Using Personal Devices to Teach Eco-Driving
Hannes Granström 2013
Bachelor of Arts Systems Science
Luleå University of Technology
Department of Computer Science, Electrical and Space Engineering
Using personal devices to teach eco-driving
Hannes Granstro m
Luleå University of Technology, Department of Computer Science, Electrical and Space Engineering August 2013
I
Preface
This report details my bachelor thesis work in Systems Sciences at Lulea University of Technology (LTU). The work has been carried out at Neava, a mobile telecommunica- tions consultancy company, located in Luleå, Sweden. The idea for the study was pre- sented to me by the company’s founder and CEO Staffan Johansson as an extension of another thesis work previously carried out at the company.
I would like to thank Neava and especially Staffan Johansson for presenting me with the idea of this work and for providing the resources to make it possible. Staffan has also provided much valuable input and ideas throughout the process. Moreover, I would like to thank Dan Harnesk and Harriet Nilsson at LTU for feedback and help with the academic aspects of this work.
Finally, I would like to thank John Viklund at Neava and my class at LTU, who helped me with valuable feedback on the report.
II
Abstract
The aim of this thesis is to evaluate if a mobile device application can be used as an aid in teaching drivers more fuel-efficient driving. This paper details the development and testing of such an application, which uses only data available through the On- Board Diagnostics port available in all modern vehicles in order to assess multiple aspects of user’s driving behaviour and appoint personal development goals. The appli- cation was tested for 10 days on a specified test course for evaluation. The results of the study showed that while the application has a good ability to assess users’ eco- driving performance and that eco-driving performance improved significantly in several areas, there was no evidence that using the application will lead to a change in long- term fuel consumption. This was caused by the fact that while the test subject im- proved in most areas, a decreased performance was seen in parameters that are highly affective on fuel consumption.
It was concluded that the application developed for the study has good potential, even though there is room for minor improvements. Moreover, it was concluded that further studies should be performed in order to more thoroughly evaluate the effects of using the application and that the results of the study could be different with a more rigorous and quantitative testing approach.
III
Table of Contents
1 Introduction ... 1
1.1 Purpose and goals ... 2
1.2 Research questions ... 3
2 Theoretical Background ... 4
2.1 Previous work ... 4
2.2 The On-Board Diagnostics System ... 5
2.3 Recommended eco-driving techniques ... 5
2.4 Characterising driver preferences ... 9
2.5 Using electronic devices as learning aids ... 9
2.6 Summary and Conclusions from Theoretical Research ... 11
3 Method ... 13
3.1 Application goals and requirements ... 13
3.2 Application architecture ... 13
3.3 Application features ... 13
3.4 Internal architecture ... 16
3.5 Recommendations and feedback ... 21
3.6 Visual design ... 22
3.7 Performance ... 22
3.8 Testing ... 23
4 Results... 28
4.1 Assessment ability ... 28
4.2 Eco driving performance vs. number of trips ... 29
4.3 Goal-based learning concept ... 30
5 Discussion ... 32
5.1 Conclusions from results ... 32
5.2 Limitations... 34
5.3 Future work ... 35
5.4 Scientific approach ... 36
5.5 Summary ... 37
IV
6 Bibliography ... 38
Appendix A. Requirements Specification ... 41
Appendix B. Application architecture... 42
Appendix C. Gathered test data ... 43
Appendix D. Correlation analysis results ... 44
Appendix E. Regression analysis results ... 45
1 Introduction
Fuel efficiency and the need for decreased energy use in cars and other vehicles is a very current and often debated matter. More efficient use of fuels is not only beneficial for the environment as it results in less emissions and less use of natural resources, it is also more economic on a personal level for drivers.
An important factor that affects the fuel efficiency in automobiles is driving behaviour where some key parameters such as average velocity, acceleration and braking behav- iour can substantially affect the fuel consumption of the vehicle. Due to this, many driving schools and institutions have begun teaching drivers to adapt a more fuel effi- cient driving behaviour, a technique commonly referred to with the term eco-driving (which is also the term that will be used in this report). Much research has been done relating to which parameters affect the fuel consumption of a vehicle, as well as how the vehicle’s driver can affect those parameters. However, relatively little research have focused on how feedback with concrete data can be used to help drivers evaluate their own driving techniques from a fuel-efficiency perspective. Most studies performed have also required the use of complex ad-hoc devices and modified vehicles in order to gather necessary data. While this approach does allow for analysing driver behaviour in a research setting, it is not practical for use in a “real-world” context. A solution that enables people to easily analyse their own driving would require a way of gathering data with far less requirements on users in terms of equipment and installations.
Most modern vehicles are equipped with an On-board diagnostics (OBD-II) port that can be used to obtain vast amounts of data from in-car sensors. The OBD-II port, also referred to as the service port, was originally developed as a way for technicians to troubleshoot engine problems, but it can also be used to gather data that can be used to analyse driving behaviour. A problem is that most tools used for reading OBD-II- data are not designed for personal use and consequently their use for the average person is very limited. Moreover, these tools are not designed for analysing fuel-efficiency and do not provide a way for users of doing so.
Until just a few years ago, only ad-hoc devices could be used to connect to the OBD- II-port and there was no way of processing the data other than sending it to a computer.
In recent years however, smartphones and other personal devices such as tablets have seen a marked increase in use and availability. According to analytics firm Gartner, 472 million smartphones were sold in 2011, and sales are expected to grow in the coming years. This kind of devices allows users to perform computer-aided tasks and calcula- tions at almost any location, for example in a vehicle. This makes them well suited for
use in the context of gathering data about, analysing and helping to improve driving behaviour. There is also a market of relatively inexpensive adapters that connect to the OBD-port and send gathered data wirelessly via the Bluetooth interface. Since most mobile devices today are equipped with Bluetooth receivers, they are well-prepared for reading the OBD-data. Developing an application that utilises this potential could therefore be a way to help drivers decrease their fuel consumption with relatively little effort and without making any changes to the vehicle.
Purpose
The purpose of this work is to investigate whether mobile devices can be used as aids in teaching drivers more fuel-efficient driving, by assessing their eco-driving perfor- mance with adaptability to external conditions and giving suggestions to improve fuel efficiency.
The major part of this work is related to the development of an Android OS application that, in combination with a third party Bluetooth OBD-II adapter, can assess and help to improve the fuel efficiency of users’ driving. As the work is an extension of a study performed by Karthik Kotari in 2012, the data gathering functionality developed by Kotari is also used in this work. Consequently, the main area of development has been functionality for analysing data gathered through the use of the functionality developed by Kotari, with some extensions.
