Journal of Building Engineering

Journal of Building Engineering
Enhancing Performance Evaluation of Low-Cost Inclinometers for the Long-Term

monitoring of buildings
--Manuscript Draft--
Manuscript Number: JBE-D-23-06911R3
Article Type: Research Paper
Section/Category: Structural health monitoring
Keywords: long-term building monitoring; low-cost sensor; inclinometer; Artificial intelligence;

neural network
Corresponding Author: Ye Xia, Ph.D.

Tongji University
Shanghai, Shanghai CHINA
First Author: Fidel Lozano Galant
Order of Authors: Fidel Lozano Galant
S. Emadi
Milad Komarizadehasl
Jesús González-Arteaga
Ye Xia, Ph.D.
Abstract: The development of low-cost structural and environmental sensors has sparked a
transformation across numerous fields, offering cost-effective solutions for monitoring
infrastructures and buildings. However, a critical challenge associated with these
devices is their long-term durability and reliability. Surprisingly, despite its interest, the
literature does not present any solutions to assure the long-term performance of low-
cost sensors. To address this gap, this paper proposes an innovative Artificial
Intelligence (AI)-based approach for evaluating the long-term performance of the low-
cost inclinometer LARA (Low-Cost Adaptable Reliable Anglemeter). This method
automatically compares the inclinations of actual on-site measurements with the
predicted ones under real environmental conditions. Over time, if the discrepancies
between both measurements surpass a statistical threshold, it may indicate potential
inaccuracies in the low-cost inclinometer or reveal a structural issue. In such cases, it
becomes imperative to consider sensor recalibration or conduct a comprehensive
structural diagnosis. The effectiveness and applicability of the proposed tool are
demonstrated through the long-term study conducted on a real steel frame located in
Spain.
Suggested Reviewers: Hong Yang

Old Dominion University
[email protected]
Dr. Yang has strong academic background in structural engineering, bridge
engineering, and transportation engineering. He has published more than 20 journal
paper in the past 3 years.
Yang Jann N.
University of California Irvine
[email protected]
Prof. Jann N. Yang is a full professor in the field of structural engineering with more
than 20 year field experience.
Opposed Reviewers:
Response to Reviewers:
Powered by Editorial Manager® and ProduXion Manager® from Aries Systems Corporation
Cover Letter
Dear Editors,
Please find the manuscript entitled ‘Enhancing Performance Evaluation of Low-

Cost Inclinometers for the Long-Term monitoring of buildings’, which we would
like to submit to ‘Journal of Building Engineering’ for potential publication without
any precedential submission or peer review.
Micro Electro-Mechanical System (MEMS) accelerometer-based inclinometers have
become increasingly prevalent in civil engineering and architectural applications due
to their cost-effectiveness. However, a critical challenge associated with these devices
is their long-term durability and reliability. Commercial inclinometers are often
employed to calibrate the performance of low-cost solutions over time, leading to
escalated monitoring costs and diminished economic advantages. To address this
limitation, this paper proposes an innovative AI-based approach for evaluating the long-
term performance of the low-cost inclinometer, LARA (Low-Cost Adaptable Reliable
Anglemeter). The proposed AI-driven tool utilizes a convolutional neural network to
predict the inclination of a monitored steel structure under real environmental
conditions. By comparing the predicted responses with actual on-site measurements,
potential inaccuracies in the low-cost inclinometer can be automatically identified, thus
indicating the need for recalibration. The effectiveness and applicability of the proposed
tool are demonstrated through a comprehensive study conducted on a real steel frame
located in Spain.
Further, the topic developed in this study as well as the results obtained are quite
relevant to the scope of the journal (Structural health monitoring, Structural testing.)
and will be of interest to the researchers, engineers and managers.
This manuscript has not been published or presented elsewhere in part or in entirety
and is not under consideration by another journal. We have read and understood your
journal’s policies, and we believe that neither the manuscript nor the study violates any
of these. There are no conflicts of interest to declare.
We appreciate your time and look forward to your response.
Sincerely,
Ye Xia
--------------------------------
Ye Xia, Ph.D.
Department of Bridge Engineering, Tongji University
E-mail: [email protected];
Website: shmc.tongji.edu.cn
Response to Reviewers
JBE-D-23-06911R2 Enhancing Performance Evaluation of Low-Cost Inclinometers for the Long-Term

monitoring of buildings
Reviewer #1:
The authors sincerely appreciate the positive comments of Reviewer 1 and are very grateful for his/her
suggestions and observations. These comments certainly have improved the quality of the paper. A detailed
answer is presented below.
Comments and Suggestions for Authors
Thank you to the authors for submitting the amended version of your article. I appreciate the effort spent
for amending the comments and suggestions. After carefully reviewing the revised the manuscript, I
recommend for eventually publishing this article with the following details:
1. Authors need consider further review to clarify with the English and grammars for the whole
manuscripts. The manuscript has been reviewed by Elsevier Language Editing Services. The
resulting certificate is provided below:
2- The references and citation need to check according to the journal format before making it final
version for this publication. As suggested, the references and citations have been adapted to the
Journal format described in Journals’ site: https://www.sciencedirect.com/journal/journal-of-
building-engineering/publish/guide-for-authors
1
Reviewer #2:
The authors sincerely appreciate the positive comments of Reviewer 1 and are very grateful for his/her
suggestions and observations. These comments certainly have improved the quality of the paper. A detailed
answer is presented below.
Comments and Suggestions for Authors
Comments are considered and answered properly, and the paper is recommended for publication. Authors
thank the Reviewer for accepting the paper.
2
Highlights
 An innovative method based on artificial intelligence has been proposed to evaluate
the long-term performance of low-cost inclinometers
 Real time automatic identification of potential issues caused by inaccuracies or
structural pathology in low-cost inclinometers
 Research on steel frames in Spain has shown their ability to accurately evaluate the
long-term performance of LARA under real-world conditions

