Positional Quality of Service using Dynamic Collection Fingerprinting

(1)

Positional Quality of Service using Dynamic Collection Fingerprinting

Fredrik Grönlund

Civilingenjör, Datateknik 2017

Luleå tekniska universitet Institutionen för system- och rymdteknik

(2)

Abstract

Positioning in environments where GPS is absent is a field currently under intensive research. Systems are currently being researched or designed for indoor use, often relying on ultra-wideband radio, ultrasound, fingerprinting or Wi-Fi. For underground mining, the problem is magnified, as installation of new equipment is expensive. Mobilaris Mining and Civil Engineering AB supplies a service, Mobilaris Min- ing Intelligence, using existing Wi-Fi infrastructure present in many mines for communication, and has developed two Wi-Fi-based positioning methods and one hybrid system, using dead reckoning and gyroscope. The first positioning method, Positioning Method 1, positions resources at the location of the strongest access point. The other positioning method, Positioning Method 4, uses signal strength values to construct an area where the tag is likely to be, similar to a Venn diagram.

This thesis proposes a Quality of Positioning system to dynamically and select the best of all available positioning systems for every object to be positioned. This should be trained automatically by

“light vehicles”, such as service pickup trucks, equipped with the hybrid positioning system acting as reference values. Testing was done at the Kristineberg Mine in Västerbotten, Sweden, using a pickup truck equipped with the hybrid positioning system and Wi-Fi personnel positioning tags. It was found that the difference between the two positioning methods was not statistically significant, and that the hybrid positioning system was insufficiently accurate to act as a reference value.

This thesis further revealed that the architecture of Mobilaris Min- ing Intelligence makes implementing a dynamic system impractical.

Although planned for, the dynamic Quality of Positioning system was not implemented due to being deemed too impractical, complex and time-consuming compared to the benefit it would have provided. A high-level description of such an implementation is however presented, should it be motivated by future studies.

(3)

1 Introduction

1.1 Background

Positioning services are of great interest to many industries, and has been for some time. There exist several methods to achieve positioning, each with their own advantages and disadvantages. Indoor positioning services, where the otherwise ubiquitous GPS system is unavailable, is a particularly popular field of research. Several Wi-Fi-based positioning methods and algorithms are currently available for a variety of applications [1][2][3].

Some industries invest research funds into developing their own positioning systems for safety of personnel [4]. In underground mining, the winding tunnels may cause poor visibility, and large, heavy machinery moves around within. There is also research to develop positioning methods to prevent accidents when moving machinery in an underground mine [5].

In open-pit mines, GPS usually suffices, making solutions easy with com- mercially available units. However, underground mines provide a challenging environment for positioning systems. Some solutions where GPS is unavailable use specialised equipment, such as Ultra-wideband (UWB) radio, which uses pulsed, wide-band radio transmissions to transmit data. This method is standardised by the IEEE , and receivers exist that include positioning [6][7]. Some attempts in using ultrasound as sonar for positioning have also been made [8]. However, underground mines tend to be vast, winding and irregularly shaped. Therefore, methods using dedicated hardware, such as UWB radios or ultrasonic emitters, are impractical to install and, most im- portantly, prohibitively expensive.

The structure of the mine shafts and the general work environment may also complicate certain methods. The walls are of varying rock composition, some of which may be partially magnetic or ferrous, potentially disrupting radio waves. Their surfaces are typically very rough, and may be moist or wet, and the rough structure of the walls may hinder signal propagation.

Tunnels also tend to be winding, reducing Line of Sight (LoS) between radio transmitters. Passing vehicles, especially large and massive ones, may also disrupt or completely block LoS to access points. Line of Sight-based positioning methods are difficult to achieve under these conditions.

Many mines use standard Wi-Fi infrastructure for radio communication and data [9]. Since the infrastructure is already installed, there is interest in using Wi-Fi for positioning. Given appropriate network architecture, a positioning system using Wi-Fi would require only software.

(7)

1.2 Mobilaris Mining Intelligence - MMI

One positioning solution tailored for use in an underground mining environment is Mobilaris Mining Intelligence (abbreviated as MMI), offered by Mo- bilaris Mining and Civil Engineering [11]. This solution is a software package that interfaces with Cisco Wireless LAN Controllers (WLC:s) in mines, and displays vehicles, resources, personnel and other tracked items superimposed on an accurate map of the mine in which it is installed.

Positions in MMI are updated in real time. All devices that are tracked are equipped with a tag, a small Wi-Fi-equipped device that broadcasts a blink signal, which is detected by nearby access points. General-purpose devices such as smartphones or laptops may also be configured to blink should the need arise. The precise frequency of these blinks depend on the type of device. General-purpose tags typically blink every fifteen seconds, while vehicle or personal tags might blink every five seconds or even more often.

These blinks are saved by the WLC in separate tables. After a short amount of time set within MMI (between 5-15 seconds), it reads all current WLC blink tables and calculates the position of every tag. MMI is intended to improve effectiveness and safety of operations by giving coordinators more information on their operations and allowing them to track personnel and resources. Since the interface is browser-based, it can be accessed from any type of device.

Maps in MMI are undirected node graphs constructed from a predefined database. All fixed resources, such as map markers, access point locations, area names, fixed equipment and others, are also stored in the database.

As the program starts, it builds a three-dimensional representation from the database. This is overlayed with a representation of the actual mine. Every mine shaft is thus represented as a series of edges approximating curves and intersections

Currently, MMI supports two different proprietary Wi-Fi-based positioning methods and one dead-reckoning based algorithm.

MMI sets the positioning method to use globally, and the program must be entirely restarted to switch which is used. Attempts at Mobilaris have been made to be able to switch while the system is running, but as of May 3, 2017, it is not possible.

For this thesis, the term positioning algorithm will be defined as the pure algorithm, as described in this Section. The term positioning method will refer to their implementation in MMI. Further, the term accuracy will refer to the normal statistical definition of accuracy, where the quantity measured is the three-dimensional distance between two points (known position of the tag, and the by MMI reported position of the tag, unless otherwise noted.).

(8)

The term precision also refers to its normal statistical definition, regarding the same quantity as with accuracy.

1.2.1 PM1 - Positioning Method 1

Positioning Method 1 (PM1) is a very simple, Wi-Fi-based algorithm. PM1 uses Received signal strength indication, RSSI, to determine distances. It operates as follows:

1. For each tag...

(a) Check which AP has the best RSSI values.

(b) Get coordinates of that AP.

(c) Set tag position to received coordinates

This algorithm is currently implemented in MMI. A representation of PM1 can be seen in Figure 1. Access points are represented as green signal icons, and the tag as a grey rectangle.

This positioning method should have a practical accuracy directly related to the distance between tag and access point, due to path loss increasing with distance. The method should thus be more accurate the closer the device is to the AP in question.