Adaptability to external conditions, such as area and traffic congestions, is of very high importance to accurately assessing the fuel efficiency of a user’s driving behaviour.
Ability to adapt to these conditions was therefore one of the major requirements for the mobile application used in the study.
It is reasonable to believe that an application that adapts to the user’s (i.e. the driver’s) preferences is less likely to cause frustrations, and thus more helpful to the user. For example, a driver who prefers quick accelerations might not appreciate recommenda- tions to use less throttle even though this would be the best advice from a fuel-efficiency perspective. Another major requirement was therefore adaptability to users’ driving preferences.
The basic requirements that was used when developing the application can be described as:
Ability to gather and utilise OBD-II data
Ability to evaluate users eco-driving performances with adaptability to driving conditions and other factors
Ability to adapt eco-driving recommendations based on user’s driving prefer- ences
Research questions
The research questions based on the purpose of this work was:
I. Is it possible to accurately evaluate a user’s eco-driving performance using only OBD-data?
II. Will using a mobile device application have a positive effect on users’ eco-driving performance, resulting in decreased fuel consumption?
2 Theoretical Background
Previous work
The previous work in the field by Karthik Kotari presented a solution to gather OBD- II data from a Bluetooth adapter onto an Android device. Much of the data transfer functionality developed by Kotari is implemented in this work. However, Kotari’s work did not aim at or provide the means of analysing the gathered data, which is the main focus of this work.
A number of studies have been performed on the use of on-board mobile devices as aids when teaching fuel-efficient driving behaviour. This includes an investigation of the feasibility of such a system (without using OBD data) (Ishikawa et. al, 2010), and evaluation of an actual implemented system, concluding that such a system indeed can have a positive effect on drivers’ fuel efficiency (Beusen et. al, 2009). However, this study did not incorporate mobile devices in the actual learning, as they were used solely for gathering data.
Other studies have examined how OBD-data can be used to estimate the fuel consump- tion of a vehicle. In most cases, these studies do not evaluate driving behaviour in terms of eco-driving. They are however still of value since fuel consumption estimation are important parts of assessing driving behaviour.
A study closely related to this examined the use of an OBD-connected Android device as an eco-driving assistant (Magan a & Organero, 2011). The work by Magan a and Organero has many similarities to this work, as the Android application used contains functionality that makes it adaptable to driving conditions, which is an important part of this work. The study also examined how drivers’ eco-driving performances were af- fected by using the application. However, one vital difference is that this work, as opposed to Magaña and Organera’s study, utilises a learning model that adapts to the user’s driving preferences.
Finally, a model for characterising and assessing driver’s eco-driving performance has been researched and proposed (Andrieu & Pierre, 2012). This study proposes a model for determining the probability that a driver is performing eco-driving on a scale of 0 to 100.
In summary, previous studies cover many areas of this work but no studies have been found that answers the research questions posed in this study.
The On-Board Diagnostics System
On-Board Diagnostics is a self-diagnostic system that gathers data about the vehicle’s state from on-board sensors. The OBD-II-system was originally developed for use by technicians, but in the later years, many interfaces that enable reading of the OBD- data have been developed.
Many of these interfaces use Bluetooth technology to transmit OBD-II data from an adapter connected to OBD-II port of the vehicle to other Bluetooth enabled devices.
Since most smartphones and similar devices today are equipped with a Bluetooth re- ceiver, data can be wirelessly from the adapter to the device.
The available data that can be retrieved are different between car models. However, there is a set of data, defined in the ISO 15031 standard that can be received from any vehicle equipped with an OBD-II port. Some of these are presented in Table 2.1. For this paper, it can be assumed that all parameters discussed are retrievable, directly or indirectly, through the standard OBD-II interface unless otherwise stated. This is im- portant as the application developed in the study is designed to be usable on any vehicle equipped with an OBD-II port.
Table 2.1. Examples of data available through the OBD-II port (ISO 15031).
Parameter Unit Description
Vehicle Speed Kilometres Per Hour The current velocity of the vehicle Engine RPM Revolutions Per Minute The current RPM of the engine
Throttle Percent The position of the accelerator pedal, where 0 is no throttle applied and 100 is full throttle applied
Engine load Percent The workload of the engine in present.
Recommended eco-driving techniques
Three important aspects of eco-driving is driving conditions, accelerations and deceler- ations. The way that these are handled by the driver can greatly affect the fuel con- sumption of the vehicle.
2.3.1 Driving conditions
Driving conditions is known to be an important factor affecting the fuel consumption of a vehicle, where fuel consumption is higher in areas that require a high amount of accelerations and decelerations. Moreover, the influence of driving behaviour on fuel
consumption are also related to driving conditions. For instance, the effect of average velocity is mostly visible in low-speed neighbourhoods and on highways, but has rela- tively little effect on fuel consumption in medium speed city driving where accelerations and decelerations have greater impact on the overall fuel consumption (Berry, 2007).
In a study of how driving conditions affect fuel consumption and emissions, the effects of five different driving conditions - very congested, congested, urban, extra urban and highway - on fuel consumption and emissions were tested (Fotouhi & Montazeri, 2012).
It was found that the average fuel consumption for a given driving behaviour is in- versely related to the average velocity. Thus, fuel consumption can be expected to be lower in areas with high average velocity, such as highways compared to areas with lower average velocity such as cities or neighbourhoods. This is illustrated in Figure 2.1.
Figure 2.1. Fuel Consumption for Different Traffic Conditions (Fotouhi & Montazeri, 2012).
Congested driving conditions are typically related to a high amount of accelerations and decelerations which are costly from a fuel-efficiency perspective. This in conjunc- tion with the correlations between average velocity and fuel consumption described by Fotouhi and Montazeri implicates it is easier to maintain a low fuel consumption in less congested driving conditions.
Variance in velocity (i.e. the difference between maximum and minimum velocity) is a factor that markedly affects fuel consumption in congested driving conditions (Fotouhi
& Montazeri, 2012). It is thereby likely that smoother accelerations lead to lower fuel consumption in these areas. On highways and other areas with comparatively little congestion, where the variance in velocity is relatively low, it is likely to be more ben- eficial to use quick accelerations in order to reach desired speed faster. Across all driving
conditions, longer decelerations and using engine braking is more fuel-efficient (Fotouhi
& Montazeri, 2012); (Berry, 2007).
2.3.2 Accelerations
Accelerations are costly in terms of fuel efficiency. As discussed by Berry (2007), an acceleration of 1 m/s2 results in a doubling of fuel consumption versus maintaining a constant speed in average. Driving in driving conditions that require a high amount of accelerations and decelerations, such as congested traffic or urban driving, tend to be more fuel-consuming than highway driving where the velocity can be kept relatively constant (Berry, 2007).
