Transfer Learning Among Data-Driven Models for Heat Exchangers Monitoring

• Grant number: 24/00140-1
• Support Opportunities: Scholarships in Brazil - Master
• Start date: April 01, 2024
• End date: July 31, 2025
• Field of knowledge: Interdisciplinary Subjects
• Principal Investigator: Samuel da Silva
• Grantee: Estênio Fuzaro de Almeida
• Host Institution: Faculdade de Engenharia (FEIS). Universidade Estadual Paulista (UNESP). Campus de Ilha Solteira. Ilha Solteira , SP, Brazil

Abstract

The continuous monitoring of heat exchanger health in the field is crucial to ensure optimal performance and safety in industrial plant operations. In this context, various state diagnostic techniques are available in the literature, especially those utilizing information from installed sensors for control actions involving variables such as pressure, flow, and temperature. However, for more accurate diagnosis, it is essential to employ a mathematical model that simulates the heat exchanger's dynamics. In practice, these models are often derived from system identification techniques, as phenomenological modeling can be prohibitively complex, requiring extensive time and a profound knowledge of the physical characteristics of the equipment that field maintenance engineers may not possess. Nonetheless, few methods prove universally effective and sufficiently generalist for a comprehensive solution to this technological challenge. One of the identified reasons for this challenge is the lack of widespread availability of historical data on the state of heat exchangers, essential for the efficient training of fault detection algorithms in these systems. Often, these datasets are incomplete, containing nearly static information with low sampling rates and are frequently scarce. In this regard, this project contributes by investigating and proposing the use of advanced transfer learning tools directly applied between data-driven models for different types of heat exchangers. The central idea is to consider a source situation with detailed information and histories of a standard heat exchanger, utilizing pressure, flow, and/or temperature variables to construct a data-driven model, such as a simple autoregressive model. Subsequently, machine learning algorithms will be employed to train and validate classifiers and regressors that provide information enabling the diagnosis of the heat exchanger's operational state. Later, sparser data from a target heat exchanger, exhibiting some discrepancies, will be used to examine how mapping and transfer learning techniques can be implemented, transferring knowledge from the source situation to the target situation. The objective is to verify, through numerical and experimental simulations, the conditions of similarity and circumstances under which it is possible to leverage data history and transfer knowledge between fault detection algorithms for heat exchangers, whether similar or completely distinct. The results of this dissertation have the potential to introduce new monitoring techniques aiming to ensure safer and more economical operation of thermal systems.