Main Manuscript File [This item is SENT to Production, so please Click here to view linked References
ensure for the latest (Clean) version]
1
1
Enhancing Performance Evaluation of Low-Cost Inclinometers for
22
3
the Long-Term monitoring of buildings
4
35 F. Lozano1, S. Emadi2, M. Komarizadehasl2, J. González-Arteaga3 and Y. Xia4,*
6
7 1
4 PhD candidate, Dept. of Civil and Building Engineering, Universidad de Castilla-La Mancha (UCLM), 13071, Ciudad Real, Spain.
8 2
5 PhD, Dept. of Civil and Environment Engineering, Universitat Politècnica de Catalunya (UPC), BarcelonaTech. C/ Jordi Girona 1-3, 08034, Barcelona, Spain.
6
9 3
Associate Professor, Geoenvironmental Group, Universidad de Castilla-La Mancha (UCLM). Av. Camilo Jose Cela s/n, 13071, Ciudad Real, Spain.
10
7 4
Associate Professor, Department of Bridge Engineering, Tongji University, 1239 Siping Rd., Shanghai 200092, China.
11
8 * Correspondence email: [email protected]
12
9
13
14
10 Abstract: The development of low-cost structural and environmental sensors has revolutionized monitoring practices across numerous fields,
15
11 enabling cost-effective solutions for infrastructure and building health assessment. However, a critical challenge associated with these sensors is
16
12 their long-term durability and reliability. Surprisingly, despite the significant interest in these low-cost devices, the literature does not present any
17
13
18 solutions for ensuring their long-term performance. To address this gap, this study proposes an innovative artificial intelligence-based approach
14
19 for evaluating the long-term performance of low-cost inclinometers using a low-cost adaptable reliable anglemeter. This method automatically
15
20 compares the inclinations of actual onsite measurements with predicted values under real environmental conditions. Over time, if the discrepancies
21
16 between both measurements surpass a predefined statistical threshold, it may signal potential inaccuracies in the low-cost inclinometer, thereby
22
17 suggesting the need for recalibration or presence of structural anomalies. The effectiveness and applicability of the proposed tool were demon-
23
18 strated through a long-term study conducted on a real steel frame in Spain.
24
25
19 Keywords: long-term building monitoring; low-cost sensor; inclinometer; artificial intelligence; multilayer perceptron; neural network
26
20
27
28
29
30
21 1. Introduction
31
32
22 Structural system identification (SSI) methods play a crucial role in architectural, engineering, and construction industries, serving
33
23 as indispensable tools for informed maintenance decision-making [1]. These methods facilitate the proactive identification of struc-
34
24 tural issues and effective risk mitigation by enabling early detection, localization, and quantification of structural damage [2]. SSI
35
25 methods rely on valuable information provided by structural health monitoring (SHM) systems [3] equipped with sensors such as
36
26 accelerometers, strain gauges, and inclinometers. These sensors capture data related to structural responses and variations, providing
37
27 insights into the presence, location, and severity of damage. The combined approach of SSI and SHM enables early detection,
38
28 localization, and quantification of damage, ultimately contributing to the overall safety of buildings.
39
29 Inclinometers, also known as tilt sensors, play a pivotal role in SHM systems by measuring the angular inclination of a structure
40
30 with respect to the vertical axis. They provide critical insights into the mechanical behavior and stability of structures. Their signif-
41
31 icance lies in their ability to detect subtle changes in the inclination, which are indicators of potential structural deformation or
42
32 damage.
43
33
44 Commercial inclinometers are essential in structural monitoring research because of their wide availability, reliability, and estab-
34
45 lished performance characteristics. A key advantage of these sensors is their well-established calibration and validation processes.
35
46 Manufacturers rigorously test and calibrate these devices to ensure accurate and consistent measurements, providing researchers
36
47 with a reliable baseline for their monitoring and identification endeavors. In fact, the documented specifications and performance
37
48 metrics associated with commercial inclinometers have contributed to their widespread acceptance in the scientific literature.
49
38 Although commercial inclinometers are documented in the literature with a wide range of technical specifications and prices, their
50
39 costs typically exceed those of other inexpensive monitoring devices, such as the low-cost adaptable reliable anglemeter (LARA)
51
40 [4]. Compared with commercial alternatives, low-cost inclinometers offer many advantages that make them highly appealing for
52
41 various applications. First, their affordability allows for large-scale deployment, enabling comprehensive monitoring of multiple
53
42 locations or structures within a given budget. This scalability is particularly beneficial when high-density sensor networks are re-
54
43 quired. Moreover, the flexibility and adaptability of low-cost solutions allow for customization and integration with existing sys-
55
44 tems, fostering innovation and opening up opportunities for tailored monitoring approaches in diverse fields. Furthermore, the rap-
56
45 idly advancing technology behind low-cost sensors is continually improving their capabilities, ensuring that they keep pace with
57
46 evolving monitoring needs. These advantages have led to increased attention in developing low-cost monitoring solutions in recent
47
58 years. A comprehensive review of the applications of low-cost solutions to building monitoring has been presented by Mobaraki et
48
59 al. [5].
60
61
62
63
64
65
49 Nevertheless, the inherent cost-effectiveness of low-cost sensors often comes with trade-offs, impacting factors such as precision,
50
1 calibration stability, and susceptibility to environmental conditions. These challenges have prompted essential considerations re-
51
2 garding the long-term performance and accuracy of low-cost sensors. Indeed, most low-cost inclinometers primarily focus on short-
52
3 term monitoring, necessitating periodic verification using commercial solutions to ensure long-term accuracy and reliability. How-
53 ever, this onsite verification process faces two challenges. First, it lacks automation, and second, it inevitably increases monitoring
4
54 costs, significantly reducing the economic advantages of utilizing low-cost inclinometers. To address these limitations, this study
5
55 proposes an innovative artificial intelligence (AI)-based approach to automate the evaluation of the long-term performance of the
56
6 low-cost inclinometer LARA. The proposed methodology uses an artificial neural network to predict the inclination of a monitored
57
7 structure under specific environmental conditions. By comparing the discrepancies between the actual inclinations onsite and the
58
8 predicted ones, potential issues arising from inaccuracies in the performance of the low-cost inclinometer can be automatically
59 9 identified, suggesting the need for sensor recalibration. Additionally, the system sends digital alarms through a dedicated Internet
10
60 of Things platform when a statistical threshold is exceeded. The proposed methodology was validated on an actual steel frame in
11
61 Spain based solely on the inclinations and environmental data (specifically, humidity and temperature) obtained from low-cost
12
62 devices. Notably, the primary limitation of the proposed methodology lies in the fact that the developed AI tool cannot be directly
13
63 employed in other case studies because it requires individualized training using onsite monitoring data for each structure, which is
14
64 a time-consuming process. To ensure that the proposed methodology is well suited for practical applications, the authors plan to
15
65 analyze different structural configurations and loading conditions in the near future.
16
66
17 The remainder of this paper is organized as follows: Section 2 presents a literature review encompassing the key topics addressed
67
18 in this study. Section 3 provides a detailed analysis of the main characteristics of LARA, including its design and key components.
68 Section 4 introduces a novel AI tool for studying the effects of environmental factors, specifically temperature and humidity, on the
19
69 inclinations measured by LARA in actual steel structures. The application of this methodology is described in Section 5, where an
20
70 actual steel frame in a building is examined to understand the influence of environmental parameters on LARA inclinations. Finally,
21
71 Section 6 presents the primary conclusions drawn from the study, summarizes the main findings, and discusses their implications.
22
72
23
24
73 2. Literature Review
25
26
74 This section begins with an overview of the primary SSI methods found in the literature. Subsequently, a detailed examination of
27
75 the key characteristics of commercial inclinometers discussed in the existing literature is presented. The section then shifts its focus
28
76 to the features and challenges associated with prominent low-cost inclinometer alternatives outlined in the literature, along with a
29
77 discussion of their primary technical challenges.
30
78
31
79 2.1 SSI
80
32
81
33 Sirca and Adeli [6] comprehensively reviewed the system identification methods in structural engineering. Their review categorizes
82
34 SSI techniques into five main approaches: (1) conventional model-based [7-10], (2) biologically inspired, (3) signal processing-
83
35 based, (4) chaos theory and fractal concept-based [11-13], and (5) multiparadigm approaches. Conventional model-based ap-
84
36 proaches typically rely on a structural model to identify the structural parameters from onsite structural responses. However, their
85
37 applicability may be limited when dealing with large real-world structures that exhibit nonlinear behaviors. Biologically inspired
86
38 approaches are based on soft computing techniques, such as neural networks [14-16], genetic algorithms [10,17,18], and particle
87
39 swarm optimization [19-21]. Signal processing-based approaches employ signal processing techniques, such as wavelets [22-24].
88
40 In an attempt to tackle complex, nonlinear problems, multiparadigm approaches integrate two or more of the previously mentioned
89
41 computing strategies. Examples of this approach are presented in [25-27]. Based on the nature of the measured structural response,
90
42 SSI methods can be categorized as static [28] or dynamic [29].
91
43
92 2.2 Inclinometers
44
93
45
94 Among the diverse SHM devices, inclinometers, also known as tilt sensors, merit special recognition owing to their extensive use
46
95 in various disciplines, particularly in the SHM of bridges [30] and buildings [31]. Huseynov et al. [32] and Hester et al. [33] utilized
47
96 rotation measurements to detect damage in bridges, whereas Lei et al. [34] employed rotations to identify damage in building beams.
48
97 The essential characteristics of commercially available inclinometers are summarized in Table 1, which includes details such as
49
98 inclinometer name, measurement range, resolution, sampling rate, price, and corresponding references.
50
99
51
52
100
53
54
55
56
57
58
59
60
61
62
63
64
65
Table 1: Characteristics of some of the commercially available inclinometers along with the low-cost alternatives
LARA and MPU9250.
1
Measurement Sampling Price
2 Model
Range (degrees)
Resolution (degrees)
Rate (Hz) (€)
Axis References
3
EL Tiltmeter SC ± 0.66° 30×10-5° - 15001 Uniaxial [35]
4
5 ZEROTR-ONIC ± 0.50° 100×10-5° 10 3950 Uniaxial [36]
6 JDI 200 ± 1.00° 10×10-5° 125 2250 Dual axis [37]
7 T935 ± 1.00° 6×10-5° 10 1696 Dual axis [38]
8
ACA 2200 ± 0.50° 10×10-5° 20 710 Dual axis [39]
9
10 HI-INC ± 15.00° 100×10-5° 100 650 Dual axis [40]
11 SCA103T ± 15.0° 120×10-5° 18 - Uniaxial [41]
12 LSM9DS0 ± 20.0° 16000×10-5° - - Uniaxial [42]
13
SCA100T ± 7.0° 250×10-5° 20 - Dual axis [43]
14
15 LARA ± 4.00° Up to 200×10-5° Up to 333 100 Dual axis [4]
16 MPU9250 ± 4.0° 5500×10-5° 333 60 Dual axis [44]
17
101
18
102 Analysis of the commercial solutions listed in Table 1 reveals significant price variations, ranging from €350 to €15001. High-
19
103 priced inclinometers (such as the EL Tiltmeter SC) usually provide a lower measurement range and higher resolution.
20
104
21
105 2.3 Low-cost inclinometers
22
106
23
107 To explore low-cost solutions for inclinometer development, a notable approach is the use of microelectromechanical system
24
108 (MEMS) accelerometers. The literature provides numerous examples of the successful employment of this technology. For instance,
25
109 Hoang et al. [41] developed an orientation system for heavy industrial applications, while Ha et al. [43] designed a low-cost incli-
110
26 nometer for estimating ground movements. Ruzza et al. [42] used MEMS sensors to measure landslide inclinations, whereas Yu et
111
27 al. [45] focused on monitoring swings in large-scale structures. Ha et al. [30] and Kumar et al. [46] proposed using wireless incli-
112
28 nometers for SHM applications. Table 1 presents the detailed characteristics of low-cost inclinometer solutions.
113
29
114
30 Despite the advantages of low-cost inclinometers, this type of monitoring system presents several inherent challenges. One primary
115
31 concern is their potentially lower accuracy than more expensive commercial alternatives, which limits their application in cases
116
32 requiring high precision. Owing to cost constraints, low-cost MEMS accelerometers may exhibit reduced precision and calibration
117
33 stability, leading to potential measurement errors. Factors such as noise, nonlinearity, and cross-axis sensitivity can introduce inac-
118
34 curacies into the measured acceleration data. Komarizadehasl et al. [4] developed a monitoring system known as LARA to address
119
35 the accuracy issues associated with low-cost inclinometers. Unlike other solutions in the literature, LARA employs five MPU9250
120
36 MEMS accelerometers to enhance its performance. Table 1 provides an overview of the main characteristics of this low-cost incli-
121 nometer, including a comparison with a single MPU9250 sensor used as an inclinometer. Analysis of the table indicates a significant
37
122 reduction in the noise error of LARA (0.002°) compared with other low-cost solutions such as xSCA100T (0.0025°), MPU9250
38
123 sensor (0.055°), and LSM9DS0 (16000×10-5°).
39
124
40
125 Furthermore, most low-cost MEMS accelerometers, including the ADXL345 sensor, are sensitive to environmental parameters such
41
126 as temperature variations, which can significantly impact their performance and reliability. Mukhopadhyay et al. [47] stated that
42
127 environmental effects can lead to drift and calibration shifts, resulting in substantial deviations between the measured and actual
43
128 acceleration values. Moreover, changes in temperature and humidity can potentially compromise both the durability and long-term
44
129 stability of these monitoring devices because they are more susceptible to mechanical wear, sensitivity degradation, or premature
45
130 failure. It is worth noting that the impact of environmental parameters is not limited to low-cost inclinometers; even certain com-
46
131 mercial inclinometers that utilize fluids may also be affected by environmental changes [48].
47
132
48
133 In the literature, the methods proposed to mitigate the adverse effects of environmental parameters on MEMS accelerometers can
49
134 be categorized into two primary groups: hardware and software methods [47]. The hardware method involves circuit control and
50
135 structural optimization to enhance device features. Examples of applications of the hardware compensation method can be found in
51
136 the studies by Wang et al. [49] and Liu and Ma [50]. The software compensation method aims to improve the mathematical equations
52
137 used to obtain measurement readings. Various mathematical approaches have been employed to compensate for the effects of tem-
138
53 perature on MEMS accelerometers. Notable examples include the use of polynomials, as demonstrated by Ruzza [51]; finite element
139
54 analysis, as employed by Zhang [52]; cubic spline interpolation, as utilized by Khankalantary [53]; and swarm optimization, as
140
55 proposed by Yu [54]. Caspani et al. [55] employed a similar software compensation approach to correct the effects of temperature
141
56 on the long-term responses of prestressed concrete bridges using a parametric equation.
142
57
143
58 Similar to accelerometers, low-cost inclinometers based on MEMS technology are susceptible to environmental factors. Several
144
59 studies, such as those conducted by Yuan et al. [56], Fontanella et al. [57,58], and Yang et al. [59], have addressed the thermal
145
60 recalibration of low-cost inclinometers. Notably, in contrast to most MEMS accelerometers and inclinometers, LARA is not sensi-
146
61 tive to variations in environmental parameters.
147
62
63
64
65
148 Environmental parameters, such as temperature and humidity, can introduce variations in material properties and considerably affect
149
1 structural performance. A clear correlation exists between temperature variation and deformation of structures. Consequently, nu-
150
2 merous researchers have examined the effects of changes in environmental parameters on the static and dynamic responses of
151
3 structures. Yao et al. [60] studied the effects of temperature and humidity changes in a wooden frame structure, whereas Cai et al.
152 [61] analyzed the impact of temperature on the natural vibration characteristics of a reinforced concrete beam. Zhao and Jin [62]
4
153 investigated the effects of temperature on the dynamic response of a steel crane, while Ribeiro et al. [63] provided a systematic
5
154 review of the effects of environmental conditions on SHM. In the literature review, no approaches similar to those proposed in this
155
6 study for automating the assessment of the long-term performance of low-cost inclinometers have been identified. This highlights
156
7 the novelty of the proposed methodology in addressing this gap in existing research.
157
8
9
158
10 3. LARA
11
159
12 LARA, which is a low-cost, adaptable, and reliable analyzer [64], represents a groundbreaking evolution of the cost-hyperefficient
160 Arduino product [65]. This device utilizes open-source prototyping tools, specifically Arduino and Raspberry Pi technologies, and
13
161 a low-cost MEMS accelerometer (MPU9250). Although this type of accelerometer is not commonly used in civil engineering and
14
162 architectural structures because of its limited precision, it addresses this issue using a pioneering approach that was first documented
15
163 in the literature. LARA incorporates the average readings of five different accelerometers placed at the same point, effectively
16
164 functioning in conjunction with a multiplexer. Figure 1 illustrates the wiring configuration of various sensors with an Arduino Due
17
165 microcontroller and its compact version. A compact version of the LARA sensor is shown in Figure 1. By processing the measure-
18
166 ments obtained from these devices, LARA compensates for individual accelerometer errors, resulting in accuracies comparable to
19
167 those of commercial-grade devices. The application of LARA for dynamic monitoring has been validated in actual buildings [66]
20
168 and bridges [67].
21
169
22
23 (a)
24 MPU9250
25
26 TCA9548A
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41 (b)
42
43
TCA9548A MPU9250
44
45
46
47
48
49
50
Figure 1: LARA sensor: (a) Wiring connection, and (b) Compact version.
51
170
52
171 In contrast to most accelerometers in the literature, LARA offers the unique capability of a remote configuration, allowing it to
53
172 function as either an accelerometer or inclinometer [4]. When operating as an inclinometer, LARA employs accelerometer gyro-
54
173 scopes along the X-, Y-, and Z-axes (forming a triaxial inclinometer) to accurately measure the orientation angle of an object in
55
174 relation to the force of gravity. This versatility distinguishes LARA from its counterparts and enhances its usefulness in various
56
175 applications.
57
176
58
177 To streamline the process, the main frame of LARA has been redesigned. This redesign enables it to automatically initiate data
59
178 acquisition when it connects to a power source and stops the process after 2 min. When the Raspberry Pi boots up, it establishes an
60
179 internet connection through Wi-Fi, LAN, or SIM cards and mounts a Google Drive on its hard drive. This configuration allows
61
180 LARA to download and store Python codes containing instructions for data acquisition tasks. Following these instructions, LARA
62
181 triggers vibration acquisition at a specified date and time. Additionally, a task is programmed into the Raspberry Pi operating system
63
64
65
182 to regularly monitor and transfer the acquired data files to Google Drive. The hard disk will not become overloaded or reach its
183
1 storage capacity without an internet connection. Any collected data will be retained on the Raspberry Pi hard drive without an
184
2 internet connection until the connection is restored or the memory card becomes full.
185
3
186 The LARA information flow is described in detail in [67]. This process is depicted in Figure 2a and comprises the following steps:
4
187 5 1- LARA initialization: When powered on, LARA is set to establish an internet connection using Wi-Fi, LAN, or SIM cards,
188 6 as long as the Wi-Fi credentials match the predefined settings within the LARA configuration.
7
