Our Work

  1. Introduction
    1. Problem Statement 

USA's aging infrastructure is currently rated a D+ by the American Society of Civil Engineers’ quadrennial “Infrastructure Report Card”. This rating has stayed constant since it’s previous publication in 2013. To combat its crumbling infrastructure, the US economy is increasingly spending hundreds of billions of dollars in the operation and maintenance of it’s infrastructure. Figure 1 below shows a noticeable increase in the share of maintenance costs from total US Infrastructure spending across multiple years.  
Figure 1: US Infrastructure Spending Trends
The goal of this project was to help streamline current maintenance paradigms and solve the following problems:
  • Emergency maintenances caused by preventable faults
  • Manually collecting and analyzing structural data
  • Lack of real-time diagnostics of  structures
  • Lack of  detailed historical diagnostics reports to help track  issues


    1. Stakeholders needs

This project is funded by The Theoretical & Applied Mechanics Group, L3 Technologies, and Noesis Analytics LLC. The stakeholders have outlined a need for an IoT Device which fits in with existing data acquisition networks, and performs user-defined real-time analytics on the collected SHM data. Figure 2 below shows a tabulated view of the stakeholder needs. The total budget allocated by our stakeholders for this project is $5000.
Figure 2: Initial Device Requirements Based on Stakeholder Needs
Due to disruption caused by the 2020 pandemic, the team was unable to complete the outdoor operations, and Over the Air update  requirements. This was caused by revoked access to Drexel University Laboratories. The rest of the requirements from Figure 2 above were met, and is discussed in the later section of this report.


    1. State of the Art 

Based on a literature review of IoT applications in Structural Health Monitoring, there are three categories of systems namely sensors to board, sensors to Cloud, sensors or boards themselves. 
Sensors to Board
A number of applications connect sensors to single-board computers in applications of IoT in SHM. Abdelgawad & Yelamarthi [2] connected piezoelectric sensors to a Pro-Trinket microcontroller which applies a mathematical model to detect damage. Afterwards, a Raspberry Pi is used to collect structural data and push it to the Cloud using Wi-Fi. Smarsly, Mthunzi, Hahn, & Planer [3] have a similar process tree with an additional remote laptop computer, which will serve as the data archive and user interface for the system.


Sensors to Cloud
Cloud computing provides a methodology to effectively store real-time data. Chang & Lin [4], developed a cloud based system to simulate real-time 3D structural models and displacements. This system is also capable of alerting users when their structure passes a design capacity or detection of possible damage. Paul et al.’s cloud based system [5] generates an alert for users through a smartphone application via Bluetooth Low Energy if the bending angle exceeds the threshold. The raw structural data can also be sent to cloud storage for future research. 


Sensors or Board Themselves
For a real-world application, it is necessary to have an effective wireless connection to transmit sensor data. Fu, Khodaei, and Aliabadi [6], developed a low-power high response wireless structural health monitoring system (WSHMS) which uses low power consumption when no impact occurs and hi performance when a hit is triggered. The system’s response time and energy consumption was measured by testing the developed printed circuit board (PCB). Tong, Yang, Wang, and Miao [7] propose a low-power wireless acceleration sensor and gateway system which was tested on the Chijing Bridge. The vibrational data was transmitted to their Cloud platform.
None of the categories presented above supports real-time analysis of large scale (Gigabytes/minute range) structural health data. 


    1. Approach 

To solve the problems stated in the introduction, we proposed a user-friendly Plug and Play (PnP) IoT device which fits in with existing data acquisition networks and is capable of performing large scale real-time data analytics. The device is built to work within an IoT framework consisting of three layers, namely the Edge, Fog and Cloud. The Edge is composed of either commercial data acquisition systems (DAQ) or by sensors without any commercial system being involved. The Edge layer is connected to the Fog - a decentralized computing layer - and consists of the proposed IoT Device. Within the Fog, real-time data is aggregated, parsed, filtered and passed through a layer of user-defined algorithms. The main goal of the algorithms used at the Fog, is to reduce the incoming data and classify it into known classes. This process allows a real-time data flow to the Cloud, as only important decision-making components of the data are propagated. The algorithms on the PnP device are trained in the Cloud layer using historical data. The Cloud is also responsible for hosting a user interface (UI) to interact with the Edge and Fog Layers. The UI enables users to control their data acquisition and visualize their analytics in (near) real-time. Figure 3 below shows the proposed system flow chart for our project.


