This research contributes to the vision of green and sustainable buildings equipped with cyber-physical substrata consisting of heating, ventilation and air conditioning (HVAC) modules, networked sensors providing information on spatial and temporal distribution of occupants, smart building management systems providing situation awareness and decision support to human operators, and improved tenant comfort.  This is achieved by devising simple, robust, generic, and scalable fault detection, diagnosis and prognosis (FDDP) of HVAC systems, while systematically considering complaints from occupants, to substantially reduce energy consumption, improve the quality of living, and reduce the logistic costs. 

      The HVAC system is a large scale cyber-physical system.  HVAC constitutes 57% of the energy used in commercial and residential buildings in US. Failure to operate the HVAC system can be due to an issue in a single subsystem or issues coupled among multiple subsystems. The number and placement of sensors is vital to fault detection and isolation capacity. The system works under uncertain weather, occupancy conditions and parameter setting is critical. All of this make the FDDP in HVAC system challenging task. The work we have done provides a novel approach for FDDP in HVAC systems. Following are the highlights of our contributions.

a. Augmented Reality-based Troubleshooting of HVAC Systems:

      Major outcomes from this effort included (1) Testability of HVAC systems should be improved significantly; (2) Sensor optimization is salient; (3) It is reasonably straightforward to create a multi-functional model of the system; (4) It is possible to remotely monitor HVAC systems using an integrated diagnostics toolset; and (5) Integration of AR technologies can redefine traditional maintenance strategies.

b. Modeling Imperfect Tests for fast and Accurate Maximum A-Posteriori (MAP) Inference

Our work on a unified test model that includes the Detection-False Alarm (DFA), the Leaky Noisy OR (LNOR) and the logistic regression (LR) model as special cases provides a general framework for accurate MAP inference in bi-partite Bayesian networks used widely for fault diagnosis.

c. A Fault Diagnosis Method for HVAC Air Handling Units Considering Fault Propagation

      A new Statistical Process Control (SPC) rule is developed to detect state transitions with low false alarm rates by considering a fault property and couplings among components.  It represents a new and effective way to diagnose dependent faults accurately and robustly.

d. Chiller Plant Fault Diagnosis Considering Fault Propagation

      Our method uses model parameters relating to faults to supplement fault indicators to provide fault distinguishing information, leading to low false alarm rates. Additionally, it integrates Coupled Hidden Markov Models (CHMM) with the Extended Kalman Filter (EKF) to identify both the failure modes and their severities.  It represents a new and effective way to accurately diagnose faults in a computationally efficient manner.

e. Fault Diagnosis Framework for Air Handling Units based on the Integration of Dependency Matrices and PCA

      Since the model parameters are considered as the potential fault indicators in our method, the faults can be diagnosed even though they occurred at the same time.  Additionally, PCA can obtain appropriate fault indicators and their normal ranges for fault diagnosis.

f. Fault Detection of Cooling Coils based on Unscented Kalman Filters and SPC

      Our method was obtained by extending a robust, replicable and scalable method to the nonlinear case.  This method was motivated by the problem of detecting faults of cooling coils, but it also applies to other systems with nonlinear models.

g. Data Analytics for Chiller System Fault Diagnosis and Prognosis:

Our work on establishing the health indicators from sensor data and tracking variation in these indicators to predict fault propagation over time has applications far beyond the chiller application considered. This work enabled us to isolate faults and estimate the severity via Deep learning techniques even when the fault is at an initial stage.

h. Highly Accurate Thermal Comfort Models

Human thermal sensation is an important factor in creating a “comfortable” work/living environment.   Our thermal comfort model achieved an accuracy of 82 − 85%, which is over twice that of the widely adopted Fanger’s model on a publicly available dataset. We believe that identifying factors that directly influence the thermal sensation of individuals, especially senior citizens, and using them to predict human thermal comfort in real-time will have enormous societal benefit.

This project helped in establishing state of the art education and training opportunities for students. The research and training opportunities in FDDP of HVAC system helped students in developing a strong understanding of an HVAC system, its functioning and modeling, thermal comfort models, sensors and fault propagations.  The overall experience provided a broader positive impact on their professional growth.  The FDDP algorithms developed are capable of detecting trends and degradations, and assessing the severity of hardware failures which, in turn, serve as an early indicator to the facility operators.  This will help them plan, prepare and prevent failures, reduce downtime and improve the thermal comfort in buildings.  

Last Modified: 10/02/2017
Modified by: Krishna R Pattipati