The ‘smart building’—or diagnosis are incr‘intelligent building’—concept has been primarily associated with the bare automation of the systems providing any service to building’s users. It extended certain characteristics of the ‘smart home’, where the technology automated several processes with schedule and remote control ^[11]. In generasingly used for extracting knowlel, the most consuming systems were the first ones to be automatized (controlled and supervised), as a whole, with building automation systems (BAS). The BAS, also known as BMS, stresses its management behavior to plan midterm and long-term strategies for the improvement of the performance of the systems. The introduction of ICT (information and communications technologies) allowed this significant advance. The advances in telecommunications enabled the use of the existing data networks for the interconnection of the elements, and even, the systems. The state of the equipment, the controlled variables and the context information, could be gathered to make better decisions about the comfort and energy savings. But it is yet a hard task to get the perfect optimization of all systems. However, the automated supervision brings benefits, such as the maintenance cost reduction and the robustness against unpredictable perturbations ^[12].

On the other hand, the industry around the smart cities is increasinge in the supervisory tasksly providing solutions in areas like energy, water, mobility, buildings or government in the short term. Energy is probably the most concerning matter because of its economic impact and social concerns. Buildings consume over 40% of the electrical energy in the most populated cities of the West ^[13]. The HVAC system is article the building service that consumes more with 32.7% of the power supply on average, while lighting requires 17.1% of the supply and the computers and appliances the 13.6% ^[14]. When buildings are public, such as offices, malls oroposes museums, the HVAC consumption is even higher, reaching 40.3% on average of the total supply.

To address this problem, researches have been autonomic cycle of data analysis tasks (ACODAseeking to optimize the consumption of the building’ services, like HVAC, lighting or elevators, applying control policies, automation and optimization. If the energy efficiency is critical, other objectives, such as improving the security & access control or people-centered policies, are gradually getting importance in smart building’s considerations. Technology evolution makes systems become ‘smarter’. Energy efficiency has been achieved with good practices in the daily operations, with social responsibility, government’s enforcement regulation or by financial departments’ cost-effectiveness pressure. Now, a smart building is broadening the response to these challenges with AI, the Cloud, Big Data, IoT or hardware parallelization, to improve the mobility, ubiquity, accuracy and interoperability.

The simplest way for controlling HVAC has been by sequencing ON–OFF orders, but these are far to meet the multi-objective building requirements ^[17]. Continuous regulation is widely performed with classic and inexpensive proportional integral derivative (PID) controllers. Nonlinearities, like partial loads incor, requiring self-tuning techniques, such as relay-autotuning, open-loop step tests and more recently fuzzy logic, Ziegler–Nichols or Cohen–Coon methods ^[4]. HVAC also requires multiple input and multiple orate dautput variables (MIMO) handling, like splitting the mechanism into several SISO subsystems—or just the PID. This in any case remains difficult to stabilize ^[4]. Complexity a mining models for exrises in applying multi-objective optimization requiring advanced control methods. Advanced control works with models that fall into three categories: ‘white box’, ‘black box and ‘gray box’ ^[17].

White box models aracting e built with direct mathematical formulations, modeling the mass balance, heat transfer, thermal momentum or flow rates with differential equations. The system analysis and generalization are simple but require deep knowledge from tin physics field. These models can be used only in simple systems, like SISO and steady-state systems, because otherwise could incur in heavy computational costs or low accuracy due to simplifications.

Black box or empirical models work with data and nee system monitoring, analyd to be built in a preliminary phase, by relating the recorded outputs to the inputs via statistic or ML methods. These data models are being implemented for real-time control, plant modeling, controller design, system performance improvement, calibration and parameterization. Once the model is learned, they are very fast, consume low computational resources, and can be used for simulating any layer components, since heat pumps up to complete subsystems, like multi-HVAC ^[1]. As a drawback, they have less capability of generalizating aon, remaining constrained to the experience learned from the actual data.

Gray or hybrid models balanormal situations and automatically identifying ace the black and white box drawbacks, improving simultaneously the accuracy and generalization capabilities. They normally use optimization like least squares, gradient descent or genetic algorithm (GA) to discover the ideal system parameters. fuzzy logic (FL) optimization shows also satisfactory performance, using simple mathematics—without formulating the physics inside- ruling robustly the systems even when they were nonlinear and complex. fuzzy adaptive network (FAN), Takagi–Sugeno fuzzy model (TS) and adaptive neuro-fuzzy inference system (ANFIS) controllers improve the accuracy of the prediction with a fast execution. However, when higher accuracy is needed, they require more grading, increasing exponentially the number of rules, and therefore, performing slower. When contextual information, such as the year season or scheduled activities are incorporated into the system’s knowledge, they are translated into fuzzy rules that shorten the training stage.

