Development of a data-driven fault detection and diagnosis algorithm for a residential propane-based heat pump

Abstract

[ES] Este trabajo compara tres algoritmos de aprendizaje automático supervisado, entrenados para la detección y diagnóstico de fallos (FDD) de una bomba de calor aire-agua residencial que funciona en modo calefacción. Los algoritmos considerados en esta tesis son redes neuronales artificiales, árboles de decisión y derivaciones, junto con máquinas de vectores de soporte. Con el propósito de superar la limitante de una compleja instrumentación de la bomba de calor para la extracción de datos, se propone un conjunto de medidas de temperatura de bajo coste. El primer objetivo de este trabajo es explorar el rendimiento alcanzable por un modelo con solo mediciones de temperatura. Para generar los datos de entrenamiento, se desarrolla un modelo de simulación en Modelica, mediante componentes de la librería Dymola, tomando como referencia una bomba de calor existente. Dentro del espectro de fallos comunes en las bombas de calor, se seleccionan el ensuciamiento del evaporador y la fuga de refrigerante. Las estrategias de modelado de fallos implican la disminución de la masa de refrigerante para el caso de fuga de refrigerante, del mismo modo, se reduce la velocidad del ventilador y el coeficiente global de transferencia de calor para emular el ensuciamiento. Las simulaciones solo tienen en cuenta estados estacionarios, sin embargo, se introducen variaciones a través de los datos meteorológicos para las simulaciones de años típicos. Los conjuntos de datos simulados se dividen en entrenamiento y prueba, extrayendo fracciones representativas del periodo invernal del primer y el último trimestre del año. El rendimiento de los algoritmos seleccionados se somete a pruebas exhaustivas, primero con los datos en estado estacionario y después con los datos del año típico, utilizando bibliotecas de optimización automatizadas para encontrar los mejores resultados. Además, los algoritmos también se entrenan con enfoques de regresión y la clasificación para explorar, dentro de los algoritmos con mejores resultados, la estrategia más adecuada para la detección y el diagnóstico de fallos en la bomba de calor.

[EN] In the context of the energy trilemma (equity, sustainability, and security) the reliance on natural gas is experiencing important drawbacks. Climate change pledges call for electrification, while the second largest consumer of this fossil fuel is the residential sector. An alternative to traditional gas boilers are heat pumps. However, the operation of heat pumps is threatened by faults, which can hinder performance, diminish efficiency, increase operation costs, and reduce equipment’s lifespan. Still, literature shows small deployment of FDD methods in the residential sector, mostly due to the high costs of a comprehensive sensor array, absence of data, wide range of models, and variability of installations. Within FDD classification, datadriven techniques, based on machine learning algorithms, stand out for their versatility and ease of deployment. These methods do not require complex models, can handle noisy data, and a reduced set of inputs. This thesis compares three supervised machine learning algorithms, trained for the fault detection and diagnosis (FDD) of a residential-sized air-to-water heat pump operating in heating mode. The algorithms considered in this thesis are artificial neural networks, decision trees and ensembles, and supported vector machines. To address the obstacle of the sensor array, a set of inexpensive temperature measurements is proposed. The first objective of this work is to explore the performance attainable by a model with only these inputs. Then, considering that real household data is scarce and comprehensive experimental trials represent a major effort, a simulation model is developed. Data is generated using Modelica Language and Dymola software, taking an existing heat pump as reference. Across the spectre of common soft faults, evaporator fouling and refrigerant leakage are selected. The fault modeling strategies involve decreasing the refrigerant mass for leakage or undercharge, similarly, the fan speed and overall heat transfer coefficient are reduced to emulate fouling. The simulations only consider steady-state, however, variance is introduced through weather data in typical year simulations. Next, simulated datasets are split in training and test sets, extracting representative fractions of the winter period from the first and last trimester of the year. The aforementioned temperature measurements are interpreted as features by the machine learning algorithms. As a result, the detection of evaporator fouling was successfully achieved through classification-based and regression-based strategies, which are developed and compared. Across all stages, neural networks exhibit the best scores, with the regression algorithms outperforming the classifiers. The resulting algorithm showed in the final evaluation with a 71% correct rate, 29% false alarm, and 0% missed detection rate.