Article

Predictive Maintenance in Building Facilities: A Machine

Learning-Based Approach
Yassine Bouabdallaoui 1,*, Zoubeir Lafhaj 1, Pascal Yim 2, Laure Ducoulombier 3 and Belkacem Bennadji 4

1 Laboratoire de Mécanique Multiphysique Multiéchelle, LaMcube, UMR 9013, Centrale Lille, CNRS,
Univ. Lille, F-59000 Lille, France; zoubeir.lafhaj@centralelille.fr
2 Centre de Recherche en Informatique Signal et Automatique de Lille, CRIStAL, UMR 9189, Centrale Lille,

CNRS, Univ. Lille, F-59000 Lille, France; pascal.yim@centralelille.fr

3 Bouygues Construction, 78280 Guyancourt, France; l.ducoulombier@bouygues-construction.com

4 Bouygues Energies et Services, 78280 Guyancourt, France; b.bennadji@bouygues-es.com

* Correspondence: yassine.bouabdallaoui@centralelille.fr

Abstract: The operation and maintenance of buildings has seen several advances in recent years.
Multiple information and communication technology (ICT) solutions have been introduced to better
manage building maintenance. However, maintenance practices in buildings remain less efficient
and lead to significant energy waste. In this paper, a predictive maintenance framework based on
machine learning techniques is proposed. This framework aims to provide guidelines to implement
predictive maintenance for building installations. The framework is organised into five steps: data
collection, data processing, model development, fault notification and model improvement. A sport
facility was selected as a case study in this work to demonstrate the framework. Data were collected
from different heating ventilation and air conditioning (HVAC) installations using Internet of
Things (IoT) devices and a building automation system (BAS). Then, a deep learning model was
Citation: Bouabdallaoui, Y.; Lafhaj, used to predict failures. The case study showed the potential of this framework to predict failures.
Z.; Yim, P.; Ducoulombier, L.; However, multiple obstacles and barriers were observed related to data availability and feedback
Bennadji, B. Predictive Maintenance
collection. The overall results of this paper can help to provide guidelines for scientists and practi-
in Building Facilities: A Machine
tioners to implement predictive maintenance approaches in buildings.
Learning-Based Approach. Sensors
2021, 21, 1044. https://doi.org/
Keywords: predictive maintenance; buildings; IoT; data; machine learning; autoencoders; HVAC
10.3390/s21041044

Academic Editor: Enrique López

Droguett
Received: 11 December 2020 1. Introduction
Accepted: 26 January 2021 In Europe, buildings are responsible for 40% of energy consumption, and approxi-
Published: 3 February 2021 mately 28% of total direct and indirect CO2 emissions [1,2]. A large part of this energy is
consumed by building installations, mainly cooling and heating systems. Indeed, faulty
Publisher’s Note: MDPI stays neu-
operations of building installations lead to significant waste, causing up to a 20–30% in-
tral with regard to jurisdictional
crease in the total building energy consumption [3]. To achieve energy efficiency goals,
claims in published maps and insti-
building maintenance and management approaches must be improved and optimised.
tutional affiliations.
Facility management (FM) teams depend on real-time, accurate and comprehensive
data to perform day-to-day maintenance activities and to provide accurate information to
top management [4]. However, the activities of inspecting facilities, assessing mainte-
Copyright: © 2021 by the authors.
nance and collecting data are labour-intensive and time consuming [5]. In addition, the
Licensee MDPI, Basel, Switzerland.
budget and resources allocated for building maintenance are limited [6], and maintenance
This article is an open access article personnel argue that their budget and resources are insufficient and below their needs
distributed under the terms and con- [7,8]. This trade-off affects the quality and the relevance of the maintenance activities and
ditions of the Creative Commons At- inspections, which leads to poor maintenance and quality management policies in facili-
tribution (CC BY) license (http://cre- ties.
ativecommons.org/licenses/by/4.0/).

Sensors 2021, 21, 1044. https://doi.org/10.3390/s21041044 www.mdpi.com/journal/sensors

Sensors 2021, 21, 1044 2 of 15

Nowadays, maintenance practices in the field of FM are mainly based on corrective

maintenance: late actions are taken following a user complaint or an unplanned failure
[9]. The lack of a budget and human resources limits the use of preventive maintenance
to the minimal level of mandatory inspections in critical installations. Moreover, the con-
cept of predictive maintenance is based on processing the operation data sent via sensors
using data analytics techniques. This approach offers the possibility to address the gap in
maintenance practices by supporting the FM teams to take early action and avoid un-
planned failures without the need for costly intensive site inspections of the installations.
Recently machine learning and data science techniques have been used in many as-
pects of modern society, including the Internet, finance, insurance, and medical applica-
tions [10]. In addition to the recent advances in machine learning, the introduction of
smart devices and the Internet of Things allowed for the connection of physical assets and
real-time data streaming at a low cost [11,12]. As a consequence, the field of operation and
maintenance management is evolving. Classic maintenance approaches and quality man-
agement methods that were all controlled by people in the past are being transformed into
predictive maintenance thanks to machine learning (ML) and artificial intelligence (AI)
[13]. The industry of facility management can benefit from the rise of machine learning
and IoT in order to improve the management of their assets and reduce waste. This paper
presents a predictive maintenance approach for the field of facility management; for this
purpose, a framework based on machine learning is proposed whilst taking into account
the specificities of the FM field.

2. Research Background
This section first presents an overview of the main maintenance approaches and de-
scribes the differences between them. Following this, an overview of the deep learning
algorithms used in this research is presented: autoencoders, recurrent neural networks
and long short-term memory (LSTM).