Figure 3: Proposed System Flow Chart
  1. Materials and Methods 

    1. Materials

The materials used were cold rolled Low-Carbon Steel Grade 1006-1026 which was milled into the compound tension testing specimens. Acrylonitrile Butadiene Styrene (ABS) was used in 3D printing our device enclosure.
    1. Equipment  

A FLIR A300 series infrared (IR) camera was used to conduct a thermal analysis on the PnP device. Two 5 megapixel (MP) Baumer TXG50 monochrome cameras were used while testing the steel. A MTS Tensile Testing machine located in Drexels Mechanical and Structural Testing Facility, was used to load our samples and run our device alongside it. Connected to the sample were two piezoelectric sensors which picked up Acoustic Emission waveforms.
    1. Software 

The software programs our team used include Pycharm, Visual basic, Notepad++, as well as MongoDB Database Management System. Additional software licenses acquired through Drexel were Fusion 360 and ARAMIS 3D deformation analysis software.
    1. Manufacturing methods

The case of the device prototype was made using a 3D printer and the test specimens were made using a CNC milling machine.


  1. Results 

    1. Constraints and Standards

Table 1: below shows the constraints considered in developing our project. Optimizing these constraints allowed the team to efficiently complete the prototyping phase.
Table 1: Constraints considered in developing the prototype
Constraint
Justification
Device throughput
We want our overall framework to be real-time
Data streaming quality
We want our data to be complete and not missing bits and pieces due to acquisition errors
Accuracy of machine-learning models
We want to have high accuracies for classification between different states of damage all while having a real-time data flow


Table 2 below shows the standards and specifications followed in completing this project. The standards used are in par with comparable commercial devices available in the market. 
Table 2: Standards and protocols used in developing the prototype
Standard/Protocol
Justification
AES, SSL, TSL
Data encryption protocols used between framework layers
SMB
Distributed file system protocol for Edge to Fog connection
UDP, TCP
Members of the internet protocol suite for communication between the Fog and Cloud
RoHS
Standard for restriction of the use of certain hazardous substances in electrical and electronic equipment
ASTM E1856 - 13
Standard Guide for Evaluating Computerized Data Acquisition Systems Used to Acquire Data from Universal Testing Machines
ASTM-E1290-08e1
Testing procedure for compound specimens


    1. Decision Matrices 

One of the first tasks of the project was finalizing the overall system architecture. For this, we narrowed down two alternative architectures as shown in Figures 4a and 4b below. Figure 4a shows a standalone Fog device configuration whereas Figure 4b shows a Multi fog device - gateway configuration. Table 3 below shows the decision matrix used to select our final architecture.
         a)                   b)
Figure 4: Architectures considered in the design phase. 4a)Architecture 1 4b)Architecture 2
Table 3: Decision Matrix for the System Architecture


Another important decision was to choose the main processing unit (device brain) of our proposed fog device. Once the processing unit was chosen, we could then pick out compatible peripherals to meet our project requirements. Table 4 below shows the decision matrix used to select our device’s processing unit.
Table 4: Decision Matrix for the Single Board Computer


    1. Detailed Design

      1. Prototype 1

The PnP device is based on a Nvidia Jetson Xavier SBC (Single Board Computer). Other peripherals include, one 512 GB Samsung 970 PRO SSD (for faster memory read/write speeds), one AC 8265 Intel Dual Band Wireless chip (for WIFI and Bluetooth), one Sierra Wireless MC7455 4G LTE Transceiver Module, and all accompanying connectors/antennas. A Multistar High Capacity 5200mAh battery was packaged with the device to provide backup power in case of power outages. Acrylonitrile Butadiene Styrene (ABS) was used in 3D printing our device enclosure. In Figure 5 below we can see our device in the failed 3D printed case.
Figure 5: Physical Prototype for the PnP IoT Device
3.3. Design Results 


The device enclosure was designed using Fusion 360. Figure 6 below shows our final enclosure design. The enclosure was not successfully printed due to travel restrictions created by the 2020 Covid-19 crisis.
Figure 6: 3D Model for device enclosure
    1. Device Throughput Performance