Other models are used to fit the system to desired trajectories based on evolutionary algorithms or statistics or linear or polynomial regressions, like nonlinear ARX (autoregressive with exogenous inputs), ARMAX (autoregressive-moving average with exogenous variables) and taking corrective actions ARIMA (autoregressive integrated moving average) models. ANNs (artificial neural networks) also contribute through the application of NNARX (neural network autoregressive with exogenous inputs), FFBP (feed forward back propagation) and RBF (radial basis function).

frequency–domain, state–space, geometric, case-based reasoning, stochastic and instantaneous methods are also applied.

The BMS is a computer-based control system that supervises and manages the building’s service, actuating in the networked electromechanical equipment. It was originally intended for monitoring the systems and improve the energy savings with the automation of control and is also known as BAS (building automation system). Nowadays, BMS usually stores data that can be analyzed for making longer-term decisions supporting the optimization of multiple objectives, such as healthier environments, pleasant indoor climate or cost reduction. The architecture has evolved from closed and standalone to an open and networked paradigm with more efficient remote procedures, providing intelligence and analytics, becoming in a cloud-based and multi-sourced architecture ^[14].

However, it is unclear in the state-of-the-art that the current technology can simultaneously optimize the multiple required objectives. The estimations about energy savings differ considerably depending on where solutions are applied, i.e., production, load or user’s behavior; the type of building; or the number of pursued objectives. For example, some authors claim that energy can be saved up to 27% working on the BMS ^[18]. Others estimate that they can save up to 20% of the energy applying control optimization in space heating and others that can be reduced up to 10% in lighting and ventilation ^[19].

Focusing on the technological challenges for the research applied to BMS, the following ones are identified ^[20]:

• The integration and usability of heterogeneous networks, technologies and applications into one single platform;

  The integration and usability of heterogeneous networks, technologies and applications into one single platform;

• The maintenance and support of smart, self-adaptive, autonomous applications and objects.

  The maintenance and support of smart, self-adaptive, autonomous applications and objects.

. The original idea published by IBM in 2001 ^[22] was that the software was sufficiently intelligent for caring of itself, similar to what the autonomic nervous system does, getting self-configuration, self-optimization, self-protection and self-healing. The data analysis tasks comprised in an autonomous cycle work together for shared common goals in the managed process, exploiting the data collected from the system to build knowledge models that describe, optimize and predict its behavior. They co-operate among them and interact with the system according to their specific roles.

ACODAT is decision-making oriented ^[5][7][8], and its tasks work together to achieve the objective of the supervised process. The tasks have different roles in the autonomous cycle, such as observing the process, analyzing and interpreting the events and making decisions to reach the objective for which the cycle was designed. This cycled solution allows solving complex problems in real time. The detailed description of the roles of each task is as follows:

• Monitoring: Tasks in charge of observing the supervised system. They capture data and information about the behavior of the system. In addition, they are responsible for the preparation of the data for the next steps (preprocessing, selecting the relevant features, etc.).

  Monitoring: Tasks in charge of observing the supervised system. They capture data and information about the behavior of the system. In addition, they are responsible for the preparation of the data for the next steps (preprocessing, selecting the relevant features, etc.).

• Analysis: Tasks in charge of interpreting, understanding and diagnosing what is happening in the monitored system. These tasks use building knowledge models of observed dynamics and behaviors, to understand what is happening in the system.

  Analysis: Tasks in charge of interpreting, understanding and diagnosing what is happening in the monitored system. These tasks use building knowledge models of observed dynamics and behaviors, to understand what is happening in the system.

• The on-demand and flexible service provision;

  The on-demand and flexible service provision;

• The size of the foreseen infrastructure with an estimation of 500 billion devices connected to the Internet by 2020, 50 billion of them via mobile wireless [21];

  The size of the foreseen infrastructure with an estimation of 500 billion devices connected to the Internet by 2020, 50 billion of them via mobile wireless ^[21];

• The machine-machine communication and multi-agent orchestration.

  The machine-machine communication and multi-agent orchestration.