2.1. Maintenance Approches

According to the British standards [14], maintenance is the combination of all tech-
nical, administrative, and managerial actions during the life cycle of an item intended to
retain it in or restore it to a state in which it can perform the required function. Approaches
of maintenance management can be grouped into three main groups as follows [15]:
 Corrective maintenance: also known as reactive maintenance or run to failure
maintenance, which consists of intervening after the failure. The equipment is al-
lowed to operate until it fails.
 Preventive maintenance: this consists of carrying out inspection and maintenance ac-
tions while the equipment is still running to reduce the probability of breakdowns.
Preventive maintenance can be either time-based via a schedule or usage-based (e.g.,
every 100 km). This approach helps to reduce the number of failures, but unnecessary
inspections are performed and unplanned failures still occur, which increases the
cost of maintenance.
 Predictive maintenance: this approach is based on using condition monitoring data
to predict the future machine health state [15]. This approach aims to predict when,
where, and which components may have potential failures.
Nowadays, maintenance practices are mainly corrective and preventive; predictive
maintenance is only applied for critical situations [16]. Traditionally in facility manage-
ment, aside from mandatory tests and inspections for critical equipment such as boilers
and chillers, the majority of maintenance activities in buildings remain mainly corrective,
simply responding to users’ complaints or to unplanned failures [9].
Previous studies have addressed the amelioration of preventive maintenance in the
FM field. The most important issue for the maintenance manager is to anticipate the suit-
able time for the effective implementation of each maintenance activity [17] within the
Sensors 2021, 21, 1044 3 of 15

limits of the maintenance budget and the available resources. With this intent, studies
were conducted to optimise the inspection scheduling in HVAC installations using opti-
misation techniques such as the Monte Carlo method [18,19]. Similar studies used data
mining techniques and time series forecasting to optimise the maintenance scheduling
based on the history of the maintenance operations in the building [17]. These studies
showed interesting results by optimising the inspection periods. However, unplanned
failures still occurred. Different studies focused on adopting industrial maintenance tech-
niques such as mechanical vibration analysis to monitor building installations using Fou-
rier transformation and fuzzy logic [20,21] or simulation techniques for fault detection
[22]; similarly, statistical models including linear and nonlinear regression were used for
fault detection and diagnostics in HVAC units [23]. However, the high cost of the model-
ling and the simulation as well as the limitation to generalise the models on similar instal-
lations have limited the use of these techniques in the FM field.
The predictive maintenance approach presents an opportunity for the FM sector to
reduce unplanned failures, reduce maintenance costs and penalties, as well as to improve
the comfort and the security of the inhabitants. However, implementing a predictive
maintenance approach presents multiple challenges, such as connecting physical assets,
extracting valuable data and developing accurate predictive algorithms. Indeed, the con-
cept of predictive maintenance is not new; multiple studies were conducted in the past
few decades, mainly focusing on statistical approaches [24]. Despite this, the deployment
of effective predictive solutions remained expensive and difficult to implement. Further-
more, the recent development in the industry has increased machine complexity, which
makes it difficult to predict failures with conventional methods [25]. Simultaneously, ma-
chine learning techniques have been gaining ground from computer vision [26] to natural
language processing [27], from medical applications [28] to games [29], including appli-
cations in predictive maintenance and anomaly detection [30–33].

2.2. Deep Learning Overview

Deep learning is a sub-field of machine learning and an approach of artificial intelli-
gence. It is a set of methods based on representation learning. It consists of representing
the learning task as an embedded hierarchy of concepts to facilitate the extraction of useful
patterns from raw data [10,34]. This hierarchy is represented as an artificial neural net-
work of several layers, with each layer containing several neurons, and each neuron per-
forming a simple but nonlinear transformation. The composition of these simple transfor-
mations allows for the learning of complex representations and solving complex tasks.
Deep learning has accelerated the advancement of multiple aspects in modern society,
from machine translation [35] to self-driving cars [36]. There are multiple deep learning
models, and they vary in terms of design and architecture. Their use depends on the na-
ture of the task (classification, prediction, clustering etc.) and the nature of the input data
(text, images, sequences, etc.). A brief review of the deep learning architectures used in
this work is presented below.

2.2.1. Autoencoders
Autoencoders are a set of deep learning architectures; they can be considered as a
special form of neural networks designed for unsupervised learning tasks [37]. The learn-
ing process is unsupervised since there is no label variable. In this type of neural network,
the output variable is set to have the same dimension as the input variable [38]. An auto-
encoder is composed of two processes—an encoder and a decoder. The encoder trans-
forms the input data trying to dig out hidden representations, while the decoder tries to
reconstruct the input data from the hidden representations [37,39]. This process of encod-
ing and decoding the input data can be seen as a learning circuit that tries to reconstruct
the inputs with the minimum amount of distortion and noise [40]. An illustration of an
autoencoder architecture is presented below in Figure 1. Autoencoders have been widely
Sensors 2021, 21, 1044 4 of 15

used for dimensionality reduction applications [41], signal reconstruction applications,

and anomaly detection applications [37,42,43].

Figure 1. Illustration of an autoencoder architecture.

2.2.2. Long Short-Term Memory (LSTM)

Long short-term memory is a variant of the artificial recurrent neural network (RNN)
architecture, which is a type of artificial neural network designed to process sequential
data as time series data and text and speech data [44]. LSTM and RNNs in general are
capable of capturing long-term dependencies in a sequence. This means they can capture
information about the past of the sequence. Thanks to this characteristic, LSTM are widely
used in multiple applications including natural language processing applications [44],
forecasting time series [35], and anomaly detection [30,45].

3. Research Objective and Methodology

3.1. Aim of the Current Study
The aim of this study is to propose a generic framework for predictive maintenance
in order to reduce unplanned failures and minimise faulty operations in building instal-
lations. However, buildings are different, in terms of size, occupancy, and use. Thus, a
generic framework should be flexible and adjustable to the differences between one build-
ing and another. For this reason, this study focuses on providing general guidelines to
implement predictive maintenance. However, some propositions in this framework, such
as the architecture of the predictive model or the collected data, can be changed depend-
ing on the building context and its environment. The objective is that the approach can be
applied to any type of building installation (HVAC, lift, electrical machinery, etc.). Later
in the study, this framework is tested using a case study of a sport facility.

