Heating, ventilation, air conditioning and refrigeration (HVACR) systems account for a large share of building energy use, making them a prime target for efficiency improvements and emissions reduction. Within HVACR, heat pumps play a central role in both residential and commercial contexts and are increasingly integrated into smart-building infrastructures that rely on dense sensor networks and automated control. In this setting, robust fault detection, diagnosis, and evaluation (FDDE) methods are essential to maintain comfort while minimizing energy consumption and operational risk. A key challenge for heat pumps is the timely identification of soft faults (such as refrigerant leakages and heat-exchanger fouling) that do not trigger an immediate shutdown but progressively degrade performance. These faults can silently reduce cooling/heating capacity and coefficient of performance (COP), leading to increased energy consumption and accelerated wear. Physics-based digital models have been used to simulate the impact of such faults and to quantify the diagnosable fault space under realistic sensor uncertainty, showing that commonly available instrumentation can detect performance penalties as small as 5%. FDDE approaches span physics-based modelling, data-driven learning, and hybrids that combine both. Data-driven methods ranging from \(k\)-nearest neighbours (KNN) and support vector machines (SVM) to artificial neural networks (ANN) and gradient-boosting ensembles have shown promising results when sufficient, representative data are available. In particular, gradient-boosting methods such as XGBoost provide strong predictive and classification performance with useful model diagnostics (e.g., feature importance), and hybrid schemes that couple clustering with boosting have achieved high detection and diagnosis accuracy on chiller applications while demonstrating good generalization. Despite substantial progress, practical deployment in commercial buildings remains limited by system heterogeneity, varying operating conditions, and transferability concerns. This thesis situates itself in this landscape and contributes along two axes: (i) a data-driven FDDE pipeline, and (ii) a Simulink-based modelling framework to generate and structure experiments, including controlled fault scenarios. The objectives are to assess soft-fault detectability under realistic sensing constraints and discuss the conditions under which such models can be transferred or adapted to new systems.
I sistemi di riscaldamento, ventilazione, condizionamento e refrigerazione (HVACR) rappresentano una quota significativa dei consumi energetici degli edifici, rendendoli un obiettivo prioritario per miglioramenti dell’efficienza e la riduzione delle emissioni. All’interno dell’ambito HVACR, le pompe di calore svolgono un ruolo centrale sia in contesti residenziali che commerciali e sono sempre più integrate in infrastrutture di smart building basate su reti dense di sensori e controlli automatizzati. In questo scenario, metodi robusti di rilevamento, diagnosi e valutazione dei guasti (FDDE) sono essenziali per mantenere il comfort riducendo al minimo il consumo energetico e i rischi operativi. Una delle principali sfide per le pompe di calore consiste nell’identificazione tempestiva dei guasti “soft” (come perdite di refrigerante o sporcamento degli scambiatori di calore), che non provocano un arresto immediato del sistema ma ne degradano progressivamente le prestazioni. Questi guasti possono ridurre in modo silenzioso la capacità di raffreddamento/riscaldamento e il coefficiente di prestazione (COP), portando a un aumento dei consumi e a un’accelerazione dell’usura. Modelli digitali basati sulla fisica sono stati impiegati per simulare l’impatto di tali guasti e per quantificare lo spazio dei guasti diagnosticabili in condizioni realistiche di incertezza dei sensori, mostrando che l’instrumentazione comunemente disponibile può rilevare penalizzazioni prestazionali anche dell’ordine del 5%. Gli approcci FDDE spaziano dalla modellazione fisica all’apprendimento basato sui dati, fino a metodi ibridi che combinano i due paradigmi. I metodi data-driven, che includono k-nearest neighbours (KNN), support vector machines (SVM), reti neurali artificiali (ANN) e modelli di boosting, hanno mostrato risultati promettenti quando sono disponibili dati sufficienti e rappresentativi. In particolare, i metodi di gradient boosting come XGBoost offrono prestazioni robuste in termini di previsione e classificazione, insieme a strumenti diagnostici utili (ad esempio, l’importanza delle caratteristiche). Schemi ibridi che combinano clustering e boosting hanno raggiunto elevata accuratezza di rilevamento e diagnosi in applicazioni con chiller, dimostrando anche buona capacità di generalizzazione. Nonostante i notevoli progressi, l’adozione pratica negli edifici commerciali rimane limitata a causa dell’eterogeneità dei sistemi, delle diverse condizioni operative e di problematiche legate alla trasferibilità dei modelli. Questa tesi si colloca in questo contesto e contribuisce lungo due direzioni: (i) lo sviluppo di una pipeline FDDE basata sui dati e (ii) la creazione di un framework di modellazione basato su Simulink per generare e strutturare esperimenti, inclusi scenari di guasto controllati. Gli obiettivi sono valutare la rilevabilità dei guasti soft in condizioni realistiche di misura e discutere le condizioni sotto le quali tali modelli possono essere trasferiti o adattati a nuovi sistemi.
Integration of Machine Learning Algorithms and Simscape Modeling for Fault Detection, Diagnosis, and Evaluation in Heat Pump Systems
DE MARCHI, FRANCESCO
2024/2025
Abstract
Heating, ventilation, air conditioning and refrigeration (HVACR) systems account for a large share of building energy use, making them a prime target for efficiency improvements and emissions reduction. Within HVACR, heat pumps play a central role in both residential and commercial contexts and are increasingly integrated into smart-building infrastructures that rely on dense sensor networks and automated control. In this setting, robust fault detection, diagnosis, and evaluation (FDDE) methods are essential to maintain comfort while minimizing energy consumption and operational risk. A key challenge for heat pumps is the timely identification of soft faults (such as refrigerant leakages and heat-exchanger fouling) that do not trigger an immediate shutdown but progressively degrade performance. These faults can silently reduce cooling/heating capacity and coefficient of performance (COP), leading to increased energy consumption and accelerated wear. Physics-based digital models have been used to simulate the impact of such faults and to quantify the diagnosable fault space under realistic sensor uncertainty, showing that commonly available instrumentation can detect performance penalties as small as 5%. FDDE approaches span physics-based modelling, data-driven learning, and hybrids that combine both. Data-driven methods ranging from \(k\)-nearest neighbours (KNN) and support vector machines (SVM) to artificial neural networks (ANN) and gradient-boosting ensembles have shown promising results when sufficient, representative data are available. In particular, gradient-boosting methods such as XGBoost provide strong predictive and classification performance with useful model diagnostics (e.g., feature importance), and hybrid schemes that couple clustering with boosting have achieved high detection and diagnosis accuracy on chiller applications while demonstrating good generalization. Despite substantial progress, practical deployment in commercial buildings remains limited by system heterogeneity, varying operating conditions, and transferability concerns. This thesis situates itself in this landscape and contributes along two axes: (i) a data-driven FDDE pipeline, and (ii) a Simulink-based modelling framework to generate and structure experiments, including controlled fault scenarios. The objectives are to assess soft-fault detectability under realistic sensing constraints and discuss the conditions under which such models can be transferred or adapted to new systems.| File | Dimensione | Formato | |
|---|---|---|---|
|
DeMarchi_Francesco.pdf
Accesso riservato
Dimensione
13.35 MB
Formato
Adobe PDF
|
13.35 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/99554