HVAC management literature has treated about supervision of control and optimization processes with predicting models, networked elements in higher hierarchical layers or orchestrated in multi-agent architectures. However, the potential of an autonomic management entity has not been fully studied and proved yet. Plain operational decisions still require manual procedures. The autonomic learning capability of the management system will probably lead to improve the system control accuracy and robustness. The ACODAT-based management for HVAC will likely improve the multi-objective based on changing fuzzy policies ^[1]

• Decision-making: Tasks in charge of defining and implementing the necessary actions based on the previous analysis, in order to improve the performance, detect failures, among other things, in the supervised system. These tasks impact the dynamics of the system to improve it. The effects of these tasks are again evaluated in the monitoring and analysis steps, restarting a new iteration of the cycle.

  Decision-making: Tasks in charge of defining and implementing the necessary actions based on the previous analysis, in order to improve the performance, detect failures, among other things, in the supervised system. These tasks impact the dynamics of the system to improve it. The effects of these tasks are again evaluated in the monitoring and analysis steps, restarting a new iteration of the cycle.

An ACODAT has a multidimensional data model that works with the data collected from different sources, to characterize the behavior of the context and transform it into knowledge. Particularly, it can work with multiple data models, like ontologies, cognitive maps, etc. It runs on a single platform that integrates the necessary tools required for the tasks to process the information. Some of these tools are of data mining, semantic mining or linked data.

Thus, the tasks specifically required for the HVAC management perform system and context monitoring, data analysis, state diagnosis and decision-making, transformed into physical signals for the actuators.

The supervision system interacts with the controller. The latter regulates the machines and/or processes and the former watches the activity to detect abnormal situations ^[2][3][23]. FDD is one of these supervision systems indicating abnormal conditions necessary to discover. Faults in coolers are usually caused by degraded installations or bad human practices. coolers’ performance degradation is hardly detected and causes 42% of the service calls and 26% of repairing expenses ^[23].

Classifying fault severity level has three steps: the detection of the fault, its isolation and its identification. First-generation FDDs were based on rules and statistics and provided simplistic knowledge with a limited set of expected faults, and thus the support of field experts was unavoidable ^[3]. Today’s generation uses ML techniques and stands out for detection and diagnosis ^[24][25]. An AFD (automatic fault detection) system continuously monitors the HVAC system’ states with fuzzy algorithms ^[26]. Recent AI models allow dynamic fault detection thresholds minimizing the number of false positives and reducing the number of missed faults, with agglomerative clustering that starts with one cluster per data point and groups them into likelihood-based clusters. Some studies associate these models with the Bayesian’s DBNs (dynamic Bayesian networks) and Markov’s HMM (hidden Markov model) techniques.

The fault prognosis (early diagnosis) by detecting equipment degradations allows keeping the optimum performance throughout the facility’s life cycle. Classical fault detection and diagnosis is based on supervised learning models, while the prognosis is based on RL ^[27]. The most common reinforced learning (RL) is implemented with MDP (Markov decision process) or its variant, POMDP (partially observable Markov decision process), necessary when the context state is not fully known. RL can be applied to a multi-agent problem, such as MARL (multi-agent reinforcement learning) or deals with the optimal coordination between cooperative or competitive agents with decentralized POMDP. When the problem applies to different tasks, TL (transfer learning) transfers knowledge once the problem is solved to support solving the next one. MTL (multi-task learning) is also based on this principle, but tasks are variants of the same problem. Another important approach is the MORL (multi-objective reinforced learning), whereas the objective is to learn multiple policies simultaneously for every objective ^[28][29].