Research shows that the lowest fuel consumption can be found between 1000 and 1500 RPM (Lee, Park, Jung, & Yoo, 2011), as illustrated in Figure 2.2. To maintain a high fuel-efficiency, it would be advisable to maximise the amount of time spent in this interval. One approach to this, promoted by the Swedish Transport Administration, is quickly accelerating in order to reach the desired velocity fast and then try to keep the vehicle at a low, constant RPM by using the highest possible gear. During acceleration, a commonly taught advice is to “skip” gears, for example by going from first to third to fifth gear. This technique also allows the driver to start cruising (maintaining con- stant speed) faster, thus saving fuel (Trafikverket, 2013).
Figure 2.2 Instantaneous Fuel Consumption vs. RPM in a Hyundai Grandeur TG Q270 (Lee, Park, Jung, & Yoo, 2011).
2.3.3 Decelerations and braking
Manual braking is undesirable from a fuel-efficiency perspective as it takes away speed that was achieved through the use of fuel. It is in other words a “waste” of momentum.
Of course, it is not possible to completely avoid braking, but it is possible for drivers to minimise the need for manual decelerations by anticipating traffic flow and avoiding accelerations into high speeds if there is a need for braking further ahead. Using engine braking instead of the car’s braking mechanism leads to smoother driving since the technique requires a long braking distance. Smoother driving and lower variance in velocity can reduce fuel consumption, as discussed in chapter 2.3.1.
This technique can for example be applied in a scenario when a driver is approaching upcoming traffic lights, where early engine braking can reduce the need of stopping and idling at a red light. In an ideal situation, there is no idling as the traffic light turns green again by the time the vehicle arrives. This reduces the amount of fuel-consuming accelerations and idling in cities (Trafikverket, 2013).
Moreover, many modern engines are designed to cut off fuel supply entirely when no throttle is applied. This is referred to as Deceleration Fuel Cut-Off, or DFCO and further lowers fuel consumption.
A summary of the best eco-driving practices identified from literature is provided in Table 2.2.
Table 2.2. Best eco-driving practices identified form literature.
Activity Best practice(s) identified from literature
Accelerations Accelerate slowly and smoothly in city traffic, but quickly when accelerating onto a highway.
Decelerations Braking is an indicator of unnecessary acceleration and should be performed as slowly as possible. The most fuel efficient practice is to use engine braking in high gears, resulting in slow decelera- tion speeds.
Gear selection ”Skipping” gears when accelerating is recommended in order to minimise acceleration time. When desired speed is achieved, the highest possible gear should be selected.
Speed Variance Maintain as constant velocity as possible yields the best fuel-effi- ciency, as the number of acceleration is minimised.
Characterising driver preferences
A person’s set of preferences of how to approach and handle various driving activities (such as accelerating or decelerating) can be regarded as the person’s driving prefer- ences (not to be confused with the driving behaviour, which refers to the actual behav- iour of the driver). Many scientists and researchers refer to driving preferences as a person’s driving style. However, in order to avoid confusion the term driving preferences is used in this study.
Driving preferences is an important factor of fuel consumption, as it affects the ap- proach taken by a driver to various driving situations. Numerous researchers have analysed driver preferences in specific driving situations and activities (Wada, Yoshimura, Doi, Youhata, & Tomiyama, 2011); (Tricot, Rajaonah, Pacaux, Anceaux,
& Popieul, 2003). Nevertheless, analysing driving preferences is a difficult matter, as it is a very wide concept. Chen, Fang & Tien (2012) suggest that ”modelling individual driver behaviours is not enough to fully understand a person’s driving style”, thus ar- guing that a person’s behaviour in a specific driving situation reflects only a small portion of the driver’s set of preferences.
Chen et al. proposes a model that is intended for general use (i.e. not tailored for characterising behaviour in specific situations). Though this model is too computa- tional-heavy for use in the context of this work, other less complex models with the same purpose exist. Proposing one such model, Contantinescu, Marinoiu & Valdiou (2010) attempts to characterise driver preferences using a set of parameters all deriva- ble through GPS data. This model clusters driving styles in four different areas; ag- gressiveness, speed, accelerations and braking.
Positive Kinetic Energy, PKE is sometimes used as a measure of the “aggressiveness”
of a driver and is calculated from the average acceleration speeds of a specific trip. A high PKE value is said to be characteristic of an aggressive driver. Even though this is a very simple technique, it fits the scope of this study since the main goal is not to characterise different drivers’ preferences. Fully understanding a person’s driving pref- erences and reasons behind it is outside the scope of this work. It is however still important to note that the model does not consider external parameters (e.g. driving conditions). Thus, the model needs to be applied over several trips in order to generate reliable results.
Using mobile devices as learning aids
Mobile learning, or M-learning, is a relatively new concept. Still, numerous studies assessing its use and effectiveness in different contexts have been performed; (Hiraoka, Terakado, Matsumoto, & Yamabe, 2009); (Ishikawa, Onda, Wantabe, Kobayashi, &
Kurihara, 2010); (Magaña & Organero, 2011). Studies have shown that the use of mo- bile devices in education can lead to improved results amongst students (Thornton &
Houser, 2004); (Riconscente, 2011).
M-learning differs from “traditional” e-learning in several aspects, and a common view is that M-learning is an evolved of e-learning (Korucu & Alkan, 2011). Most im- portantly, M-learning includes being able to study and learn independent of location, thus opening up new possibilities in terms of learning concepts and methodologies.
Few studies comparing the relative effectiveness of different implementations of M- learning have been performed. However, a common way of designing this kind of ap- plications seems to be employing a game-style concept. Studying effects on school re- sults for children in 5th grade given access to an iPad mathematical fractions learning game, an average increase of 15% in test performance was observed for the children who used the game (Riconscente, 2011). Also notable, the children given access to the game showed a higher liking of fractions at the end of the test period.
However, designing an m-learning application is not an easy task. In order for a game- style learning application to be effective, the balance between play and learning is important. Simply adding teaching elements into a game might cause boredom rather than improve learning among users (Ma, Chen, Hwang, & Ding, 2012). Research has shown that usefulness as well as ease of use are important factors in terms of user adoption (Cheon, 2012), and that the potential benefits of using mobile devices in education can be turned into disadvantages or frustrations should the technology not be working as intended or be difficult to use (Gikas & Grant, 2013). In short, to max- imise the potential of a mobile learning application, it is important that both educa- tional, motivational, and usability factors are considered.