Mobilaris approximate accuracy to ±100 m, due to relative frequency of access points. The current implementation of PM1 is simple, and the method is believed by Mobilaris to be robust in environment with poor Wi- Fi reception.

(9)

(a) RSSI values of visible APs as received by the tag.

(b) Reported position of tag versus actual position

Figure 1: Visualisation of Positioning Method 1.

1.2.2 PM4 - Positioning Method 4

Positioning Method 4 (PM4)¹ is another Wi-Fi RSSI-based positioning algorithm. In short, it constructs several "confidence zones", areas where a tag could be, and overlays them on the node graph. The tag is then placed in the area where the confidence zones intersect.

In more detail, it operates as follows:

1. For each Access Point (AP)...

(a) Acquire a list of all APs within signal range.

(b) Calculate the expected RSSI values to each AP in the list, based the path between this and it.

This calculations assumes RSSI values decrease by a predictable value per metre and per degree of curvature, with all such changes being cumulative.

(c) For each AP in the list, compare the expected RSSI to the received RSSI.

Then, calculate a distance factor by dividing received RSSI with expected RSSI.

1Positioning Methods 2 and 3 were internal developmental versions that never reached deployment.

(10)

(d) For every edge in the graph, calculate the predicted values for the distance.

This is done by applying the distance factor to the expected RSSI and through that also to distance from AP.

2. For each tag...

(a) Aquire RSSI for each visible AP.

(b) For each visible AP, calculate confidence area where the tag may be.

This is the distance from AP calculated by the predicted value.

(c) Calculate an intersection area, where the confidence area of every AP intersect.

(d) If previous tag position is within this area, do not move it. Oth- erwise, move it to the middle of the intersection area.

In implementation, there are some computational considerations such as discarding APs that are too weak or other edge cases. A visualisation of the steps of PM4 is presented in Figure 3. This represents a small section of a mine, with access points represented in green and tag as a grey rectangle.

PM4 is currently implemented in MMI. The accuracy of this method is approximated by its creator at Mobilaris to be ±50 m. This number is not confirmed from other sources.

(11)

(a) Expected RSSI values based on path

to APs. (b) Received RSSI and distance factor

(c) Confidence area for AP #1

(d) All confidence areas, with intersection area highlighted in yellow.

Figure 3: Visualisation of Positioning Method 4.

(12)

1.2.3 Hybrid - Vehicle-based dead-reckoning positioning

In the Kristineberg Mine, Mobilaris is currently testing a newly-developed hybrid positioning system (henceforth referred to as hybrid positioning or hybrid system), combining sensors and Wi-Fi positioning methods. It was developed as a student thesis in cooperation with Umeå University [10]. This system mounts onto a vehicle (a pickup truck, known as a light vehicle) and consists of a small on-board microcomputer. This device carries a gyro, a Wi- Fi module and a module for interfacing with the on-board vehicle diagnostics (using the ODB2 interface). It is powered by direct connection to the car battery. Furthermore, it uses the default positioning method (as set globally by MMI) to determine its position, and using the vehicle speed from the on-board diagnostics, it calculates its position with dead-reckoning. This is done by taking its current position, speed and acceleration into account and calculating the approximate speed travelled. Additionally, whenever a turn is detected by the gyro, the algorithm searches the node graph for curves in a breadth-first manner for a set distance to find a matching turn in the map. It then resets its position and heading to that turn. Should a turn be detected when no intersections are within reasonable range, the algorithm resets to its lost state and tries to locate itself with the regular positioning methods.

The method is currently in development, but preliminary tests by Mobi- laris and the results of the thesis suggest an accuracy approaching ±5 m [10].

This testing version does not create a separate entity but instead bootstraps onto an existing tag MAC address and overrides the normal positioning for that device. This allows it to circumvent the normal limit of only using the globally set positioning method.

(13)

1.2.4 Positioning method accuracy & use

Mobilaris has recently developed the PM4 and Hybrid systems. As such, they have no statistical test data on their actual performance in practice. There are as such no data on the performance of either method in any particular environment, except for limited manual testing performed by the company.

PM1 was recently swapped for PM4 in some mines, and the Hybrid system is only available in one mine as of writing. However, there are some question regarding the accuracy of each positioning system. There may be times where PM1 is more robust or more accurate and would be preferable to the more advanced PM4. Given that the accuracy error of PM1 should be proportional to the proximity of the AP, and that the accuracy error of PM4 should not be, there may be locations where PM1 has better accuracy than PM4. For example, in a section densely populated with AP:s, PM1 might prove more accurate than PM4. A section with poor signal strength could favour PM1 or PM4, depending on AP placements. If MMI could have access to the accuracy of different positioning methods, it could select and use whichever method is preferable in a particular area as a tag moves. This would require an additional system making this selection and some way to collect and store this data.

Making assessments on the relative accuracy of different positioning methods by hand would likely be a difficult task. Mines can often be large and busy, making manual measurements difficult. However, it may be possible to use the hybrid positioning method as a reference value, a position which is assumed accurate which all other positioning methods are compared to.

A system could be build which collects accuracy data, and would preferably operate automatically.

It is possible this system could be expanded in the future to handle other algorithms or entirely separate technologies for positioning.

1.3 Purpose of the project

Development of accurate and precise positioning algorithms are of great im- portance both from a commercial and academic point of view. Accuracy and precision of these vary depending on method and technology. However, verification of accuracy within such an algorithm can be difficult, especially in a very widespread and demanding application such as a mine. To manually check and verify the performance of a certain method is unfeasible for a system as large and varied as a mine. Therefore, this thesis aims to construct a Quality of Position (QoP) system for use in conjunction with MMI to dynamically assess the accuracy of available positioning methods.

(14)

This QoP system would not only be of use when developing future algorithms, but might also inform the mine operations coordinator how trustwor- thy positioning data is. This could be used to improve security and productivity. Productivity could be increased by ensuring that traffic is moderated and controlled, tools and personnel are at required positions at the correct times and that erroneous positions can be predicted or handled.

As an addition to the positioning system, a dynamic switching system is to be constructed. The positioning system would retain the evaluated accuracy of each system and use whatever algorithm is most appropriate.

This should allow for the best possible position for all objects, given that accuracy data for their particular area exists.

Future developments in Wi-Fi algorithms or completely new positioning methods could also be evaluated automatically by equipping them to a dead- reckoning-equipped vehicle, giving clear information on its usefulness.

The QoP system could also be adapted for other positioning systems, given they have a method that is accurate enough to act as a “control” but too expensive or complicated to see widespread use.

1.4 Overview

This paper is structured as follows:

Section 1 presents an introduction, background and purpose of the subject. Section 2 addresses the problem and provides delimitations and goals.