3.2. Quantitative Research

Since the implementation of predictive maintenance and machine learning ap-
proaches in buildings is still new, there is little empirical research on this topic. To gather
data and complete the literature research, the authors used face-to-face interviews with
experts operating in facility management. Interviews were conducted with six experts.
The goals of the interviews were: (1) to identify the actual tools used in the building envi-
Sensors 2021, 21, 1044 5 of 15

ronment, (2) to identify the available data, and (3) to assess their perspectives of the frame-
work. The results of the interviews were used alongside the literature research to identify
the list of the data sources in the building environment. These interviews helped us to
design the fault detection and the feedback modules in the framework found in Section 4.
Interviewees were asked questions such as: how do you (the user) expect the interface of
the fault detection? Would you like to control the number of alerts? How would you like
to formulate and send your feedback in the application?

4. Framework Design
4.1. Defining Data Sources
The first step in this study was to define the data sources available in the building
environment. Below, a list of data sources in the building environment are identified.
 Building automation systems (BAS): BAS are largely used in modern buildings to
control and monitor the different installations via real-time data [46]. BAS contain
both numerical and categorical data. Typical examples of numerical data are meas-
urements such as temperature, energy consumption, and air and water flow rate, etc.,
whereas categorical data consist of time, alerts, and the binary state (ON/OFF),
etc.[37]
 IoT devices and sensors: connected sensors and IoT devices have been introduced in
buildings in recent years in order to collect information on the building and its sur-
roundings [47–49]. These devices can be used to collect multiple types of information;
they can be deployed on the installations (air handling unit (AHU), lift, chiller, etc.)
to extract data such as temperature and vibration [50]. They can be used to collect
human behaviour data such as occupation or mobility [48,51]. They can also be used
to collect indoor and outdoor environmental measurements such as CO2 levels [52].
 Computerised maintenance management systems (CMMS): CMMS are used to man-
age daily maintenance activities. Functionalities of CMMS include: receiving emer-
gency work orders and users’ requests, scheduling preventive maintenance activi-
ties, recording the history of maintenance activities, and inventory control, etc.
[53,54]. Thus, CMMS represents an important data source for predictive mainte-
nance.
 Building information modelling (BIM): the BIM model provides architectural 3D vis-
ualisation and standardisation of building information exchange between the stock-
holders along the construction project life cycle [55]. In recent years, several studies
have been carried out to implement the BIM in the operation and maintenance phase
[56–58]. BIM can be used to support FM teams while operating maintenance activities
[59], to monitor energy efficiency in buildings [57] and to provide visual analytics for
maintenance activities [60].
 Other data sources: building energy management system (BEMS), computer aided
facility management (CAFM) and integrated workplace management system
(IWMS) are examples of other data sources that can be found in the building envi-
ronment. However, their use is limited to some specific facilities.

4.2. The Framework Architecture

The proposed framework in Figure 2 represents a machine learning approach
adapted to the building context. The framework is composed of five steps: data collection,
data processing, model development, fault notification and model improvement. All the
steps are discussed and detailed below.
Sensors 2021, 21, 1044 6 of 15

Figure 2. The proposed framework architecture.

4.3. Data Collection, Storage and Processing

This part consists of preparing the data flow for the learning process. It starts with
collecting data from the sources then storing it before applying necessary data cleaning
and transformation.
 Data collection: the first step of the framework aims to collect data from the available
sources in the building environment. It involves defining the data sources in the
building then connecting them to extract the necessary data. The data sources were
defined in this work in Section 4.1. However, data collection methods are not speci-
fied in this framework, since they depend on the user preference and the available
ICT infrastructure.
 Data storage: this consists of storing the data after collecting them in a storage me-
dium. There are different storage methods (cloud, local, etc.) which depend on the
preference and the infrastructure of the user. in this framework, data storage is not
discussed.
 Data pre-processing: the purpose of this step is to transform the raw data into a struc-
tured dataset ready for the training process. This step is comprised of two main parts:
o Data cleaning: this consists of cleaning the data entries by removing irrelevant
entries, replacing Nan values, and treating outliers.
o Data transformation: in this study, only two transformations were proposed:
normalisation of numerical features and encoding of categorical features. How-
ever, other transformations and feature engineering can be used [61,62].

4.4. Model Development

In this framework, the dataset is split into a training set and a testing set. The training
set is used while training the model to learn the anomaly patterns, while the testing set is
used to validate the model and to tune its parameters such as the anomaly threshold. In
this study, an autoencoder architecture was chosen for the machine-learning model.
Sensors 2021, 21, 1044 7 of 15

4.4.1. The Autoencoder Model

The proposed model has the architecture of an autoencoder: the encoder part consists
of two LSTM layers followed by a reshaping layer to reshape the vector into the right
dimension. The decoder part is similar and is composed of two LSTM layers and a reshap-
ing layer. The resultant vector in the output layer has the same dimension as the input
vector (Figure 3).

Figure 3. Illustration of the autoencoder model.

The choice of the autoencoder is justified by the model’s nature as an unsupervised

learning algorithm [63]. The autoencoder does not require labelled datasets (data explic-
itly tagged with fault/normal labels) which are generally not available in buildings. Fur-
thermore, autoencoders are flexible methods that require less hand engineering work and
can be adapted to several applications. Moreover, the choice of LSTM layers is justified
because they are designed to process sequential data such as time-series [44], which is the
case of the majority of the installations monitoring data (temperature, vibration, and en-
ergy consumption, etc.).