189 2- Google Drive: Once internet connectivity is established, LARA gains entry into a predetermined Google Drive repository,
8
190 where all monitoring information is stored.
9
191
10 3- Operational instructions: LARA retrieves the open-source Python code from the designated Google Drive to commence
192
11 data acquisition. This adaptable code allows users to define settings, such as scheduled vibration acquisition and data
193
12 acquisition duration, without directly reprogramming the Raspberry Pi for every specific task.
13
194 4- Data collection: Upon completion of data acquisition, the files obtained are transferred to a dedicated folder on a linked
14
195 Google Drive. This design guarantees the unburdened state of LARA's local hard drive, as long as the internet connection
15
196 is consistently maintained. Notably, in the event of internet connectivity loss, the gathered data will be retained on LARA's
16
197 local hard drive until the connection is reestablished. The option of employing external hard drives to expand the storage
17
198 capacity is also possible.
18
199
19 5- Data processing: During this phase, the data acquired from LARA are postprocessed with the primary objective of estimat-
200
20 ing the inclination based on the compiled information. Moreover, this stage allows the integration of additional programs
201
21 or processes to increase estimation precision or execute specific computations. This phase demonstrates adaptability, which
202
22 can be tailored to meet the unique requirements of monitoring applications.
23
203 The information flow during the postprocessing phase is displayed in Figure 2b. This flow includes the following steps:
24
204
25 1- Input data for the AI tool: The Y-axis inclinations measured by LARA over time and onsite humidity and temperature
205
26 values are introduced into the AI tool.
27
206 2- Estimation of intelligent (I) LARA inclinations: The AI tool calculates the X-axis inclinations based on the input data over
28
207 time.
29
208 3- Integration of data into the IoT platform: The X-axis rotations estimated by the AI tool are introduced into an IoT cloud,
30
209 along with the values of the X-axis directly measured by LARA and the threshold of the Mean Square Error (MSE), which
31
210 is obtained through a comparison between LARA and a commercial inclinometer. The MSE values, obtained by comparing
32
211 the AI tool’s estimated X-axis rotations with those directly measured by LARA, are plotted in a graph along with the MSE
33
212 threshold on an online IoT platform hosted on the ThingSpeak website.
34
213
35 4- Analysis of MSE errors over time: Within the ThingSpeak platform, a MATLAB code is incorporated to facilitate the
214
36 assessment and in-depth analysis of data originating from channels hosted on ThingSpeak. The channel data can be ac-
215
37 cessed within the MATLAB analysis section of the ThingSpeak website by supplying the channel ID and associated Ap-
216
38 plication Programming Interface key. A specialized alert program has been developed for online monitoring and task au-
217
39 tomation. Its principal objective is to continuously monitor the MSE values to ensure that they remain within acceptable
218
40 limits. When the deviation between the LARA X-axis value and I-LARA X-axis estimation exceeds the predefined thresh-
219
41 old, the program automatically triggers an electronic alert.
42
220 5- Alert: The algorithm integrated into the ThingSpeak platform operates systematically and sends electronic alerts via email
43
221 when the MSE exceeds the threshold. An example of these notification alerts is shown in Figure 2c.
44
222
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
223
1 (a)
2
3
4
5
6
7
8
9 (b)
10
11
12
13
14
15
16
17
18
19 (c)
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34 Figure 2: (a) LARA setup, (b) Flow of information of the proposed methodology, and (c) Alert message.
224
35
36 The general and primary technical specifications of LARA working as an accelerometer and inclinometer are summarized in Tables
225
37 2 and 3, respectively.
226
227
38
228
39 Table 2: General specifications of LARA.
40 LARA
41
Operating temperature -40° to +85°
42
43 Communication Ethernet or USB
44 Power Source Ethernet or USB
45 Power Consumption 200mA @ 5V
46
Waterproof IP67
47
48 Dimensions 40 x 40 x 5 mm
229
49
50
230 Table 3: Technical specifications of LARA working as an accelerometer and inclinometer.
51
LARA accelerometer LARA inclinometer
52
53 Sensor type Triaxial Sensor type Triaxial
54 Measurement range ±2.0 g Measurement range 0-4°
55 Noise density 51µg/√Hz Static precision Up to 0.002°
56
Resolution 31µg Dynamic precision Up to 0.02°
57
58 Frequency response 0-61 Hz Sampling rate 333 Hz
59 Sampling rate 333 Hz
60
231
61
62
63
64
65
232 Based on the information in Table 3, LARA functioning as an inclinometer can reach a precision of up to 0.002° while working in
233
1 the static mode for a measurement range between 0° and 4 °.
234
2
235 3 4. Al tool
4
236 5 Literature reviews, such as that of Sun et al. [68], have demonstrated a growing interest in utilizing AI tools, including machine
237 6 learning [69], deep learning with recurrent neural networks (RNNs) [70], convolutional neural networks (CNNs) [71], and image
238 7 identification [72]. RNN and CNN tools have been employed by various scholars, such as Mukhopadhyay et al. [47], Wang [73],
239 8 Qi [74], and Pan [75], to compensate for the temperature effects in low-cost MEMS accelerometers. Hoang et al. [76] used AI tools
240 9 to improve the accuracy of low-cost MEMS accelerometers.
241
10
242
11 In this section, we introduce a novel AI tool designed to understand the impact of environmental factors, specifically temperature
243
12 and humidity, on the inclinations measured by LARA in actual steel structures. The foundation of this tool is a multilayer perceptron
244 (MLP) neural network. The following subsections describe the key characteristics, architecture, and training methodology of the
13
245 proposed AI tool.
14
246
15
247 4.1. MLP
16
248
17
249 The proposed AI tool is based on an MLP [77], which is an artificial neural network comprising multiple layers of neurons arranged
18
250 sequentially. MLP networks are distinguished by their feedforward architecture, which ensures a unidirectional flow of information
19
251 from the input layer, through hidden layers, and finally to the output layer, without any loops or cycles. With an adequate number
20
252 of hidden neurons and correct training data, MLP networks can approximate any continuous function with a specified degree of
21
253 precision.
22
254
23
255 These characteristics make the MLP a powerful and flexible neural network model capable of learning complex patterns and per-
24
256 forming predictions based on input data. More importantly, the MLP requires low-level processing power and time compared to
25
257 deeper learning models, which makes it a perfect match for application in LARA real-time measurements. In other words, LARA
26
258 has limited power for processing data. In fact, this AI tool has been extensively employed across a diverse range of applications,
27
259 including prediction of material properties [78,79], comfort assessment [80], life prediction [81], evaluation of structural perfor-
28
260 mance [82,83], and energy consumption prediction in buildings [84].
261
29
262
30 The MLP used in this study was programmed using the scikit-learn package, which is an open-source machine-learning library for
263
31 Python. This library has several key features. First, it equips researchers with robust tools for data preprocessing, encompassing
264
32 essential tasks, such as feature scaling, categorical variable encoding, handling missing values, and feature extraction. In addition,
265
33 it seamlessly integrates with other Python libraries, enabling efficient data manipulation and visualization throughout the model
266
34 deployment process. The benefits of the scikit-learn library have led to its widespread use in various applications. Ren et al. [85]
267
35 used the scikit-learn package to determine the optimal composition of cement grouting materials, while Jain et al. [86] applied it to
268
36 forecast data-driven energy consumption in residential buildings. Günaydın et al. [87] utilized scikit-learn to evaluate concrete
269
37 strength, and Sanzana et al. [88] employed it to predict energy storage in commercial buildings. Moreover, Riu et al. [89] used it to
270
38 estimate the consumption of air-conditioning systems in eco-buildings.
271
39
272 4.2 MLP architecture
40
273
41
274 The MLP models were designed to acquire the ability to accurately predict inclinations along the X-axis using Y-axis rotations,
42
275 temperature, and humidity data obtained from low-cost sensors. Consequently, the trained model can be employed to identify anom-
43
276 alous measurements by comparing its predicted results with actual rotations on the X-axis measured by LARA. The lack of statistical
44
277 correlation between the predicted and measured rotations may indicate a calibration problem, an external load not being considered,
45
278 or structural damage.
46
279
47
280 Figure 3 depicts a schematic representation of the proposed MLP. The MLP comprises the following components:
48
281 - Input layer, including three input neurons representing the normalized values of the Y-axis rotations, temperature, and
49
282 relative humidity measured onsite. To normalize the included data outliers, the “RobustScaler” feature is used. The Y-axis
50
283 rotations are directly measured by LARA, whereas the latter two parameters are measured using a low-cost sensor (SHT22).
51
284 This device has extensive applications in various fields. For instance, Mobaraki et al. [90] employed SHT22 sensors for
52
285 ambient monitoring, while Saini et al. [91] extensively explored their use for monitoring indoor air quality. Furthermore,
53
286 Xie et al. [92] utilized them for thermal comfort sensing.
54
287 - Two hidden layers denoted as i and j, with m and n hidden neurons, respectively.
55
288 - Output layer, consisting of a single neuron representing LARA's X-axis rotation.
289
56
290
57 The configuration of the hidden layers in this architecture was customized based on the data of the case study, as detailed in the
291
58 subsequent section.
292
59
293
60
61
62
63
64
65
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18 Figure 3: Typical architecture of the proposed MLP.
19
294
20
295
21
296
22 5. Case study: Steel Gable Frame
23
297 In this section, the characteristics of the structure under study are described, along with the main features of the monitoring program.
24
298 Subsequently, an application of the AI tool detailed in the previous section is presented. Finally, the results obtained from the
25
299 application of the tool are analyzed.
26
300
27
301 5.1 Description of the structure and monitoring program
28
302
29
303 The monitored structure was a steel gable frame located at the Institute of Technology of the University of Castilla-La Mancha,
30
304 Cuenca, Spain. It measured 11.5 m in length and 6 m in height, as illustrated in Figure 4a. The construction consists of HEB 300
31
305 columns and IPE 400 beams. Figure 4b shows the cloud point, while Figure 4c presents the building information model for this
32
306 structure.
33
307
34
308 This structure was selected based on its advantageous characteristics for mitigating wind effects, which were attributed to its strate-
35
309 gic location and presence of wind-protective parapets on its roof, as displayed in Figure 4d. Furthermore, roofs experience occasional
36
310 live loads associated with maintenance activities, limiting their impact. Consequently, structural variations in this frame are primar-
37
311 ily influenced by environmental factors such as temperature and humidity.
38
312
313
39 This frame has been continuously monitored since its construction in 2008 to investigate the impacts of environmental variables
314
40 and foundation settlement. This long-term monitoring program comprised 27 measurement points for soil water content, four incli-
315
41 nometers, four thermometers, a weather station, and 22 points for topographical leveling. The detailed information on the sensor
316
42 characteristics can be referred to in the study by González-Arteaga et al. [93].
317
43
318
44 Among the monitoring solutions installed in this structure, emphasis was placed on those incorporated into the gable frame under
319
45 study. These solutions utilize the inclinometer model EL Tiltmeter SC [35] from Durham, which measures the one-axle inclination
320
46 and temperature. Figure 5a shows a photograph of the sensor installed onsite along the direction of the inclination. The key technical
321
47 specifications of the device are listed in Table 4.
322
48
323
49 Table 4: Technical features of the commercial inclinometer.
50 EL Tiltmeter SC
51
Sensor type Uniaxial electrolytic tilt sensor
52
53 Resolution inclinometer 1 arc second.
54 Range inclinometer ±40 arc minutes
55 Resolution thermometer 30×10-5°
56
Operating temperature -20° to +50°
57
58 Dimensions 124 x 80 x 59 mm
59 Price 15001 €
60
324
61
62
63
64
65
325 As depicted in Figure 5b, a commercial inclinometer was positioned at the junction between the beam and column on the east façade.
326 1 This inclinometer recorded readings at 15-minute intervals. The measurements from these sensors cannot be accessed wirelessly;
327 2 instead, they must be manually extracted using a USB. The sensors were connected directly to the electrical power supply of the
328 3 building.
329 4
5 (a) (b)
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20 (c) (d)
21
22
23
24
25
26
27
28
29
30
31
32
33
Figure 4: Monitored gable frame: (a) Picture of the real structure, (b) Cloud points model, (c) BIM model, and (d) Picture of the wind para-
34 pets on the roof.
35
330
36 (a) (b)
37
38
39
40 X
41 LARA and EL Tilmeter SC
42
43
44
45
46
47
48
49
50
51 Figure 5: Commercial inclinometer EL Tiltmeter SC: (a) Picture, and (b) Sensor location in the gable frame.
331
52
332
53 The low-cost monitoring solution incorporated in this structure consists of two sensors: (1) LARA, which determines the inclinations
333
54 along three axes, and (2) DHT22, which measures the ambient temperature and humidity. Both sensors were installed at the same
334
55 location as the commercial inclinometer. Figure 6a displays pictures of the low-cost sensors installed onsite, with the rotational axes
335
56 considered by LARA (X, Y, and Z) highlighted. Information from both sensors was transmitted wirelessly over the internet and
336
57 displayed on the ThingSpeak IoT platform, as shown in Figure 6b. The reading interval for both sensors was set to 15 min to
337
58 facilitate comparison with the commercial inclinometer. Both sensors were connected to the electrical power supply of the building.
338
59
339
60
61
62
63
64
65
(a) (b)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18 Figure 6: Low-cost sensors in the gable frame: (a) Picture of the installed LARA sensor, and (b) IoT platform
19 to access the measurements wirelessly.
340
20
341
21 5.2 Architecture of the AI tool
342
22
343
23 The architecture of the MLP tool described in Section 4.2 was tailored for the case study. The main characteristics of the input layer,
344
24 hidden layers, activation functions, output layer, and solver used in the proposed MLP are summarized as follows:
345
25
346
26 - Hidden layers: Determining the optimal numbers of hidden layers and neurons is often an empirical process involving
347
27 experimentation. In this application, the numbers of hidden layers and neurons were determined through experimental
348
28 analysis of the model performance using cross-validation datasets. The collected data were randomly divided into three
349 distinct datasets: (1) training dataset, comprising 70% of the data; (2) cross-validation dataset, covering 15% of the data;
29
350 and (3) test dataset, encompassing the remaining 15% of the data. A similar methodology was employed by Tanhadoust et
30
351 al. [94] to predict the stress-strain behavior of concrete. The network performances for various architectures were analyzed
31
352 using a cross-validation dataset for different numbers of hidden layers (ranging from 1 to 10) and neurons (ranging from 1
32
353 to 6). Based on the results of this empirical process, we determined that an MLP architecture with three hidden layers, each
33
354 containing five neurons, delivered the best performance. After investigating the learning rate, the initial value of this pa-
34
355 rameter was fixed at 0.0001 to optimize the model performance.
35
356
36 - Activation functions: The activation function chosen for the different neurons was the rectified linear activation function
357
37 (ReLU) from the scikit-learn library. This selection helps prevent exponential growth in the computations required to op-
358
38 erate a neural network. Moreover, this activation function provides additional advantages, including simplicity and efficient
359
39 empirical performance. The ReLU function has found extensive applications in various fields, such as mathematics [95,96],
360
40 SSI [97,98], and material science [99].
41
361 - Solver: The Adam solver, which utilizes the stochastic gradient-based optimizer proposed by Kingma et al. [100], was
42
362 used. Adam works efficiently on relatively large datasets (with thousands or more training samples) in terms of both train-
43
363 ing time and validation. It has been used in several studies such as [101].
44
364
45
365
46 The model variables for the MLP are listed in Table 5.
366
47
367
48 Table 5: Properties of the MLP model.
49 Hidden layers 3
50 Neurons per layer 5
51
Initial learning rate 0.0001
52
53 Epochs 200
54 Activation function ReLU
55 Solver Adams
56
368
57
369 5.3 Results
58
370
59
371 The monitoring data of both low-cost sensors were collected from their installation between January 21 and April 1, 2023, resulting
60
372 in 3328 records. The measurement results, illustrated in Figures 7a-d, depict the variations in temperature, humidity, and inclinations
61
373 along the X- and Y-axes.
62
63
64
65
374
375
1 The correlations between temperature, humidity, and Y-axis inclinations with respect to the X-axis inclinations were studied using
376
2 Pearson’s correlation coefficient (PCC). The obtained coefficients indicated a moderately strong linear correlation for temperature
377
3 (PCC = 0.61), a weak negative linear correlation for humidity (PCC = -0.21), and a strong positive linear correlation for Y-axis
378 inclinations (PCC = 0.80).
4
379
5
(a) (b)
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20 (c) (d)
21
22
23
24
25
26
27
28
29
30
31 Figure 7: Measurements of the low-cost sensor from January 21st to April 1st 2023: (a) Temperature, (b) Humidity, (c), Inclinations in the X-
32 axis, and (d) Inclinations in the Y-axis.
380
33
381
34 Figure 8 shows the inclinations and temperatures measured using the commercial inclinometer on February 20, 2023. Analysis of
382
35 the information in this figure highlights the significant effect of temperature on structure rotation. On this specific day, the beam
383
36 rotations in the structure varied from -0.01º to 0.002º, which were attributed to temperature changes ranging from 15ºC to 26ºC.
6
384
37
38 0,025 30
39 0,02
4 25
40
Temperature (ºC)
0,015
41
(º)
20
42 0,01
Inclination
2
43 0,005 15
44 0
0
45 10
-0,005 0:13 2:43 5:13 7:43 10:13 12:43 15:13 17:43 20:13 22:43
13
19
25
31
37
43
49
55
61
67
73
79
85
91
97
1
7
103
109
115
121
127
133
139
145
151
157
163
169
175
181
187
193
199
205
46 5
2 47 -0,01
48 -0,015 0
49 Time (h)
LARA
4
50
51 Inclination
EL Tilmeter SC Temperature
6 52 Figure 8: Example of the effects of the temperature on the inclination of the beam measured by the commercial sensor on a specific day (Feb-
53 Number of Data
ruary 20th, 2023).
54
385
55
386 Figure 9 compares the inclinations measured by the commercial sensor with those recorded by LARA on the X-axis over two
56
387 different periods: two months from January 23 to March 23, 2023 (Figure 9a) and three days from February 20 to 22, 2023 (Figure
57
388 9b). The wider range of measurements recorded by LARA in Figure 9a indicates its lower accuracy compared with that of the
58
389 commercial device. Moreover, Figure 9b shows that both sensors display similar trends in capturing the influence of environmental
59
390 parameters over time. It is essential to emphasize that the role of the commercial inclinometer in this study is to validate the proper
60
391 calibration of LARA readings during the training of the proposed MLP. The standard deviation and mean squared error (MSE)
61
392 obtained from comparing LARA and the commercial sensor over time were 0.02408 and 0.00058, respectively.
393
62
63
64
65
394
(a) 0,08
1
2
0,04
3
Degrees (º)
4
0
5
6
-0,04
7
LARA
8 EL Tiltmeter SC
9 -0,08
1/23/2023 2/7/2023 2/22/2023 3/9/2023 3/24/2023
10 Day
395
11
396
12
(b) 0,06
13
14
15 0,04
16
17
18 0,02
19
Degrees (º)
20 0
21
13
19
25
31
37
43
49
55
61
67
73
79
85
91
97
1
7
103
109
115
121
127
133
139
145
151
157
163
169
175
181
187
193
199
205
22
23 -0,02
24
25 LARA
-0,04
26
27 EL Tilmeter SC
28 -0,06
29 Number of Data
30 Figure 9: Comparison between the inclinations measured by the commercial inclinometer and LARA X-axis across two different periods: (a)
31 a two-months span from January 23rd to March 23rd, 2023, and (b) a three-day span from February 20th to 22nd, 2023.
32
397
33
398 Figure 10 presents the errors between the X-axis rotations predicted by the MLP tool and actual measurements obtained by LARA
34
399 from the test dataset.
35
400
36
401
37
38 0.06
39
0.04
40
41
0.02
42
Error (º)
43 0.00
44
45 -0.02
46
47 -0.04
48 0 100 200 300 400 500

