TY  - THES
A3  - Obermaisser, Roman
AB  - The building sector and its embedded control systems, especially the Heating, Ventilation, and Air-Conditioning (HVAC) systems, consume a considerable part of the global energy and produce gaseous emissions such as CO2. On the other hand, the air exchange based on natural ventilation is a cost-efficient method to improve indoor air quality, dilute indoor CO2concentration and odors, or remove pollutants or airborne virus particles (e.g., Covid-19) from the building zones. This air exchange during the cold seasons accounts for a heating load for the heating system that causes an increase in energy consumption. Therefore, optimization of HVAC systems to decrease harmful emissions considering potential energy saving is vital. Moreover, if the CO2 generated by human metabolism is not correctly controlled to some limits, it can degrade indoor air quality, reduce the occupants’ efficiency, lead to severe mental problems, or considerably impair the thinking ability. Thus, implementing a robust ventilation control system for the buildings particularly crowded office buildings is momentous. Demand-Controlled Ventilation (DCV) systems are promising solutions that control and optimize the ventilation rates based on thermal comfort and indoor air quality demands with a high potential in energy saving. Many researchers in the literature study DCV systems or adaptive thermal control separately while a comprehensive model containing both DCV and thermal control strategies is missing. Therefore, this thesis contains the combination of the DCV and heating systems with embedded sensors and actuators with the fault injection capabilities in a simulation framework to study such a complex system due to its numerous functions, inputs, and outputs for an in-depth assessment of the involved components’ functionality and effective parameters, especially in case of component failures. Indoor air quality and comfort parameters in an office building can be monitored and controlled in real-time for various architectures based on a high-level specification of the building characteristics. The developed model is scalable based on the modular composability scheme. The user can generate different types of buildings with various architectures with many rooms and floors. The system model, fault injection capabilities, and diagnostic modules are automatically extended.
The high complexity of the DCV and heating systems with their many components makes them error-prone, more susceptible to faults, and more fragile. Faults in system components such as sensors and actuators can result in different types of failures and severe implications on efficiency with discomfort and performance degradation of occupants, energy waste, shortened component lifetime, and increased maintenance costs. Failure detection and fault diagnosis is the combination of system failure detection, which is the implication of the fault in a component of a system, with fault diagnosis that is finding the type, severity, time of occurrence, and locality of faults. The state-of-the-art of fault diagnosis methods for building energy systems, e.g., HVAC systems, contains data-driven and knowledge-driven diagnostic methods with corresponding strengths and shortcomings. The knowledge-driven methods are mainly based on expert knowledge and simulate the diagnostic thoughts of domain experts with the argumentation of uncertainties, diagnosis of different fault severities, and understandability. But they need a higher and time-consuming effort to deeply understand the causal relationships among system inputs, faults, and symptoms. Moreover, the knowledge-based methods still lack automatic strategies to improve efficiency and they are less accurate than the data-driven methods. The data-driven methods, on the other hand, depend on similarities and patterns with high sensitivity to any change of patterns and more accuracy than the knowledge-driven methods. However, the data-driven methods require a huge amount of data for training the neural network for fault classification and they cannot provide the reason behind the results. In addition, the data-driven strategies indicate black boxes with low understandability.
The research gap filled by this thesis is therefore the combination of knowledge-driven and data-driven fault diagnosis in DCV and heating systems to gain advantages from both categories. The diagnostic method presented in this thesis involves an automatic strategy with low expert effort without necessitation of in-depth understanding of the causal relationships compared to existing knowledge-driven methods with high understandability and high accuracy compared to the existing data-driven methods. The fault diagnosis method in this thesis combines a data-driven classifier with knowledge-driven inference, e.g., fuzzy logic and a Bayesian Belief Network (BBN) to provide an automatic diagnostic classifier that can diagnose any stuck-at or constant-valued faults in sensors and actuators. The combination of BBN and fuzzy logic itself analyzes the dependencies of the system signals based on the mutual information theory. In offline mode, a Relation-Direction Probability (RDP) table for each fault class is computed and stored in an offline fault library. The online mode determines the similarities between the real case RDP in the runtime and the offline library’s RDPs. On the other hand, a data-driven strategy is specifically established using deep neural networks to compare and evaluate the performance of the presented composed diagnostic classifier. The data-driven classifier uses observed signals from faulty and healthy operations of the system to train and evaluate the performance of the designed neural network model. The diagnostic technique in this thesis is independent of the historical data, independent of the expert knowledge, and computing-resource efficient. For the evaluation, four types of stuck-at faults at different components such as temperature sensor, CO2 sensor, heater actuator, and damper actuator with various fault values at different instants of time were investigated. A fault injection framework artificially injects the faults to serve the diagnostic classifiers, e.g., training the models and evaluations. Results show the combined classifier introduced in this thesis has comparable performance to the data-driven method while advantaging the strengths of knowledge-driven methods.
AU  - Behravan, Ali
DA  - 2021
DO  - 10.25819/ubsi/10075
KW  - Deep learning
KW  - Diagnostic Classifiers
KW  - Fault Diagnosis
KW  - Fuzzy Bayesian Belief Network
KW  - Demand Controlled Ventilation
KW  - Deep Neural Network
LA  - eng
PY  - 2021
TI  - Diagnostic classifiers based on fuzzy Bayesian belief networks and deep neural networks for demand-controlled ventilation and heating systems
TT  - Diagnoseklassifikatoren auf Basis von Fuzzy Bayesian Belief Networks und Deep Neural Networks für bedarfsgesteuerte Lüftungs- und Heizungsanlagen
UR  - https://nbn-resolving.org/urn:nbn:de:hbz:467-21541
Y2  - 2024-11-22T04:12:19
ER  -