4.4.2. Anomaly Score

Anomaly score is an evaluation metric to calculate the distance between the input
vector and the output vector . There are multiple evaluation metrics such as root
mean square error (RMSE), mean absolute error (MAE), etc.
In this study, RMSE was used. It is defined as below (Equation (1)).
Let = ( , , … . , ) and = ( , , … . , )

∑ ( − )
RMSE , = (1)

A threshold is defined depending on the training algorithm and depending on the

targeted accuracy. If the anomaly score at a given time is higher than the defined thresh-
old, the model sends an anomaly alert. This process is illustrated in Figure 4.

Figure 4. An Illustration of the process of calculating the anomaly score.

Sensors 2021, 21, 1044 8 of 15

4.5. Fault Notification

In this step, the model is deployed online after the training and the validation steps
are done. Data are collected and processed before they are streamed to the fault detection
model. If the anomaly score is bigger than the defined threshold, a notification is dis-
played in the CMMS then sent to the facility manager via SMS. The notification contains
the name of the equipment, its location, and the time of the alert event.

4.6. Feedback and Continuous Improvement

Machine learning models need to be regularly updated in order to improve the accu-
racy of the results. For this reason, collecting feedback after the deployment is crucial to
improve the fault prediction model. The feedback is collected from the facility manage-
ment team (the users) where they can report false alerts or undetected failures. This feed-
back is collected and stored via the CMMS. After the feedback is collected, a procedure of
error evaluation is carried out, where the errors of the model are inspected. Following
this, the model is updated using new training data, and its parameters are tuned to reduce
the errors ratio.
The improvement of the model is not a systematic approach; the procedure of updat-
ing the model using the collected feedback should be done via a proper schedule and by
a machine learning specialist. In order to give the FM team a quicker response, the anom-
aly threshold was designed as an external parameter where the user can directly change
the setting without a need for a total update to the model.

4.7. Model Implementation

The framework can be applied to the different installations in the building including
HVAC, lifts, and pumps. After defining the installations that will be involved, data are
collected following the guidelines in Section 4.1; these data can also be completed by the
use of IoT sensors such as vibration sensors, temperature sensors, and energy consump-
tion meters, as illustrated later in the case study. As highlighted previously, the FM teams
are free to choose how to send and store the data depending on the available ICT infra-
structure in the building; for example, in the case study below, the Lora network was used
to send the data that were stored and processed in the cloud. The first implementation of
the model requires a period of collecting data to create the training and validation da-
tasets. The development step in the framework is based on the classic ML validation meth-
odology by splitting the dataset into a training set and a validation set. The proposed au-
toencoder model is flexible and can be deployed with different types of time series data.
After the validation phase is completed, the model can be deployed online in the facility,
the framework proposes to integrate the FM team in order to set the alert threshold de-
pending on the availability of the team to check the alerts and on the criticality of each
installation. The framework proposes integrating the FM team into the model improve-
ment process. The FM team can continuously change the threshold alert to improve the
accuracy of the model. Their feedback is also collected for future updates of the model.
Since the predictive model is based on a deep learning model, the predictive model can
be reused on similar installations in different facilities by using transfer learning [64],
which can allow one to cut the development cost and reduce the implementation time.

5. Case Study: Predictive Maintenance for HVAC Installations in Sport Facility Build-
ings
5.1. The Facility Characteristics
The case study in this paper was conducted at a sport facility in the Paris region,
France. It is composed of two principal buildings covering an area of 15,000 m2. The facil-
ity contains multiple installations. However, for accessibility and privacy reasons, this
case study focused only on a selected group of HVAC installations that includes: two
AHUs, three boilers, and three double pumps. The facility is equipped with a building
Sensors 2021, 21, 1044 9 of 15

automation system (BAS) that monitors and controls the different installations in the fa-
cility.

5.2. Data and Model Characteristics

According to the guideline defined previously in the framework, data sources were
identified. The available sources are presented below.
 The building automation system (BAS) was connected to a web server. Each installa-
tion is monitored via the BAS through one or multiple variables such as temperature,
energy consumption, water consumption, and air or water flow rate, etc. A report
from the BAS is uploaded every hour. The report contains the date and the time, the
name of the variable and its value at that time.
 An extract from the CMMS that contains a part of the maintenance record.
 Vibration device: an IoT device was installed on the surface of the equipment; it is
used to collect the vibration measurements on the installation. The data reported by
the device are the acceleration measures on the three axis (x, y and z-axis), the fre-
quency of vibration of the three axis, a binary variable (ON/OFF, which detects if the
machine is enabled) and the temperature in the surface of the equipment. An exam-
ple is presented in Figure 5 that shows a vibration device attached to a double pump.
 Electric meter device: an IoT meter installed to collect the electric energy data which
includes the following measures: electric current intensity and voltage as well as the
temperature on the surface of the equipment.

Figure 5. A photo of a vibration device installed on a double pump.

The IoT devices were only attached to the surface of the installations so that no dete-
rioration or harm happens to the installations. The IoT devices are connected to the inter-
net via the Lora network [65]. They were programmed to upload the measures in a 1 h
cycle. The data are uploaded and stored on the web platform of Objenious: a French com-
pany (By Bouygues Telecom) specialising in IoT solutions and development.
Table 1 presents the group of the HVAC installations used in this study and the IoT
device attached to each installation. The vibration device was attached to each installation.
Due to accessibility issues, the electric meter device was only attached to the two AHUs.
Sensors 2021, 21, 1044 10 of 15

Table 1. The HVAC installations and the attached IoT devices.

