Optimal Constrained Wireless Emergency Network Antennae Placement

With increasing number of mobile devices, newly introduced smart devices, and the Internet of things (IoT) sensors, the current microwave frequency spectrum is getting rapidly congested. The obvious solution to this frequency spectrum congestion is to use millimeter wave spectrum ranging from 6 GHz to 300 GHz. With the use of millimeter waves, we can enjoy very high communication speeds and very low latency. But, this technology also introduces some challenges that we hardly faced before. The most important one among these challenges is the Line of Sight (LOS) requirement. In the emergent concept of smart cities, the wireless emergency network is set to use millimeter waves. We have worked on the problem of efficiently finding a line of sight for such wireless emergency network antennae in minimal time. We devised two algorithms, Sequential Line of Sight (SLOS) and Tiled Line of Sight (TLOS), both perform better than traditional algorithms in terms of execution time. The tiled line of sight algorithm reduces the time required for a single line of sight query from 200 ms for traditional algorithms to mere 1.7 ms on average.

Engineering for providing access to the hardware required for this project.Special tanks to Dr. Teng Moh for teaching me approximation algorithms which were useful in methods devised in this project.
We would also like to thank Department of Computer Science for help in setting up this project defense.We would also like to express our thanks to the library for giving required books and materials to learn different aspects and concepts of our project and extended study hours for students.
Last but not the least, I would also like to heartfully thank my friends and family for their moral support, encouragement, and blessings without which I could not able to finish my project successfully.

Introduction
In case of natural disaster or an emergency, the ability to communicate with outside world is of the highest consequence to resolve the situation.A large floor or an earthquake may damage power cables, telephone lines or cellular antennae resulting in loss of traditional communication lines such as the landline phones or the cellphones.
In such situations, emergency responders still must have a reliable communication with each other and outside world to effectively mitigate the situation.To address this need, Office of Emergency Services (OES) of City of San José and similar offices in other major cities are deploying Wireless Emergency Networks (WENs).These networks will allow fire department, police department, hospitals, shelters, and food banks to communicate securely with each other and the outside world.For deploying these WENs, the most important question is to identify suitable locations to install wireless access points (APs) within the limits of cities.The WENs will be constructed using millimeter wave antennae.There are multiple reasons for moving from traditional microwave to millimeter wave technology.The prime reason among them is congestion of microwave frequency spectrum, 3 kHz to 6 GHz.There are a wide variety of applications that already use wireless communications and almost all of them use microwave frequency spectrum.Microwave frequency ranges are reserved for these specific usages and applications.Some of these applications include aviation, military and government use, television broadcasting, radio broadcasting, global positioning systems (GPS), RADARs, and even microwave ovens [1].The reserved frequencies of some of these applications are listed in Table 1.
Hence, for wireless emergency networks, the obvious choice of frequency range was outside microwave, i.e. millimeter frequency spectrum, 6 GHz to 300 GHz.In the millimeter wave spectrum, frequency range 60 GHz to 70 GHz is reserved for missioncritical services.These services include healthcare, communications for self-driving cars, smart city infrastructures and similar mission-critical services.This allows wireless emergency networks, a mission-critical service to gain easily reserved spectrum in millimeter frequency.In this project, we created a framework which allows a user to identify the best locations for wireless APs in the city limits.The locations are generated given multiple constraints.Generating such optimal AP locations is a multi-constrained optimization problem.Even though network antennae placement problems are studies for many years, the problem of placing WENs is unique due to use of millimeter waves.In this case, there are several additional constraints that we never faced before in microwave antennae placement.A wireless emergency network is created using several access points that use high-speed millimeter waves.So, existing algorithms are insufficient for optimally finding locations for these APs.The constraints posed include physical limitations of antennae such as transmission range, point to point transmission or sector transmission and many soft constraints.We consider following constraints on the overall network for finding optimal locations: • Antennae Type: Millimeter wave antennae are generally point-to-point antennae, but even these antennae can be bidirectional or omnidirectional.In some cases, WENs can use sector antennae, but it reduces the range of transmission.
• Range: Depending on the range of antennae, we need to place the different number of antennae in the same area.
• Line of Sight (LOS): The line of sight constraint is the most important difference between millimeter waves and microwaves.The millimeter waves cannot penetrate most of the solid subjects like terrain, buildings, bridges, etc.Hence, there must be a clear line of sight between transmitting and receiving antennae.
• Minimum Degree: To ensure safety and reliability of communication during times of disaster, every AP must have the ability to communicate with more than one AP.
• Mobile Access Points: In congested areas or in areas without reliable communication, mobile access points can be deployed in form of a cell on wheels (CoW) or small cells on drones.
In this project, we focus on identifying the line of sight (LOS) between two locations in the city.The rest of the report is organized as follows.Section 2 describes related works and their limitations.Section 3 gives an overview of component parts of our proposed framework.Section 4 includes two algorithms for calculating line of sight between two locations within the city.Section 5 and 6 describe the experimental setup and experimental result respectively.Section 7 concludes the report.

