The International Energy Agency (IEA) established that heating and cooling needs represent 30% of world energy consumption and 26% of global 02 emission. Following the aim provided by the Net Zero Emission (NZE) international agreement for the year 2050, the demand for heat pumps is constantly growing. This entailed greater attention to the environmental impact of those systems with more strict regulation. In this context, predictive maintenance has become a fundamental strategy for monitoring the health state of a heat pump, preventing faults, improving reliability and limiting costs. Among all possible FDDE (Fault Detection Diagnosis and Evaluation) strategies, three stand out: black- white and gray box. Black box techniques are based on data and can be divided into supervised or non-supervised methods. Due to the lack of large labeled datasets related to malfunction conditions, this thesis work proposes the development of unsupervised data driven methods based on nominal operating data. The data are acquired from a single- circuit air/water heat pump with two on-off compressors and inverter fans, serving offices. The implemented digital models allow for the identification of anomalies by comparing the acquired data with the model predictions. In addition, to simulate failures and validate the models, tests were conducted by imposing an offset on selected sensors. The distinction between fault and new operating conditions, and consequently the management of the so called ”false positive” is one of the main challenges. Three different FDD approaches were implemented, which share a validation procedure and a data elaboration based on knowledge of the physical domain. Each algorithm has been produced in two distinct versions: one for winter and one for summer. The first one uses Principal Component Analysis, the second one Recurrent Neural Network, and the last one is based on Graph neural networks. All developed strategies are capable of recognizing whether the heat pump is working correctly and detecting anomalies, even if all have peculiar characteristics. In particular, the PCA-based method is characterized by computational simplicity and integrated control of operating conditions; however, it does not have high diagnostic capabilities. The RNN approach is more computationally burdensome than the preview approach, but offers precise predictions of the operating features, making it much easier to detect the failure. Although the GNN technique is the most computationally complex, it provides the best diagnostic predictions thanks to the relationships expressed by its graph structure. This thesis confirms the importance of the data validation process, which is fundamental to detecting potential inconsistencies that could compromise the reliability of a data-driven model. Ultimately, a generalizable workflow is identified that relies exclusively on data from sensors already present in the heat pump, eliminating the need for additional ones.
L'Agenzia Inìnternazionale per l'Energia (IEA) ha stabilito che il fabbisogno di riscaldamento e raffreddamento degli edifici rappresenta il 30% del consumo energetico globale e il 26% delle emissioni globali di CO2. In linea con gli obiettivi di decarbonizzazione previsti dall'accordo internazionale NZE (Net Zero Emission) per il 2050, la domanda di pompe di calore è in forte crescita. Questo ha portato a una maggiore attenzione sull'impatto ambientale si tali sistemi, con regolamentazioni più stringenti. Le nuove leggi F-gas (2024/573) vietano l'uso di gas sintetici ad alto GWP (Global Warming Potential) e promuovono l'utilizzo di gas naturali, i quali però comportano nuove sfide tecniche e di sicurezza. In questo contesto, la manutenzione predittiva emerge come strategia fondamentale per monitorare lo stato operativo delle pompe di calore e prevenire guasti, migliorando affidabilità e riducendo costi. Tra tutte le possibili metodologie utilizzate per il FDDE (Fault Detection,Diagnosis and Evaluation) si distinguono due macro-gruppi: quelle basate sulla fisica e quelle basate sui dati. Le tecniche “data-driven” si suddividono ulteriormente in metodi supervisionati e non supervisionati. Questo lavoro di tesi propone lo sviluppo di metodologie data-driven non supervisionate, basate su dati relativi al funzionamento nominale, acquisiti da una pompa di calore aria/acqua monocircuito, con due compressori on-off e ventilatori inverter, a servizio di uffici. Inoltre, sono state condotte delle prove imponendo un offset a sensori scelti, per simulare malfunzionamenti.Una delle principali sfide nella manutenzione predittiva riguarda la distinzione tra anomalia e nuova condizione operativa e, di conseguenza, la gestione di eventuali “falsi positivi”. Sono stati implementati tre approcci differenti