Quantity Installation Attached IoT Devices

Vibration device
2 AHU
Electric meter device
3 Boilers Vibration device
3 Double pump Vibration device

In order to build the training dataset, data were collected for a period of around 3
months. The predictive model used in this case study has the same architecture presented
in Section 4.4.1. The root mean square error (RMSE) was used as an anomaly score (Section
4.4.2). Following this, the alert threshold was in accordance with the user to limit the num-
ber of alerts below an acceptable number. As a result of this process, the threshold was set
to 0.0040, considering that the anomaly score (RMSE) varies from 0 to 0.0125 (dimension-
less measure), as illustrated in Figure 6.

Figure 6. Distribution of the prediction error.

5.3. Results and Analysis

The model was tested for a period of 45 days (from 10 April 2020 to 25 May 2020).
The results are presented in Table 2. There are three possibilities: (1) confirmed failure or
true positive (the algorithm truly predicted a failure, then it is confirmed by the technician
in the site). (2) Not confirmed failure or false positive (false alarm and no failure was re-
ported in the site). (3) Failure not detected (the algorithm fails to report a failure). As
shown in Table 2, four alerts were issued; two of them were confirmed as a true positive,
and the other two were reported as false positive. During this period, a failure happened
in “AHU 1” but the model failed to detect it. The true positive alerts were issued two days
before they occurred on site.
Sensors 2021, 21, 1044 11 of 15

Table 2. Summary of faults alerts during the test period.

Installation Alert date Feedback

Boiler 2 14 April 2020 Confirmed Failure
Boiler 2 15 April 2020 Confirmed Failure
AHU 2 25 April 020 Not Confirmed
Boiler 2 12 May 2020 Not Confirmed
AHU 1 Not detected Failure not detected

This demonstrates that the algorithm can predict failures in advance. False positives
are accepted since they can correspond to small anomalies that have not been reported as
failures. However, the algorithm needs to be improved and to be tested on a bigger set of
data.
For instance, Figure 7 illustrates the anomalies detected in “Boiler 2”. The anomalies
are projected on the vibration graph of “Boiler 2”, where the x-axis represents the time,
and the y-axis represents the vibration (unit: 1 mGal = 1 × 10 m/s2).

Figure 7. Projection of the anomalies detected in ‘Boiler 2’ on the vibration graph of the ‘Boiler 2’.

The first observation from this case study is the low number of alerts and failures
during this period. This is due to the COVID-19 pandemic lockdown, which coincided
with this period. Indeed, the sport facility was closed, and the HVAC installations were
operating at their minimum regime, making failures less likely to happen.
One of the limitations of this case study is the small duration of data collection. This
period does not take into consideration the changes in the HVAC operations related to
change in seasons or the change in occupancy levels in the facility (big events). These
changes in the regime are the cause of the majority of breakdowns and failures in the
HVAC system according to the FM team in the facility. However, this trial phase was
necessary and required by the FM team to give first feedback and results about the model.
In accordance with the user, data and feedback will be collected for a period of one
year to take into consideration the different changes. This aims to give an accurate evalu-
ation of the model and helps to improve the model for further implementation.

6. Discussion and Conclusions

The primary purpose of this study was to develop a generic framework for the pre-
dictive maintenance for buildings. For this, this study incorporated a literature review and
face-to-face interviews with FM experts. Before designing the framework, identifying data
sources in the building environment was necessary since data are the essence of the ap-
proach. A not exhaustive list was defined containing the most frequent sources in the
buildings. Then the framework was organised into five steps: (1) data collection, (2) data
processing, (3) model development, (4) fault notification, and (5) model improvement. A
case study was addressed to demonstrate the implementation of the framework. This case
Sensors 2021, 21, 1044 12 of 15

study has multiple limitations, mainly related to the duration of the study and the small
size of the collected data. However, this case study illustrates the process to implement
the framework. It also revealed that the topic of predictive maintenance for building in-
stallations presents multiple opportunities as well as multiple challenges. Below, we reit-
erate some observations from the case study.
 Data between diversity and scarcity: data in the building environment are diverse in
terms of sources and in terms of nature. They are generated from the human activity
indoors, from the diverse installations in the building (mechanical, electrical, elec-
tronic etc.), and from the building itself. However, the majority of data are not col-
lected and not stored. Moreover, unlike other industries, there is a lack of open data-
bases containing building data, except some databases mainly focusing on building
energy consumption [66]. As a result, building predictive maintenance has become a
hard and a costly task.
 Return on investment: predictive maintenance strategy offers the facility manager
the possibility to take early action to prevent failures, which improves the lifespan of
the installations and improves the comfort of the inhabitants. However, the imple-
mentation of predictive maintenance may take a significant time to build an effective
model. This presents an important barrier for the facility managers to invest in solu-
tions that can take a significant time before it starts getting profitable.
 Each building is unique: unlike other industries, such as the manufacturing industry,
where multiple installations are the same, each building has a different use, different
architecture, and different occupancy. Thus, two same AHUs are not the same any-
more once they are installed in different buildings. This unicity of buildings presents
multiple opportunities and a wide market for predictive maintenance; moreover, it
reveals several challenges to developing effective and affordable solutions.
In order to address the limitations of the case study, data will be collected for a one-
year period to improve the training dataset. Further work will also focus on implementing
the framework in different buildings to test the limits of scaling over multiple buildings.

