**1. Introduction**

Commercial buildings constitute 18% of the U.S. primary energy consumption [1] and account for \$149 billion in annual energy expenditures [2]. Much of this consumption is due to operational waste, representing a tremendous potential for savings. The literature indicates that median whole-building savings of 16% are achieved by commissioning existing buildings [3] and that 5–30% of commercial building energy use is wasted due to problems associated with controls [4–9].

Commercially available fault detection and diagnostics (FDD) tools provide a means of monitoring-based commissioning, through which instances of operational inefficiency can be continuously identified, isolated, and surfaced for resolution by operations and maintenance staff. Today's FDD technology has been documented to enable whole building savings of 8% on average, across users [10]. These technologies integrate with building automation systems (BASs) or can be implemented as retrofit add-ons to existing equipment, and continuously analyze operational data streams across many system types and configurations. This is in contrast to the historically typical variants of FDD that are delivered as original equipment manufacturer-embedded equipment features or handheld FDD devices that rely upon temporary field measurements.

Figure 1 represents an idealized architecture of a BAS, adapted from American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE) Guideline 13 [11]. Field devices

(and controllers) connect to the sensors and actuators in the field. Network controllers typically provide supervisory control capabilities, scheduling, alarms, trending, local data storage, and user interfaces, in addition to some security features. Modern versions of these controllers have the ability to communicate via a BACnet (a data communication protocol for building automation and control network) IP over an IP network. When such functionality is not available, a common integration strategy employs "integration gateways" (e.g., Niagara JACE) that translate from proprietary protocols to standard protocols, such as BACnet IP. For larger installations and campuses, the controllers or gateways are also connected to a BAS server that provides configuration and management, long-term data storage (i.e., databases) and visualization tools. FDD tools can be installed in the local IP network, run from the cloud, or have a combination of cloud and local components. Integration with the BAS is typically implemented through a one-way interface using one of these three FDD–BAS integration pathways:


Through these interfaces, system-level operational data are made available to the FDD software. Meter data are often also included. Data are continuously analyzed and detected faults are presented to operational staff through a graphical user interface. Since the BAS is the primary source of data, the FDD is most commonly focused on Heating, Ventilation, and Air Conditioning (HVAC) equipment. However, today's technologies offer extensive libraries of FDD logic and algorithms, and therefore can be applied to lighting and other building end-use systems for which data are available [12].

**Figure 1.** Schematic illustration of the integration of building automation system (BAS) data into fault detection and diagnostics (FDD) products. (BACnet MS/TP - BACnet Master Slave Token Passing protocol).

Although FDD tools are being used to enable cost-effective energy savings, there remains an opportunity to advance the state of the technology. In practice, the need for human intervention to fix faults once they are identified often results in delay or inaction, resulting in additional operations and maintenance (O&M) costs or lost opportunities. Traditionally, FDD generates recommendations

and follow-up actions which are implemented by service technicians or other sta ff. An emerging capability comprises the integration of FDD outputs with facility managemen<sup>t</sup> "work order" or CMMSs (computerized maintenance managemen<sup>t</sup> systems). While this makes it possible to automatically generate work orders from the FDD system, human intervention is still required to implement the corrective action. Therefore, this work seeks to develop automated fault-correction approaches and integrate them with commercial FDD technology o fferings, thereby closing the loop between the passive diagnostics and active control. This is possible by converting the one-way BAS interface into a two-way interface, as is done with supervisory predictive control technologies that are emerging in the market.

It is not possible to automate the correction of mechanical faults such as failed actuators; however, there is nonetheless a compelling set of operational problems that are detectable in today's FDD offerings and are correctable through the software-based manipulation of the BAS parameters that can be exposed to external applications via BACnet. For example, Fernandez et al. [6] assessed control problems in commercial buildings, as well as their prevalence and whole-building energy impact for key commercial building sectors. Among the most common faults that relate to biased sensors, improper control parameter settings and ine fficient schedules have significant impact and high prevalence rates. Automating the correction of these types of faults can increase the savings realized through the use of FDD tools and reduce the extent to which savings are dependent upon human intervention.