Previous work and research on the subject is presented in Section 3. The methodology of this paper is discussed in Section 4. Evaluation of algorithm data and method is done in Section 5. Section 6 presents a discussion on experimental data collected. A description of the dynamic collection and selection system is presented in Section 7. Concluding thoughts and future work is presented in Section 8.

Finally, Appendix A contains descriptions of measurement points.

(15)

2 Problem Description & Delimitations

Thesis statement

This experimental thesis will investigate the feasibility of using vehicle-based dead reckoning positioning to dynamically evaluate the accuracy of two Wi- Fi-based positioning methods in a mining environment.

2.1 Approaches & goals

This section will describe the approach taken to fulfil the thesis statement.

The approach is divided into a number of separate Goals, each pertaining to a separate requirement of the thesis. Each Goal may be divided into Subgoals, which can be fulfilled independently. These Goals will be described below.

In this section, the term QoP system will be used to refer to the entirety of the system in this thesis, including collection, MMI integration, visualisation, switching system and extendability.

2.1.1 Positioning method accuracy

To investigate if a QoP system, as mentioned in Section 1.2.4, is feasible, the accuracy and precision of PM1, PM4 and the hybrid positioning must be determined. If the difference in accuracy between PM1 and PM4 is too small, a QoP system would not be able to improve the overall accuracy.

Likewise, if the hybrid system is insufficiently accurate, it cannot be used as a reference position. For statistical correctness, confidence intervals will be used to judge these factors. The confidence level selected is P ≤ 0.05.

These should show notable differences to motivate a dynamic QoP system.

To judge this accuracy, it is suggested a manual measuring method be used, using measuring points with known positions compared to positions reported by MMI. More on that in Section 4.

Hybrid positioning should be significantly more accurate than PM1 and PM4. According to predictions by Mobilaris, hybrid positioning should be around ten times more accurate than either of PM1 and PM4. To accurately judge either of the other two, it is suggested that hybrid positioning should maintain a median accuracy equal to or less than one fifth of the lower median accuracy of either PM1 or PM4.

The goals therefore become:

(16)

Goals:

1. The difference between PM1 and PM4 should be statistically significant using confidence intervals.

2. Hybrid positioning should maintain a median accuracy equal to or lower than one fifth the median accuracy of either PM1 or PM4.

2.1.2 QoP-MMI Integration

For this section, the term QoP data will refer to accuracy measurements, containing the accuracy and precision of PM1 and PM4 each.

As the QoP system would be handling positioning methods, which are a key component of MMI, it would need to be integrated with MMI. The map screen is always built when booting the program and is kept in memory afterwards. As it is possible that map data changes while MMI is online, all changes are also be saved into a database as they are performed or on shutdown. This collection of QoP data is henceforth referred to as the QoP database. Something similar must be accomplished for QoP data points gathered, and since the QoP data is local to some space in the mine, it would seem logical to store it in the same database as the map itself. Therefore, it is suggested the original database is appended with QoP data. Since the graph consists of edges and nodes, the logical data structure to attach QoP data to the edges of the map, as nodes have no size or length.

Therefore, the goals for this section are:

Goals:

3. QoP data should be integrated into the internal structures of MMI.

(a) This QoP data should be attached to the relevant edges of the graph, and contain the accuracy and precision of available positioning methods, i.e. PM1 and PM4, in metres.

2.1.3 QoP Data Collection

In order for the QoP data to cover any significant area, it must be automatically collected. Manually collecting data for every area of every mine is unfeasible. Since the hybrid method is proposed as a reference value, only vehicles equipped with it can update QoP data. The updating process should require as little input from the vehicle driver as possible, as they are likely busy with driving. Collection of QoP data should not interfere with regular operations either, so the position of the vehicle must not be disrupted by the

(17)

collection and the integration of data into MMI should be performed by some central server. To cover more area in a shorter time, the system should be able to handle several hybrid-equipped vehicles collecting QoP data simultaneously, and each vehicle should be able to evaluate all available positioning methods at once.

As the environment may change, integration into MMI should take past measurements into account, to dampen the effect of random errors.

The goals for this section are:

Goals:

4. QoP data integration into MMI should be performed in an automatic fashion.

(a) A hybrid positioning-equipped vehicle should be able to collect QoP data as it travels without manual input.

(b) The collection should not interfere with the positioning of the vehicle within MMI.

(c) The collection should collect data from both PM1 and PM4.

(d) Integrating QoP data into MMI should be done automatically, either continuously or when the vehicle is stationary with sufficient Wi-Fi reception.

(e) The integration of QoP data into MMI should take earlier measurements into account to better handle random errors.

(f) The collection algorithm should be able to handle several hybrid- equipped vehicles updating the same database.

(18)

2.1.4 Dynamic Switching System

To take advantage of the the QoP data, a system for this selection should be constructed (henceforth referred to as the dynamic switching system). It should select the positioning method most suitable for every tag for every update.

Goals:

5. A dynamic switching system shall be developed to, for each tag, dynamically select the currently best position algorithm for that particular tag, position and time.

(a) The dynamic switching system should select whichever positioning method provides the highest accuracy for each tag, given the existence of QoP data.

2.1.5 QoP Data Visualisation

It should be clear to the operator what QoP data exists and what positioning method any tag uses at any particular time, as well as the estimated accuracy of that positioning method.

Some general overview should be possible, displaying an overlay on the map clearly highlighting the accuracy of available positioning methods. A

“heat map”-like solution is suggested.

Goals:

6. Which positioning method is currently used and QoP data, if any, should be presented in a clear way to an operator of the system.

(a) Whenever a tag is selected in MMI, its currently used positioning method should be clearly visible to the operator.

(b) There should be an option overlay all QoP data for a particular positioning method over the map screen.

i. These should be separated for ease of reading

ii. Areas where the positioning method accuracy is high should be displayed as red, areas where the accuracy is low should be displayed as blue.

iii. Values between these two extremes should be illustrated as a coloured gradient between the two extreme colours.

iv. Areas with no or invalid QoP data should not be coloured.

(19)

2.1.6 QoP System extension

Although only two general-purpose positioning methods are in use today, there may be more in the future. The QoP system could be used to evaluate new methods, whether Wi-Fi based or something completely different.

Goals:

7. The QoP system should be extendable, such that future positioning systems can be added easily.

2.2 Delimitations

This thesis will not take sensor fusion into account. The Hybrid positioning system discussed in Section 1.2.3 can be seen as a sensor fusion system, but analysing its function is not within the scope of this thesis. Possibility of using other fusion techniques in concert with the available Wi-Fi-based systems is disregarded due to time constraints.

This thesis will not build a positioning method from QoP data gathered.

Such an algorithm is acknowledged to be possible, but its complexity leads it to be better represented in a separate study.