Related Work
There is very limited research present to address the placement of millimeter wave antennae.Many academic works are focused on automatic adjustment of receiver antennae in order to capture maximum signal for transmitting antennae.There are also studies on effects of obstructions and increasing or decreasing distance on strength of the signal.There are few studies which work on millimeter waves with similar constraints, but these studies were specifically focused on small range antennae up to . Palizban et al. [5] and Szyszkowicz et al. [6] only considered only the single type of obstacles, buildings.They assumed these buildings as only 2dimensional polygon shape on a flat map at an equal altitude as potential obstacles for signal transmission.In addition, none of these studies addressed additional constraints presented by wireless emergency networks discussed in section 1.
Many manufacturers of millimeter wave antennae have published online tools that help their customers deciding the locations for placing the antennae.These tools address the problem of the line of sight to some extent, but these are very primitive tools and have many limitations.Some of these limitations are man-made objects like buildings.Some tools can only address the line of sight problem for terrain, some actually consider buildings as an obstacle.SCADACore RF -Line of Sight tool [7] and AirLink -Outdoor Wireless Link Calculator [8] only consider natural topologies such as mountains.On the other hand, tools like HeyWhatsThat Path Profiler [9] and Solewise Surface Elevation Tool [10] consider man-made obstacles, but these tools do have limitations considering certain objects such as small objects like traffic signals or organic objects like trees as a potential hindrance.

Proposed Framework
In this section, we discuss the components of our proposed framework.These components include wireless emergency network planning, i.e. what aspects need to be considered to efficiently plan WENs, millimeter wave links and their properties, and how can improve estimates of the height of any objects in the city.

Wireless Emergency Network Planning
In conventional network planning, we are concerned about sectoral coverage of the network.In other words, we are mainly focused on how we can serve more people with a minimum number of antennae.But in the problem like wireless emergency network planning, there is a very limited audience to receive these signals.Hence, rather than focusing on local coverage, we have to focus on high-speed connectivity between pairs of important locations which are at a considerably higher distance.If we have such high-speed connectivity between distant locations, the signal can be distributed using traditional techniques such as Wi-Max or Wi-Fi.To ensure reliable long-distance communication, point-to-point antennae are used.Figure 1 shows multiple point-topoint antennae hosted on a single tower.
There are two components for solving optimal constrained wireless emergency network antennae placement problem in the proposed framework.Component one provides an efficient way of estimating the line of sight between given points in terms of latitude and longitude on a map.This component considers all types of topographies, man-made structures and organic vegetation that can interfere wireless communication.
The related systems discussed in sections 2 use Google Maps or in some cases Google Earth data for estimating heights.These sources as not specifically designed for the task at hand can be highly inaccurate for a long-distance line of sight estimations.Most of such sources assume the earth is flat and approximate the heights of small unimportant buildings.Component two will take a list of pairs of locations which have a clear line of sight and are suitable for emergency situations and produce the list of locations where the antennae must be placed.This component will use min-k degree multi-constrained optimization method to solve the optimization problem.The method ensures reliable k degrees of communication links between all access points.The component two algorithm will be based on recommender system techniques.If we have a list of selected access points which are optimal, next access point can be added to this list based on following the top recommendation.This recommendation will be ranked based on most similar access point.The access points having most reliable locations and highest possible communication speed based on the constraints discussed in section 1 will be ranked higher.The algorithm can be repeated till optimal size of the network is reached.Even though this problem is not convex, some overall good solutions can be obtained if recommender systems based on tensor factorization and matrix, because they may capture inherent structures of this problem.

