger Monitoring

• Grant number: 25/07225-5
• Support Opportunities: Scholarships abroad - Research Internship - Scientific Initiation
• Start date: July 20, 2025
• End date: August 19, 2025
• Field of knowledge: Engineering - Mechanical Engineering
• Principal Investigator: Samuel da Silva
• Grantee: Vitória Batista Godoy
• Supervisor: Francesco Coletti
• Host Institution: Faculdade de Engenharia (FEIS). Universidade Estadual Paulista (UNESP). Campus de Ilha Solteira. Ilha Solteira , SP, Brazil
• Institution abroad: Brunel University, England
• Associated to the scholarship: 23/13794-7 - Data-Driven Bayesian Modeling for Heat Exchanger Monitoring, BP.IC

Abstract

Fouling is one of the main factors degrading the performance of thermal systems, particularly boilers, heat exchangers, and piping networks. Monitoring such effects is typically carried out using plant control signals, with models developed through physics-based approaches or data-driven techniques. While the application of artificial intelligence has been extensively explored in the literature, implementing effective monitoring solutions in industrial settings remains limited. The main challenges are the lack of commissioning data (i.e., data from clean equipment), poorly calibrated sensors, inadequate sampling rates and data storage, and significant inherent uncertainties in these systems. These practical limitations hinder the development of robust monitoring tools. Thus, the current project proposes using a Bayesian approach to quantify uncertainties associated with monitoring heat exchangers. The central idea is to work directly with the available data from industrial control systems-even if incomplete or noisy-to calibrate and identify a reduced-order model capable of functioning as a digital twin or, more precisely, a digital shadow. This "shadow model" would replicate the essential behavior and performance of the equipment and enable the prediction of key output signals. Relevant features can be extracted, and the equipment's current condition can be classified using a data-oriented Bayesian framework. With probabilistic confidence, the model can then estimate whether the equipment is operating in a clean state or under a level of fouling that significantly compromises performance. This 7-week BEPE internship, hosted at Brunel University London, aims to apply the proposed methodology using real industrial data provided by Hexxcell, a UK-based company specializing in advanced monitoring solutions for heat exchangers. This plan outlines the expected contributions and original results of the research, the collaborative exchange between the research group at UNESP and the teams at Hexxcell and Brunel University, and the project's benefits for the scientific development of the scholarship holder, Vitória Godoy. (AU)