2.5.1 Goal-based learning
It has been argued that learning games employing concepts with personal goals are more likely to have positive cognitive effects on its users (Tanes & Cho, 2013). One likely reason for this, as noted by Tanes & Cho, is that achieving or completing a goal often requires many attempts, and so the concept is highly repetitive by nature. Also, fulfilling tasks is known to cause positive feelings amongst individuals, something that
is likely to increase the chance that users remain committed to the game and conse- quently the learning process.
So called motivational constructs have been argued to be vital components in self- learning concepts (Code, MacAllister, Gress, & Nesbit, 2006). Motivational constructs, as implied by the name, refers to techniques used to keep a person motivated and committed to a learning process. Applying goals to the learning model is a commonly used motivational construct for self-learning. It not only works to ensure that a person stays motivated to the process, but also provides the ability to track progress and achievements over time. Though few studies exist that compares goal-based e-learning to other e-learning models, goal-oriented approaches are very common in other areas, mainly organizational management, where researchers have found positive correlations to learning (Payne, Youngcourt, & Beaubien, 2007).
Together with goal-orientation, self-assessment has been shown to have positive effects on teaching driving behaviour (Wada, Yoshimura, Doi, Youhata, & Tomiyama, 2011);
(Hiraoka, Terakado, Matsumoto, & Yamabe, 2009). In a study of how an electronic learning aid system adaptive to driver’s skill could be used to affect eco-driving behav- iour, a noticeable decrease in fuel consumption was seen in subjects’ driving over a five day-period (Wada, Yoshimura, Doi, Youhata, & Tomiyama, 2011). Important to note is that this study did not use mobile devices, but utilised a purposely built driving simulator in which test subjects were instructed to follow a lead car. Nevertheless, the study shows that drivers are likely to adapt their driving behaviour if instructed to do so, even without the help of a human instructor.
Based on theories covered in this chapter, using a learning model that incorporates self- learning together with goal orientation can be argued to be the most suitable approach.
Summary and Conclusions from Theoretical Research
Correctly assessing driving behaviour is not an easy task. Numerous external factors such as driving conditions, vehicle type, road condition and more all play important roles on the fuel consumption of a vehicle. It is therefore important to make a clear distinction between a driver’s effort to minimise fuel consumption and the de facto measured fuel consumption. Not only does the recommended best practice differ de- pending on the context, but also the relative difficulty of actually following this prac- tice. Consequently, an assessment model must not only be general and widely appli- cable, but also adaptable to varying short- and long-term conditions.
Prior studies in the field mostly focus on determining an estimation of the user’s eco- driving performance (Andrieu & Pierre, 2012); (Magaña & Organero, 2011). However, as the aim of this study is not only to assess drivers’ eco-driving performance but also teach more fuel-efficient driving techniques, the model used must have ability to dis- tinguish different parts of the driver’s behaviour in order to find improvable areas.
Success in terms of an actual ability to affect users’ driving behaviour towards lower fuel consumption does not only depend on giving accurate advice, but also on main- taining user dedication throughout the process. This is another difficult matter to han- dle, due to several aspects, with the most important being:
Interaction between the user and the application is likely very low, since users must focus on driving instead of using the application.
A person’s driving behaviour is in many cases something that has been built over many years and therefore difficult to affect.
Though the use of personal devices as eco-driving learning aids is a largely unexplored field, one can presume that an application that gives correct advice as well as encour- ages and motivates users to follow the given advice will be able to affect users’ driving behaviour towards lower fuel consumption. An ideal implementation of such an appli- cation seems to be a concept that unites work with play, and uses a concept employing personal goals as the main learning model.
3 Method
Application goals and requirements
A number of goals and requirements conforming to the overall project goals was defined for the application. The requirements were based on the project goals as well as im- portant factors identified from literature. Non-functional requirements were not ob- tained through a requirements gathering process, but were instead based on well-rec- ognised best practices. A complete specification of requirements is provided in Appen- dix A.
Fulfilling the specified requirements was the main objective when developing the ap- plication. It should be noted that visual design aspects, such as user interface, have had a lower priority than the functional requirements. Still, considerable time was devoted to make the application aesthetically pleasing. The visual design of the appli- cation is described in more detail in section 3.6.
Application architecture
The application was developed for the Android platform, which uses the Java program- ming language. The Android platform was suitable for this work, as it is widely used and cross-compatible for a wide range of mobile devices, such as smartphones and tablets. An alternate solution would have been to develop the application for Apple’s iOS operating system. However, the interface functionality developed by Karthik Ko- tari is written in Java for Android. Thus, developing the application for Android elim- inates the need of porting the interface functionality for iOS.
Application features
Based on the found correlations between driving behaviour and fuel consumption sum- marised in Table 2.2, the application has ability to analyse and assess four key aspects relating to driving behaviour and fuel consumption; acceleration behaviour, decelera- tion behaviour, gear selection and velocity variance. The application also detects and adjusts assessment based on the current type of area.
The user is given a score in the range of 0 to 100, where 100 is the best possible behaviour in the current situation. The user is also presented scores for each individual parameter measured, both in real-time and after completing a trip. In order to project simplicity and clearness, scores are illustrated as five-star bars that are filled to the percentage of the score, rather than as the actual score value. For example, a score of
75 would be represented as a 75% filled bar of stars. After completing a trip, the user is also presented with a chart that displays speed and score over time.
After completing the first trip, the user is given a goal to improve a specific aspect of their driving, such as engine braking or highway accelerations. This goal is generated based on which assessed parameter that was given the lowest score. A goal requires the user to reach a certain score on the specified parameter for five times. The required score is based on the user’s original score for that parameter. Once a goal is completed, the application will generate a new goal based on the user’s current “weakest point” in terms of eco-driving. If a user completed a goal but is still performing better on other parameters, the same goal may be given again, but with higher requirements. The user can track progress over time, as well as accessing an overview of the current goal directly from the main menu in the application.
In addition to the above described features, the user can access information about the best practices used in the application, as well as instructions of how to set up and use the application. A specific part of the application allows for configuring parameters and preferences.
The different parts of the application are illustrated in Figure 3.1
Figure 3.1 Application visual design. 1: Main Menu, 2: Instructions, 3: Settings, 4: Profile overview, 5: Real-time feedback, 6-7: Post-drive feedback
In order to use the application, it is required that a coupling between the device and the Bluetooth adapter has been set up. This is made outside of the application, in the operating system. It is also required that the user has specified a vehicle to use in the application’s configuration.