Antennae Links
The most important components of wireless emergency network planning are antennae links.The millimeter waves antennae emit ultra-high-frequency directed beams of waves to transmit data between two fixed points.Figure 2   Smooth terrains, water reservoirs, or any other reflective surfaces in the transmission path can reflect the signal in the unintended direction.This sometimes causes the signal to never arrive at receiving antenna RX.In some cases, the signal does arrive at RX, but it is out of phase.This causes significant degradation of the signal

Line of sight (LOS)
Fresnel zone TX RX [12].Physicist Augustin-Jean Fresnel invented concentric ellipsoidal volumes in the transmitted radiation pattern that help visualize the reflection potential of the radio waves [13].These ellipsoidal volumes are named as Fresnel zones, demonstrated in Figure 2. If Fresnel zone is obstructed more than 40%, transmission signal gets severely degraded.Hence, while predicting the line of sight, it is also important to consider Fresnel zone in addition to the direct path.

Estimating Line of Sight
Obstructions in communication path degrade the efficiency of the wireless network.
Millimeter waves in wireless communication get reflected off buildings.When we choose a pair of locations for antennae placement, we have to make sure that there is a clear line of sight for optimal utilization of Fresnel zone ensuring maximum bandwidth.
In this section, we discuss our two algorithms that may be used to effectively estimate the line of sight between two arbitrary locations within the city.

(a) Building C is not tall enough to obstruct the LOS between buildings A and B. (b) Building C obstructs the LOS between buildings A and B.
Our simple baseline algorithm of the line of sight estimation queries the GIS database for all buildings C between two selected antennae locations A and B. These buildings are OpenStreetMap polygons intersecting the path between the locations.
Each returned building is then verified.Each building's height must be lower than the intersecting point on the slope from the location A to location B. Figure 3 shows two scenarios for one such building C. In the first case (a), it is clear that building C will not obstruct LOS between buildings A and B. The three buildings start at the same altitude, and buildings A and B are both taller than building C. In the second case (b), it is not as clear, as building B is not as tall as building C. Note, however, that LOS would be clear, for example, if building C and B were the same height.Specifically, the height of building C must be lower than the height of the transmission line (connecting the tops of the A and B buildings) at building C.

Figure 4 Potential Fresnel obstruction
One potential problem with the SLOS method is the case when the transmission line passes narrowly between two tall structures.As noted in Section II-B, the path between the TX and RX antennas should be clear not only on the direct line between them but also within at least 60% of the Fresnel zone of the two antennas.Figure 4 shows one such potential example.To account for this scenario, we execute two additional queries, for parallel lines situated at  m on each side of the transmission line.
SLOS then checks the height of buildings that were not present in the original query result to ensure they will not create obstacles in the transmission path.The SLOS method is formulated in Algorithm 1.

Tiled Line of Sight
In the sequential line of sight, we noticed some discrepancies like uneven building polygon size and missing buildings polygon in OpenStreetMap data.If a certain polygon is of a very big area, having a single height for that entire area reduces the accuracy of the Line of Sight queries.On the contrary, very small polygon structure does not efficiently reduce the time required for the query.To address these issues, we decided to draw our own artificial polygons of even sizes.We call this data structure as Tile.Unlike OSM polygons, each Tile is adjacent to another Tile.So, there is very less room for error due to non-polygon structures like area immediately outside a building.
The Tile data structure also addresses the issue of overlapping polygons.Hence, we will not consider the same area twice in this data structure.These Tiles can then further be aggregated to higher level data structures.Figure 5 shows Tile representation and pooling [14].These tiles are constructed and populated using Algorithm 2. Method for finding a Tile ID in this algorithm is derived from Bresenham's line algorithm for Computer Graphics [15].
In the case of a clear line of sight, the SLOS algorithm has to check each building that intersects the transmission line to ensure they will not obstruct transmission.We propose an efficient data structure that can significantly reduce the number of buildings whose height must be retrieved and considered, based on a hierarchical tiling of the city's surface area.Figure 5 shows an initial decomposition of the surface area into equally sized tiles.Note that the number of tiles may be different between levels.In the example, green tiles have a maximum height lower than the minimum intersecting point of the transmission line slope, while red tiles exceed that height.Given the number of tiles per level, the tiling algorithm is straight-forward.For each tile, we pre-compute and store the highest elevation of any building in the tile.Note that the maximum heights of lower-level tiles can be easily aggregated to find the maximum height of a higher-level tile.The highest elevation in a higher-level tile is the maximum elevation among all child tiles.Table 3. Algorithm Tile Creation When predicting LOS, we first check the maximum height of tiles intersecting the transmission path.If a tile's maximum height does not obstruct the transmission path, we can skip checking all the buildings within that tile.On the other hand, in the case of an obstruction, we can dig into the next lower level, retrieving child tiles that intersect the transmission path.At the lowest (leaf) level, data is stored without tiling and we determine the line of sight using the baseline SLOS approach.This method will significantly reduce the number of calculations required to solve the LOS prediction problem.