Acoustic emission data was acquired at the edge (using a DAQ - PCI2) and aggregated on a network shared folder in the PnP IoT Device. An online k-means clustering algorithm was performed on the real-time data. Table 1 below shows the PnP device’s throughput performance results.
Table 5: PnP Device Throughput Test Results
Total Data Size (MB)
Preprocessing Time (s)
Cloud Upload Time (s)
Preprocessing Throughput (MB/s)
Cloud Upload Throughput (MB/s)
1530
24.44
35.98
62.6
42.51
1530
21.28
33.83
63
44.2
1530
20.73
33.41
73.78
45.58
1530
20.72
33.65
73.84
45.45


Avg Preprocessing Throughput (MB/s)
Avg Cloud Upload Throughput (MB/s)
Overall Throughput (MB/s)
Upload Speed At Time of Test (MB/s)
68.3
44.4
44.4
252.83


Acoustic emission data collected from compound tension aluminium specimens were used to classify between crack and no crack states. Table XX - XX below show the results achieved from classification using the Support Vector Machine, Decision Tree and Random Forest algorithms respectively. These algorithms are just a subset of algorithms that can be run on the Fog device. User defined algorithms can also be loaded from the client side web interface.  More data was planned to be collected in the Spring quarter, but the 2020 pandemic caused a change in our plans. This would have helped us train the algorithms better and achieve higher accuracies. 

Table 6. Classification Accuracy for the Support Vector Machine Model on the Fog device  


Table 7. Classification Accuracy for the Decision Tree (DT) on the Fog device  
 


Table 8. Classification Accuracy for the Random Forest (RF) on the Fog device


    1. Website User Interface



The Figure # and # below show the client-side web interface. More about the web interface is elaborated on the discussion section below. 
Figure 7: Login page of user interface


Figure 8:


    1. Additional Analyses 

The goals reached since the Winter term are as follows:
  • Crack detection algorithms implemented for compact tension Aluminium specimens
  • Client side web interface completed
  • Full system test (done with simulated sensor data)


  1. Discussion 

    1. Thermal Analysis

A thermal analysis was done on each individual component of our proposed PnP Device. Based on the results obtained, Acrylonitrile Butadiene Styrene (ABS), which is a member of the thermoplastic family of polymers, was chosen as the 3D print material for the device casing. ABS has a melting point of approximately 105 ºC, which is considerably higher than the maximum temperature recorded in the IR analysis. Figure 9 below shows images from the IR analysis.
(a)(b)
Figure 9 : Sample infrared pictures from IR analysis. (a) System on Module (b) Heat sink.
    1. Code Description

The current code implementation is done using three scripts. All three scripts are run in parallel to achieve concurrency. Below is a short description of each script. 
Script 1: Jetson Edge Stream:The Fog device reads in the chunked .TXT saved by the DAQ/s from a locally shared folder. This process is governed by utilizing a custom Watchdog. The purpose of the Watchdog is to delegate the incoming data and make sure no data is left unread (No Data Loss). The Watchdog keeps on processing the same .TXT file repeatedly until a new .TXT file is available in the shared folder. Once a new file emerges, it signals the Watchdog that no new data will be appended to the old file. As a redundancy routine, the watchdog checks the old .TXT file once again (for modifications) before moving on to the new file. And this process continues.  
Once a file passes the Watchdog, all new data (data not available in previous iterations) is parsed. Each new file gets an object (from object-oriented programming) of its own, and within the object different states (last line read, rows to skip) pertaining to the file is stored. As a new file emerges, the old file’s object is deleted (garbage collection) to avoid memory buildup.  The parsed data (Pandas Dataframe) is then passed through on-line clustering/classification models and saved in a binary file format known as Feather. Feather is chosen as the file dump format since it is a fast, lightweight, and easy-to-use binary file format for storing data frames and can be directly read back in as a data frame in concurrent/future processes.  
Script 2: Jetson Mongo Push:The Jetson Mongo Push script imposes another (less complex) Watchdog for the Feather binary files saved by script 1. This script reads in each successive Feather file and uploads it to a destination of choice in our Cloud MongoDB database. During the upload process, the Feather data frames are converted to a dictionary format resembling BSON. A bulk insertion method is then used to upload the rows of the data frame as individual records. The link between the Cloud layer and the Fog device is achieved using an Internet connection. TLS and SSL encryption are used by MongoDB in this link. 
Script 3: Jetson Plot :The Jetson Plot script is responsible for plotting chosen features saved in the Feather files.  
    1. Web Interface