References

1. Aguilar, J.; Garces-Jimenez, A.; Gallego-Salvador, N.; De Mesa, J.A.G.; Gomez-Pulido, J.M.; Garcia-Tejedor, A.J.; Garcis-Jiminez, A.; De Mesa, J.G.; Garcma-Tejedor, A. Autonomic Management Architecture for Multi-HVAC Systems in Smart Buildings. IEEE Access 2019, 7, 123402–123415. [Google Scholar] [CrossRef]
2. Kiluk, S. Diagnostic information system dynamics in the evaluation of machine learning algorithms for the supervision of energy efficiency of district heating-supplied buildings. Energy Convers. Manag. 2017, 150, 904–913. [Google Scholar] [CrossRef]
3. Tamani, N.; Ahvar, S.; Santos, G.; Istasse, B.; Praça, I.; Brun, P.-E.; Ghamri, Y.; Crespi, N.; Becue, A. Rule-Based Model for Smart Building Supervision and Management. In Proceedings of the 2018 IEEE International Conference on Services Computing (SCC), San Francisco, CA, USA, 2–7 July 2018; pp. 9–16. [Google Scholar]
4. Escobar, L.M.; Aguilar, J.; Garcés-Jiménez, A.; De Mesa, J.A.G.; Gómez-Pulido, J.M.; Morales, L. Advanced Fuzzy-Logic-Based Context-Driven Control for HVAC Management Systems in Buildings. IEEE Access 2020, 8, 16111–16126. [Google Scholar] [CrossRef]
5. Aguilar, J.; Sanchez, M.; Cordero, J.; Valdiviezo, P.; Guamán, L.R.B.; Chamba-Eras, L. Learning analytics tasks as services in smart classrooms. Univers. Access Inf. Soc. 2017, 17, 693–709. [Google Scholar] [CrossRef]
6. Morales, L.; Ouedraogo, C.A.; Aguilar, J.; Chassot, C.; Medjiah, S.; Drira, K. Experimental comparison of the diagnostic capabilities of classification and clustering algorithms for the QoS management in an autonomic IoT platform. Serv. Oriented Comput. Appl. 2019, 13, 199–219. [Google Scholar] [CrossRef]
7. Aguilar, J.; Cordero, J.; Buendía, O. Specification of the Autonomic Cycles of Learning Analytic Tasks for a Smart Classroom. J. Educ. Comput. Res. 2017, 56, 866–891. [Google Scholar] [CrossRef]
8. González-Eras, A.; Buendia, O.; Aguilar, J.; Cordero, J.; Rodriguez, T.; Peñalvo, F.J.G.; Lagos-Ortiz, K.; Alcaraz-Mármol, G.; Del Cioppo, J.; Vera-Lucio, N.; et al. Competences as Services in the Autonomic Cycles of Learning Analytic Tasks for a Smart Classroom. In International Conference on Technologies and Innovation; Springer: Cham, Switzerland, 2017; pp. 211–226. [Google Scholar]
9. Lopez, C.P.; Santorum, M.; Aguilar, J. Autonomous Cycles of Collaborative Processes for Integration Based on Industry 4.0. In Information Technology and Systems; Rocha, A., Ferrás, C., Paredes, M., Eds.; Springer: Berlin, Germany, 2019; pp. 1771–1786. [Google Scholar]
10. Vizcarrondo, J.; Aguilar, J.; Exposito, E.; Subias, A. MAPE-K as a service-oriented architecture. IEEE Lat. Am. Trans. 2017, 15, 1163–1175. [Google Scholar] [CrossRef]
11. Snoonian, D. Smart buildings. IEEE Spectr. 2003, 40, 182–183. [Google Scholar] [CrossRef]
12. Žáček, M.; Janošek, M. SBOnto: Ontology of Smart Building. Far East J. Electron. Commun. 2017, 17, 1101–1109. [Google Scholar] [CrossRef]
13. Khatib, H. IEA World Energy Outlook 2011—A comment. Energy Policy 2012, 48, 737–743. [Google Scholar] [CrossRef]
14. Minoli, D.; Sohraby, K.; Occhiogrosso, B. IoT Considerations, Requirements, and Architectures for Smart Buildings—Energy Optimization and Next-Generation Building Management Systems. IEEE Internet Things J. 2017, 4, 269–283. [Google Scholar] [CrossRef]
15. Staff. Global Revenue for Energy Efficiency Commercial Building Retrofits Is Expected to Exceed $100 Billion in 2025, According to Navigant Research. Available online: https://www.businesswire.com/news/home/20160714005110/en/Global-Revenue-Energy-Efficiency-Commercial-Building-Retrofits (accessed on 10 April 2020).
16. Mohamed, N.; Al-Jaroodi, J.; Lazarova-Molnar, S. Energy Cloud: Services for Smart Buildings. In Sustainable Cloud and Energy Services; Springer Science and Business Media LLC: Berlin, Germany, 2017; pp. 117–134. [Google Scholar]
17. ASHRAE. ASHRAE Handbook–HVAC Applications; American Society of Heating, Refrigerating and AIR-Conditioning Engineers, Inc.: Atlanta, GA, USA, 2003. [Google Scholar]