Author Contributions: Y.B. designed the framework of the model and carried out the experiment
in this case study, which is a part of his PhD thesis work. Z.L. is the thesis director; he supervised
the thesis work including the work in this paper and he also revised the paper on several occasions.
P.Y. is the co-director of the PhD thesis. He supervised the model development and the different
data processing in this paper. L.D. was responsible for the data acquisition and she also revised the
paper on several occasions. B.B. is a FM expert with over a decade of work experiences in the FM
field. He contributed to this work by managing the relationship with Bouygues Energies and Ser-
vices Company. All authors have read and agreed to the published version of the manuscript.
Funding: This research was carried out as part of the industrial research chair Construction 4.0 at
Centrale Lille Institute, funded by Bouygues Construction, The Métropole Européenne de Lille
(MEL) and the European Regional Development Fund (ERDF).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The data presented in this study are available on request from the
corresponding author. The data are not publicly available due to confidentiality agreement with
the company.
Acknowledgments: The authors are grateful to Bouygues Energies Et Services Facility Mainte-
nance’s Team for facilitating and cooperating in this research.
Conflicts of Interest: The authors declare no conflict of interest.
Sensors 2021, 21, 1044 13 of 15

References
1. Hong, T.; Koo, C.; Kim, J.; Lee, M.; Jeong, K. A review on sustainable construction management strategies for monitoring, diag-
nosing, and retrofitting the building’s dynamic energy performance: Focused on the operation and maintenance phase. Appl.
Energy 2015, 155, 671–707, doi:10.1016/j.apenergy.2015.06.043.
2. Seyedzadeh, S.; Pour Rahimian, F.; Glesk, I.; Roper, M. Machine learning for estimation of building energy consumption and
performance: A review. Vis. Eng. 2018, 6, 5, doi:10.1186/s40327-018-0064-7.
3. Li, Y.; O’Neill, Z. A critical review of fault modeling of HVAC systems in buildings. Build. Simul. 2018, 11, 953–975,
doi:10.1007/s12273-018-0458-4.
4. Matarneh, S.T.; Danso-Amoako, M.; Al-Bizri, S.; Gaterell, M.; Matarneh, R.T. BIM for FM: Developing information requirements
to support facilities management systems Facilities 2019, 38, 378–394.
5. Zhan, J.; Ge, X.J.; Huang, S.; Zhao, L.; Wong, J.K.W.; He, S.X. Improvement of the inspection-repair process with building infor-
mation modelling and image classification. Facilties 2019, 37, 395–414, doi:10.1108/f-01-2018-0005.
6. Lee, H. H. Y.; Scott, D. Overview of maintenance strategy, acceptable maintenance standard and resources from a building
maintenance operation perspective, J. Build. Apprais. 2009, 4, 269–278.
7. T. P. W., Maintenance practices in Hong Kong and the use of the intelligent scheduler J. Qual. Maint. Eng. 2002, 8, 369–380.
8. Pitt, T.J. Data requirements for the prioritization of predictive building maintenance. Facilties 1997, 15, 97–104,
doi:10.1108/02632779710160612.
9. Gunay, B.; Shen, W.; Yang, C. Text-mining building maintenance work orders for component fault frequency. Build. Res. Inf.
2018, 47, 518–533, doi:10.1080/09613218.2018.1459004.
10. LeCun, Y.; Bengio, Y.; Hinton, G. Deep learning. Nature 2015, 521, 436.
11. Boyes, H.; Hallaq, B.; Cunningham, J.; Watson, T. The industrial internet of things (IIoT): An analysis framework. Comput. Ind.
2018, 101, 1–12, doi:10.1016/j.compind.2018.04.015.
12. Mourtzis, D.; Vlachou, E.; Milas, N. Industrial Big Data as a Result of IoT Adoption in Manufacturing. Procedia CIRP 2016, 55,
290–295, doi:10.1016/j.procir.2016.07.038.
13. Lee, S.M.; Lee, D.; Kim, Y.S. The quality management ecosystem for predictive maintenance in the Industry 4.0 era. Int. J. Qual.
Innov. 2019, 5, 4, doi:10.1186/s40887-019-0029-5.
14. BSI Standards Publication. Maintenance Terminology; BS EN 13306:2017; BSI Standards Publication: London, UK 2017; ISBN 978-
0-580-90370-0.
15. Schmidt, B.; Wang, L. Cloud-enhanced predictive maintenance. Int. J. Adv. Manuf. Technol. 2016, 99, 5–13, doi:10.1007/s00170-
016-8983-8.
16. ACachada, A.; Barbosa, J.; Leitño, P.; Gcraldcs, C.A.S.; Deusdado, L.; Costa, J.; Teixeira, C.; Teixeira, J.; Moreira, A.H.J.; Moreira,
P.M.; Romero, L. Maintenance 4.0: Intelligent and Predictive Maintenance System Architecture. In Proceedings of the 2018 IEEE
23rd International Conference on Emerging Technologies and Factory Automation (ETFA), Turin, Italy, 4–7 September 2018,
pp. 139–146, doi:10.1109/ETFA.2018.8502489.
17. Gholami, P.; Hafezalkotob, A. Maintenance scheduling using data mining techniques and time series models. Int. J. Manag. Sci.
Eng. Manag. 2017, 13, 100–107, doi:10.1080/17509653.2017.1314201.
18. Kwak, R.-Y.; Takakusagi, A.; Sohn, J.-Y.; Fujii, S.; Park, B.-Y. Development of an optimal preventive maintenance model based
on the reliability assessment for air-conditioning facilities in office buildings. Build. Environ. 2004, 39, 1141–1156,
doi:10.1016/j.buildenv.2004.01.029.
19. Halim, T.; Tang, L.-C. A graphical approach for confidence limits of optimal preventive maintenance cycles. Qual. Reliab. Eng.
Int. 2009, 25, 199–213, doi:10.1002/qre.962.
20. Sandeepan, C.; Dubey, K. Mechanical Vibration Analysis of HVAC system and Its Optimization Techniques, Adv. Res. Electr.
Electron. Eng. 2015, 2, 77–82.
21. Ning, M.; Zaheeruddin, M.; Chen, Z. Fuzzy-Set Based HVAC System Uncertainty Analysis. In Proceedings of the NAFIPS
2006—2006 Annual Meeting of the North American Fuzzy Information Processing Society, Montréal, QC, Canada, 3–6 June
2006; pp. 229–234.
22. Wang, L.; Hong, T. Modeling and Simulation of HVAC Faulty Operation and Performance Degradation due to Maintenance
Issues. In Proceedings of the ASIM 2012—1st Asia conference of International Building Performance Simulation Association,
25–27 November 2014.
23. Mattera, C.; Quevedo, J.; Escobet, T.; Shaker, H.R.; Jradi, M. A Method for Fault Detection and Diagnostics in Ventilation Units
Using Virtual Sensors. Sensors 2018, 18, 3931, doi:10.3390/s18113931.
24. Saxena, L.M.A., Jr.; Knapp, G.M. Statistical-based or condition-based preventive maintenance? J. Qual. Maint. Eng. 1995, 1, 46–
59, doi:10.1108/13552519510083156.
25. Peng, Y.; Dong, M.; Zuo, M.J. Current status of machine prognostics in condition-based maintenance: A review. Int. J. Adv.
Manuf. Technol. 2010, 50, 297–313, doi:10.1007/s00170-009-2482-0.
26. Krizhevsky, A.; Sutskever, I.; Hinton, G.E. Imagenet classification with deep convolutional neural networks. In Proceedings of
the 25th International Conference on Neural Information Processing Systems, Lake Tahoe, NV, USA, 3–6 December 2012; pp.
1097–1105.
27. Otter, D.W.; Medina, J.R.; Kalita, J.K. A Survey of the Usages of Deep Learning for Natural Language Processing. IEEE Trans.
Neural Netw. Learn. Syst. 2020, 1–21, doi:10.1109/tnnls.2020.2979670.
Sensors 2021, 21, 1044 14 of 15