The implementation of the actual positional algorithms will not be improved or altered for this thesis. They will be kept in their current, working implementation (as of May 3, 2017).

Interference of vehicles other than the one used for testing operating in the testing area will not be explored and will be avoided as far as possible.

It is acknowledged that particularly heavy vehicles may interfere with Wi- Fi signals. However, such vehicles are usually not available for testing by external parties and may require special training or licenses to operate, and are such left out for practical purposes.

(20)

3 Previous Work

3.1 Positioning & Wi-Fi

Positioning algorithms for indoor purposes have been the subject of many studies. Comparative studies on the subject have been performed, evaluating both different mediums and algorithms, including infrared, ultrasonic, and WLAN (Wi-Fi). Wi-Fi methods are split into types of algorithms, including but not limited to Time of Arrival, calculating distance from signal travel time, Time Difference of Arrival, measuring differences between arrival of synchronised transmitters, and Received Signal Strength Indication (RSSI), an index often integrated into Wi-Fi signals [12][13]. The paper by M. Yassin and E. Rachid surveys existing positioning techniques and technologies [13].

Various Wi-Fi algorithms are discussed, and they propose that RSSI-based solutions should improve with an increased amount of access points within range. The literature review and survey by K. Al Nuaimi and H. Kamel provide further surveys of positioning algorithms, noting an accuracy of 2 m for WLAN-based algorithms in a one-story building [21].

As noted in Section 1.1, the mine environment introduces many error sources and possible sources of noise. A paper by N. Hakem, G. Delisle and Y. Coulibaly suggest that 2.4 GHz signals have similar path loss in tunnels to open air, but suffer from variances in signal strength [14].

Attempts have been made to accurately simulate and model path loss of radio signals in a mining environment. A paper by M. Boutin, A. Benzakour, C. Despins and S. Affles explores propagation of both 2.4 and 5.8 GHz using a statistical model and find that [15]

Radio propagation characteristics obtained in this confined underground environment have proved to be fairly different from what has been found in conventional indoor environments.

It concludes with a statement that

It is clear that wireless propagation in underground mine tunnels can be a challenge to model accurately in view of the hostile nature of the environment. (...)

This gives insight into the behaviour of wireless signals in mines, general accuracy of Wi-Fi-based positioning algorithms, and the difficulty of mod- elling path loss in mines.

(21)

3.2 Fingerprinting & Dynamic Collection

A promising method of positioning is fingerprinting. Fingerprinting is a positioning technique where the signal levels in a particular spot, caused by the unique relations between sources and static interference (i.e. walls and furniture), is used to position devices by matching the signals detected with known signal levels of a certain location.

The first step a of fingerprint positioning is the training phase. In this phase, a number of known positions are determined, typically a grid in a office building or similar. A Wi-Fi receiver is placed in one of these position and records received RSSI values from all visible access points, either once or over a period of time. This data is recorded in a database and is considered the “fingerprint” of that location, defining this database as the fingerprint database. The procedure is then repeated for every single point of the grid.

After the training phase is completed, the online phase begins, where the system can be used for positioning. A device to be positioned records RSSI values of all visible access points and sends them to a central server handling the fingerprint database. The fingerprint database takes the received values and searches for a close match. If the match is close enough, the device should be located at that reference point, and the position has been determined.

Although fingerprinting suffer from the multipath issues and random disruptions of Wi-Fi RSSI, methods to make it more robust solutions have been explored [16][17].

In a mining environment, the training phase of a fingerprint system is normally unfeasible. The large distances and time to set up and measure makes it immensely impractical. Additionally, mines are constantly changing and expanding as galleries are created, exploited and closed, and the measuring would likely interfere with regular mining. To manually train a fingerprint database for an entire mine is hence impossible in practice. However, J.

Kim, M. Ji, Y. Cho, Y. Lee and S. Park developed and evaluated a system for dynamic fingerprint training [18]. In their paper they describe a method in which a device is brought along a predetermined path at a constant speed, continuously gathering RSSI values. Since speed and path is known, values of RSSI for positions could be correlated over time.

This dynamic collection serves as the inspiration for the QoP collection system mentioned in Section 2.1.3, specifically the suggestion of using a reference position to judge all other positioning methods by. However, the path need not be decided beforehand and no extrapolation is needed as the known locations and times are already known.

It may be possible to build an entirely new positioning method using fingerprinting methods as trained by the hybrid positioning system. However,

(22)

as noted in Section 2.2, this is an exercise better left to a separate thesis.

3.3 Fusion methods

All sensor systems have their own inherent strengths and weaknesses, which vary greatly depending on the algorithms and technologies involved Combin- ing data from various sources and sensors is typically referred to as sensor fusion. When using methods involving sensor fusion, different input data can be combined in such a way that their combined accuracy and precision is greater than the mere sum of their data.

An example of a kind of sensor fusion system can be found in the Hy- brid positioning presented in Section 1.2.3, as it uses Wi-Fi, gyro and speed sensors (type varying between vehicles) to determine position.

Sensor fusion and related research is currently very popular. In most mines serviced by Mobilaris, only Wi-Fi is available for the vast majority of devices. With no other sensor data to use, sensor fusion becomes irrelevant for this thesis other than as an aspect of hybrid positioning.

(23)

4 Methodology

4.1 Introduction

In this section, available methods of analysis will be presented for this thesis.

After available methods are discussed, a summary is presented in which the chosen method is described.

4.2 Available methods

4.2.1 Mathematical derivation

A possible approach to evaluate the accuracy, precision and availability is to construct a mathematical model of the mine gallery, the positions of the access points and all other available factors.

This presents a series of clear advantages. Since factors would be known and controllable, they could be changed at any time for new test cases. The solution would also be generic for any and all mine tunnels, given enough knowledge of relevant factors. This extends to implementation as well, as a mathematical model would be agnostic to the platform or method in which it is calculated. Finally, it is possible to mathematically prove the validity of such a system.

However, such a model would be incredibly complicated to construct. It would need to take into account every physical factor capable of affecting Wi- Fi signals in any way. Many of these are difficult to accurately measure and model, and human interaction and movement might be nearly impossible.

Path loss, when used in calculations, is typically simplified as a path loss exponent, usually ranging from 2 to 4. This can be predicted by mathematical models or approximated using statistical methods.

Mathematical models are beyond the scope of this thesis. The author is unfamiliar with the complexity of math required, and, as seen in the articles by M. Boutin, A. Benzakour, C. Despins and S. Affle, and N. Hakem, G.

Delisle and Y. Coulibaly presented in Section 3.1, “(...) wireless propagation in underground mine tunnels can be a challenge to model accurately in view of the hostile nature of the environment”.