49 Number of Data
50
51 Figure 10: Errors between the X-axis rotations predicted by the MLP tool and those obtained by LARA from January 21st to April 1st 2023.
402
52
403 The root MSE (RMSE), MSE, PCC, and mean absolute error (MAE) were calculated based on the errors displayed in Figure 10.
53
404 The obtained values are listed in Table 6. These results reveal the high predictive performance of the developed tool, as evidenced
54
405 by an MSE value of 0.002, and a strong positive linear correlation, as indicated by a PCC of 0.87.
55
406
56
407
57 Table 6: Comparison between MLP and LARA.
58 Root Mean Squared Error (RMSE) 0.0127
59 Mean Squared Error (MSE) 0,0002
60
Pearson Correlation Coefficient (PCC) 0.87
61
62 Mean Absolute Error (MAE) 0.0098
63
64
65
408
409
1 The probability density of errors between the rotations measured by LARA and predicted values with the fitted normal distribution
410
2 was characterized by a mean of 0.00032 and a standard deviation of 0.0127. This function is illustrated in Figure 11.
411
3
412
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21 Figure 11: Probability density of errors with fitted normal distribution in the test set.
22
413
414
23 The calculated MSE values over a predetermined environmental cycle (such as one day) indicate the potential necessity for LARA
415
24 recalibration. This need arises when the acquired errors between the X-axis inclinations measured by LARA and values predicted
416
25 by the MLP exceed the accuracy threshold obtained by comparing LARA with the commercial inclinometer. Figure 12 displays the
417
26 MSE between the X-axis inclinations of LARA and values predicted by the MLP for the period from March 1 to April 1, 2023. This
418
27 figure also incorporates the MSE value of 0.00058 obtained from the comparison with the commercial inclinometer, labeled as the
419
28 MSE threshold. Analysis of the results confirms the reliability of the obtained data, as the error threshold was not exceeded. How-
420
29 ever, if the MSE threshold is exceeded in the future, it may indicate a need for recalibrating the low-cost sensors. To notify this
421
30 scenario, an email alert will be automatically sent.
422
31
32
33
34
35
36
37
MSE
38 MSE threshold
39
40
41
42
43
44
45
46
47 Figure 12: Example of the MSE of the inclinations measured by LARA and the predicted values by the MLP on specific days (April
48 2023) compared with the MSE threshold.
49
50
423
51
424
52 6. Conclusions
53
425
54 Developing low-cost structural and environmental sensors has led to significant advancements in various fields, particularly in the
426
55 cost-effective monitoring of infrastructure and buildings. However, it is critical to ensure the long-term durability and reliability of
427
56 these sensors. The existing literature lacks solutions for guaranteeing the sustained performance of low-cost sensors.
428
57
429
58 This study introduces a novel AI-based methodology that enables evaluating the long-term performance of low-cost sensors. This
430
59 evaluation is achieved without high-cost monitoring equipment and relies solely on structural and environmental data (specifically
431 humidity and temperature) obtained from low-cost devices. Furthermore, the system automatically sends digital alarms through a
60
432 dedicated IoT platform to identify low-cost sensors that require recalibration.
61
433
62
63
64
65
434 The application of the proposed AI tool was validated using the low-cost inclinometer LARA, which relies on an automatic com-
435
1 parison between the inclinations of actual onsite measurements and those predicted under real environmental conditions. If the
436
2 discrepancies between these measurements consistently exceed a predefined statistical threshold over time, it may indicate potential
437
3 inaccuracies in the low-cost inclinometer, suggesting the need for recalibration. The results of the long-term monitoring of a real
438 frame located in Spain underscored the effectiveness and practicality of this approach as a reliable and cost-effective solution for
4
439 automating the long-term performance assessment of low-cost structural inclinometers.
5
440
6
441 7 7. Acknowledgements
8
442 9 This work was supported by the projects PID2021-126405OB-C31 and PID2021-126405OB-C32 funded by FEDER funds through
443
10 the program “A Way to Make Europe” and by the Spanish Ministry of Economy and Competitiveness
444
11 MICIN/AEI/10.13039/501100011033/. It also received support from the National Natural Science Foundation of China (Grant No.
445
12 52278313).
446
13
447
14 8. References
15
16
448 [1] X. Peng, Q. Yang, Damage detection in beam-like structures using static shear energy redistribution, Frontiers of Structural
17
449
18 and Civil Engineering. 16 (2022) 1552-1564, https://doi.org/10.1007/S11709-022-0903-4/METRICS.
19
450 [2] M. Nogal, J.A. Lozano-Galant, J. Turmo, E. Castillo, Numerical damage identification of structures by observability techniques
20
21
451 based on static loading tests, Structure and Infrastructure Engineering. (2015) 1216–1227,
22
452
23 https://doi.org/10.1080/15732479.2015.1101143.
24
453 [3] Q. Zhou, Q.S. Li, X.L. Han, J.W. Wan, K. Xu, Horizontal displacement estimation of high-rise structures by fusing strain and
25
26
454 acceleration measurements, Journal of Building Engineering. 57 (2022) 104964, https://doi.org/10.1016/J.JOBE.2022.104964.
27
455
28 [4] S. Komarizadehasl, M. Komary, A. Alahmad, J.A. Lozano-Galant, G. Ramos, J. Turmo, A Novel Wireless Low-Cost Incli-
29
456 nometer Made from Combining the Measurements of Multiple MEMS Gyroscopes and Accelerometers, Sensors. 22 (2022) 5605,
30
31
457 https://doi.org/10.3390/S22155605.
32
458
33 [5] B. Mobaraki, F. Lozano-Galant, R.P. Soriano, F.J.C. Pascual, Application of Low-Cost Sensors for Building Monitoring: A
34
459 Systematic Literature Review, Buildings. 11 (2021) 336, https://doi.org/10.3390/BUILDINGS11080336.
35
36
460 [6] G.F. Sirca, H. Adeli, System identification in structural engineering, Scientia Iranica. 19 (2012) 1355–1364,
37
461
38 https://doi.org/10.1016/J.SCIENT.2012.09.002.
39
462 [7] D. Tomàs, J.A. Lozano-Galant, G. Ramos, J. Turmo, Structural system identification of thin web bridges by observability
40
41
463 techniques considering shear deformation, Thin-Walled Structures. 123 (2018) 282–293,
42
464
43 https://doi.org/10.1016/J.TWS.2017.11.017.
44
465 [8] J.A. Lozano-Galant, M. Nogal, E. Castillo, J. Turmo, Application of Observability Techniques to Structural System Identifi-
45
466
46 cation, Computer-Aided Civil and Infrastructure Engineering. 28 (2013) 434–450, https://doi.org/10.1111/MICE.12004.
47
467
48 [9] S. Emadi, H. Ma, J.A. Lozano-Galant, J. Turmo, Simplified Calculation of Shear Rotations for First-Order Shear Deformation
49
468 Theory in Deep Bridge Beams, Applied Sciences. 13 (2023) 3362, https://doi.org/10.3390/app13053362.
50
469
51 [10] Z. Gu, W. Lu, Y. Fan, Y. Gao, Automated simplified structural modeling method for megatall buildings based on genetic
52
470 algorithm, Journal of Building Engineering. 77 (2023) 107485, https://doi.org/10.1016/J.JOBE.2023.107485.
53
54
471 [11] G. Saad, R. Ghanem, S. Masri, Robust system identification of strongly non-linear dynamics using a polynomial chaos based
55
472
56 sequential data assimilation technique, Collection of Technical Papers - AIAA/ASME/ASCE/AHS/ASC Structures, Structural
57
473 Dynamics and Materials Conference. 6 (2007) 6005–6013, https://doi.org/10.2514/6.2007-2211.
58
59
474 [12] K. Hijikata, A. Mita, Damage detection of wooden buildings using chaos analysis and system identification, Smart Structures
60
475
61 and Materials. 6174 (2006), https://doi.org/10.1117/12.658954.
62
63
64
65
476 [13] H. Li, D. Tao, Y. Huang, Y. Bao, A data-driven approach for seismic damage detection of shear-type building structures using
1
477 2 the fractal dimension of time–frequency features, Structural Control and Health Monitoring. 20 (2013) 1191–1210,
478
3 https://doi.org/10.1002/STC.1528.
4
479 5 [14] H.S. Park, H. Adeli, Distributed Neural Dynamics Algorithms for Optimization of Large Steel Structures, Journal of Struc-
6
480 7 tural Engineering. 123 (1997) 880–888, https://doi.org/10.1061/(ASCE)0733-9445(1997)123:7(880).
481 8 [15] J. Shin, S. Park, Optimum retrofit strategy of FRP column jacketing system for non-ductile RC building frames using artifi-
9
10
482 cial neural network and genetic algorithm hybrid approach, Journal of Building Engineering. 57 (2022) 104919,
11
483
12 https://doi.org/10.1016/J.JOBE.2022.104919.
13
484 [16] H. Adeli, H.S. Park, A neural dynamics model for structural optimization-Theory, Computers & Structures. 57 (1995) 383–
14
485
15 390, https://doi.org/10.1016/0045-7949(95)00048-L.
16
486
17 [17] Z. Chen, L. Zhang, K. Li, X. Xue, X. Zhang, B. Kim, C.Y. Li, Machine-learning prediction of aerodynamic damping for
18
487 buildings and structures undergoing flow-induced vibrations, Journal of Building Engineering. 63 (2023) 105374,
19
488
21
489
22 [18] G.S. Wang, Application of hybrid genetic algorithm to system identification, Structural Control and Health Monitoring. 16
23
490 (2009) 125–153, https://doi.org/10.1002/STC.306.
24
491
25 [19] S. Xue, H. Tang, J. Zhou, Identification of Structural Systems Using Particle Swarm Optimization, Journal of Asian Archi-
26
492 tecture and Building Engineering. 8 (2009) 517–524, https://doi.org/10.3130/JAABE.8.517.
27
28
493 [20] V. Plevris, M. Papadrakakis, A Hybrid Particle Swarm—Gradient Algorithm for Global Structural Optimization, Computer-
29
494
30 Aided Civil and Infrastructure Engineering. 26 (2011) 48–68, https://doi.org/10.1111/J.1467-8667.2010.00664.X.
31
495 [21] A.R.M. Rao, G. Anandakumar, Optimal placement of sensors for structural system identification and health monitoring using
32
33
496 a hybrid swarm intelligence technique, Smart Materials and Structures. 16 (2007) 2658, https://doi.org/10.1088/0964-
34
497
35 1726/16/6/071.
36
498 [22] Y.Y. Chen, W. Te Su, S. Tesfamariam, Z.C. Qian, W. Zhao, C.Y. Shen, F. Lin Zhou, Experimental testing and system identi-
37
38
499 fication of the sliding bistable nonlinear energy sink implemented to a four-story structure model subjected to earthquake exci-
39
500
40 tation, Journal of Building Engineering. 61 (2022) 105226, https://doi.org/10.1016/J.JOBE.2022.105226.
41
501 [23] P. Wei, Q. Li, M. Sun, J. Huang, Modal identification of high-rise buildings by combined scheme of improved empirical
42
43
502 wavelet transform and Hilbert transform techniques, Journal of Building Engineering. 63 (2023) 105443,
44
503
46
504 [24] O. Yazdanpanah, B. Mohebi, M. Yakhchalian, Seismic damage assessment using improved wavelet-based damage-sensitive
47
48
505 features, Journal of Building Engineering. 31 (2020) 101311, https://doi.org/10.1016/J.JOBE.2020.101311.
49
506
50
51
507
52
53
508
54
509
55 [25] H. Adeli, X. Jiang, Dynamic Fuzzy Wavelet Neural Network Model for Structural System Identification, Journal of Struc-
56
510 tural Engineering. 132 (2006) 102–111, https://doi.org/10.1061/(ASCE)0733-9445(2006)132:1(102).
57
58
511 [26] X. Jiang, H. Adeli, Dynamic Wavelet Neural Network for Nonlinear Identification of Highrise Buildings, Computer-Aided
59
512
60 Civil and Infrastructure Engineering. 20 (2005) 316–330, https://doi.org/10.1111/J.1467-8667.2005.00399.X.
61
62
63
64
65
513 [27] X. Jiang, H. Adeli, Pseudospectra, MUSIC, and dynamic wavelet neural network for damage detection of highrise buildings,
1
514 2 International Journal for Numerical Methods in Engineering. 71 (2007) 606–629, https://doi.org/10.1002/NME.1964.
515
3 [28] S. Emadi, J.A. Lozano-Galant, Y. Xia, G. Ramos, J. Turmo, Structural system identification including shear deformation of
4
516 5 composite bridges from vertical deflections, Steel and Composite Structures. 32 (2019) 731–741,
6
517 7 https://doi.org/10.12989/SCS.2019.32.6.731.
518 8 [29] T. Peng, M. Nogal, J.R. Casas, J.A. Lozano-Galant, J. Turmo, Constrained observability techniques for structural system
9
10
519 identification using modal analysis, Journal of Sound and Vibration. 479 (2020) 115368,
11
520
12 https://doi.org/10.1016/J.JSV.2020.115368.
13
521 [30] D.W. Ha, H.S. Park, S.W. Choi, Y. Kim, A Wireless MEMS-Based Inclinometer Sensor Node for Structural Health Moni-
14
522
15 toring, Sensors. 13 (2013) 16090-16104, https://doi.org/10.3390/S131216090.
16
523
17 [31] J. Zhang, D. Liu, Z. Liu, L. Mo, X. Xiang, Y. Wang, Rotational behavior of bolted post-to-beam glulam connections with
18
524 friction damped knee brace, Journal of Building Engineering. 76 (2023) 107215, https://doi.org/10.1016/J.JOBE.2023.107215.
19
525
20 [32] F. Huseynov, C. Kim, E.J. OBrien, J.M.W. Brownjohn, D. Hester, K.C. Chang, Bridge damage detection using rotation
21
526
22 measurements – Experimental validation, Mechanical System and Signal Processing. 135 (2020) 106380,
23
527 https://doi.org/10.1016/J.YMSSP.2019.106380.
24
528
25 [33] D. Hester, J. Brownjohn, F. Huseynov, E. Obrien, A. Gonzalez, M. Casero, Identifying damage in a bridge by analysing
26
529 rotation response to a moving load, Structure and Infrastructure Engineering. 16 (2019) 1050–1065,
27
28
530 https://doi.org/10.1080/15732479.2019.1680710.
29
531
30 [34] J. Lei, M. Nogal, J.A. Lozano-Galant, D. Xu, J. Turmo, Constrained observability method in static structural system identi-
31
532 fication, Structural Control and Health Monitoring. 25 (2018) 1–15, https://doi.org/10.1002/STC.2040/ABSTRACT.
32
33
533 [35] EL Tiltmeter - DGSI. https://durhamgeo.com/products/el-tiltmeter/, 2023 (accessed 01 Aug. 2023).
34
534
35 [36] Excellent inclination sensors with digital ouput and high resolution. https://www.wylerag.com/en/products/inclination-meas-
36
535 uring-sensors/zerotronic-inclination-sensor/#specifications, 2023 (accessed 06 Jul. 2023).
37
38
536 [37] JDI-100/200 Digital MEMS Inclinometer Series - Jewell Instruments. https://jewellinstruments.com/products/inertial-tilt-
39
537
40 sensors/inclinometer/mems-i/jdi-100-200-digital-mems-inclinometer-series/, 2023 (accessed 06 Jul. 2023).
41
538 [38] T935 Series ±1° to ±90° - Sherborne Sensors. https://www.sherbornesensors.com/product/t935-series-%C2%B11-
42
43
539 to-%C2%B190/, 2023 (accessed 06 Jul. 2023).
44
540
45 [39] ACA2200-High Accuracy Digital Type Dual-Axis Inclinometer with Full Temperature Compensation SPECIFICATIONS.
46
541 http://www.jca.kr/img_up/shop_pds/jcaauto3/product/aca2200t-canopenspecificationdownload.pdf, 2023 (accessed 06 Jul. 2023).
47
48
542 [40] Wireless IoT Inclinometer Sensor | 2.4GHz sensors | Overview - BeanAir. https://www.beanair.com/wireless-iot-inclinome-
49
543
50 ter-sensor-2.4ghz-overview.html, 2023 (accessed 06 Jul. 2023).
51
544 [41] M.L. Hoang, A. Pietrosanto, A Robust Orientation System for Inclinometer with Full-Redundancy in Heavy Industry, IEEE
52
53
545 Sensors Journal. 21 (2021) 5853–5860, https://doi.org/10.1109/JSEN.2020.3040374.
54
546
55 [42] G. Ruzza, L. Guerriero, P. Revellino, F.M. Guadagno, A Multi-Module Fixed Inclinometer for Continuous Monitoring of
56
547 Landslides: Design, Development, and Laboratory Testing, Sensors. 20 (2020) 3318, https://doi.org/10.3390/S20113318.
57
58
548 [43] D.W. Ha, J.M. Kim, Y. Kim, H.S. Park, Development and application of a wireless MEMS-based borehole inclinometer for
59
549
60 automated measurement of ground movement, Automation in Construction. 87 (2018) 49–59,
61
550 https://doi.org/10.1016/J.AUTCON.2017.12.011.
62
63
64
65
551 [44] Measure acceleration, angular rate, magnetic field, and temperature along axes of MPU-9250 sensor - Simulink - Math-Works
1
552 2 España. https://es.mathworks.com/help/supportpkg/arduino/ref/mpu9250imusensor.html, 2023 (accessed 06 Jul. 2023).
553