28. Nguyen, P.; Tran, T.; Wickramasinghe, N.; Venkatesh, S. Deepr: A Convolutional Net for Medical Records. IEEE J. Biomed. Health
Inform. 2017, 21, 22–30.
29. Silver, D.; Huang, A.; Maddison, C.J.; Guez, A.; Sifre, L.; Driessche, G.V.D.; Schrittwieser, J.; Antonoglou, I.; Panneershelvam,
V.; Lanctot, M.; et al. Mastering the game of Go with deep neural networks and tree search. Nature 2016, 529, 484–489,
doi:10.1038/nature16961.
30. Chalapathy, R.; Chawla, S. Deep learning for anomaly detection: A survey. arXiv 2019, arXiv:1901.03407v2.
31. Tehrani, M.M.; Beauregard, Y.; Rioux, M.; Kenne, J.-P.; Ouellet, R. A Predictive Preference Model for Maintenance of a Heating
Ventilating and Air Conditioning System. IFAC Pap. 2015, 48, 130–135, doi:10.1016/j.ifacol.2015.06.070.
32. Mahamad, A.K.; Saon, S.; Hiyama, T. Predicting remaining useful life of rotating machinery based artificial neural network.
Comput. Math. Appl. 2010, 60, 1078–1087, doi:10.1016/j.camwa.2010.03.065.
33. Pham, M.T.; Kim, J.-M.; Kim, C.H. Deep Learning-Based Bearing Fault Diagnosis Method for Embedded Systems. Sensors 2020,
20, 6886, doi:10.3390/s20236886.
34. Ian Goodfellow and Yoshua Bengio and Aaron Courville, Deep Learning; MIT Press Cambridge, 2016.
35. Gasparin, A.; Lukovic, S.; Alippi, C. Deep Learning for Time Series Forecasting: The Electric Load Case. arXiv 2019,
arXiv:1907.09207.
36. Hadsell, R.; Sermanet, P.; Ben, J.; Erkan, A.; Scoffier, M.; Kavukcuoglu, K.; Muller, U.; LeCun, Y. Learning long-range vision for
autonomous off-road driving. J. Field Robot. 2009, 26, 120–144, doi:10.1002/rob.20276.
37. Fan, C.; Xiao, F.; Zhao, Y.; Wang, J. Analytical investigation of autoencoder-based methods for unsupervised anomaly detection
in building energy data. Appl. Energy 2018, 211, 1123–1135, doi:10.1016/j.apenergy.2017.12.005.
38. Morán, A.; Alonso, S.; Prada, M.A.; Fuertes, J.J.; Díaz, I.; Domínguez, M. Analysis of Parallel Process in HVAC Systems Using
Deep Autoencoders. Progr. Ing. Nat. 2017, 744, 15–26, doi:10.1007/978-3-319-65172-9_2.
39. Li, C.; Ding, Z.; Zhao, D.; Yi, J.; Zhang, G. Building Energy Consumption Prediction: An Extreme Deep Learning Approach.
Energies 2017, 10, 1525, doi:10.3390/en10101525.
40. Baldi, P. Autoencoders, Unsupervised Learning, and Deep Architectures. JMLR Workshop Conf. Proc. 2012, 27, 37–49.
41. Wang, Y.; Yao, H.; Zhao, S. Auto-encoder based dimensionality reduction. Neurocomputing 2016, 184, 232–242, doi:10.1016/j.neu-
com.2015.08.104.
42. Murphree, J. Machine learning anomaly detection in large systems. IEEE Autotestcon 2016, 1–9, doi:10.1109/autest.2016.7589589.
43. Araya, D.B.; Grolinger, K.; El Yamany, H.F.; Capretz, M.A.M.; Bitsuamlak, G. An ensemble learning framework for anomaly
detection in building energy consumption. Energy Build. 2017, 144, 191–206, doi:10.1016/j.enbuild.2017.02.058.
44. Bouabdallaoui, Y.; Lafhaj, Z.; Yim, P.; Ducoulombier, L.; Bennadji, B. Natural Language Processing Model for Managing Mainte-
nance Requests in Buildings. Buildings 2020, 10, 160, doi:10.3390/buildings10090160.
45. Su, Y.; Zhao, Y.; Niu, C.; Liu, R.; Sun, W.; Pei, D. Robust Anomaly Detection for Multivariate Time Series through Stochastic
Recurrent Neural Network. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery &
Data Mining, Anchorage, AK, USA, 4–8 August 2019; Volume 1485, pp. 2828–2837.
46. Fan, C.; Xiao, F.; Yan, C. A framework for knowledge discovery in massive building automation data and its application in
building diagnostics. Autom. Constr. 2015, 50, 81–90, doi:10.1016/j.autcon.2014.12.006.
47. Moreno, M.V.; Dufour, L.; Skarmeta, A.F.; Jara, A.J.; Genoud, D.; Ladevie, B.; Bézian, J.-J.; Cano, M.V.M. Big data: The key to
energy efficiency in smart buildings. Soft Comput. 2015, 20, 1749–1762, doi:10.1007/s00500-015-1679-4.
48. 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,
doi:10.1109/jiot.2017.2647881.
49. Schmidt, M.; Åhlund, C. Smart buildings as Cyber-Physical Systems : Data-driven predictive control strategies for energy effi-
ciency. Renew. Sustain. Energy Rev. 2018, 90, 742–756.
50. Carli, R.; Cavone, G.; Ben Othman, S.; Dotoli, M. IoT Based Architecture for Model Predictive Control of HVAC Systems in
Smart Buildings. Sensors 2020, 20, 781, doi:10.3390/s20030781.
51. Akkaya, K.; Guvenc, I.; Aygun, R.; Pala, N.; Kadri, A. IoT-based occupancy monitoring techniques for energy-efficient smart
buildings. In Proceedings of the 2015 IEEE Wireless Communications and Networking Conference Workshops (WCNCW), New
Orleans, LA, USA, 9–12 March 2015; pp. 58–63, doi:10.1109/WCNCW.2015.7122529.
52. Bourdeau, M.; Zhai, X.; Nefzaoui, E.; Guo, X.; Chatellier, P. Modeling and forecasting building energy consumption: A review
of data-driven techniques. Sustain. Cities Soc. 2019, 48, doi:10.1016/j.scs.2019.101533.
53. Gunay, B.; Shen, W.; Newsham, G. Data analytics to improve building performance: A critical review. Autom. Constr. 2019, 97,
96–109, doi:10.1016/j.autcon.2018.10.020.
54. Marmo, R.; Nicolella, M.; Polverino, F.; Tibaut, A.; Marmo, R. A Methodology for a Performance Information Model to Support
Facility Management. Sustainability 2019, 11, 7007, doi:10.3390/su11247007.
55. Zhang, J.; Seet, B.-C.; Lie, T.T. Building Information Modelling for Smart Built Environments. Buildings 2015, 5, 100–115,
doi:10.3390/buildings5010100.
56. Lu, Q.; Chen, L.; Lee, S.; Zhao, X. Activity theory-based analysis of BIM implementation in building O&M and first response.
Autom. Constr. 2018, 85, 317–332, doi:10.1016/j.autcon.2017.10.017.
57. Gerrish, T.; Ruikar, K.; Cook, M.J.; Johnson, M.; Phillip, M.; Lowry, C. BIM application to building energy performance visual-
isation and management: Challenges and potential. Energy Build. 2017, 144, 218–228, doi:10.1016/j.enbuild.2017.03.032.
Sensors 2021, 21, 1044 15 of 15