4.2.2 Simulation

An approach of evaluating algorithms is to construct a digital simulation model. This could be achieved using a mathematics or physics-focused software such as MATLAB to build the tunnel in a digital environment, and

(24)

include all relevant factors. A simulation is in many ways similar to a mathematical model discussed in Section 4.2.1, although each separate element could be modelled separately and allow the software to model their interac- tions. Factors may also be simplified or omitted entirely for ease of computation.

Simulations would allow for parameters and settings to be dynamically changed. It could be run independently faster than reality, simulating quicker than reality if built sufficiently well. Various setups could also be automated to run after each other, to test which is closest.

However, building the simulation would be time-consuming and complicated. Much like in the mathematical model, many factors are difficult to determine. It would also be limited to the area which it models, as it cannot indicate anything beyond that. Depending on the method used, simulations may require significant hardware to run at acceptable speeds. Ray tracing, for example, is used by some companies for urban environments and is known for a very computation complexity [19][20].

Constructing such a model is best left to a future study.

(25)

4.2.3 Empirical measurements

A method of finding the accuracy of a system is to measure it in practice. This can be done by setting up checkpoints with a known position and measuring what each algorithm reports at that time. These tests must be repeatable, to gather a large enough data set to draw conclusion from. Ideally, the location should be representative of the problem in general, and secluded enough to prevent unplanned outside interference.

Data gathering via measurement is relatively easy. There is no need to calculate path loss, multipathing effects, refraction or other factors as data is recorded instead. Measured data is also inherently representative of the location it was taken, given proper experimental setup.

However, empirical data can only be guaranteed to represent the particular situation and time in which the data was collected. Although it might provide a “rule of thumb” for behaviours in general, it cannot be proven accurate for any other place or time outside the specific environment, as noted by M. Boutin, A. Benzakour, C. Despins and S. Affes [15]:

Results thus show that underground multipath characteristics are quite specific and vary considerably depending upon the gallery dimensions and the transmitter-receiver distance.

4.2.4 Statistical analysis

Statistical analysis provide a useful tool for data management. Data gathered from other methods may be useless without some way of analysing its content.

Statistical analysis therefore is more a complement to other data gathering methods.

4.3 Chosen Test Method & General Information

Given the difficulty of constructing theoretical models of a mine shaft, either in pure mathematical form or in software, an empirical-based version was chosen for this thesis.

The term measurement point, abbreviated MP, denotes a real-world position whose position is easy to find and measure at, and whose coordinates within MMI is well known. This is used for determining accuracy and precision of positioning methods, as noted below.

(26)

4.4 Test setup

4.4.1 Location

All tests are performed in the Kristineberg Mine, Västerbotten, Sweden.

The area used is a unused mine shaft in the mine known as A4. This is due to A4 being a reasonable representation of the mine in general, and that it is currently unused but fully equipped with Wi-Fi access points. A4 is approximately 1680 m long. No other traffic uses the tunnel and it is otherwise free from obstruction.

Fourteen measurement points (denoted as MP 0 to 13) are set up along this stretch. These are set up at different locations with different distances to APs. This is to allow investigation on the effect of proximity to APs. Every MP is also marked clearly on the mine shaft wall so that it is clearly visible to occupants of the vehicle while driving.

A4 is a dead end. Moving from the entrance toward the end is referred to as going “down” A4. The opposite, moving from the end toward the entrance is referred to as going “up” A4.

Figure 4: Image of the A4 mine shaft, taken from MMI. Access points are highlighted as bright green icons, MP:s are faint blue orbs.

4.4.2 Equipment Configuration

Tests are performed in a vehicle equipped with four ekahau tags, a common type of positioning tag in Kristineberg as seen in Figure 5a, and one comm phone, worn on the vest of the driver. The comm phone is an Ascom i62 handset, commonly used for comm radio over Wi-Fi (hence the nickname), and can be seen in Figure 5b. Ekahau tags are configured to blink every fifteen seconds, and comm phones update every second, as is standard.

(27)

(a) An ekahau tag. (b) An Ascom comm phone.

Figure 5: Mounting positions of measurement on test car.

These tags were affixed to the front bumper, rear bumper, roof and centre console of the vehicle. The positions were chosen to determine if the mass of the car would cause any significant disturbances due to blocking direct signals or absorb reflections off of the floor. All tags have the same orientation to eliminate directional properties of the antennas.

Most Wi-Fi access points are placed in the gallery ceiling. Images of these can be found in Figure 6 (Tag position inside car not pictured).

The vehicle carries a driver and a tester, with the tester sitting in the passenger’s seat.

A software is created for recording measurement data. It loads a series of MAC addresses from a list and can request the position of these devices from MMI. The response is comprised of a UNIX time stamp and the X, Y and Z coordinates of every tag in the MAC address list. This is recorded in a log file, which is later analysed.

(28)

(a) An ekahau tag mounted on front bumper.

(b) An ekahau tag mounted on rear bumper.

(c) An ekahau tag mounted on car roof.

(d) Onboard computer used for in Hy- brid positioning.

Figure 6: Mounting positions of measurement on test car.

4.4.3 Procedure

In order to fulfil Goal 1 and 2, analysis of available positioning methods must be performed. Since the positions of MPs are known, measurements can be taken when the vehicle is at the location of an MP, querying MMI of the position of the experimental tags. MMI will return measurements, including the position (in the form of three-dimensional coordinates) of every tag used in the experiment (including the hybrid system) as well as the current time as a Unix timestamp. The position retrieved from MMI this way will be referred to as the reported position, abbreviated RP.

The Euclidean distance between two positions will be denoted as the deviation. This is typically between the MP (the “real” position) and the RP (the “perceived” position), but can also be used to refer to the distance between two other positions such as between an MP and AP. Deviation is a single measurement, and is associated with a single place, time and positioning method (if a tag is relevant). The unit will be meters. It will be used to analyse the general accuracy and precision of the positioning method in question.

(29)

Procedure for testing is as follows:

• The driver drives the vehicle until the MP mark is directly beside them and stops.

• The tester requests a measurement from MMI after 15 seconds have passed

This pause was chosen so that every tag with certainty has blinked and that MMI has performed its update cycle.

• The driver then proceeds to the next checkpoint, repeating the cycle.

The tests can be performed both moving up and down A4. A test going up A4 only needs to inverse its MP order.

4.5 Summary & Method of Analysis

Statistical analysis will be used for data analysis, as mentioned in Sec- tion 2.1.1.

Values to examine are:

• Deviance of RP and MP.

• Deviance of RP and the physically closest AP.

• Deviance of MP and the physically closest AP.

Metrics to be used for data analysis are:

• Mean average.

• Median.

• Maximum & minimum values.

• Upper & lower quartiles.

• Standard deviation.