3 [45] Y. Yu, J. Ou, J. Zhang, C. Zhang, L. Li, Development of wireless MEMS inclination sensor system for swing monitoring of
4
554 5 large-scale hook structures, IEEE Transactions on Industrial Electronics. 56 (2009) 1072–1078, https://10.1109/TIE.2009.2012469.
6
555 7 [46] J. Kumar, R.P. Bajpai, Application of Mems in Bridge Structures Health Monitoring, Science and Technology of Advanced
556 8 Materials. 16 (2015) 3, https://10.1088/1468-6996/16/3/034607.
9
10
557 [47] S. Mukhopadhyay, N. Cennamo, M.J. Deen, J. Lee, S. Morais, G. Guo, B. Chai, R. Cheng, Y. Wang, Temperature Drift
11
558
12 Compensation of a MEMS Accelerometer Based on DLSTM and ISSA, Sensors. 23 (2023) 1809,
13
559 https://doi.org/10.3390/S23041809.
14
560
15 [48] S.M. Khan, N. Qaiser, M.M. Hussain, An inclinometer using movable electrode in a parallel plate capacitive structure, AIP
16
561
17 Advances. 9 (2019), https://doi.org/10.1063/1.5092146/1076331.
18
562 [49] Y. Wang, J. Zhang, Z. Yao, C. Lin, T. Zhou, Y. Su, J. Zhao, A MEMS Resonant Accelerometer with High Performance of
19
563
20 Temperature Based on Electrostatic Spring Softening and Continuous Ring-Down Technique, IEEE Sensors Journal. 18 (2018)
21
564
22 7023–7031, https://doi.org/10.1109/JSEN.2018.2852647.
23
565 [50] Y. Liu, T. Ma, Parasitic resistance-based high precision capacitive MEMS accelerometer phase shift and its usage for tem-
24
566
25 perature compensation, IEEE Sensors Journal. 18 (2018) 629–634, https://doi.org/10.1109/JSEN.2017.2777864.
26
567 [51] G. Ruzza, L. Guerriero, P. Revellino, F.M. Guadagno, Thermal Compensation of Low-Cost MEMS Accelerometers for Tilt
27
28
568 Measurements, Sensors. 18 (2018) 2536, https://doi.org/10.3390/S18082536.
29
569
30 [52] H. Zhang, X. Wei, Y. Gao, E. Cretu, Analytical Study and Thermal Compensation for Capacitive MEMS Accelerometer with
31
570 Anti-Spring Structure, Journal of Microelectromechanical Systems. 29 (2020) 1389–1400,
32
33
571 https://doi.org/10.1109/JMEMS.2020.3011949.
34
572
35 [53] S. Khankalantary, S. Ranjbaran, S. Ebadollahi, Simplification of calibration of low-cost MEMS accelerometer and its tem-
36
573 perature compensation without accurate laboratory equipment, Measurement Science and Technology. 32 (2021) 045102,
37
38
574 https://doi.org/10.1088/1361-6501/ABD0BF.
39
575
40 [54] X.D. Yu, J.L. Li, S.K. Yan, G. Wei, G. Li, Temperature Drift Compensation Based on Artificial Fish Swarm Algorithm for
41
576 Quartz Flexible Accelerometer, Applied Mechanics and Materials. 513–517 (2014) 4030–4034,
42
43
577 https://doi.org/10.4028/WWW.SCIENTIFIC.NET/AMM.513-517.4030.
44
578
45 [55] V.F. Caspani, D. Tonelli, F. Poli, D. Zonta, Designing a Structural Health Monitoring System Accounting for Temperature
46
579 Compensation, Infrastructures. 7 (2022) 5, https://doi.org/10.3390/INFRASTRUCTURES7010005.
47
48
580 [56] B. Yuan, Z. Tang, P. Zhang, F. Lv, Thermal Calibration of Triaxial Accelerometer for Tilt Measurement, Sensors. 23 (2023)
49
581
50 2105, https://doi.org/10.3390/S23042105.
51
582 [57] R. Fontanella, D. Accardo, R.S. Lo Moriello, L. Angrisani, D. De Simone, MEMS gyros temperature calibration through
52
53
583 artificial neural networks, Sensors and Actuators A: Physical. 279 (2018) 553–565, https://doi.org/10.1016/J.SNA.2018.04.008.
54
584
55 [58] R. Fontanella, D. Accardo, E. Caricati, S. Cimmino, D. De Simone, An extensive analysis for the use of back propagation
56
585 neural networks to perform the calibration of MEMS gyro bias thermal drift, Proceedings of the IEEE/ION Position, Location and
57
58
586 Navigation Symposium, PLANS. (2016) 672–680, https://doi.org/10.1109/PLANS.2016.7479760.
59
587
60 [59] H. Yang, B. Zhou, L. Wang, H. Xing, R. Zhang, A Novel Tri-Axial MEMS Gyroscope Calibration Method over a Full Tem-
61
588 perature Range, Sensors. 18 (2018) 3004, https://doi.org/10.3390/S18093004.
62
63
64
65
589 [60] D. Yao, H. Xing, S. Xue, Dynamic characteristics of a wood frame structure: (I) the effect of temperature and relative hu-
1
590 2 midity, Advanced Material Research. 671–674 (2013) 897–901, https://doi.org/10.4028/WWW.SCIENTIFIC.NET/AMR.671-
591
3 674.897.
4
592 5 [61] Y. Cai, K. Zhang, Z. Ye, C. Liu, K. Lu, L. Wang, Influence of Temperature on the Natural Vibration Characteristics of Simply
6
593 7 Supported Reinforced Concrete Beam, Sensors. 21 (2021) 4242, https://doi.org/10.3390/S21124242.
594 8 [62] L. Zhao, H. Jin, Effect of temperature on natural frequencies of bridge structure of metallurgical crane, IOP Conference Series:
9
10
595 Materials Science and Engineering. 1043 (2021), https://doi.org/10.1088/1757-899X/1043/4/042047.
11
596
12 [63] D. Ribeiro, S. Czarnecki, A. Cury, A. Keshmiry, S. Hassani, M. Mousavi, U. Dackermann, Effects of Environmental and
13
597 Operational Conditions on Structural Health Monitoring and Non-Destructive Testing: A Systematic Review, Buildings. 13 (2023)
14
598
15 918, https://doi.org/10.3390/BUILDINGS13040918.
16
599
17 [64] S. Komarizadehasl, F. Lozano, J.A. Lozano-Galant, G. Ramos, J. Turmo, Low-Cost Wireless Structural Health Monitoring
18
600 of Bridges, Sensors. 22 (2022) 5725, https://doi.org/10.3390/S22155725.
19
601
20 [65] S. Komarizadehasl, B. Mobaraki, H. Ma, J.A. Lozano-Galant, J. Turmo, Development of a Low-Cost System for the Accu-
21
602
22 rate Measurement of Structural Vibrations, Sensors. 21 (2021) 6191, https://doi.org/10.3390/S21186191.
23
603 [66] E. Atencio, S. Komarizadehasl, J.A. Lozano-Galant, M. Aguilera, Using RPA for Performance Monitoring of Dynamic SHM
24
604
25 Applications, Buildings. 12 (2022) 1140, https://doi.org/10.3390/BUILDINGS12081140.
26
605 [67] S. Komarizadehasl, P. Huguenet, F. Lozano, J.A. Lozano-Galant, J. Turmo, Operational and Analytical Modal Analysis of a
27
28
606 Bridge Using Low-Cost Wireless Arduino-Based Accelerometers, Sensors. 22 (2022) 9808, https://doi.org/10.3390/S22249808.
29
607
30 [68] L. Sun, Z. Shang, Y. Xia, S. Bhowmick, S. Nagarajaiah, Review of Bridge Structural Health Monitoring Aided by Big Data
31
608 and Artificial Intelligence: From Condition Assessment to Damage Detection, Journal of Structural Engineering. 146 (2020)
32
33
609 04020073, https://doi.org/10.1061/(ASCE)ST.1943-541X.0002535.
34
610
35 [69] T. Huang, T. Liu, Y. Ai, Z. Ren, J. Ou, Y. Li, N. Xu, Modelling the interface bond strength of corroded reinforced concrete
36
611 using hybrid machine learning algorithms, Journal of Building Engineering. 74 (2023) 106862,
37
38
39
613
40 [70] Y. Jiang, W. Zuo, C. Yuan, G. Xu, X. Wei, J. Zhang, W. She, Deep learning approaches for prediction of adiabatic tempera-
41
614 ture rise of concrete with complex mixture constituents, Journal of Building Engineering. 73 (2023) 106816,
42
43
44
616
45 [71] L. Wang, D. Xie, L. Zhou, Z. Zhang, Application of the hybrid neural network model for energy consumption prediction of
46
617 office buildings, Journal of Building Engineering. 72 (2023) 106503, https://doi.org/10.1016/J.JOBE.2023.106503.
47
48
618 [72] G. Ye, J. Qu, J. Tao, W. Dai, Y. Mao, Q. Jin, Autonomous surface crack identification of concrete structures based on the
49
619
50 YOLOv7 algorithm, Journal of Building Engineering. 73 (2023) 106688, https://doi.org/10.1016/J.JOBE.2023.106688.
51
620 [73] S. Wang, W. Zhu, Y. Shen, J. Ren, H. Gu, and X. Wei, Temperature compensation for MEMS resonant accelerometer based
52
53
621 on genetic algorithm optimized backpropagation neural network, Sensors and Actuators A: Physical. 316 (2020) 112393,
54
622
55 https://doi.org/10.1016/J.SNA.2020.112393.
56
623 [74] B. Qi, S. Shi, L. Zhao, J. Cheng, A Novel Temperature Drift Error Precise Estimation Model for MEMS Accelerometers
57
58
624 Using Microstructure Thermal Analysis, Micromachines. 13 (2022) 835, https://doi.org/10.3390/MI13060835.
59
625
60 [75] Y. Pan, L. Li, C. Ren, H. Luo, Study on the compensation for a quartz accelerometer based on a wavelet neural network,
61
626 Measurement Science and Technology. 21 (2010) 105202, https://doi.org/10.1088/0957-0233/21/10/105202.
62
63
64
65
627 [76] M.L. Hoang, A. Pietrosanto, New Artificial Intelligence Approach to Inclination Measurement Based on MEMS Accelerom-
1
628 2 eter, IEEE Transactions on Artificial Intelligence. 3 (2022) 67–77, https://doi.org/10.1109/TAI.2021.3105494.
629
3 [77] S. Emadi, S. Emadi, Analyzing Cost and Time Objectives in the Construction Projects Using Artificial Neural Network,
4
630 5 International Review of Civil Engineering (IRECE). 13 (2022) 91–98, https://doi.org/10.15866/IRECE.V13I2.21124.
6
631 7 [78] Z. Chen, M.N. Amin, B. Iftikhar, W. Ahmad, F. Althoey, F. Alsharari, Predictive modelling for the acid resistance of ce-
632 8 ment-based composites modified with eggshell and glass waste for sustainable and resilient building materials, Journal of Building
9
10
633 Engineering. 76 (2023) 107325, https://doi.org/10.1016/J.JOBE.2023.107325.
11
634
12 [79] Y. El Khessaimi, Y. El Hafiane, A. Smith, C. Peyratout, K. Tamine, S. Adly, M. Barkatou, Machine learning-based predic-
13
635 tion of compressive strength for limestone calcined clay cements, Journal of Building Engineering. 76 (2023) 107062,
14
636
16
637
17 [80] L. Shi, J. Qiu, R. Zhang, Y. Li, Z. Yang, X. Qi, L. Tao, S. Li, W. Liu, An intelligent optimization method of exercisers’ visual
18
638 comfort assessment in gymnasium, Journal of Building Engineering. 76 (2023) 107135,
19
639
21
640
22 [81] A. Michel, V. Marcos-Meson, On the application of neural networks for service life prediction of steel fibre-reinforced con-
23
641 crete, Journal of Building Engineering. 76 (2023) 107286, https://doi.org/10.1016/J.JOBE.2023.107286.
24
642
25 [82] K. Daneshvar, M. J. Moradi, M. Khaleghi, M. Rezaei, V. Farhangi, and H. Hajiloo, Effects of impact loads on heated-and-
26
643 cooled reinforced concrete slabs, Journal of Building Engineering. 61 (2022) 105328,
27
28
29
645
30 [83] V.V. Degtyarev, K.D. Tsavdaridis, Buckling and ultimate load prediction models for perforated steel beams using machine
31
646 learning algorithms, Journal of Building Engineering. 51 (2022) 104316, https://doi.org/10.1016/J.JOBE.2022.104316.
32
33
647 [84] J. Kim, Y. Kwak, S.H. Mun, J.H. Huh, Electric energy consumption predictions for residential buildings: Impact of data-
34
648
35 driven model and temporal resolution on prediction accuracy, Journal of Building Engineering. 62 (2022) 105361,
36
37
38
650 [85] J. Ren, H. Zhao, L. Zhang, Z. Zhao, Y. Xu, Y. Cheng, M. Wang, J. Chen, J. Wang, Design optimization of cement grouting
39
651
40 material based on adaptive boosting algorithm and simplicial homology global optimization, Journal of Building Engineering. 49
41
652 (2022) 104049, https://doi.org/10.1016/J.JOBE.2022.104049.
42
43
653 [86] R.K. Jain, T. Damoulas, C.E. Kontokosta, Towards Data-Driven Energy Consumption Forecasting of Multi-Family Resi-
44
654
45 dential Buildings: Feature Selection via The Lasso, Computing in Civil and Building Engineering - Proceedings of the 2014 Inter-
46
655 national Conference on Computing in Civil and Building Engineering. (2014) 1675–1682,
47
48
656 https://doi.org/10.1061/9780784413616.208.
49
657
50 [87] O. Günaydın, E. Akbaş, A. Özbeyaz, K. Güçlüer, Machine learning based evaluation of concrete strength from saturated to
51
658 dry by non-destructive methods, Journal of Building Engineering. 76 (2023) 107174, https://doi.org/10.1016/J.JOBE.2023.107174.
52
53
659 [88] M.R. Sanzana, M.O.M. Abdulrazic, J.Y. Wong, T. Maul, C.C. Yip, Charging water load prediction for a thermal energy-
54
660
55 storage air-conditioner of a commercial building with a multilayer perceptron, Journal of Building Engineering. 75 (2023) 107016,
56
57
58
662 [89] R. Liang, T. Le-Hung, T. Nguyen-Thoi, Energy consumption prediction of air-conditioning systems in eco-buildings using
59
663
60 hunger games search optimization-based artificial neural network model, Journal of Building Engineering. 59 (2022) 105087,
61
62
63
64
65
665 [90] B. Mobaraki, S. Komarizadehasl, F.J. Castilla Pascual, J.A. Lozano-Galant, Application of Low-Cost Sensors for Accurate
1
666 2 Ambient Temperature Monitoring, Buildings. 12 (2022) 1411, https://doi.org/10.3390/BUILDINGS12091411.
667
3 [91] J. Saini, M. Dutta, and G. Marques, Sensors for indoor air quality monitoring and assessment through Internet of Things: a
4
668 5 systematic review, Environmental Monitoring and Assessment. 193 (2021) 1–32, https://doi.org/10.1007/S10661-020-08781-
6
669 7 6/TABLES/4.
670 8 [92] J. Xie, H. Li, C. Li, J. Zhang, M. Luo, Review on occupant-centric thermal comfort sensing, predicting, and controlling,
9
10
671 Energy and Buildings. 226 (2020) 110392, https://doi.org/10.1016/J.ENBUILD.2020.110392.
11
672
12 [93] J. González-Arteaga, J. Alonso, M. Moya, O. Merlo, V. Navarro, Á. Yustres, Long-term monitoring of the distribution of a
13
673 building’s settlements: Sectorization and study of the underlying factors, Engineering Structures. 205 (2020) 110111,
14
674
15 https://doi.org/110.1016/J.ENGSTRUCT.2019.110111.
16
675
17 [94] A. Tanhadoust, T.Y. Yang, F. Dabbaghi, H.K. Chai, M. Mohseni, S.B. Emadi, S. Nasrollahpour, Predicting stress-strain
18
676 behavior of normal weight and lightweight aggregate concrete exposed to high temperature using LSTM recurrent neural net-work,
19
677
20 Construction and Building Materials. 362 (2023) 129703, https://doi.org/10.1016/J.CONBUILDMAT.2022.129703.
21
678
22 [95] D. Dũng, V.K. Nguyen, and D.T. Pham, Deep ReLU neural network approximation in Bochner spaces and applications to
23
679 parametric PDEs, Journal of Complexity. 79 (2023) 101779, https://doi.org/10.1016/J.JCO.2023.101779.
24
680
25 [96] Q. Gong, W. Kang, F. Fahroo, Approximation of compositional functions with ReLU neural networks, Systems & Control
26
681 Letters. 175 (2023) 105508, https://doi.org/10.1016/J.SYSCONLE.2023.105508.
27
28
682 [97] T.T. Truong, D. Dinh-Cong, J. Lee, T. Nguyen-Thoi, An effective deep feedforward neural networks (DFNN) method for
29
683
30 damage identification of truss structures using noisy incomplete modal data, Journal of Building Engineering. 30 (2020) 101244,
31
32
33
685 [98] J. Chen, G. Shang, Localization and imaging of internal hidden defects in concrete slabs based on deep learning of vibration
34
686
35 signals, Journal of Building Engineering. 76 (2023) 107087, https://doi.org/10.1016/J.JOBE.2023.107087.
36
687 [99] R. Kazemi, M.Z. Naser, Towards sustainable use of foundry by-products: Evaluating the compressive strength of green con-
37
38
688 crete containing waste foundry sand using hybrid biogeography-based optimization with artificial neural networks, Journal of
39
689
40 Building Engineering. 76 (2023) 107252, https://doi.org/10.1016/J.JOBE.2023.107252.
41
690 [100] D.P. Kingma, J.L. Ba, Adam: A Method for Stochastic Optimization, 3rd International Conference on Learning Represen-
42
43
691 tations, ICLR 2015 - Conference Track Proceedings. https://arxiv.org/abs/1412.6980v9, 2014 (accessed 01 Aug. 2023).
44
692
45 [101] S. Emami, G. Martinez-Munoz, Sequential Training of Neural Networks with Gradient Boosting, IEEE Access. 11 (2023)
46
693 42738–42750, https://doi.org/10.1109/ACCESS.2023.3271515.
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
Declaration of Competing Interest (use
"https://declarations.elsevier.com/home")
Declaration of interests
☒The authors declare that they have no known competing financial interests or personal relationships
that could have appeared to influence the work reported in this paper.
☐The authors declare the following financial interests/personal relationships which may be considered
as potential competing interests:
Author Statement (please upload in Doc format only))
Author statement
F. Lozano (First Author): Conceptualization, Methodology, Software, Investigation,