58. Dong, B.; O’Neill, Z.; Li, Z. A BIM-enabled information infrastructure for building energy Fault Detection and Diagnostics.
Autom. Constr. 2014, 44, 197–211, doi:10.1016/j.autcon.2014.04.007.
59. Chen, W.; Chen, K.; Cheng, J.C.; Wang, Q.; Gan, V.J. BIM-based framework for automatic scheduling of facility maintenance
work orders. Autom. Constr. 2018, 91, 15–30, doi:10.1016/j.autcon.2018.03.007.
60. Motamedi, A.; Hammad, A.; Asen, Y. Knowledge-assisted BIM-based visual analytics for failure root cause detection in facilities
management. Autom. Constr. 2014, 43, 73–83, doi:10.1016/j.autcon.2014.03.012.
61. Nargesian, F.; Samulowitz, H.; Khurana, U.; Khalil, E.B.; Turaga, D. Learning Feature Engineering for Classification. In Pro-
ceedings of the Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, Melbourne, Australia,
19–25 August 2017; pp. 2529–2535.
62. Zheng, A.; Casari, A. Feature Engineering for Machine Learning; O’Reilly Media, Inc.: Sebastopol, CA, USA, 2018.
63. Sagheer, A.; Kotb, M. Unsupervised Pre-training of a Deep LSTM-based Stacked Autoencoder for Multivariate Time Series
Forecasting Problems. Sci. Rep. 2019, 9, 1–16, doi:10.1038/s41598-019-55320-6.
64. Zhuang, F.; Qi, Z.; Duan, K.; Xi, D.; Zhu, Y.; Zhu, H.; Xiong, H.; He, Q. A Comprehensive Survey on Transfer Learning. Proc.
IEEE 2021, 109, 43–76.
65. Mekki, K.; Bajic, E.; Chaxel, F.; Meyer, F. A comparative study of LPWAN technologies for large-scale IoT deployment. ICT
Express 2019, 5, 1–7, doi:10.1016/j.icte.2017.12.005.
66. Miller, C.; Arjunan, P.; Kathirgamanathan, A.; Fu, C.; Roth, J.; Park, J.Y.; Balbach, C.; Gowri, K.; Nagy, Z.; Fontanini, A.D.; et al.
The ASHRAE Great Energy Predictor III competition: Overview and results. Sci. Technol. Built Environ. 2020, 26, 1427–1447,
doi:10.1080/23744731.2020.1795514.