An important requirement is that the application can be used in real-time as well as asynchronously, since the user may want, or be required by regulations, to focus atten- tion fully on the actual driving. The application has been designed such that the anal- ysis process can operate without any interaction as soon as a connection to the Blue- tooth adapter has been established. Should the user wish to receive real-time infor- mation from the application, a user interface (shown as part 5 of Figure 3.1) has been implemented.
Internal architecture
The application’s internal architecture utilises a number of modules responsible for gathering data and analysing different aspects of the user’s driving. A high-level over- view of the architecture, with the currently utilised modules, is illustrated in Appendix B.
The main analysis functionality has been implemented as a background service, rather than in the main application process. This allows the application to work in the back- ground for long periods of time and prevents the data gathering and analysis from being interrupted should the user perform other activities with the device, such as placing phone calls or opening up other applications. This not only ensures that data gathering and analysis is active throughout a complete trip, but also decreases the invasiveness of the application, in conformance with requirement 2.3. The service es- tablishes a connection to the Bluetooth adapter and is automatically started once the user activates real-time analysis. In case the user has not paired the device with a Bluetooth adapter, the service will shut down and display an error message. Conse- quently, the application is not usable without an active connection to the Bluetooth adapter.
The main data stream processing functionality is based on the free to use event pro- cessing engine Esper. Esper provides Java functionality for advanced analysis of real time or asynchronous events, which is very suitable for this kind of application. Ac- cording to the official documentation, the engine ”Enable[s] rapid development of ap- plications that process large volumes of incoming messages or events, regardless of whether incoming messages are historical or real-time in nature” (EsperTech, 2013).
The main benefit of using the Esper engine is that it simplifies handling of data streams, as it allows for easy extraction of data within customisable time windows. There are also built-in functions that perform basic mathematical operations, for example com- puting the standard deviation of a value within a given timespan.
The application uses independent ”analysis modules” for processing and analysing in- formation relating to certain aspects of the user’s driving. For instance, one module analyses the driver’s behaviour when accelerating and decelerating, and another module monitors the variance in speed. The architecture is scalable and allows for adding of new modules in order to expand the functionality of the application. Each module scores different aspects of the user’s driving behaviour and modules are only active when needed. For example, the user’s speed variance is not measured when driving in
congested traffic situations, as drivers in these situations are essentially forced to follow the flow of the traffic, which is not always the best from an eco-driving perspective.
A number of static helper classes coordinates and manages the separate modules de- scribed above. This includes distributing timestamps, activating and deactivating mod- ules and calculating weighted scores based on each module’s respective current score.
A simplified, high level overview of the application architecture is illustrated in Figure 3.2
Figure 3.2 Application architecture
Although the data gathering architecture is based on the Esper engine, all algorithms and analytical processes have been custom made and coded into the application. The functionality and algorithms used are unique for each module. However, as stated pre- viously, all modules employ the Esper framework.
3.4.1 Data gathering functionality
Data gathering is based on the functionality developed by Karthik Kotari, with some modifications. Since this functionality is extensively detailed in Kotari’s paper, only a brief description will be given here. This functionality is used to send OBD-II-request commands to the Bluetooth adapter, which returns the value of the requested sensor as a response. By default, commands are issued with an interval of 250ms. The com- mands used are specified in Table 3.1.
Service
Data gathering
Data analysis User
Interface
Table 3.1. Data measured for analysis.
Sensor Command Return type
Vehicle speed 010D Vehicle speed in km/h
Throttle 0111 Applied throttle in per-
cent.
RPM 010C Revolutions Per Minute
Engine Load 0101 Engine load in percent
The received data is sent to the static handler classes which in turn send it to the analysis modules for processing.
The modifications to this functionality were mainly related to porting and formatting data into the application’s object architecture, and streamlining and optimising re- source use.
3.4.2 Data analysis functionality
As described previously, a set of modules is responsible for analysing different param- eters of the user’s driving behaviour. The parameters analysed by the application is based on the best practices described in Table 2.2. As data is received from the Blue- tooth adapter, a static handler class sends the raw data to the Esper engine. Each module contains its own Esper clauses, which means that they will extract differing data from the engine. Each module extract the minimum amount of data required to perform its operations.
Each module has access to a static class which keeps updated information about the current state of the vehicle, as well as parameters such as driving conditions, time etc.
This is necessary since many of the algorithms used are dependent of the vehicle’s state.
The modules can both read from and alter the current state, should a change be de- tected.
If a module detects a relevant fuel consumption-related activity it will generate a score and send it back to the static handler class. After all modules have finished processing the current round of OBD-data, the handler class will calculate an average of each module’s score. Since the effect in terms of actual fuel consumption varies between the different parameters measured, each module is assigned a weight used when calculating the average score. When calculating the weighted average, only the weights of the currently measured parameters are included.
The average score is logged and sent to the user interface to be presented to the user.
Figure 3.3 illustrates the complete process that occurs with a 250ms interval.
Figure 3.3 Data gathering and analysis cycle
This process is repeated until the user manually chooses to stop the analysis service, or the contact with the adapter is lost.
3.4.2.1 Determining area
Area and driving conditions highly affect fuel consumption. It is therefore important that the application has the ability to adapt assessment techniques based on the current driving conditions. The current area of the vehicle is determined by calculating an average of the recorded velocity over the last 10 seconds. If the average speed is below 70 km/h, the vehicle is considered to be in a city area, otherwise in a highway area.
The current area is reported to the state handler class, so that other modules can use the information.
3.4.2.2 Assessing accelerations
Accelerations and decelerations are measured using the same module. This is because the data used is almost identical as technically, decelerations are negative accelerations.
As the vehicle enters a state of acceleration, the speed and time of acceleration start are stored in memory. When the vehicle exits the acceleration state, either by entering deceleration or keeping a constant speed for a certain amount of time, the acceleration is considered to be completed. Accelerations are assessed differently depending on the area. Slow and smooth accelerations are most fuel efficient except when accelerating on to a highway (i.e. from low to high speed). Thus, the acceleration module uses
Gather data
Send to handler
Calculate parameter
scores
Process scores Display on
UI Wait
different algorithms for assessing the driver’s performance depending on the current area. When accelerating within one area, the user’s score will improve with lower ac- celeration speeds in m/s2. When accelerating from a city area on to a highway, accel- eration speed is not measured since acceleration speeds vary between vehicles. Instead, this scenario measure the driver’s use of throttle and rewards high throttle percentage.
3.4.2.3 Assessing decelerations
Since the best practice for decelerations is the same in every context (long, smooth decelerations and as much use of engine braking as possible), all decelerations are eval- uated in the same way.
