Digital Twin Enabled Facilities Maintenance Management through the Integration of Artificial Intelligence and Sensory-Level Data
Restricted (Penn State Only)
- Author:
- Shakerian, Shahrad
- Graduate Program:
- Architectural Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- February 16, 2022
- Committee Members:
- Bill Sitzabee, Co-Chair & Dissertation Advisor
Somayeh Asadi, Major Field Member
Houtan Jebelli, Co-Chair & Dissertation Advisor
Asok Ray, Outside Unit & Field Member
Somayeh Asadi, Professor in Charge/Director of Graduate Studies - Keywords:
- Facility maintenance
digital twins
machine learning
signal processing
failure prediction
early failure detection
anomaly detection - Abstract:
- Facility life cycle cost (LCC) includes four primary stages of construction, operation, maintenance, and destruction. The construction cost of a facility corresponds only to 15-20% of the total expenses, while facility maintenance develops the high costs, accounting for over 80% of the entire cost of the facility life cycle. Facility maintenance is defined as a combination of managerial, administrative, and technical activities during facility operation intended to maintain or restore a facility to the required functional condition. Such a critical process requires constant and systematic monitoring as a possible facility breakdown can put a heavy toll on the organization and dramatically increase the operating costs. These critical failures adversely affect system quality, occupants' satisfaction, additional repair costs, safety, and environmental integrity. These operational and retainment concerns demand cohesive and effective management. Facility maintenance management is defined as the managerial activities that determine maintenance responsibilities, objectives, control, and planning and improve maintenance activities and economies. Previous investigations have built up layers of evidence demonstrating the efficacy and viability of facility maintenance management in the context of LCC. These studies have classified facility maintenance management into three main categories: preventive maintenance, corrective maintenance, and predictive maintenance. Preventive maintenance can be described as all maintenance tasks carried out on a periodic, planned, and specific schedule to maintain facilities in the operational condition through the process of checking and restoring. Corrective maintenance corresponds to procedural operations required to detect, isolate, and resolve a failure. Predictive maintenance attempts to reduce facility failures by collecting equipment's useful data to diagnose future faults. Preventive and corrective maintenances are critical to the continuation and operability of the facilities. However, the underlying expenses for replacement parts, skilled laborers, and system downtime can bring the organizations to their knees. Further, corrective maintenances entail reactive operations that repair the system after the occurrence of the failure, which again can be costly and inefficient in the long run. These concerns urge organizations to infuse other affordable techniques for the maintenance process and the facility life cycle. Predictive maintenance is an emerging area of research that aims to warn facility managers about a probability of a failure in the system based on the historical information gathered from previous failure events. Although the number of studies in the predictive maintenance domain is limited, there is an increasing trend towards the use of artificial intelligence and sensory data to obtain the underlying patterns of failure in the facilities. Integrating sensory-level data and artificial intelligence is the key motivation to propose this dissertation to tackle the limitations mentioned above and enhance facility maintenance innovatively. As such, the primary goal of this study is to develop and examine a digital twin-based predictive facility maintenance approach that minimizes the system failures and maintenance cost by capturing a high volume of sensory data from the embedded sensors and continuously processing and classifying them into various levels of system health conditions through advanced machine learning algorithms. This dissertation aims to achieve the mentioned goal by performing four main objectives as the research steps: 1) As the first objective, this research attempts to identify critical sensory data that can be used to monitor facility performance, which can assist in detecting and predicting facility failures. Accordingly, the sensory data should be clearly identified, continuously collected, and thoroughly analyzed. Such operational data contains valuable information about how the system was working before and after the failure occurred. Therefore, the proper selection of sensory data, data processing methods, and predictive analysis are of utmost prominence. To identify the useful critical sensory data, various operational data (e.g., zone humidity, zone temperature, fan speed, and fan power consumption) of an HVAC system are retrieved from embedded sensors. To assess the potential of operational data in detecting and predicting failures of HVAC systems, measurable metrics in time-domain and frequency-domain such as mean frequency and kurtosis were calculated both from failure and non-failure data. Then, the Pearson product-moment correlation (PPMC) test showed a strong correlation (correlation coefficient greater than 0.6) between the calculated metrics and the working condition of the system. Results demonstrated the feasibility of applying operational data from the embedded sensors to early detect failures in HVAC systems. This potential of the embedded sensors will set the stage for assistive mechanisms whereby the HVAC operators can prognose the faults before they emerge by using predictive models. 