18. Reynolds, J.; Rezgui, Y.; Kwan, Y.; Piriou, S. A zone-level, building energy optimisation combining an articial neural network, a genetic algorithm, and model predictive control. Energy 2018, 151, 729–739. [Google Scholar] [CrossRef]
19. Wang, F.; Yoshida, H.; Ono, E. Methodology for optimizing the operation of heating/cooling plants with multi-heat-source equipments. Energy Build. 2009, 41, 416–425. [Google Scholar] [CrossRef]
20. Foteinos, V.; Kelaidonis, D.; Poulios, G.; Vlacheas, P.; Stavroulaki, V.; Demestichas, P. Cognitive Management for the Internet of Things: A Framework for Enabling Autonomous Applications. IEEE Veh. Technol. Mag. 2013, 8, 90–99. [Google Scholar] [CrossRef]
21. Organisation for Economic Co-Operation and Development. No. 192. In OECD Digital Economy Papers; Organisation for Economic Co-Operation and Development (OECD): Paris, France, 2010. [Google Scholar]
22. Kephart, J.; Chess, D. The vision of autonomic computing. Computer 2003, 36, 41–50. [Google Scholar] [CrossRef]
23. Namburu, S.M.; Azam, M.S.; Luo, J.; Choi, K.; Pattipati, K.R. Data-Driven Modeling, Fault Diagnosis and Optimal Sensor Selection for HVAC Chillers. IEEE Trans. Autom. Sci. Eng. 2007, 4, 469–473. [Google Scholar] [CrossRef]
24. Bravo, C.; Aguilar, J.; Rivas, F. Diseño de una Arquitectura de Automatización Industrial basada en Sistemas Multi-agentes. Ciencia e Ingeniería 2004, 25, 75–88. [Google Scholar]
25. Araujo, M.; Aguilar, J.; Aponte, H. Fault detection system in gas lift well based on artificial immune system. In Proceedings of the International Joint Conference on Neural Networks, Portland, OR, USA, 20–24 July 2003; Volume 3, pp. 1673–1677. [Google Scholar] [CrossRef]
26. Wall, J.; Guo, Y.; Li, J.; West, S. A Dynamic Machine Learning-based Technique for Automated Fault Detection in HVAC Systems. ASHRAE Trans. 2011, 117, 449–456. [Google Scholar]
27. Lo, C.; Chan, P.; Wong, Y.; Rad, A.B.; Cheung, K. Fuzzy-genetic algorithm for automatic fault detection in HVAC systems. Appl. Soft Comput. 2007, 7, 554–560. [Google Scholar] [CrossRef]
28. Neto, A.H.; Fiorelli, F.A.S. Comparison between detailed model simulation and artificial neural network for forecasting building energy consumption. Energy Build. 2008, 40, 2169–2176. [Google Scholar] [CrossRef]
29. Aguilar, J. Definition of an energy function for the random neural to solve optimization problems. Neural Netw. 1998, 11, 7317–7337. [Google Scholar]
30. Mason, K.; Grijalva, S. A review of reinforcement learning for autonomous building energy management. Comput. Electr. Eng. 2019, 78, 300–312. [Google Scholar] [CrossRef]
31. Kim, W.; Katipamula, S. A review of fault detection and diagnostics methods for building systems. Sci. Technol. Built Environ. 2017, 24, 3–21. [Google Scholar] [CrossRef]
32. Deshmukh, S.; Glicksman, L.; Norford, L. Case study results: Fault detection in air-handling units in buildings. Adv. Build. Energy Res. 2018, 1–17. [Google Scholar] [CrossRef]

Kim et al. review studies about the automated FDD (AFD) since 2004 for commercial buildings ^[30]. They categorize AFD’s methods in three groups and analyze several to understand their strengths and weaknesses. Deshmukh et al. present analytical methods embodied in useful software tools to identify and evaluate some building system faults, which cause large building energy inefficiencies ^[31]. They define the target faults, such as the imbalanced airflows within several large air-handling units. The experiments show that embracing uncertainty with an HVAC’s fault detection system is paramount to a good fault inference. Deshmukh, continuing his study, considers algorithms for faults like stuck dampers and leaking dampers ^[32]. These damper’s fault detection algorithms can be applied to both outdoor and return air dampers. They combine expert-rule based fault detection models with the first principles of thermodynamics, for fault detection with minimal non-intrusive measurements. The algorithms focus on detecting faults with minimal data in a large monitored academic building. The experiment used the data collected from the BEMS (building energy management system) of an academic building in Boston.