Formal Analysis, Writing - Original Draft;
S. Emadi: Validation, Investigation, Visualization, Writing - Review & Editing
M. Komarizadehasl: Data Curation, Methodology
J. González-Arteaga: Investigation, Formal Analysis
Ye Xia (Corresponding Author): Conceptualization, Funding Acquisition,

Resources, Supervision, Project administration, Writing - Review & Editing.
Revised Manuscript with Changes Marked (This item is NOT Click here to view linked References
sent to production - for evaluation process only)
1 Abstract: The development of low-cost structural and environmental sensors has sparked revolutionized monitoring practices a transformation
2 across numerous fields, offering enabling cost-effective solutions for monitoring infrastructures and building health assessment. However, a critical
3 challenge associated with these devices sensors is their long-term durability and reliability. Surprisingly, despite thisitsthe significant interest in
4 these low-cost devices, the literature does not present any solutions for ensuringto assure their long-term performance of low-cost sensors. To
5 address this gap, this paper study proposes an innovative Artificial artificial Intelligence intelligence(AI)-based approach for evaluating the long-
6 term performance of athe low-cost inclinometers withLARAusing a (lLow-cCost aAdaptable rReliable Anglemeteranglemeter). This method au-
7 tomatically compares the inclinations of actual on-siteonsite measurements with the predicted valuones under real environmental conditions. Over
8 time, if the errors discrepancies between both measurements surpass a predefined statistical threshold, it may indicate signal potential inaccuracies
9 in the low-cost inclinometer, thereby suggesting the need for recalibration, or presence of structural pathologiesanomalies. The effectiveness and
10 applicability of the proposed tool are were demonstrated through athe long-term study conducted on a real steel frame located in Spain.
11 Keywords: long-term building monitoring; low-cost sensor; inclinometer; artificial intelligence; multilayer perceptron; neural network
12
13 1. Introduction
14 Structural System system Identification identification (SSI) methods play a crucial role in Architecturalarchitectural, Engineerin-
15 gengineering, and Construction construction industries (AEC), serving as indispensable tools for informed maintenance decision-
16 making [1]. These SSIThese methods facilitate the proactive identification of structural issues and effective risk mitigation by ena-
17 bling the early detection, localization, and quantification of structural damage [2]. SSI methods rely on valuable information pro-
18 vided by Structural structural Health health Monitoring monitoring (SHM) systems [3], equipped with sensors such as accelerome-
19 ters, strain gauges, and inclinometers. These sensors capture data related to structural responses and variations, providing insights
20 into the presence, location, and severity of damages. The combined approach of SSI and SHM enables an early detection, localiza-
21 tion, and quantification of damages, ultimately contributing to the overall safety of buildings.
22 Inclinometers, also known as tilt sensors, play a pivotal role in SHM systems by measuring the angular inclination of a structure
23 with respect to the vertical axis. They provide critical insights into the mechanical behavior and stability of structures. Their signif-
24 icance lies in their ability to detect subtle changes in the inclination, which can indicateare indicators of potential structural defor-
25 mations or damages.
26 Commercial inclinometers hold a significant place in theare essential literaturein structural monitoring research becausethanks ofto
27 their wide availability, reliability, and established performance characteristics. A key advantage of this these type of sensors is their
28 well-established calibration and validation processes. Manufacturers rigorously test and calibrate these devices to ensure accurate
29 and consistent measurements, providing researchers with a reliable baseline for their monitoring and identification endeavors. In
30 fact, the documented specifications and performance metrics associated with commercial inclinometers have contributed to their
31 widespread acceptance in the scientific literature.
32 AlthoughWhile commercial inclinometers are documented in the literature with a wide range of technical specifications and prices,
33 their costs typically exceed those of other alternative low-costinexpensive monitoring solutionsdevices, such as the LARA (lLow-
34 cCost aAdaptable rReliable Anglemeteranglemeter (LARA) [4]. Compared withto commercial alternatives, low-cost inclinometers
35 offer a range ofmany advantages that make them highly appealing for various applications. First, their affordability allows for larger-
36 scale deployment, enabling the comprehensive monitoring of multiple locations or structures within a given budget. This scalability
37 is particularly beneficial when high-density sensor networks are required. Moreover, the flexibility and adaptability of low-cost
38 solutions allow for customization and integration with existing systems, fostering innovation, and opening up opportunities for
39 tailored monitoring approaches in diverse fields. Furthermore, the rapidly advancing technology behind low-cost sensors is contin-
40 ually improves improving their capabilities, ensuring that they keep pace with the evolving monitoring needs. These advantages
41 have led to increased attention in developing low-cost monitoring solutions in recent years. A comprehensive review of the appli-
42 cations of low-cost solutions to building monitoring was has been presented by Mobaraki et al. [5].
43 Nevertheless, the inherent cost-effectiveness of low-cost sensors often comes with trade-offs, impacting factors such as precision,
44 calibration stability, and the susceptibility to environmental conditions. These challenges have prompted essential considerations
45 regardingconcerning the long-term performance and accuracy of low-cost sensors. Indeed, most low-cost inclinometers primarily
46 focus on short-term monitoring, necessitating periodic verification using commercial solutions to ensure long-term enduring accu-
47 racy and reliability. However, this on-siteonsite verification process faces two challenges. Firstly, it lacks automation, and secondly,
48 it inevitably increases monitoring costs, significantly reducing the economic advantages of utilizing low-cost inclinometers. To
49 address these limitations, this studypaper proposes an innovative Artificial artificial Intelligence intelligence (AI)-based approach
50 to automate the evaluation of the long-term performance of the low-cost inclinometer LARA. The proposed methodology uses an
51 artificial neural network to predict the inclination of a monitored structure under specific environmental conditions. By comparing
52 the errors discrepancies between the actual inclinations on-siteonsite and the predicted ones, potential issues arising from inaccura-
53 cies in the performance of the low-cost inclinometer can be automatically identified, suggesting the need for sensor recalibration.
54 Additionally, the system also sends digital alarms through a dedicated Internet of Things, IoT, platform when a statistical threshold
55 is surpassedexceeded. The proposed methodology wais validated on an actual steel frame in Spain based solely on the inclinations
56 and environmental data (specifically, humidity and temperature) obtained from low-cost devices. It is essential to highlight that-
57 Notably, the primary limitation of the proposed methodology lies in the fact that the developed AI tool cannot be directly employed
58 in other case studies, because it requiresas it needs individualized training using on-siteonsite monitoring data for each structure,
59 which is a time-consuming process. To ensure that the proposed methodology is well- suited for practical applications, the authors
60 plan to analyze different structural configurations and loading conditions in the near future.
61
62 The remainder of this paper is organized as follows:. Section 2 providesents a literature review encompassing the key topics ad-
63 dressed in thise studypaper. Section 3 provides a detailed analysis of the main characteristics of the LARA, including its design and
64 key components. Section 4 introduces a novel AI tool for studyingto study the effects of environmental factors, specifically temper-
65 ature and humidity, on the inclinations measured by the LARA in actual steel structures. The application of this methodology is
66 illustrated described in Section 5, where an actual steel frame in a building is examined to understand the influence of environmental
67 parameters on LARA inclinations. Finally, Section 6 presents the primary conclusions drawn from the study, summarizes the main
68 findings, and discusses their implications.
69
70 2. Literature Review
71 This section begins with an overview of the primary structural system identificationSSI methods found in the literature. Subse-
72 quently, a detailed examination of the key characteristics of commercial inclinometers discussed in the existing literature is presented
73 follows. The section then shifts its focus to the features and challenges associated with prominent low-cost inclinometer alternatives
74 outlined in the literature, along with a discussion of their primary technical challenges.
75
76 2.1 Structural System IdentificationSSI
77
78 Sirca and Adeli [6] comprehensively reviewed the system identification methods in structural engineering. This Their review cate-
79 gorizes SSI techniques into five main approaches: (1) Conventional conventional model-based [7-10], (2) biologically inspired, (3)
80 signal processing-based, (4) chaos theory and fractal concept-based [22-24], and (5) multiparadigm approaches. Conventional
81 model-based approaches [7-10], which typically rely on a structural model to identify the structural parameters from the on-siteonsite
82 structural responses. However, their applicability may be limited when dealing with large real-world structures that exhibit a non-
83 linear behaviors. (2) Biologically- inspired approaches: These methods are based on soft computing techniques, such as neural
84 networks [11-13], genetic algorithms [10,14,15], and particle swarm optimization [16-18], (3). Signal processing-based approaches:
85 These methods employ signal processing techniques, such as wavelets [19-21], ]. (4) Chaos theory and fractal concepts [22-24], and
86 (5) Multi-paradigm approaches: In an attempt to tackle complex, nonlinear problems, multiparadigm approaches these methods
87 integrate two or more of the previously mentioned computing strategies. Examples of this approach areinclude those presented in
88 [25-27]. Based on the nature of the measured structural response, SSI methods can be categorized as static [28] or dynamic [29].
89 2.2 Inclinometers
90
91 Among the diverse SHM devices, inclinometers, also known as tilt sensors, merit special recognition owingbecause toof their ex-
92 tensive use in various disciplines, particularly in the SHM of bridges [30] and buildings [31]. For instance, Huseynov et al. [32] and
93 Hester et al. [33] utilized rotation measurements to detect damage in bridges, whereas Lei et al. [34] employed rotations to identify
94 damage in building beams. The essential characteristics of commercially available inclinometers are summarized in Table 1, which
95 includes details such as inclinometer name, measurement range, resolution, sampling rate, price, and corresponding references.
96 -
97
98 The aAnalysis of the commercial solutions listed in Table 1 reveals a significant price variations, ranging from €350 to €15001.
99 High-priced iInclinometers with higher prices (such as the EL Tiltmeter SC) usually provide a lower measurement range and higher
100 resolution.
101
102 2.3 Low-cost inclinometers
103
104 ToIn exploreing low-cost solutions for inclinometer development, a notable approach is the use ofto utilize microelectromechanical
105 system (MEMS) accelerometers. The literature provides numerous examples of the successful employment of this technology. For
106 instance, Hoang et al. [41] developed an orientation system for heavy industrial applications, while, and Ha et al. [43] designed a
107 low-cost inclinometer for estimating ground movements. Ruzza et al. [42] used MEMS sensors to measure landslidthe inclinations
108 of landslides, whereas Yu et al. [45] focused on monitoring the swings inof large-scale structures. Ha et al. [30] and Kumar et al.
109 [46] proposed using wireless inclinometers for Structural Health Monitoring (SHM) applications. Table 1 presents the detailed
110 characteristics of low-cost inclinometer solutions.
111
112 Despite the advantages of low-cost inclinometers, this type of monitoring system presents severalparticular inherent challenges.
113 One primary concern is their potentially lower accuracy than more expensive commercial alternatives, which limitsing their appli-
114 cation in cases requiring high precision. Due Owing to cost constraints, low-cost MEMS accelerometers may exhibit reduced pre-
115 cision and calibration stability, leading to potential measurement errors. Factors such as noise, nonlinearity, and cross-axis sensitiv-
116 ity can introduce inaccuracies into the measured acceleration data. Komarizadehasl et al. [4] developed a monitoring system known
117 as athe low-cost adaptable reliable angle meter (LARA) to address the accuracy issues associated with low-cost inclinometers.
118 Unlike other solutions in the literature, the LARA employs five MPU9250 MEMS accelerometers to enhance its performance. Table
119 1 provides an overview of the main characteristics of this low-cost inclinometer, including a comparison with a single MPU9250
120 sensor used as an inclinometer. The aAnalysis of the table illustrates indicates a significant reduction in the noise error of LARA
121 (0.002°) compared with other low-cost solutions such as the xSCA100T (0.0025°), MPU9250 sensor (0.055°), and LSM9DS0
122 (16000×10-5°).
123
124 Furthermore, most low-cost MEMS accelerometers, such asincluding the ADXL345 sensor, are sensitive to environmental param-
125 eters, such as temperature variations, which can significantly impact their performance and reliability. Mukhopadhyay et al. [47]
126 stated that environmental effects can lead to drift and calibration shifts, resulting in significant substantial deviations between the
127 measured and actual acceleration values. Moreover, changes in temperature and humidity can potentially compromise both the
128 durability and long-term stability of these monitoring devices, because they are more susceptible to mechanical wear, sensitivity
129 degradation, or premature failure. It is worth noting that the impact of environmental parameters is not limited to low-cost incli-
130 nometers; even certain commercial inclinometers that utilize fluids may also be affected by environmental changes [48].
131
132 In the literature, the methods proposed to address mitigate the adverse effects of environmental parameters on MEMS accelerome-
133 ters can be categorized into two primary groups: hardware and software methods [47]. The hardware method involves circuit control
134 and structural optimization to enhance device features. Examples of the applications of the hardware compensation method can be
135 found in the studies by Wang et al. [49] and Liu and Ma [50]. The software compensation method aims to improve the mathematical
136 equations used to obtain measurement readings. Various mathematical approaches have been employed to compensate for the effects
137 of temperature on the MEMS accelerometers. Notable examples include the use of polynomials, as demonstrated by Ruzza [51];
138 finite element analysis, as employed by Zhang [52];, cubic spline interpolation, as utilized by Khankalantary [53]; and swarm opti-
139 mization, as proposed by Yu [54]. Caspani et al. [55] employed a similar software compensation approach to correct for the effects
140 of temperature on the long-term responses of prestressed concrete bridges using a parametric equation.
141
142 Similar toLike accelerometers, low-cost inclinometers based on microelectromechanical systems (MEMS) technology are suscep-
143 tible to environmental factors. Several studies, such as those conducted by Yuan et al. [56], Fontanella et al. [57,58], and Yang et
144 al. [59], have been presented in the literature to addressed the thermal re-recalibration of low-cost inclinometers. HoweverNotably,
145 unlike in contrast to most MEMS accelerometers and inclinometers, the LARA is not sensitive to variations in environmental pa-
146 rameters.
147
148 Environmental parameters, such as temperature and humidity, can introduce variations in material properties and significantly con-
149 siderably affect structural performance. There is aA clear correlation exists between the temperature variations and the deformation
150 of the structures. Consequently, numerous researchersscholars have examined the effects of changes in environmental parameters
151 on the static and dynamic responses of structures. Yao et al. [60] studied the effects of temperature and humidity changes in a
152 wooden frame structure, whereas Cai et al. [61] analyzed the effects impact of temperature on the natural vibration characteristics
153 of a reinforced concrete beam. Zhao and Jin [62] examined investigated the effects of temperature on the dynamic response of a
154 steel crane, and while Ribeiro et al. [63] provided a systematic review of the effects of environmental conditions on Structural Health
155 Monitoring (SHM). In the literature review, no similar approaches similar to thosethe one proposed in this work study for automating
156 the assessment of the long-term performance of low-cost inclinometers have been identified. This highlights the novelty and unique-
157 ness of the proposed methodology in addressing this gap in existing research.
158
159 3. Low-cost Adaptable Reliable Anglemeter (LARA)
160 LARA, which is a low-cost, adaptable, and reliable analyzer [64], represents a groundbreaking evolution of the cost-hyper efficient
161 Arduino product (CHEAP) [65]. This device utilizes open-source prototyping tools, specifically Arduino and Raspberry Pi technol-
162 ogies, and a low-cost MEMS accelerometer (MPU9250). Although this type of accelerometer is not commonly used in civil engi-
163 neering and architectural structures because of its limited precision, itthe LARA addresses this issue usingthrough a pioneering
164 approach, that wase first documented in the literature. The LARA incorporates the average readings of five different accelerometers
165 placed at the same point, effectively functioning in conjunction with a multiplexer. Figure 1 illustrates the wiring configuration of
166 the various sensors with anthe Arduino Due microcontroller and its compact version. A compact version of the LARA sensor is
167 shown in Figure 1. . By processing the measurements obtained from these devices, LARA compensates for individual accelerometer
168 errors, resulting in accuracies comparable to those of commercial-grade devices. The application of LARA for dynamic monitoring
169 has been validated for in monitoring actual buildings [66] and bridges [67].
170 -
171
172 In contrast to most accelerometers in the literature, the LARA offers the unique capability of a remote configuration, allowing it to
173 function as either an accelerometer or an inclinometer [4]. When operating as an inclinometer, LARA employs accelerometer gy-
174 roscopes along the X-, Y-, and Z- axes (forming a triaxial inclinometer) to accurately measure the orientation angle of an object in
175 relation to the force of gravity. This versatility distinguishes the LARA from its counterparts and enhances its usefulness in various
176 applications.
177
178 To streamline the process, the main frame of the LARA was has been redesigned. This redesign enables it to automatically initiate
179 data acquisition when it connects to a power source and stops the process after two 2 minutes. When the Raspberry Pi boots up, it
180 establishes an Internet internet connection through Wi-Fi, LAN, or SIM cards and mounts a Google Drive on its hard drive. This
181 configuration allows LARA to download and store the Python codes containing instructions for data acquisition tasks. Following
182 these instructions, the LARA triggers vibration acquisition at a specified date and time. Additionally, a task was is programmed into
183 the Raspberry Pi operating system to regularly monitor and transfer the acquired data files to Google Drive. The hard disk will not
184 become overloaded or reach its storage capacity without an Internet internet connection. Any collected data will be retained on the
185 Raspberry Pi hard drive without an Internet internet connection until the connection is restored or the memory card becomes full.
186
187 The LARA information flow is described in detailed in [67]. This process is illustrated depicted in Figure 2a. and comprises the
188 following steps:
189 1- LARA Initializationinitialization: When powered on, LARA is set to establish an Internet internet connection using Wi-Fi,
190 LAN, or SIM cards, as long as the Wi-Fi credentials match the predefined settings within the LARA configuration.
191 2- Google Drive: Once internet connectivity is established, LARA gains entry into a predetermined Google Drive repository,
192 where all monitoring information is stored.
193 3- Operational Instructionsinstructions: LARA retrieved retrieves the open-source Python code from the designated Google
194 Drive to commence data acquisition. This adaptable code allows users to define settings, such as scheduled vibration ac-
195 quisition and data acquisition duration, without directly reprogramming the Raspberry Pi for every specific task.
196 4- Data Collectioncollection: Upon completion of data acquisition, the files obtained were are transferred to a dedicated folder
197 on a linked Google Drive. This design guarantees the unburdened state of LARA's local hard drive, as long as the Internet