2) As the second objective, this research tries to develop machine learning models for early detection of failures in facilities. To that end, this study proposes a framework for early failure detection of HVAC facilities based on collected operational data (airflow pressure, temperature, humidity, and carbon dioxide). The proposed framework applies signal-processing methods to remove signal artifacts. Then, it extracts metrics of the signals (features) and applies feature-selection algorithms to select the most relevant from the collected operational signals. Several supervised learning algorithms are applied to train an early failure-detection model to predict failures with a higher probability of occurrence in HVAC systems. To examine the performance of the proposed framework, the authors collected the operational data of HVAC facilities at the Chemical and Biomedical Engineering building at Pennsylvania State University. Data for recorded failures over six months was used to train and test the early fault detection models. The proposed framework resulted in 90% accuracy, 92% recall, and 87% precision in early failure detection of the targeted HVAC facilities. This study confirms the feasibility of the proposed framework to continuously assess the performance of the HVAC facilities and early detection of failures based on operational data. The proposed framework should increase the lifespan and utility of a building by optimizing maintenance management and reducing the failure rate of HVAC facilities through the early detection of faulty units. 3) As the third objective, it is attempted to use memory-based deep neural network algorithms such as recurrent neural network (RNN) and long-short term memory (LSTM) network to develop predictive models. These algorithms are designed for time-series data and can learn the patterns of failures and apply them to predict future failures. Thus, one year of operational data of an air handling unit in an HVAC system was applied as the training data set to train an LSTM model. Then, the authors developed a self-monitoring system that consists of three main steps of updating the model, continuous prediction, and continuous monitoring. When the model is trained by the initial dataset, including one year of operational data, the model is continuously updated every week based on the newly collected data of the targeted facility. This process improves model performance and accuracy in predicting the patterns of the normal operation of the system. As the second step, the developed self-monitoring system attempts to predict the patterns of operational data for the next 12 hours. This process employs the real-time streaming data by using an application programming interface (API) provided by OnBoard Data to feed the developed machine learning model and predict the pattern of time-series data. When the pattern of time series data for the next 12 hours of the facility operation is predicted, as the next step, an LSTM encode decode model forecasts the probable failures in the predicted pattern of data. Finally, to automatically monitor facility operation, the real-time operational data and the predicted data are compared every two hours based on the mean absolute error, called the anomaly score. The model was able to detect failures in the predicted patterns of operational data with an accuracy of 87%. The results of this objective make considerable contributions to the current body of literature. First, it shows the potential of using operational data captured from embedded sensors of buildings' installations to continuously assess the facilities' working conditions and predict probable system failures. Second, the results of this research provide new insights on predictive maintenance of building facilities by using real-time and recorded operational data and applying machine learning algorithms. 4) For the last objective, this research attempts to design a web-based digital platform for monitoring facility operations by analyzing the retrieved real-time operational data, which can assist in developing a feedback system to notify the facility managers and practitioners about the probable failures in the future. To that end, this step of the dissertation aims to develop and examine a framework for continuous monitoring of facility systems by using their digital twin models. Digital twins, besides the machine learning algorithms, internet of things (IoT), and data mining technologies, propose great potential in renewing the current paradigm of traditional facility management toward intelligent facility management. The proposed framework develops digital twin models for facility management based on the collected operational data from embedded sensors, machine learning, and a feedback loop for information transmission between the physical components of the facility and the virtual digital model to optimize maintenance management. The proposed framework integrates the developed machine learning models from the previous steps, as well as real-time operational data to train and improve the digital twin models. It can support facility management and maintenance to respond to changes in the pattern of operational data while the system is working, which improves controlling the system and prevents potential failures. The proposed digital twin framework implements practical solutions for machine learning difficulties in facility maintenance management, including the high number of operational data collected from various embedded sensors, time lags and alignment among time series data, and large, imbalanced datasets recorded from the system operations. To examine the performance of the proposed framework, the developed digital twin models are applied for an HVAC system of a building at the University Park campus of the Pennsylvania State University. The digital twin models are trained via the operational data collected from the embedded sensors of the air handling unit and used to enhance intelligent management based on real-time operational data. The case study represents the effectiveness of this framework in facility management by resulting in more than 85% accuracy in detecting failures before they actually occur in the real system. The proposed framework is expected to improve intelligent facility management and enhance the productivity and efficiency of the facility managers' performance in maintaining system operations.