Decelerations are analysed by calculating two separate scores, one score for the actual deceleration speed, and one score for engine braking utilisation. A combination of both these scores is the user’s total deceleration score. In order to determine whether the driver is engine braking or not, the application monitors data feeds from throttle and engine load sensors. As long as the vehicle is decelerating, the application monitors the use of engine braking. When the deceleration is complete, the percental use of engine braking is combined with the deceleration speed score to create a total deceleration score.
3.4.2.4 Assessing velocity variance
Velocity variance is only measured on highways, as it is practically difficult for a driver not to follow the traffic flow in congested traffic situations. The variance analysis algo- rithm monitors the range between the vehicle’s highest and lowest speed in intervals of 30 seconds. If the range is very high, it is assumed that the user has performed a major acceleration (such as on to or off a highway), and the measurement is discarded.
3.4.2.5 Assessing gear selection
As previously described, the standard OBD-II interface does not support reading the current gear of the vehicle. This makes it difficult to accurately determine whether the user is driving with the optimal gear for the current velocity. The application works around this problem by calculating the ratio between engine RPM and velocity at the highest gear of the vehicle. Since this ratio varies between vehicles, the application needs to ”learn” each vehicle’s gear ratio at the highest gear.
This is done by monitoring the ratio between RPM and velocity at high speeds, where users are likely to use the highest gear of the vehicle. However, it is not unlikely that lower gears might be used in these velocities for shorter periods of time, for instance when performing an acceleration. To ensure that the driver actually uses the highest gear of the vehicle, the lowest ratio observed during an interval of 20 seconds is used.
If the driver slows down to lower speeds before the specified 20 second time limit the ratio is disregarded, otherwise the ratio is stored so that it can be used for any future trips.
This technique does not make it possible to derive exactly what gear the driver is using.
However, since the relationship between RPM and velocity is linear for any given gear, it is assumed that if the ratio at any given point is higher than the observed ratio for the highest gear, it is possible for the driver to change up at least one gear.
Once this ratio has been established, the application monitors gear selection and gen- erates a score based on the vehicle’s current rpm/speed ratio versus the “optimal” ratio detailed above at any time when the user is driving in highway conditions and not accelerating or decelerating.
Since it is not always advisable to use the highest gear, this measurement is only per- formed if the vehicle is not accelerating. Furthermore, gear selection analysis is deac- tivated if the user has defined that the vehicle used has an automatic gearbox.
3.4.3 Driver preferences characterisation functionality
The application characterises user’s driving preferences by using the ”aggressiveness factor”, i.e. average acceleration speed in m/s2 approach discussed in section 2.4. Results gathered through this approach are interpreted into three different driving preferences;
Careful Moderate and Aggressive.
The recommendations used by the application are matched to one or more driving preferences, so that the user is only given recommendations matched to their respective preferences. It is possible to deactivate this feature, in which case the application will disregard the user’s preferences.
Recommendations and feedback
Recommendations are given to the user in the form of goals to achieve. This conforms to the game-style concept of the application and makes sure that the user has only one aspect of their driving behaviour to focus on at a time. Goals are generated based on which aspect of the user’s driving behaviour was given the lowest score. In case the feature is activated, the goal will be matched to the user’s driving style preferences.
Table 3.2 summarises the different goals that can be generated by the application.
Table 3.2. Goals/Recommendations used in application.
Goal Cause Matched driver prefer-
ences
Decrease acceleration speeds
Too high acceleration speeds when not accelerat- ing from low speed to high speed
Careful, moderate
Increase acceleration speeds
Too low use of throttle when accelerating from low to high speeds.
Moderate, aggressive
Use higher gear Too high RPM/velocity ratio when not accelerat- ing
Careful, moderate, aggres- sive
Use longer decelerations and utilise engine braking
Too fast decelerations or low engine braking per- centage
Careful, moderate
Keep a more constant speed
Too high speed variance in highway conditions
Careful, moderate, aggres- sive
Visual design
Though visual design has had a relatively low priority compared to the actual features of the application, considerable time has been put into designing the application’s visual elements. Similar to the application requirements, design principles are based on rec- ognised best practices rather than pre-study research.
The main goal of the visual design is to communicate the application’s recommenda- tions to the user. The visual elements have therefore been designed to give clear and quick access to information, which is important should the user want to monitor feed- back while driving. In conformance with the established application requirements, sim- plicity has also been an important design goal. Moreover, the visual elements of the application are designed to communicate ”green” thinking, by employing a green and white theme throughout the application.
Performance
Since the application is designed for use on handheld devices with limited battery life and processing power compared to many other computing devices, resource use and performance optimisation is important.
In order to minimise the application’s resource use, all features and algorithms have been designed to be as lightweight computationally as possible. The architecture has
also been designed to reuse as much resources, such as UI elements or methods, as possible instead of utilising ad-hoc implementations.
Moreover, the application contains functionality for adjusting resource use according to the situation. For example, the interval of how often data is written to the device’s memory is dynamic, with increasing intervals with trip time in order to decrease memory use. This means that a short trip will have a shorter interval between data log entries than a longer trip. The application also has functionality for freeing up memory space, should so be needed.
Testing
Tests were performed by supplying a test subject with an ELM-327 Bluetooth adapter, produced by ELM technologies as well as a copy of the application. The ELM-327 adapter is relatively inexpensive compared to other similar devices with a market price at around SEK 400 or EUR 50. The driver was tasked with using the application over a period of 10 days in April 2013, with the aim being to evaluate real-world effects on the test subject’s fuel consumption from using the application, as well as determining if there is a correlation between the scores generated by the application and the meas- ured fuel consumption
3.8.1 Test setup
The main objective of the test was to analyse and be able to answer the research questions posed in the study. The test subject was instructed to use the application for a routine trip one time per day for a total of 10 trips. The subject was not explicitly instructed to follow the recommendations or to attempt achieving goals generated by the application. However, in order to evaluate the application in terms of user appre- ciation, the subject was instructed to actively use and give feedback on the application’s different features.
The test course spanned a distance of 15 kilometres within the town and suburban areas of Luleå, Sweden. Trips were performed at approximately the same time of day each time in order to minimise impact of external factors such as traffic congestions, temperature, etc. Figure 3.4 illustrates the trip course.
Figure 3.4 Test course overview.
This course consists of both low- and high speed areas, with high speed-areas making up for the major part. Table 3.3 shows the different speed limits of the trip path, it should however be noted that the speed limits over the course only serves as a reference.
The test subject was not given instructions to maintain specific speeds.
Table 3.3. Test course speed limits.