198 internet connection is consistently maintained. Notably, in the event of Internet internet connectivity loss, the gathered data
199 will be retained on LARA's local hard drive until the connection is re-established. The option of employing external hard
200 drives to expand the storage capacity is also possible.
201 5- Data Processingprocessing: During this phase, the data acquired from the LARA are post-processed with the primary ob-
202 jective of estimating the inclination based on the compiled information. Moreover, this stage allows the integration of
203 additional programs or processes to increase estimation precision or execute specific computations. This section phase
204 demonstrates adaptability, whichand can be tailored to meet the unique requirements of monitoring applications.
205 The information flow during the postprocessing phase is summarized displayed in Figure 2b. This flow, which is based onincludes
206 the following steps::
207 1- Input data for the AI tool: The Y-axis inclinations measured by the LARA over time and the on-siteonsite humidity and
208 temperature values were are introduced into the AI tool.
209 2- Estimation of Intelligent intelligent (I) LARA inclinations: The AI tool calculates the X-axis inclinations based on the input
210 data over time.
211 3- Integration of data into the IoT platform: The X-axis rotations estimated by the AI tool are introduced into an IoT cloud,
212 along with the values of the X-axis directly measured by the LARA and the threshold of the MSR, which is obtained
213 through a comparison between the LARA and athe commercial inclinometer. The MSR values, obtained by comparing the
214 AI-tool tool’s estimated X-axis rotations with those directly measured by the LARA, are plotted in a graph along with the
215 MSR threshold on an online IoT platform hosted on the ThingSpeak website.
216 4- Analysis of MSR errors over time: Within the ThingSpeak platform, a MATLAB code was is incorporated to facilitate the
217 assessment and in-depth analysis of data originating from the channels hosted on ThingSpeak. The channel data can be
218 accessed within the MATLAB analysis section of the ThingSpeak website by supplying the channel ID and associated
219 Application Programming Interface key. A specialized alert program was has been developed for online monitoring and
220 task automation. Its principal objective is to continuously monitor the MSR values to ensure that they remain within ac-
221 ceptable limits. Suppose When the deviation between the LARA X-axis value and the I-LARA X-axis estimation exceeds
222 the predefined threshold. In that case, the program automatically triggers an electronic alert.
223 5- Alert: The algorithm integrated into the ThingSpeak platform operates systematically and sends electronic alerts via email
224 when the MSR exceeds the threshold is surpassed. An example of these notification alerts is shown in Figure 2c.
225 -
226
227 The general and primary technical specifications of LARA working as an accelerometer and inclinometer are summarized in Tables
228 2, and 3, respectively.
229 -
230
231 The aBased onnalysis the information in Table 3, shows that the LARA working functioning as an inclinometer can reach a precision
232 of up to 0.002° while working in the static mode for a measurement range between 0° and 4 °.
233
234 4. Artificial IntelligenceAI tool
235 Literature reviews, such as that of Sun et al. [68], have demonstrated a growing interest in utilizing Artificial Intelligence (AI) tools,
236 including machine learning [69], deep learning with Recurrent recurrent Neural neural Networks networks (RNNs) [70], Convolu-
237 tional convolutional Neural neural Networks networks (CNNs) [71], and image identification [72]. RNN and CNN tools have been
238 employed by various scholars, such as Mukhopadhyay et al. [47], Wang [73], Qi [74], and Pan [75], to compensate for the temper-
239 ature effects in low-cost MEMS accelerometers. Hoang et al. [76] used AI tools to improve the accuracy of low-cost MEMS accel-
240 erometers.
241
242 In this section, we introduce a novel Artificial Intelligence (AI) tool designed to understand the impact of environmental factors,
243 specifically temperature and humidity, on the inclinations measured by LARA in actual steel structures. The foundation of this tool
244 is a Multilayer multilayer Perceptron perceptron (MLP) neural network. The following subsections describe the key characteristics,
245 architecture, and training methodology of the proposed AI tool.
246
247 4.1. Multilayer PerceptronMLP
248
249 The proposed AI tool is based on an Multilayer Perceptron (MLP) [77], which is an artificial neural network comprising multiple
250 layers of neurons arranged sequentially. MLP networks are distinguished by their feedforward architecture, which ensuresing a
251 unidirectional flow of information from the input layer, through hidden layers, and finally to the output layer, without any loops or
252 cycles. With an adequate number of hidden neurons and the correctright training data, MLP networks canhave the capability to
253 approximate any continuous function with a specified degree of precision.
254
255 These characteristics make the MLP a powerful and flexible neural network model capable of learning complex patterns and making
256 performing predictions based on input data. More importantly, the MLP requires low-level processing power and time compared to
257 the deepest deeper learning models, which makes it a perfect match for application in LARA real-time measurements. It is to say
258 thatIn other words, LARA has limited power for processing data. In fact, this artificial intelligenceAI tool has been extensively
259 employed across a diverse range of applications, including, among many others, the prediction of material properties [78,79], com-
260 fort assessment [80], life prediction [81], evaluation of structural performance [82,83], and energy consumption prediction in build-
261 ings [84].
262
263 The MLP used in this study was programmed using the scikit-learn package, which is an open-source machine-learning library for
264 Python. This library has several key features. First, it equips researchers with robust tools for data preprocessing, encompassing
265 essential tasks, such as feature scaling, categorical variable encoding, handling missing values, and feature extraction. In addition,
266 it seamlessly integrates with other Python libraries, enabling efficient data manipulation and visualization throughout the model
267 deployment process. The benefits of the scikit-learn library have led to its widespread use in various applications. For example, Ren
268 et al. [85] used the scikit-learn package to determine the optimal composition of cement grouting materials, while Jain et al. [86]
269 applied it to forecast data-driven energy consumption in residential buildings, . Günaydın et al. [87] used utilized scikit-learn it to
270 evaluate concrete strength, and Sanzana et al. [88] used employed it to predict energy storage in commercial buildings. Moreover,
271 Riu et al. [89] used scikit-learn it to estimate the consumption of air- conditioning systems in eco-buildings.
272
273 4.2 MLP architecture
274
275 The MLP models were designed to acquire the ability to accurately predict inclinations along the X-axis using Y-axis rotations,
276 temperature, and humidity data obtained from low-cost sensors. SubsequentlyConsequently, the trained model can be employed to
277 identify anomalous measurements by comparing its predicted results with the actual rotations on the xX-axis measured by LARA.
278 The lack of a statistical correlation between the predicted and measured rotations maymight indicate a calibration problem, an
279 external load not being considered, or structural damage.
280
281 Figure 3 depicts a The schematic representation of the proposed Multilayer Perceptron (MLP) is depicted in Figure 3, . The MLP
282 encompassincomprisesg the following components:
283 - Input layer, including three input neurons representing the normalized values of the on-site Y-axis rotations, temperature,
284 and relative humidity measured on-siteonsite. To normalize the included data outliers included, the “RobustScaler” feature
285 was is used. The Y-axis rotations were are directly measured by LARA, whereas. In contrast, the latter two parameters
286 were are measured using athe low-cost sensor (SHT22). This device has extensive applications in various fields. For in-
287 stance, Mobaraki et al. [90] employed SHT22 sensors for ambient monitoring, while Saini et al. [91] extensively explored
288 their use for monitoring indoor air quality, and. Furthermore, Xie et al. [92] utilized them for thermal comfort sensing.
289 - Two hidden layers denoted as i and j, with m and n hidden neurons, respectively.
290 - Output layer, consisting of a single neuron representing LARA's X-axis rotation.
291
292 The configuration of the hidden layers in this architecture was customized based on the data of the case study, as detailed in the
293 subsequent section.
294 -
295
296 5. Case studyStudy: Steel gable Gable frameFrame
297 In this section, the characteristics of the structure under study are described, along with the main characteristics features of the
298 monitoring program. NextSubsequently, we present anthe application of the AI tool detailed in the previous section is presented.
299 Finally, the results obtained from the application of the tool were are analyzed.
300
301 5.1 Description of the structure and monitoring program
302
303 The monitored structure was a steel gable frame located at the Institute of Technology of the University of Castilla-La Mancha,
304 Cuenca, Spain. It measured 11.5 meters in length and 6 m in height, as shown illustrated in Figure 4a. The construction consists of
305 HEB 300 columns and IPE 400 beams. Figure 4b shows the cloud point, and while Figure 4c presents the Building building Infor-
306 mation information Model model (BIM) forof this structure.
307
308 ThisThe selection of this structure was selected based on its advantageous characteristics forin mitigating wind effects, which were
309 attributed to its strategic location and the presence of wind-protective parapets on itsthe roof, as shown displayed in Figure 4d.
310 Furthermore, roofs experience occasional live loads associated with maintenance activities, limiting their impact. Consequently,
311 structural variations in this structure frame are primarily influenced by environmental factors such as temperature and humidity.
312
313 Its structureThis frame has been continuously monitored since its construction in 2008 to investigate the impacts of environmental
314 variables and foundation settlement. This long-term monitoring program comprised 27 measurement points for soil water content,
315 four inclinometers, four thermometers, a weather station, and 22 points for topographical leveling. For The detailed information on
316 the sensor characteristics of the sensors, please refer tocan be referred to in the study by González-Arteaga et al. [93].
317
318 Among the installed monitoring solutions installed in this structure, emphasis was placed on those incorporated into the gable frame
319 under study. These solutions utilize the inclinometer model EL Tiltmeter SC [35] from Durham, which measures the one-axle
320 inclination and temperature. Figure 5a shows a photographpicture of the sensor installed on site along the direction of the inclination
321 direction. The key technical specifications of the device are listed in Table 4.
322 -
323
324 As shown depicted in Figure 5b, a commercial inclinometer was positioned at the junction between the beam and the column on the
325 east façade and. This inclinometer recorded readings at 15-minute intervals. TheIt is essential to note that the measurements from
326 these sensors cannot be accessed wirelessly; instead, they must be manually extracted using a USB. The sensors were connected
327 directly to the electrical power supply of the building.
328 -
329
330 The low-cost monitoring solution incorporated in this structure consists of two sensors: (1) LARA, which measures determines the
331 inclinations along three axes, and (2) DHT22, which measures the environmental ambient temperature and humidity. Both sensors
332 were installed at the same location as a the commercial inclinometer. Figure 6a displays pictures of the low-cost sensors installed
333 onsite, with the rotational axes considered by LARA (X, Y, and Z) highlighted. Information from both sensors was transmitted
334 wirelessly over the Internet internet and displayed on the ThingSpeak IoT platform, as shown in Figure 6b. The reading interval for
335 both sensors was set to 15 min to facilitate comparison with a the commercial inclinometer. Both sensors were connected to the
336 electrical power supply of the building.
337 -
338
339 5.2 Architecture of the AI Tooltool
340
341 TheIn this section the architecture of the MLP tool described in Section 4.2 was tailored for the case study. The main characteristics
342 of the input layer, hidden layers, activation functions, output layer, and solver used in the proposed MLP are summarized as follows:
343
344 - Hidden layers: Determining the optimal numbers of hidden layers and neurons is often an empirical process involving
345 experimentation. In this application, the numbers of hidden layers and neurons were determined through an experimental
346 analysis of the model performance using cross-validation datasets. The collected data were randomly divided into three
347 distinct datasets as follows: (1) tTraining dataset, comprising 70% of the data;, (2) cCross-validation dataset, comprising
348 covering 15% of the data;, and (3) tTest dataset, encompassing the remaining 15% of the data. A similar methodology was
349 employed by Tanhadoust et al. [94] to predict the stress-strain behavior of concrete. The network performances for various
350 architectures were analyzed using a cross-validation dataset for different numbers of hidden layers (ranging from 1 to 10)
351 and neurons (ranging from 1 to 6). Based on the results of this empirical process, we determined that anthe MLP architec-
352 ture with three hidden layers, each containing five neurons, delivered the best performance. After investigating the learning
353 rate, the initial value of this parameter was fixed at 0.0001 to optimize the model performance of the model.
354 - Activation functions: The activation function chosen for the different neurons was the rectified linear activation function
355 (ReLU) from the scikit-learn library. This selection helps prevent exponential growth in the computations required to op-
356 erate a neural network. In additionMoreover, this activation function provides additionalfurther advantages, including sim-
357 plicity and efficient empirical performance. The ReLU function has found extensive applications in various fields, such as
358 mathematics [95,] and [96], structural system identificationSSI [97,98], and material science [99].
359 - Solver: The "Adam" solver, which was used. This solver is based onutilizes the stochastic gradient-based optimizer pro-
360 posed by Kingma et al. [100], was used. Adam works efficiently on relatively large datasets (with thousands or more of
361 training samples or more) in terms of both training time and validation. It and is has been used in several studies such as
362 [101].
363
364 The model variables for the MLP are listed in Table 5.
365 -
366
367 5.3 Results
368
369 The monitoring readings data of both low-cost sensors were collected from their installation between January 21 st and April 1st,
370 2023, resulting in 3328 records. The measurement results, illustrated in Figures 7a-d, depict the variations in temperature, humidity,
371 and inclinations along the X- and Y-axisaxes-, and inclinations along the Y-axis.
372
373 The correlations between temperature, humidity, and Y-axis inclinations in relationwith respect to the X-axis inclinations was were
374 studied using Pearson’s Correlation correlation Coefficient coefficient (PCC). The obtained coefficients indicated a moderately
375 strong linear correlation for temperature (PCC = 0.61), a weak negative linear correlation for humidity (PCC = -0.21), and a strong
376 positive linear correlation for Y-axis inclinations (PCC = 0.80).
377 -
378
379 Figure 8 shows the inclinations and temperatures measured using the commercial inclinometer on February 20th, 2023. The aAnal-
380 ysis of the information in this figure highlights the significant effectinfluence of the temperature on structure rotationthe rotation of
381 the structure. On this specific day, the beam rotations in the structure varied from -0.01º to 0.002º, which were attributed to temper-
382 ature changes ranging from 15ºC to 26ºC.
383 -
384
385 Figure 9 compares the inclinations measured by the commercial sensor with those recorded by the LARA oin the X- axis over two
386 different periods: two months from January 23rd to March 23rd, 2023 (Figure 9. a) and three days from February 20th to 22nd, 2023
387 (Figure 9. b). The wider range of measurements recorded by LARA in Figure 9.a indicates its lower accuracy compared to with that
388 of the commercial device. Moreover, Figure 9.b showsdemonstrates that both sensors displayed similar trends in capturing the
389 influence of environmental parameters over time. It is essential to emphasize that the role of the commercial inclinometer in this
390 study was is to validate the proper calibration of the LARA readings during the training of the proposed MLP (Multi-Layer Percep-
391 tron). The standard deviation and mean squared error (MSE) obtained from comparing the LARA and the commercial sensor over
392 time were 0.02408 and 0.00058, respectively.
393 -
394
395 Figure 10 presents the errors between the X-axis rotations predicted by the MLP tool and the actual measurements obtained by
396 LARA from the test dataset.
397 -
398
399 The Root root Mean Squared ErrorMSE (RMSE), Mean Squared Error (MSE), Pearson’s Correlation Coefficient (PCC), and Mean
400 mean Absolute absolute Error error (MAE) were calculated based on the errors presented displayed in Figure 10. The obtained
401 values are listedsummarized in Table 6. These results highlight reveal the high predictive performance of the developed tool, as
402 evidenced bywith an MSE value of 0.002, and a strong positive linear correlation, as indicated by a PCC of 0.87.
403 -
404
405 The probability density of errors between the rotations measured by LARA and the predicted values with the fitted normal distribu-
406 tion was characterized by a mean of 0.00032 and a standard deviation of 0.0127. This function is illustrated in Figure 11.
407 -
408
409 The calculated MSE values over a predetermined environmental cycle (such as a one day) can indicate the potential necessity for
410 LARA recalibration. This need arises when the acquired errors between the X-axis’ inclinations measured by LARA and the values
411 predicted by the MLP surpass exceed the accuracy threshold obtained by comparing the LARA with the commercial inclinometer.
412 Figure 12 displays the MSE between the LARA’s X-axis inclinations of LARA and the values predicted by the MLP for the period
413 from March 1st to April 1,st 2023. This figure also incorporates the MSE value of 0.00058 obtained from the comparison with the
414 commercial inclinometer, labeled as the MSE threshold. AThe analysis of the results illustrates confirms the reliability of the ob-
415 tained data, as the error threshold was not surpassedexceeded. In the futureHowever, if the MSE threshold is exceeded in the future,
416 it may indicate the necessitya need for recalibrating the low-cost sensors. To notify this scenario, an email alert will be automatically
417 sent.
418 -
419 6. Conclusions
420 Developing low-cost structural and environmental sensors has led to significant advancements in various fields, particularly in the
421 cost-effective monitoring of infrastructure and buildings. However, it is critical to ensureing the long-term durability and reliability
422 of these sensors is a critical issue. The existing literature lacks solutions for guaranteeingto guarantee the sustained performance of
423 low-cost sensors.
424
425 This studypaper introduces a novel Artificial-Intelligence (AI)-based methodology that enables the evaluation evaluating of the
426 long-term performance of low-cost sensors. This evaluation is achieved without high-cost monitoring equipment and relies solely
427 on structural and environmental data (specifically humidity and temperature), specifically humidity and temperature, obtained from
428 low-cost devices. Furthermore, the system automatically sends digital alarms through a dedicated IoT platform to identify when the
429 low-cost sensors that require recalibration.
430
431 The application of the proposed AI tool is was validated using the low-cost inclinometer LARA, which relies on anrelying on the
432 automatic comparison between the inclinations of actual on-siteonsite measurements and those predicted ones under real environ-
433 mental conditions. If, over time, the errors discrepancies between these measurements consistently exceed a predefined statistical
434 threshold over time, it may signal indicate potential inaccuracies in the low-cost inclinometer, indicating suggesting the need for
435 recalibration. The results of the long-term monitoring of a real frame located in Spain underscored the effectiveness and practicality
436 of this approach. Consequently, the proposed approach offers as a reliable and cost-effective solution for automating the assessment
437 of the long-term performance assessment of low-cost structural inclinometers.
438
439 7. Acknowledgements
440 The authors are indebted toThis work was supported by the projects PID2021-126405OB-C31 and PID2021-126405OB-C32
441 funded by FEDER funds through the program —“A Way to Make Europe” and by the Spanish Ministry of Economy and Compet-
442 itiveness MICIN/AEI/10.13039/501100011033/. , and It also received support from the National Natural Science Foundation of
443 China (Grant No. 52278313).
444
445 REFERENCES HAVE BEEN UPDATED TO THE JOURNAL FORMAT