The user interface is one of the major components in creating a user-friendly environment for our clients. It has been developed using Javascript, python, Ajax, CSS, HTML and Nodejs. A wide range of actions have been embedded in each of the tabs to control the settings of the PnP IoT Device and the algorithms being run. The dashboard tab allows users to select individual DAQ systems and check their status in real-time. The data format settings tab can be used to adjust the type and formatting of the incoming data from each individual DAQ. The algorithm selection tab includes preset algorithms to be run in our device. Currently, the create algorithm and data setting tabs are inactive and require backend adjustments. Lastly, a LED panel has been placed to notify the user about the overall health of their structure. Users will receive an immediate notification about the status of their system via mail or text if the light turns red. Figure 6  in the results section shows the developed online user interface.
    1. Data Analysis

Historical test data saved in the cloud is used to  train machine learning models, that are then sent back to the Fog device for real-time operations. Features extracted from the datasets collected are first normalized and then compared using a Pearson coefficient correlation matrix, as shown in Figure 10 below. In this example,  out of the initial twenty-five features, twenty features were kept after the feature selection process was applied, by removing features with greater than 90% correlation. 
Figure 10: Pearson coefficient correlation matrix


The next step is to perform feature reduction, which decreases the feature space by choosing the optimum number of principal components (PCs) based on finding a point where the residuals of the next principal component do not provide additional information to the principal component space. A user-defined threshold of 95% variance was used to determine the number of PCs to describe the reduced feature space. The cut-off point for the training data is exhibited in Figure 6. As shown, the 10th PC does not significantly vary in the principal component space and is eliminated along with every component after it, giving a reduced feature space of 9 components. 


Figure 11: Variance per Principal Component 


These principal components are then used in the GMM which was performed iteratively using 600 iterations testing the classification results achieved by attempting to group data in one of n clusters. Where n is the number of damage states we plan on classifying later. These clusters are then used as labels to the supervised algorithms discussed in the results section. 
    1. Crack Initiation in Compound Tension Aluminum specimen 

Figure 12: a. Applied loading profile correlated with a scatter plot of the amplitude of all AE signals recorded during testing. b. Evolution in time of the peak frequency of all AE signals, representative AE waveform acquired. c. prior to crack initiation and d. upon crack initiation.
After the data analysis is conducted, we are left with charts and graphs that are identical to the Fig. 5. These graphs are easier to read and pinpoint crack initiation. When the load is noted to suddenly drop accompanied by a noticeable increase in the AE amplitude as shown in Fig. 5(a). This observed increase in the AE amplitude coincided with the visually confirmed onset of cracking on the surface of the specimen. It could be also noted that there was some discrete increases in the AE amplitude prior to crack initiation. Consequently, these discrete increases in the AE amplitude could serve as crack precursors and therefore act as early signs of impending damage. It is expected that the AE waveform would have a short rise time and short duration characterized by a burst type signature which can be easily detected in graphs. The synchronous use of full field optical metrological techniques in conjunction with the AE method provided a cross validated experimental mechanics approach to filter the AE signals associated with the onset of cracking in aluminum alloys. 


  1. Context and Impact 

    1. Economic analysis 

Initial prototyping costs will be significantly different from full scale manufacturing and production. Current prototyping cost lies around $1500 with shipping and handling costs in mind. Full scale manufacturing of the device is estimated around $600 per device. This reduced cost is mainly estimated based on the expected economies of scale. Furthermore, Printed Circuit Boards will be used to replace most of the components used in the prototype. All known SHM monitoring, testing and analysis are performed in labs under ideal or simulated conditions. Supported by the Institute of Electrical and Electronics Engineers (IEEE), Prognostics and Health Management (PHM) Conferences are held pertaining to SHM and topics explored in this paper. PHM goes into the potential cost avoidances which include: “failure avoidance thrown away when performing scheduled maintenance, logistics footprint reduction, repair cost reduction through improved diagnosis and avoidance of collateral damage, reduction in no-fault-founds, reduction in redundancy, warranty verification, waste stream reduction, and improvements in design and qualification of future systems.” 
    1. Environmental impact analysis