Correlations to positioning method accuracy, given as distance between RP and MP, to be examined at p = 0.05 are:

• Deviation from MP to nearest AP.

• Deviation from RP to nearest AP.

• Amount of APs within 400m.

(30)

The distance 400m was chosen as a theoretical maximum range, although improbable. This correlation is to determine if PM4 is improved by many access points within (theoretical) range. Correlation P -value was chosen to P ≤ 0.05 as a standard value, as suggested in Section 2.1.2.

5 Algorithm Evaluation

Data for these tests were gathered on two occasions. On the 3rd of March 2017, only PM1 was tested. On the 22nd of March 2017, all three methods were tested.

In this and following section, a distant data point is a data point 1.5 times the interquartile distance under or over the first and third quartile, respectively. In box plots, these are represented as red plus symbols. These are defined to help illustrate any occurrence of data points of particularly large distance from the other data.

5.1 Tag Placement Comparison

5.1.1 Data

The effects of the tag placement on the vehicle should be determined. Col- lected deviation data, divided on a tag-by-tag basis can be seen in Figure 7.

5.1.2 Analysis

Before any analysis on the effectiveness of positioning methods is done, the effect of the vehicle body must be determined. The box diagrams for both PM1 and PM4 show no visible patterns in accuracy depending on the position of the tag. In general, median, upper and lower quartile values shift very little between tag positions, and the box whiskers are of similar length or coincide with distant data points of similar value.

Since the car alternates direction of travel, it alternates which tag points up or down the gallery. Proximity and LoS to APs should therefore be evenly distributed over several measurements.

The tag mounted on the vehicle roof and the one mounted in the cabin showed no discernible differences in accuracy. Both show similar box plots in Figure 7, with similar median, upper and lower quartiles and whiskers. The same seem to hold for both PM1 and PM4. Whatever influence the cabin has on accuracy can be considered negligible.

(31)

(a) Box plot showing deviation of PM1, by tag.

(b) Box plot showing deviation of PM4, by tag.

(c) Box plot showing deviation of PM1, by tag, cropped to 400m for ease of comparison.

(d) Box plot showing deviation of PM4, by tag, cropped to 400m for ease of comparison.

Figure 7: Graphs showing deviation of PM1 and PM4, divided by tag. Dis- tant data points represented as red plus-signs.

Since positioning of the tags do not influence the positioning methods in any noticeable way, any influences of the vehicle mass can be neglected.

The exception to this is the comm phone for PM1. It shows much lower median values and all values belonging to it are much smaller. The exact cause of this is uncertain, as only two factors differentiate it from the ekahau tags; its update frequency and its antenna. Update frequency should have no effect on PM1, since the measurement time takes the slower update frequency into account. However, the better antenna could be an explanation, as it could receive signals more clearly. It still has many distant data points around and above ≈ 150 m, which could point to some kind of noise. The phone will however be included in the general average for PM1.

(32)

As a potential issue, it should be mentioned that these tags are placed some meters apart on the car. This, however, was quickly discovered to be a non-issue, as the average positioning accuracy was at least a magnitude greater than the distance between the tags.

In conclusion, the positioning of the tags in this setup have little impact on accuracy, and data from any tag can be treated as equivalent to data from any other tag.

Therefore, as described in Section 2, the extra tags on the vehicle can assist in determining precision, as they can be used as simultaneous testing runs. This allows for collection of a large data selection on PM1 and PM4.

Hybrid positioning cannot be “repeated”, since only one was available for testing, meaning every round of testing yields a much smaller data set.

5.2 Positioning Method 1

For Positioning Method 1, a total of six (6) series of measurements were performed for a total of 84 data points. Each of these data points contain four (4) positions, one for each tag placed on the vehicle.

The results of these measurements can be seen in Table 3. Measurements tabulated are deviation between MP and RP, in metres. These are used to calculate correlation and statistical data in Table 4.

Visualisations to clarify correlations can be seen in Figures 8, 9 and 10.

These illustrate correlation of factors described in Section 4.5. Plotting scat- ter plots with two factor suspected to be correlated as X and Y-axes can visually show correlation if one exists. The data points would group in a shape approximating the function of correlation. The trend lines shown in Figures 8, 9 and 10 are linear regression lines calculated by the least-squared method. No clear correlation pattern is seen in these Figures.

Correlation coefficients, tabulated, can also be found in Table 1.

Table 1: Correlation coefficients of PM1

Correlation to Distance between RP and MP Correlation coefficient Deviation from RP to nearest AP 0.4174186124

Deviation from MP to nearest AP 0.7990944971 Number of APs within 400m of MP 0.2070652778

(33)

Figure 8: Deviation RP to MP plotted against deviation RP to nearest AP.

Figure 9: Deviation RP to MP plotted against deviation MP to nearest AP.

(34)

Figure 10: Deviation RP to MP plotted against amount of APs within 400m.

(35)

5.3 Positioning Method 4

For Positioning Method 4, a total of seven (7) series of measurements were performed for a total of 96 data points. Two data points gathered were discarded tue to testing equipment malfunctions. 55 of these data points contain five (5) positions, one for each tag placed on the vehicle and one for the comm phone carried by the vehicle driver. The remaining 41 did not have access to the position of the comm phone, as it was used by the hybrid positioning as described in Section 1.2.3.

The results of these measurements can be seen in Table 3. Measurements tabulated are the Euclidean distance between the by MMI reported position to the actual position of the tags (i.e. the position of the current MP), in metres.

Visualisations to clarify correlations can be seen in Figures 11, 12 and 13.

These illustrate correlation of factors described in Section 4.5. Plotting scat- ter plots with two factor suspected to be correlated as X and Y-axes can visually show correlation if one exists. The data points would group in a shape approximating the function of correlation. The trend lines shown in Fig- ures 11, 12 and 13 are linear regression lines calculated by the least-squared method. No correlation pattern is clear in these Figures.

Correlation coefficients, tabulated, can also be found in Table 2.

(36)

Figure 11: Deviation RP to MP plotted against deviation RP to nearest AP.

Figure 12: Deviation RP to MP plotted against deviation MP to nearest AP.

(37)

Figure 13: Deviation RP to MP plotted against amount of APs within 400m.

Table 2: Correlation coefficients of PM4

Correlation to Distance between RP and MP Correlation coefficient Deviation from RP to nearest AP 0.0674694848

Deviation from MP to nearest AP 0.2183978284 Number of APs within 400m of MP 0.083693808

(38)

Table 3: Comparison table of statistics for all positioning methods, in metres.