per effettuare FD, che condividono una procedura di validazione dei dati e un'elaborazione basata sulla conoscenza del dominio fisico. Il primo sfrutta reti neurali ricorrenti (RNN), il secondo l’analisi delle componenti principali (PCA) e , infine, il terzo è incentrato su reti neurali a grafo.Tutte le strategie sviluppate riescono con successo a distinguere un funzionamento corretto della macchina da uno anomalo, pur presentando ciascuna aspetti peculiari. In particolare, il metodo basato su PCA e cluster è caratterizzato da semplicità computazionale ed effettua previsioni a parità di condizione operativa; tuttavia,le cause vengono individuate solo tramite un’analisi a posteriori. L’approccio RNN risulta invece computazionalmente più oneroso, ma offre previsioni per ogni grandezza acquisita, permettendo così di risalire più facilmente alla causa del malfunzionamento. La tecnica GDN si distingue per l’elevata complessità computazionale, ma offre la possibilità di effettuare una diagnosi esplicitando le dipendenze tra i parametri attraverso l’uso del grafo. Questa tesi conferma l'importanza del processo di validazione dei dati e quanto sia fondamentale evidenziare eventuali incoerenze, che comprometterebbero l’affidabilità di un modello “data-driven”. Infine, si individua una metodologia generalizzabile per la rilevazione di anomalie, basata esclusivamente sui dati relativi a sensori già presenti nella macchina, senza la necessità di doverne aggiungere alcuno.
Sviluppo di modelli di Machine Learning per la manutenzione predittiva di pompe di calore
DITURI, LUCA
2024/2025
Abstract
The International Energy Agency (IEA) established that heating and cooling needs represent 30% of world energy consumption and 26% of global 02 emission. Following the aim provided by the Net Zero Emission (NZE) international agreement for the year 2050, the demand for heat pumps is constantly growing. This entailed greater attention to the environmental impact of those systems with more strict regulation. In this context, predictive maintenance has become a fundamental strategy for monitoring the health state of a heat pump, preventing faults, improving reliability and limiting costs. Among all possible FDDE (Fault Detection Diagnosis and Evaluation) strategies, three stand out: black- white and gray box. Black box techniques are based on data and can be divided into supervised or non-supervised methods. Due to the lack of large labeled datasets related to malfunction conditions, this thesis work proposes the development of unsupervised data driven methods based on nominal operating data. The data are acquired from a single- circuit air/water heat pump with two on-off compressors and inverter fans, serving offices. The implemented digital models allow for the identification of anomalies by comparing the acquired data with the model predictions. In addition, to simulate failures and validate the models, tests were conducted by imposing an offset on selected sensors. The distinction between fault and new operating conditions, and consequently the management of the so called ”false positive” is one of the main challenges. Three different FDD approaches were implemented, which share a validation procedure and a data elaboration based on knowledge of the physical domain. Each algorithm has been produced in two distinct versions: one for winter and one for summer. The first one uses Principal Component Analysis, the second one Recurrent Neural Network, and the last one is based on Graph neural networks. All developed strategies are capable of recognizing whether the heat pump is working correctly and detecting anomalies, even if all have peculiar characteristics. In particular, the PCA-based method is characterized by computational simplicity and integrated control of operating conditions; however, it does not have high diagnostic capabilities. The RNN approach is more computationally burdensome than the preview approach, but offers precise predictions of the operating features, making it much easier to detect the failure. Although the GNN technique is the most computationally complex, it provides the best diagnostic predictions thanks to the relationships expressed by its graph structure. This thesis confirms the importance of the data validation process, which is fundamental to detecting potential inconsistencies that could compromise the reliability of a data-driven model. Ultimately, a generalizable workflow is identified that relies exclusively on data from sensors already present in the heat pump, eliminating the need for additional ones.| File | Dimensione | Formato | |
|---|---|---|---|
|
Dituri_Luca.pdf
Accesso riservato
Dimensione
12.72 MB
Formato
Adobe PDF
|
12.72 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/20.500.12608/94274