Speed limit (km/h) Distance (km) Percent of total trip distance
40 1 6,6
50 1,2 8,0
70 2 13,2
100 10,9 72,2
Total: 15,1 100
.
The vehicle used was a 2007 model Saab 95 Bio Power. This car can be driven on both gasoline and ethanol. However, in order to eliminate variance caused by fuel type, the car was driven solely on ethanol during the test period.
The test subject was informed about the purpose of the study prior to the test start.
However, in order to minimise the risk that the subject adapted a more fuel-efficient driving style in order to perform well in the study, no information about what is good eco-driving practice was given. Nevertheless, as described in section 3.3, a set of in- structions describing fuel-efficient techniques is available within the application. No instructions or feedback, other than help with equipment setup, was given to the sub- ject during the test.
The data gathered during the test are described in Table 3.4.
Table 3.4. Data measured for application effects testing.
Id Data Unit Frequency
1 Baseline average fuel consumption l/10 km Before test start
2 Average fuel consumption l/10 km Trip 5, 10
3 Acceleration score Score, unitless, 0 - 100 Each trip 4 Deceleration score Score, unitless, 0 - 100 Each trip 5 Speed variance score Score, unitless, 0 - 100 Each trip 6 Gear selection score Score, unitless 0 - 100 Each trip 7 Total score Score, unitless, 0 - 100 Each trip
Data id 1 and 2 were measured in order to evaluate effects of using the application.
Other parameters were measured in order to compare the scores generated by the ap- plication to the de facto measured fuel consumption.
Measuring exact fuel consumption was not feasible, since the vehicle used was not equipped with sensors that monitor real-time fuel flow. However, most modern cars (including the one used for this test) feature on-board computers that allow drivers to monitor fuel consumption in terms of litres per 100 km. Fuel consumption was meas- ured with this technique as the moving average of the fuel consumption of trip 1 – 5 and trip 6 – 10 respectively. This method was chosen in order to minimise non-random variation of fuel consumption. Since the car was not solely used for purposes related to this test, the test subject was asked to reset the fuel consumption meter before and after each test trip, thus ensuring that fuel consumption was only measured for trips on the specified test course. A simple procedure to follow for each measured trip was established in order to ensure that each trip was measured in the same way. This procedure is detailed below:
Reset fuel consumption meter
Perform drive on test course
Report data described in Table 3.1.
Prior to beginning the test, the subject was instructed to perform the procedure detailed above for 5 trips on the test course. An average of the fuel consumption of these trips was calculated in order to establish a reference pre-test fuel consumption. This reference consumption was used for comparison and evaluation of the effects of using the appli- cation.
During the test period, a specific functionality for automatic sending of measured data via e-mail was implemented in the application. This functionality automatically gath- ered and formatted the required test data directly from the application. The feature was implemented in an effort to decrease the risk that the test subject accidentally reported incorrect or irrelevant data.
3.8.2 Data Analysis
To answer the research questions posed in the study, the gathered data was analysed in different steps. The first step involved performing a linear regression analysis be- tween the test subject’s fuel consumption, the scores generated by the application, and the number of times of application use. Since the fuel consumption was measured as a moving average of intervals of five trips, the regression analysis was designed to com- pare the fuel consumption to the moving average score for the same five trips. In the next step, a correlation analysis was performed to find any relationships that existed between the parameters measured.
An additional analysis was performed in order to evaluate if the goal-based learning approach used in the application and study is likely to give better results than an application that uses a pure self-assessment concept. The analysis method for the re- search questions is summarised in Table 3.5.
Table 3.5. Data analysis summary.
Research question Evaluation method Measures
I Correlation analysis (parameter
scores vs. measured fuel con- sumption).
Driving behaviour parameter scores as generated by applica- tion, measured fuel consump- tion.
II Regression analysis (Eco-driving score vs. number of trip, fuel consumption vs. number of trips)
Total weighted average score, measured fuel consumption.
Goal-based learn- ing concept
Regression analysis (goal pa- rameter score vs. number of trips, average score vs. number of trips)
Driving behaviour parameter scores
The goal-based learning concept was tested by comparing the correlation between the parameter included in the test subject’s personal goal against the score progress for the weighted average. It is important to note that in order to be able to reliably evaluate the goal-based learning concept, research question I must be concluded to be true. This is because the measures used need to be accurate, and their accuracy is validated in research question I.
Analysis was performed using Microsoft’s Excel Software, with the Analysis Toolpak plugin, as well as IBM’s SPSS software for statistical analysis.
4 Results
The application is able to successfully connect to and gather data from vehicles equipped with an OBD-II Bluetooth adapter. Even though no formal performance tests have been made, the application has not been perceived as having a noticeable effect on the performance of the device on which it is used. No severe errors (i.e. bugs) oc- curred during the test period, and the test subject was generally pleased with the ex- perience of using the application. The average fuel consumption prior to test start was measured to be 1.27 litres per 10km. This average was based on a total of 5 trips on the test course.
Assessment ability
For the regression analyses performed (measured fuel consumption vs. average score, fuel consumption vs. trip number, average score vs. trip number), the results are shown in Table 4.1.
Table 4.1. Correlation between Fuel Consumption and Trip Number, Average Score.
Moving avg. Fuel con- sumption Vs.
R B
TripNumber 0,444 0,001
Moving average score 0,952 -0,020
The Pearson correlation coefficient (also referred to as the R value) represents how strong the relationship is between two variables, ranging from -1 for a strong negative relationship, to 1 for a strong positive relationship. The B-coefficient represents the regression model.
The correlation between measured fuel consumption and the scores generated by the application was strong, with a Pearson relation coefficient of 0,952, indicating a very strong likelihood that fuel consumption is related to the total average score generated by the application. The B-coefficient of the correlation indicates that the fuel consump- tion is expected to decrease with 0,02 l/10 km if the score goes up one value, resulting in a difference of -0,02 * 100 = -2,0 l/10 km between a score of 0 and a score of 100. It is important to note that this is only an estimate, based on a vehicle driven solely on ethanol.
Eco driving performance vs. number of trips
Correlations between each measured parameter and the number of times of application use, hereafter referred to as trip number, is shown in Table 4.2. For a full representation of the correlations between the measured parameters, see Appendix D.
Table 4.2. Parameter correlations to number of times of application use.