Hybrid,

(m) PM1 PM4 Hybrid trimmed

Average deviation 92.2 106.4 73.0 51.6 Median deviation 64.3 72.9 33.1 27.8 Maximum deviation 360.0 1348.6 478.6 478.6 Minimum deviation 2.5 0.6 1.1 1.1 Standard deviation 96.0 137.3 98.0 81.3 Lower quartile 4.22 30.5 16.7 13.0 Upper quartile 124.9 134.4 90.1 56.7

5.4 Hybrid positioning

For Hybrid positioning, a total of three (3) series of measurements were performed for a total of 41 data points. One data point gathered was discarded due to testing equipment malfunctions. These 41 data points contained one (1) position each.

For testing purposes, the hybrid system was initialised manually. This sets both starting position and heading, skipping the initial locating step using regular Wi-Fi positioning. This was done to save time and to isolate the hybrid positioning from other systems.

The results of these measurements can be seen in Table 3. Measurements tabulated are the Euclidean distance between the by MMI reported position to the actual position of the tags (i.e. the position of the current MP), in metres.

An unidentified malfunction on the last measurement series caused significant drift when testing. This manifested as an apparently incorrect value for vehicle speed, as it would move visibly slower in the map screen. Both the unchanged data set and one where these data points have been removed (for a total of 36 data points in that set) are included in Table 3.

5.5 Deviation Comparison

Statistical data on all positioning methods are presented in Table 3. This data shows that PM1 has higher average and median deviation, lower maximum and smaller standard deviation.

(39)

Table 4: Deviation of positioning methods, by MP.

MP PM1

Avg. Deviation

PM4

Avg. Deviation

Hybrid

Avg. Deviation

Distance,

MP to closest AP

0 15.6 56.5 66.8 10.7

1 64.3 72.9 52.3 104.7

2 41.2 29.8 38.3 77.2

3 2.93 32.7 35.9 7.3

4 51.7 51.9 26.0 47.3

5 64.3 47.7 3.9 64.6

6 70.2 53.7 2.0 147.2

7 4.2 72.3 10.7 10.2

8 18.1 145.7 20.4 5.6

9 38.6 25.7 5.6 9.7

10 172.8 133.2 68.0 10.1

11 173.8 141.1 14.6 83.0

12 332.6 268.4 78.7 256.1

13 139.5 48.6 12.3 6.4

(a) Box plot of deviation data, by positioning method.

(b) Box plot of deviation data, by positioning method, cropped to 400m.

Figure 14: Box plots of deviation data. Hybrid positioning data is of unal- tered version.

Statistical data is also presented as box plots shown in Figure 14. This illustrates that the many distant data points could cause the high averages.

A breakdown of deviation data for each positioning method can be found in Table 4. An illustration of the three positioning methods and their confidence intervals can be seen in Figure 15, with significance level of P ≤ 0.05

(40)

(as defined in Section 4.5).

Figure 15: Confidence intervals of PM1, PM4 and hybrid positioning, displaying their median values with P ≤ 0.05.

5.6 Analysis

As noted in the paragraphs above, the collected data indicate PM1 is the more accurate positioning method, contrary to expectations laid out in Sec- tion 1.2. The difference in median deviation, 8.6 m, is not very large in a mining environment. In Figure 15, we also see some overlap between PM1 and PM4. This means we cannot definitely say that the difference is statistically significant. Hybrid positioning, although accurate after turns, has a median deviation of 27.8 m, even when malfunctions are removed.

(41)

6 Evaluation & Discussion

6.1 Algorithm Evaluation

6.1.1 PM1

Data gathered is presented in Section 5. Table 3 shows very interesting results. In this, PM1 seem to function better when the MP is in closer proximity to the AP. This is partly supported by the correlation coefficient for MP proximity to AP presented in Table 1. A correlation coefficient of only

≈ 0.8 is poor, however, although the value is high enough to warrant some weight. Accuracy is close to predictions by Mobilaris, which also matches up with expectation given the distances between MP and AP. Precision is lower than expected, however. The accuracy of the phone, as displayed in At times, the reported position would not remain at a different AP even at times where the MP is in very close proximity to the AP. This is likely due to the inherent instability of Wi-Fi signals seen in research, even in ideal conditions, not helped by the unfavourable conditions of a mine [15][22]. Despite this, the simplicity of PM1 causes it to behave somewhat predictably, as it has few distant data points in Figure 14.

6.1.2 PM4

Deviation data for PM4 from Table 3 is interesting as it deviates heavily from predictions laid out in Section 1.2.2. The maximum value is very high, as is the standard deviation. This points to a very “swingy” data set, with many distant data points in the upper limits. The high lower quartile value also point to a value with few very low values and high variance in the upper end.

Observing Figure 14a, the large amount of distant data points support this.

Table 2 show no noticeable correlation between amount of APs in the vicinity to accuracy. According to the specifications of PM4, more APs within range should improve the algorithm result, but no such pattern is apparent. The cause of the very large maximum value is unknown, as it matches no known disruptions when testing and data has been recorded accurately. It is possible this unusually large deviation is an error in implementation of the algorithm, but the error has not yet been reproduced in a controllable manner and seem quite rare.

The general lack of accuracy in PM4 can be explained in the same way as with PM1; random instability. Fluctuations in Wi-Fi signals impact a simple algorithm like PM1 very little. PM4 is considerably more complex, and any disturbances in RSSI values from environmental factors could have a large effect due to its Venn Diagram-like approach. Should a detected

(42)

tag signal be smaller due to unlucky disturbance, the confidence area would be of improper size, either being overly exact or overly generous, with both possibilities degrading the end accuracy. With the system relying on several different signal sources, all errors in each path compound in the end result.

The results might be different in a less challenging environment.

6.1.3 Hybrid

Data from the Hybrid positioning can be seen in Table 3 in comparison to the other positioning methods. From this, we can observe that the Hybrid positioning has better deviation values than both PM1 and PM4, as expected.

However, the need to trim values due to malfunctioning equipment is worry- ing for an autonomous dynamic collection algorithm as suggested. Observing average deviation per PM as outlined in Table 4, we can conclude that MP:s 5, 6 and 9 have remarkably good accuracy. These three MP:s are situated close to corners in the tunnel², which would trigger the gyro to match car position to a matching turn on the map. This is intended behaviour, but there still exists some drift, which could be attributed to poor accuracy of the speedometer in the car. The exact method of calculating speed is unknown and would vary on what the vehicle used. It is possible that whichever method is used is not entirely accurate on purpose. A speed sensor based on wheel axle rotations would report incorrect values if tires are changed (which is an occurring repair/upgrade in the Kristineberg Mine). Additional drift could come from small adjustments in velocity or course that are not accounted for in the one-dimensional representation of mine tunnels in MMI.

Driving styles could also have an effect on the hybrid system. There is no data on if rapid acceleration, aggressive braking or other things which could cause minor wheel spin has any noticeable effect on the accuracy. Other possible error sources could be wheel spin from poor traction, bumps causing the tires to lose contact with the ground, and possibly others.

