Fault diagnosis services and realistic fault models for HVAC systems

Heating, Ventilation, and Air-Conditioning (HVAC) systems are large-scale distributed systems comprising distributed components, including controllers, sensors, and actuators that must be coordinated to establish the intended behavior. Therefore, HVAC systems are subject to single and multiple faults affecting the electronics, potentially causing high energy consumption, occupant discomfort, degraded indoor air quality, thermal conditions, and risk to critical infrastructures. In addition, in large-scale critical infrastructures, HVAC systems serve an essential role in emergencies. Emergency reactions demand realtime response, consistency, and fault tolerance. Fault tolerance is essential for both operational faults and design faults. In the development phase of fault-tolerant systems, simulation is a common technique to obtain insights into system functionality, performance, and dependability. It saves time, reduces cost and avoids risks of carrying out tests in the presence of faults in real-world systems. As a result, fault injection in simulation environments is an effective experimental method to validate and evaluate the dependability of HVAC systems. Fault injection in a simulation offers high controllability and observability. It is thus ideal for an early dependability analysis and fault-tolerance evaluation. HVAC systems in critical infrastructures are safety-relevant systems that should guarantee adequate ventilation and air conditions for occupants.
Accordingly, in this thesis, a simulation-based fault injection framework with a combination of two techniques, simulator command and simulation code modification with realistic fault patterns is proposed and introduced as a generic and extendable framework. The fault-injection framework is integrated and connected to simulation models of other electronic components via the connection of ports. The fault injection framework is developed in a component-based structure, implemented and simulated in MATLAB/Simulink using Stateflow diagrams with healthy and faulty system states. To determine the fault attributes and the fault location, an automated fault injection algorithm is proposed and integrated with a system-model generation algorithm. The system structure is adaptable and its parameters such as the number of floors and the number of rooms on each floor are defined based on the system requirements. An automated single/multiple fault injection algorithm triggers faults and supports a comprehensive range of faults with corresponding fault attributes including the fault type, time, location, persistence, duration, interarrival time and occurrence incidence. To validate the fault injection framework, a scenario-based approach is used to study the system impact and quality of the services. Each scenario consists of multiple events and subevents that result in multiple fault injections. The fault injection framework considers a realistic fault model adding white noise with Gaussian distribution as signal uncertainties and it supports reproducibility for a set of specific fault scenarios and for random fault injection scenarios. The framework incorporates a multi-dimensional fault model and provides compatibility to a wide range of other simulation components. The experimental results of single and multiple fault injection components show the correctness, the system behavior, accuracy, and other system parameters, such as the heater energy consumption and heater duty cycle in the presence of different fault cases. The experimental results serve as a quantitative evaluation of key performance indicators (KPI) such as energy efficiency, air quality, and thermal comfort. For example, combining a CO2 sensor fault with a heater actuator fault impacts energy consumption significantly by more than 70%.
Furthermore, in this thesis a novel and generic fault diagnostic technique based on the Fuzzy Bayesian Belief Network (FBBN) construction is integrated with a simulated system model as a monitoring approach to determine the causes of faulty operations based on system observations and measurements. A data-driven classifier algorithm is also proposed to be combined with knowledge-driven methods, including fuzzy theory and Bayesian belief networks, enabling accurate fault diagnosis in HVAC systems. In this thesis, the data-driven approach reduces time consumption through automation and classification based on automated ranking methods. The fuzzy theory relies on reasoning about the uncertainties and divides the system attributes into several subdomains to facilitate the probability calculations for continuous system attributes via proper likelihood membership functions based on the system specifications. The probabilities are used to construct the Bayesian belief network based on the correlations of the fuzzified system attributes using mutual information theory. Mutual information for all pairs of fuzzified subdomains must be calculated and a positive value of the mutual information is an indicator of a strong dependency between two subdomains.
Eventually, fault injection supports the fault diagnosis technique to define different fault cases and produce the faulty output data as a time series, including all healthy and faulty system measurements. The FBBN algorithm specifies the stringent relations, direction, and probability features of all fuzzified subdomains using the produced time-series by injecting the different fault cases. The hybrid fault diagnostic technique uses a data-driven classifier in combination with fuzzy logic theory and a Bayesian Belief Network in offline and online modes. Offline mode trains an offline library based on relation-direction-probability relationships of subdomains. Online mode determined the most similar faults in the offline library with actual fault cases based on the correlation of system attributes and the ranking method. The results show high accuracy of diagnosing permanent stuck-at fault in different HVAC system components.