Supplementary Material
Click here to access/download

Supplementary Material
Language_editing_certificate.pdf

Journal of Building Engineering

Uploaded by

Copyright:

Available Formats

Journal of Building Engineering

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Journal of Building Engineering

Uploaded by

Copyright:

Available Formats

Journal of Building Engineering

Enhancing Performance Evaluation of Low-Cost Inclinometers for the Long-Term

Manuscript Number: JBE-D-23-06911R3

Article Type: Research Paper

Section/Category: Structural health monitoring

Keywords: long-term building monitoring; low-cost sensor; inclinometer; Artificial intelligence;

Corresponding Author: Ye Xia, Ph.D.

First Author: Fidel Lozano Galant

Order of Authors: Fidel Lozano Galant

Suggested Reviewers: Hong Yang

Please find the manuscript entitled ‘Enhancing Performance Evaluation of Low-

JBE-D-23-06911R2 Enhancing Performance Evaluation of Low-Cost Inclinometers for the Long-Term

 An innovative method based on artificial intelligence has been proposed to evaluate

the long-term performance of low-cost inclinometers

 Real time automatic identification of potential issues caused by inaccuracies or

structural pathology in low-cost inclinometers

long-term performance of LARA under real-world conditions

48 0 100 200 300 400 500

F. Lozano (First Author): Conceptualization, Methodology, Software, Investigation,

S. Emadi: Validation, Investigation, Visualization, Writing - Review & Editing

M. Komarizadehasl: Data Curation, Methodology

J. González-Arteaga: Investigation, Formal Analysis

Ye Xia (Corresponding Author): Conceptualization, Funding Acquisition,

445 REFERENCES HAVE BEEN UPDATED TO THE JOURNAL FORMAT

Click here to access/download

You might also like