As a note on hybrid drift, P. Hansson notes in his thesis that while distance along straight roads only drift around 1% over a 1.5 km strech of road, the gyroscope can drift up to 152.67 m in 60 s when used for inertial naviga- tion [10].

Important to note is that direction and position was reset at the start of every measurement run. If this was not handled manually, and the vehicle was simply turned around, it would misinterpret this as a regular turn and set the vehicle position wildly wrong.

2See Appendix A

(43)

6.2 General Evaluation

As noted in the paragraphs above, the collected data indicate PM1 is the more accurate positioning method, contrary to expectations laid out in Sec- tion 1.2. The difference in average deviation, 14.2 m, is not very large in a mining environment. Figure 15 is clear that there is overlap between the confidence intervals of PM1 and PM4. Therefore, no conclusion can be drawn on their statistical significance in relation to each other. Thus, Goal 1 was not met. Hybrid positioning, although accurate after turns, has an average deviation of 51.6 m, even when malfunctions are removed. As defined in Goal 2, hybrid positioning should be at least as accurate as the difference in accuracy between PM1 and PM4. The Goal was hence not met.

Discussion on a program structure is presented in Section 7.

(44)

7 Dynamic Collection & Selection System

An approach for a dynamic collection and selection system, satisfying the conditions laid out Section 2, is described in concept in this section.

7.1 Database structure

As noted in Section 1.2, the map used by MMI is a node graph. To satisfy Goal 3 the database would ideally store the accuracy of each positioning method within the same data structure as the map itself. The intuitive way of storing this would be to associate a spatially dependent measurement (the accuracy of a particular method at a particular place) with a spatial type Since measurements are ideally taken everywhere continuously, the logical data type to store accuracy in would be the map edges, as they are the only element that spans distance.

The system maintains its map internally while running including all nodes and edges, but only nodes are permanently stored in the database. Therefore, any stored data would have to be connected to the nodes rather than the edges.

The suggested approach to this is to divide each connecting edge into two equal parts, and let the accuracy of the node be the average of all values collected on the halves of edges connected to it. This is suboptimal in many ways. Tunnels connected to the same “crossing” in reality, or node in the map, could have wildly different signal levels and thus different levels of accuracy.

It is also more complex in implementation.

The alternative would be to implement a edge type in the database, which is with all likelihood a worse idea. Such an edge entry would not only need to be synchronised with the node entries (who already handle their connections and thus edges), but would also contain very little other useful information.

Maintaining such a database would be difficult, and the alternative to that maintenance is a major restructuring, which is both beyond the scope of this thesis and impractical.

This would satisfy Subgoal 3a, although for nodes instead of edges.

Examples of both methods can be seen in Figure 16.

For debugging and time-dependencies, the accuracy measurements should always be associated with a timestamp (standard UNIX timestamp is suggested).

(45)

(a) Map structure with accuracy associated with edges. Node size is exaggerated (nodes have no size).

(b) Map structure with accuracy associated with edges. Node size is exaggerated (nodes have no size).

Figure 16: Visualisations of accuracy storage methods. Accuracy levels are mock-ups. Colour indicate overall accuracy, with red being most accurate, blue least accurate and all other levels fitting in the spectrum between. No colour means no accuracy data available

(46)

7.2 Data collection

Data collection would be performed by a specially equipped light vehicle.

This would have the hybrid positioning system installed, as well as one device per positioning system to evaluate (i.e. one Wi-Fi tag per positioning method), to satisfy Subgoal 3d. A software module running on a central server would gather the positions of the hybrid and all other tags registered to that vehicle. For accuracy to be measured, the hybrid is used as reference value, and all other methods calculate the Euclidean distance from their respective reported positions to the position reported by the hybrid. The reference position and all calculated accuracy values are then passed back to MMI. Since all handling of the data is managed centrally, Subgoals 3a and 3b should be satisfied by this. Given the correct implementation, multiple light vehicles should not be an issue, fulfilling Subgoal 3f

A module in MMI then parses these and store them in the node data structure. This change is then represented in the external database. This would satisfy Subgoal 3d.

To account for history, each node would store an appropriate number of its previous accuracy measurements. A First-In-First-Out (FIFO) would store a set amount of these, discarding the oldest value when a new is received.

All elements in this queue is evaluated when determining the accuracy at the node. Older entries are given less weight in this calculation. This would let it fulfil Subgoal 3e.

To allow for simultaneous collection of data from both PM1 and PM4, as specified in Subgoal 3c, a new module would have to be created, due to MMI does not natively support multiple positioning methods to be run simultaneously. There also lacks

Given all of these implementations, Goal 4 should be satisfied.

7.3 Selection algorithm

The selection algorithm would be simple. Every time a tag position is to be determined, their current position is cross-referenced with the accuracy database. Whichever method at the tag’s current position (before update) has the highest accuracy is used to determine the next position. Although somewhat inflexible and reliant on previous methods, this should satisfy the basic requirements of Goal 5.

Preferably, this system has a default method set, should no accuracy data for that location exist.

A possible issue could occur when PM1 is selected as the method of choice.

Given a scenario where a tag is very close to an AP, the system selects PM1

(47)

for that tag. Its position is now the same as the AP. The tag then moves away from the AP to a point some distance away where PM4 is the most appropriate positioning method. The system then observes the last position of the tag, determines that for that position, PM1 is the most appropriate and maintains that method, even if it is not the optimal solution. No easy solution to this problem was apparent to the author of this thesis.

7.4 Accuracy Representation

There are several ways to represent data, as outlined in Goal 6. In its most basic form, representing the accuracy could be as simple as a text field in the descriptions of the tag. A mockup of this is presented in Figure 17, and should satisfy Goal 5a.

Figure 17: Mockup displaying positioning method and accuracy in the existing tag data window of MMI.

(48)

Displaying all accuracy data for a particular positioning method, as by Subgoal 5b could be done by modifying the “heatmap”-function already present in MMI. This is normally used to colour paths in accordance to their RSSI values, but it could be modified to take accuracy values instead. A drop-down menu of all available positioning methods would switch which positioning method is currently shown.

7.5 Extendable structure

Maintaining an internal structure modular enough to add further positioning methods, as specified in Goal 7 should be doable.

One approach is to make each individual positioning method a subclass of a posmet-class, which could represent generic positioning method classes, implementing standard functions required for positioning. These could be added to a vector, which is iterated over to evaluate each. With a type defined such, a function could be added which just adds a method to that vector. Thus, to add a new method would require:

• Write the subclass and implement its functions

• Install equipment on the vehicles

• Note MAC addresses of all equipment

• Adding the subclass to the vector.

Positional Quality of Service using Dynamic Collection Fingerprinting