The raw materials used to manufacture the proposed device are in high demand and constantly recycled and utilized universally for most computer-based applications. Certain components of the device are cheaper to produce in house and can be produced at a faster rate inhouse. Other components with basic functionalities will be procured from other established companies. Because in-house production is required, jobs will be created from this decision. With dedicated facilities and staff created for the manufacturing the device can easily be modified or upgraded. Packaging may also be automated and because of the proposed size of the device and its components, storage and delivery will be on a need/client basis. Distribution of the product will be on a need/client basis. Because the interface installed on the devices will be very versatile in its applications, the device will be used widespread and possibly recycled or implemented in other designs/equipment being utilized. 
    1. Social impact analysis
The first notable impact is the availability of jobs. The device comes off as another automated device capable of replacing a human engineer. This is not the intended idea. This device is to be implemented alongside the current methods of determining material lifespans. The application of this device is real time and does help nigate issues that can be experienced as described in the background of this project but shall be used as a third line of defence in safety. Jobs may be created in hopes that engineers will take up the programs within the device and how to use them to make complete safety assessment reports of materials and operations. Explored in PHM’s paper, health monitoring and this device “represents a shift away from routine inspections, which are currently scheduled based on ensemble statistics obtained on similar structures, to inspections that are scheduled based on the actual state of the structural health of individual structures.” Depending on the sensors and monitoring conducted, a variety of unthought issues can be addressed while utilizing on staff engineers with knowledge of the structure.
The sensors capable of connecting to the device are noninvasive to the structures being monitored. There will be no impact to the functionality of the structure being monitored unlike other methods of health monitoring. Outside of internal company infrastructures is the general public who utilizes said structures. Current methods of analyzing and fixing bridges causes major inconvenience and traffic delays. With the help of routine noninvasive monitoring, engineers can avoid trial and error analysis and work on the core issue in a timely manner.
    1. Ethical analysis

Typical to any engineering project, there are two types of approaches taken. These approaches are based on knowledge that the engineer has access to, Theoretical and Experimental. A theoretical approach uses proven equations and constants to design while experimental uses data collected from previous alike designs and using real-time data to make analysis and designs. A large scale project known for primarily using a theoretical approach for design was the Tacoma Narrow Bridge in Washington State. In 1940 the bridge collapsed and “investigations around that time conclude that the probable cause was self-induced vibrations driven by vortex shedding as the wind passed around the solid plate girders.”[9] To avoid devestations as such, current engineers use a mix of theoretical and experimental knowledge with access to computer programs such as finite elements to determine bridge design. 


  1. Project Management Update 

    1. Team Organization 

The team is made up of three Electrical and Computer Science and two Mechanical Engineers, and is under the direction of Dr. Antonios Kontsos. The Electrical and Computer Science Engineers have been working on the web interface, translation and processing of data within the device along with its internal composition. The two Mechanical Engineers have applied mechanical theory and knowledge to the raw data to narrow down the data acquired from the sensors as to how it should be read and how changes shall be applied. Along with the basic composition of the device and its aesthetic.
    1. Schedule and Milestones

The schedule for this project was to first determine the system architecture and the state of the art analysis. With these completed, the next step was the procurement of the system components and samples for testing the prototype. After all the parts of the device arrived the components of the device were put together and measured to design the casing. Once the device setup was completed multiple lab test runs to ensure accuracy of the system’s data acquisition/analysis were planned, but were not completed because of COVID 19. After passing the lab tests the device was to be deployed in the field for further testing and verification of the system and its capabilities. Once all the tests were completed and calibrated, a final layer of security was to be added to safeguard the data during transfer. 
    1. Project Budget

The budget for this project was a total of $5,000. The funds were provided from our stakeholders TAMG, L3Harris Technologies and Noesis. Key items considered for the overall prototype include the Jetson, SSD, RAM, Wi-Fi and 4G LTE Modules and Router Module. These items are estimated at $426-$695 without shipping and handling.