Figure 5 Surface area tiling in the TLOS algorithm. Green tiles have a maximum height lower than the minimum intersecting point of the transmission line slope, while red tiles exceed that height
This method is detailed in Algorithm 3.

Experimental Setup
In order to determine the effectiveness of our baseline method, we compared its ability to detect clear or obstructed LOS, which is solved as a classification problem.
We use accuracy to measure the effectiveness of our method, which is defined as the ratio between the number of correctly predicted samples and the total number of samples predicted.Additionally, we report precision, recall, and F1-measure for the prediction.Precision is defined as the fraction of samples that are relevant in the prediction, recall is the fraction of relevant samples that were successfully predicted, and F1-measure is the harmonic mean of the precision and recall scores.Note that precision, recall, and F1-measure depend on which class is considered relevant and produce different results in an imbalanced binary classification scenario.We measure efficiency in the time (wall-clock) taken by each query, in milliseconds.In following subsections, details about execution environment and technologies used are explained.

Execution environment
Both algorithms SLOS and TLOS were executed on a server with dual-socket 12core 2.5 GHz Intel Xeon E52680 v3 (Haswell) processors and 384 GB RAM.We also made sure that there are no other programs running on the server when these experiments were carried out.We used PostgreSQL version 9.6.1 to host our GIS database.As a proof of concept, we used a subset of OSM and LiDAR data covering the city of San José, which takes up 685GB of storage, including all indexes.While our system has 24 cores available, the PostgreSQL engine used only one core for executing each query.In all experiments, we set  = 6.
Algorithm 4 details the PostGIS methods and technique used to populate the heights of the buildings in our database.

qGIS
qGIS [22] is the utility we used to visualize the information in PostgreSQL database.
All the building structures in the city of San José can be seen in Figure 6. Figure 7 shows the height data points superimposed on building geometries.
In some locations where the LiDAR data was not available, we used Google Maps Elevation API to fill the missing information [11].After this process, our database was ready to carry of the experiments.

Test data
For our experiments, we used public FCC tower locations dataset [23] as true positive test data.This data is stored in form of Google fusion table.For the negative sample, we generated random pairs of points in the city of San Jose and tested them using online utilities.The negative results were stored as true negatives.For our efficiency tests described in section 6, we again used randomly generated 20000 data point pairs.But, in this case, the emphasis was on efficiency rather than the accuracy of the results.