The academic and technical literature has extensively covered the development of automated FDD applied to HVAC and lighting systems [13,14]; however, very little has been published on the automatic correction of the identified faults via the actual control system. One set of relevant papers stem from the vast literature on rule-based FDD algorithms. Fernandez et al. [15,16] developed passive and proactive fault auto-correction algorithms for various HVAC components and systems. The methods proposed to correct some faults which include biased air-handler unit (AHU) mixed air (MA), outside air (OA), and return air (RA) temperature and humidity sensors; damper control hunting; minimum outdoor air damper too open/closed; and manual overrides in large HVAC systems. Using the same approach, Brambley et al. [17] extended Fernandez et al. [15,16] by adding correction routines for the biased AHU supply air (SA) temperature and flow rate sensors and the biased variable-air-volume (VAV) box discharge air temperature and flow rate sensors. This project implemented and tested a subset of these algorithms (sensor bias and minimum outdoor air damper position) in a laboratory experiment. This research stopped short of validating the developed solutions in physical buildings or integrating them with existing BAS and commercial FDD products.

Related to the concept of fault correction is a body of work in the building control literature that focuses on fault tolerant control. The purpose of a fault tolerant controller is to the maintain proper operation of a system despite the presence of faults [18,19]. These approaches have been widely adopted in other industries for safety-critical systems such as nuclear power plants, spacecraft and aircraft. In the context of buildings, Padilla et al. [20] developed a model-based strategy which aims to replace defective sensors in AHUs [20] with "virtual sensors." The signal generated from these "virtual sensors" can be used in the AHU control system when the actual physical sensors behave abnormally. Supply air temperature and pressure sensor faults are e ffectively corrected by using the proposed algorithms. Wang et al. [21] developed a supervisory control scheme that adapts to the presence of a measurement error in an outdoor air flow rate. The method uses neural network models to estimate the correct behavior of the faulty sensor and to maintain indoor air quality while minimizing energy use. Hao et al. [22] employed principal component analysis to develop fault-tolerant control and data recovery in the HVAC monitoring system. Bengea et al. [23] developed a fault-tolerant optimal control strategy for an HVAC system integrating FDD and model predictive control. The output of the FDD algorithm is used to continuously update the model's predictive control algorithm parameters. The approaches described in these papers o ffer innovations to the state of the art, ye<sup>t</sup> they are not readily implemented in today's buildings control systems. This is because they comprise strategies

that are not supported by traditional BAS capabilities. Similarly, while the literature focuses on the development of these advanced controllers, it does not explore their integration with existing FDD technologies. An additional practical challenge is that a large volume of non-faulty data under various operational conditions is typically needed to train the models employed in these solutions.

This paper complemented and extended previous work in three ways. (1) It developed a comprehensive set of fault auto-correction algorithms designed to be integrated with commercial FDD tools. These algorithms target incorrectly programmed schedules, manual control "lock out," sensor bias, control hunting, rogue zone, and suboptimal setpoints/setpoints setback. Typically, commercial FDD tools are developed as a software layer on top of the existing BAS. There exists a natural separation of roles in this arrangement, in which the BAS actively controls the building and the FDD tool observes its operation and provides insights and recommendations to the building manager. The new auto-correction algorithms a fford the FDD technology a certain degree of control capability. (2) It conducted preliminary testing and performance validation during which two auto-correction routines were deployed in a commercial FDD tool and tested on two AHUs in a real building. The enhanced FDD tool was able to correct faults successfully. (3) It presented the challenges of the integration of developed auto-correction algorithms into commercial FDD tools along with the solutions through work with three industry partners. New insights were gained by implementing the pseudo-code developed by the research team in real systems and real buildings. Sections 2–4 present the auto-correction algorithms, preliminary testing and the implementation changes and solutions, respectively. Section 5 concludes the paper and describes future work.