  1. Summary and Conclusions

The past year spent behind bringing our senior design proposal to life has been an enlightening experience for everyone on the team. Everyone grew as a person, and moreover as Engineers. Going through the different phases in designing and prototype has given us acute insight into what it takes to make high-performance commercial products. Navigating senior year, making professional connections to aid our project, and especially making changes to our plans to accommodate disruptions caused by the pandemic has helped everyone involved expand their personal toolsets. The most challenging part of this project was getting our multi-disciplinary team of Engineers on the same page . But when everyone did get on the same page (regarding the design philosophy), the eclectic group of majors in the team also became one of our biggest strengths. We were able to effectively modularize our project into the mechanical and electrical subcomponents, which increased the efficiency of our work and led to a physical prototype by the end of the winter quarter. Overall, the senior design project was successful. We have successfully produced a device prototype which allows real-time analysis of large scale structural data. Furthermore, we also developed a web based user interface that allows users to control analysis parameters (such as the model to be run) and view real-time statistics about the health of their structures.
    1. Future Work

Due to COVID-19, we were unable to conduct more testing on samples other than aluminum. We would have prefered using steel as it is a commonly used material in structures.

Testing Plan

To be able to implement  macro-scale real time data analytics in applications, lab-scale analysis will be performed. In this respect, Mode 1 - Tension Crack Growth Tests will be  achieved using cold rolled Low-Carbon Steel Grade 1006-1026. Specimen geometries are designed to localize crack initiation, shown in Fig. 7 and are extracted using Electrical Discharge Machining  (EDM) in Drexel University Engineering Machine Shop. Figure 8 displays the steel machined to our dimensions. ASTM-E1290-08e1 testing procedure will be followed. The specimen will be clamped to the MTS Tensile Testing machine via clevis clips and then will be preloaded at a rate of 0.25mm/min until 150N is reached. Then a continuous load rate of 0.5 mm/min will be applied until 2mm of crack growth is physically observed using Digital Image Correlation (DIC). The DIC equipment was manufactured by GOM Optical Measuring Techniques and consisted of two 5 megapixel (MP) Baumer TXG50 monochrome cameras providing imagery to the commercially available ARAMIS 3D deformation analysis software. The cameras had a maximum acquisition rate of 7 Hz at full (5 MP) resolution and 29 Hz with binning at reduced resolution (2.5 MP). Each camera has a 2/3” Charged Coupled Device (CCD) sensor, and an image resolution of 2448  2050 pixels (width x height). A field of view (FOV) of 65x55 mm2 will be chosen for the experiment. AE monitoring was performed during loading using two PICO sensors having a frequency range from 150 to 750 kHz. AE activity was recorded using a PCI-2 data acquisition board and all signals were pre-amplified by 40 dB using a 2/4/6 preamplifier. Hits were recorded at 10 MSPS to avoid aliasing of the signals with a Peak Definition Time (PDT), Hit Definition Time (HDT), and Hit Lockout Time (HLT) of 50, 300, and 300 μs respectively to minimize the loss of hits from a large HLT while still recording the full signal. All AE equipment is produced by MISTRAS Group. All data will be collected by our device and filtered to display the specific information we desire in an aesthetic fashion.


Figure 13: Mild/Low Carbon Steel Compound Tension Specimen Dimensions
Figure 14: Mild/Low Carbon Compound Tension Specimen


  1. References

[2]Abdelgawad, A., & Yelamarthi, K. (2016). Structural health monitoring: Internet of things application. Paper presented at the Circuits and Systems (MWSCAS), 2016 IEEE 59th International Midwest Symposium on.
[3]Smarsly, K., Mthunzi, E., Hahn, O., & Planer, J. (2019). Validation of an ultra-low-cost wireless structural health monitoring system for civil infrastructure. 
[4]Chang, H.-F., & Lin, T.-K. (2019). Real-time Structural Health Monitoring System Using Internet of Things and Cloud Computing. arXiv preprint arXiv:1901.00670. 
[5]C. M. Bishop, Pattern recognition and machine learning. springer, 2006. 


[5]Paul, P., Dutta, N., Biswas, B. A., Das, M., Biswas, S., Khalid, Z., & Saha, H. N. (2018). An Internet of Things (IoT) Based System to Analyze Real-time Collapsing Probability of Structures. Paper presented at the 2018 IEEE 9th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON).
[6]Fu, H., Khodaei, Z. S., & Aliabadi, M. F. (2018). An event-triggered energy-efficient wireless structural health monitoring system for impact detection in composite airframes. IEEE Internet of Things Journal, 6(1), 1183-1192. 


[7]Tong, X., Yang, H., Wang, L., & Miao, Y. (2019). The development and field evaluation of an IoT system of low-power vibration for bridge health monitoring. Sensors, 19(5), 1222. 

Comments

Popular posts from this blog