Results & Discussion
We executed two experiments in order to test the efficiency and effectiveness of our baseline SLOS method and improved TLOS method.In the first experiment, we compared the ability to detect clear or obstructed LOS on 1506 our generated dataset.
We selected LOS queries by choosing two buildings to form these sets from the city of San José and predicting the link between them.As we can see the results in Table 6, among the 1506 pairs of queries, 47 were misclassified, resulting in the accuracy of 96.87% for the sequential line of sight (SLOS) and only 18 were misclassified, resulting in the accuracy of 98.80% for the tiled line of sight (TLOS).These results were encouraging when compared to our last set of the experiment [14], as this dataset is balanced, and we had very good results in both cases clear and obstructed line of sight queries.As shown in    On average, the distance between selected points in the queries was approximately 17 km.The classification line in Figure 6 shows the distribution of execution time among the 200 queries in our effectiveness experiment.Note that the y-axis is logscaled.Most queries took between 60 and 100 ms to execute for SLOS and between 1 and 2 ms for TLOS.The time improvement was also due to the parallel implementation of the queries in both the methods using multi-threading in Java.We also made sure to keep the threads alive after the execution and they could pick up next task in threadsafe blocking queues.This improved the performance by almost 20% than creating new threads for every query reducing the overhead of creation and destruction of threads.Without the multi-threading architecture, the SLOS method tool 120 ms on average to execute and TLOS method took 17 ms on average.As a way to further investigate the efficiency of our method, we executed a set of 20,000 random LOS queries, recording the execution time for each.The random line in Figure 9 shows the distribution of execution times among these queries, which almost exactly matches that of the classification experiment, with the exception of the extreme start of the distribution.A few points in our classification experiments took 150-217 ms to execute in SLOS.In TLOS, the behavior was somewhat irregular.At the end of experiments, there were some spikes in execution time.Similarly, a few of the points in our random query experiment took 500-2567 ms to execute for SLOS, but in TLOS they took around 100-159 ms.However, we could not see a clear correlation between-point distance and query execution time among these few points in either experiment.We conclude that the high execution times in these few queries are likely due to the GIS system loading certain indexes from disk in case of SLOS.To better understand the relationship between query execution time and distance in our experiments, we

Conclusion
In this paper, we first presented a general framework for solving the optimal constrained wireless network antenna placement problem and then detailed two algorithms for efficiently identifying whether there is a clear line of sight between two locations in the city, which is a critical component in the framework.Our first baseline algorithm, SLOS, uses off-the-shelf GIS aware database systems and open-source data to effectively solve the problem, resulting in 96.87% accuracy in our initial experiments.While its average execution time of 60 ms may impede it from solving the optimization problem in a reasonable amount of time, our Tiling Line of Sight (TLOS) method delivered much higher efficiency by drastically reducing the number of structures whose height must be checked to ensure it does not obstruct LOS.TLOS promises the execution time of 1.7 ms on average at accuracy close to 98.80% which definitely is the best contender to be used in the constrained optimization system.

ACKNOWLEDGMENTI
would like to express my honest gratitude to my project advisor Dr. Melody Moh for her continuous guidance, time, and assistance during this project.I am grateful, and I thank my committee members Dr. David C. Anastasiu and Dr. Teng Moh for their thoughts, time, and suggestions.Special thanks to Dr. David C. Anastasiu and Department of Computer

Figure 1
Figure 1 Tower with multiple point-to-point antennae1 demonstrates an example of such communication.The transmission antenna (TX) modulates the signal to encode transmission data.The antenna then sends the signal in the form of a directed beam in the atmosphere.As the waves propagate through the free space, some signal is deteriorated or gets lost.But on a clear day, such transmission can reach up to 65 km distance.The receiving antenna (RX) collects the signal transmitted by the antenna (TX) and re-translates it into binary form.

Figure 2
Figure 2 Communication between a transmitting antenna (TX) and receiving antenna (RX)Even though millimeter waves can transmit huge amounts of data at almost the

Figure 3
Figure 3 LOS obstruction scenarios.(a) Building C is not tall enough to obstruct the LOS between buildings A and B. (b) Building C obstructs the LOS between buildings A and B.

Figure 6 Figure 7
Figure 6 qGIS visualization of building structures in city of San José

Figure 8 FCC
Figure 8 FCC tower location links data on Google Fusion table

Figure 9
Figure 9 Distribution of execution time in our effectiveness experiment between the two variables.The 200 sample classification experiment results showed a correlation of 0.49, while the 20,000 sample random experiment results showed a correlation of 0.21.

Figure 10
Figure 10 Time-Distance co-relation of line-of-sight queries using SLOS and TLOS algorithmsFigure10shows a scatter plot of the random experiment results, denoting the

Table 1 .
Frequency Applications and Allocations in U.S.

Table 5 .
Algorithm Populating building heights

Table 6 .
Accuracy comparison for SLOS and TLOS