As shown in Table 4.2, eco-driving performance for most driving behaviour parameters showed a positive correlation to the trip number. An exception is the test subject’s acceleration performance which showed a negative correlation, caused by decreasing city acceleration performance. Highway accelerations showed a positive correlation, but since the amount of highway accelerations required on the test course was very low — the test course only required one acceleration from low to high speed — compared to the amount of city accelerations, there was very little impact from highway acceleration score on the total acceleration score. Most correlations were relatively weak, with R values in the range of -0,2 to 0,2. Due to the fact that some parameters showed positive and some showed negative R values, the average score had a fairly weak correlation to the trip number with a correlation coefficient of 0,150. This parameter, together with fuel consumption, had the weakest correlation to the trip number out of all the param- eters measured. The parameter with the strongest correlation was engine braking, which was the parameter that the test subject was given a goal to improve by the application.
Due to the layout of the test course, with long distances where the test subject could use a single gear, the gear selection was not heavily tested, and was given maximum scores for every measured trip. Consequently, no correlation between this parameter
Parameter R B
TotalAccelerationScore -0,412 -0,909
CityAccelerations -0,461
HwyAccelerations 0,291
TotalDecelerationScore 0,913 0,297
DecelerationSpeed 0,185 0,115
EngineBraking 0,938 1,152
SpeedVariance 0,426 0,333
WeightedAvgScore 0,150 0,062
FuelConsumption 0,444 0,001
and number of times of application use could be distinguished. Another parameter to receive consistent high scores (>90) due to the same reason was speed variance and consequently, no trend could be distinguished in this parameter either.
Goal-based learning concept
After the first trip, the test subject was given a goal to improve on engine braking as this was the parameter that was given the lowest score. The goal set by the application was to achieve a score above 20 on engine braking for 5 consecutive trips.
The test subject’s behaviour in this parameter improved significantly, and scores in- creased more than 100% over the course of the test. The average score increase was approximately 5% per trip. In other words, after 10 trips, the test subject utilised engine braking more than twice as much as in the beginning of the test. The increased use of engine braking also led to an increase in overall deceleration performance, which was the area of lowest performance for the test subject. The test subject’s progression in engine braking performance is illustrated in Figure 4.1.
Figure 4.1 Engine braking score progress
The correlation between engine braking and trip number was very strong, with a cor- relation coefficient of 0,938, more than six times stronger compared to the respective values for the correlation between the total average score and the trip number. The parameter with the second strongest correlation (not counting total deceleration score, which is directly related to the engine braking score) was speed variance with a corre- lation coefficient 0,426. These results strongly suggest that drivers will perform better in parameters included in the goal than other parameters.
7 10 10 12 13 14 17 18
15 18
0 5 10 15 20
1 2 3 4 5 6 7 8 9 10
Score
Trip Number
Engine Braking Score
Engine Braking Linjär (Engine Braking)
The strong correlation between average score and fuel consumption described previ- ously suggest that the risk that parameters are not representative of the fuel consump- tion is small. Consequently, it is likely that the test subject’s improvement in engine braking scores leads to less fuel consumption.
5 Discussion
Conclusions from results
Based on the fact that an increase in scores generated by the application has a very strong correlation to lower fuel consumption, it is very likely that the application has the ability to - at least to some degree – assess eco-driving performance using solely OBD-retrievable data, and that a driver who receives high scores by the application will consume less fuel than a driver who receives lower scores. This does not however mean that the techniques and algorithms used are fault-free. For instance, it is plausible that some parameters require less effort than others in order to achieve a high score.
An example of this is the fact that the test subject’s scores varied highly between different parameters. All trips were given gear selection scores of 100, whereas deceler- ation-related parameters were given scores below 20. While it is possible that the test subject performed better at gear selection and speed variance than on other parameters, it is highly likely that it is easier to receive high scores on these parameters.
The fact that fuel consumption was measured as an average of accumulated trips, resulting in only three data points infers that one must be careful with interpreting the regression analysis results. These results indicated a difference of 2 litres per 10 km between a score of 0 and 100, but this is possibly an overestimate. Moreover, the effect from external factors on the measured fuel consumption cannot be disregarded, even though much care has been taken in order to minimise their occurrence and impact on the results.
While most parameters showed positive correlations to the trip number, accelerations in city traffic showed a fairly strong negative correlation. The test subject improved on highway accelerations, but since the amount of highway accelerations was very small compared to the amount of city accelerations for the specified route, this did not affect the total acceleration score much. The reasons for this negative correlation are difficult to answer and can only be speculated in. One possibility is that the test subject, when given a goal to improve on deceleration, focused more on improving this aspect of driving and concentrated less on other activities, leading to a declining performance on the parameters not covered by the goal. Another possibility is that the eco-driving instructions provided in the application was unclear and that the test subject conse- quently was confused about what practices that are most fuel efficient. This possibility is not unlikely, as acceleration behaviour is measured differently depending on the sit- uation. As no instructions were given to the test subject apart from what is available
in the application, it is possible that the test subject utilised the same technique for accelerations in all areas, which is not optimal and will lead to low scores.
A third possibility is that the application is not fully adaptable to the traffic situation.
This can be derived from the fact that the scores decreased specifically in city traffic, where drivers are less free to follow the best eco-driving practices.
Finally, since the test course covered a path routinely travelled by the test subject, there is a possibility that the test subject was used to driving the course in a specific way, making it more difficult to cause a change in driving behaviour. This is something that may have affected the results of the study and it could be debated whether choos- ing a test course that the test subject was not familiar with would have generated different, and perhaps more concrete results.
The negative progress in acceleration scores resulted in a very weak correlation between fuel consumption and number of trips. Consequently, it is difficult to answer whether using the application actually has an effect on drivers’ long-term fuel consumption. The answer to research question II is therefore inconclusive. A question brought up by analysing the results is whether accelerations are measured and weighted correctly, as the decreased score in this single parameter had a very high impact on the total aver- age. However, since the correlation between scores and fuel consumption vas very strong, it is difficult to argue that the scores are not representative of the driver’s eco- driving performance.
Another factor contributing to the weak correlation between the two parameters is the fact that the fuel consumption measurements only consisted of three data points. It is likely that a longer test with more data would have resulted in a stronger correlation between the two parameters. Furthermore, the way that fuel consumption was meas- ured was not very accurate, resulting in an error probability that is relatively high.
However, utilising precise fuel flow instruments to measure fuel consumption is proba- bly redundant for the used testing methodology.
The goal oriented concept of the application seems to have been beneficial, at least for improving the parameters covered by the goal. In the test, engine braking (which was the parameter given to improve), increased more than 100%. One could argue that this increase cannot be - fully or at all - attributed to the fact that it was the goal parameter, and that the progress would have been achieved with a pure self-assessment concept as well. Another aspect that should be investigated is whether using the goal-based ap- proach has a negative impact on parameters not included in the goal. This possibility
