Towards Data-driven Services in Vehicles

Milan Koch

, Hao Wang

, Robert B

urgel

and Thomas B

ack

BMW Group, Munich, Germany

Leiden University, Leiden, The Netherlands

Keywords:

Data-driven Service, Machine Learning, Damage Prediction, Connected Car, Vehicle Network, Online

Learning.

Abstract:

Numerous recent studies show the prosperous future of data-driven business models. Some key challenges

have to be dealt with when moving towards the development of data-driven car services. In this paper, a new

data-driven customer service is proposed for the settlement of vehicle low speed accidents. Beyond that, we

present a more general approach towards the development of data-driven car services. We point out its main

challenges and suggest a method for developing new customer-oriented data-driven services. This approach

illustrates key points in developing a practical service, from a technical and business related perspective. Such

data-driven services are developed mostly on a small number of initial test data, which results often in a limited

prediction performance. Therefore, based on an optimized CRISP-DM approach, we propose a methodology

for developing initial prediction models with limited test data and stabilizing the models with newly gained

data after deployment by online learning. On-board and off-board services are discussed with the result that

especially off-board running services offer a large potential for future data-driven business models in a digital

ecosystem. The ﬂexibility of such an ecosystem depends on the degree of the integration of the vehicle in the

ecosystem - in other words, the car needs to be enabled to deliver data on demand according to GDPR and to

any applicable regional law and in cooperation with the customer. The presented method, together with the

ecosystem, enables fast developments of various data-driven services.

1 INTRODUCTION

Robotics and transportation have been underpinned

by artiﬁcial intelligence since its early beginning. In

1969, Nilsson discussed the use of artiﬁcial intelli-

gence in integrated robot systems (Nilsson, 1969). In

the late 1970s pioneering discussions were made on

the ﬁrst autonomous vehicles with artiﬁcial intelli-

gence (Tsugawa et al., 1979). Across the end of the

1970s to the 1990s, ﬁrst prototypes were developed

by different scientists and organizations (Schmidhu-

ber, 2018). Such technical progresses continued until

2000 and the autonomous driving was feasible for the

ﬁrst time, sparking major developments in both re-

search and industry (Stone et al., 2018; Huber et al.,

2008; Aeberhard et al., 2019; Ardelt et al., 2012).

In autonomous driving, data from different sensors

are combined by computers deployed in the car (Liu

et al., 2017). Using methods of artiﬁcial intelligence

(speciﬁcally deep learning techniques), these comput-

ers predict the car actions that are required to han-

dle situations. Due to the large data volume, those

artiﬁcial intelligence models are mainly deployed on

on-board-systems (embedded) in the car (Aeberhard

et al., 2019). Beyond autonomous driving functional-

ities, certain types of car data, especially the one re-

lated to self-driving, are combined with a car internet

interface and a robust internet connection, offering a

new era of data-driven services. Most of these ser-

vices require no additional car hardware and operate

only with the vehicle data that is available. As such

services are mainly driven by small data volumes, the

data set used can be transferred to a back-end sys-

tem, complying with data protection regulations and

customer’s consent. This enables running the data-

driven service outside of the car (off-board). Hence,

new services do not require any changes in the hard-

ware, which signiﬁcantly simpliﬁes the service devel-

opment. It allows for the continuous and faster cre-

ation of new services, even within the lifetime of cars.

Therefore, off-board running services are much more

powerful than in-car computations: A car interface

sends data on request to a back-end system, which

uses data-driven models for a prediction and sends the

output, e.g., back to the car. This is combined with

full transparency and involvement of the customer re-

Koch, M., Wang, H., Bürgel, R. and Bäck, T.

Towards Data-driven Services in Vehicles.

DOI: 10.5220/0009458700450052

In Proceedings of the 6th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2020), pages 45-52

ISBN: 978-989-758-419-0

garding certain data. The interconnection of vehicle

and back-end system builds a so-called digital ecosys-

tem, which integrates all aforementioned methods to

provide car services. It enables faster service devel-

opment and deployment, even within the lifetime of

cars.

The revenue from mobility services and connected

car services are projected to reach USD 1,087 billion

by 2030 (Seiberth and Gruendinger, 2018). This is

not only a large business for OEMs (Original Equip-

ment Manufacturer), but also for suppliers as well

as ecosystem developers and other parties involved

(Seiberth and Gruendinger, 2018). Data availabil-

ity, its protection and privacy of an open (e.g. to

third parties) digital ecosystem, is of key importance

to integrate cars more seamlessly into our lives with

more digital services. This paper mainly illustrates

the technical approach with its challenges for devel-

oping new data-driven customer services, going from

the idea to a running service.

The remainder of the paper is organized as fol-

lows: First, existing work that is related to our ap-

proach is discussed in section 2. Second, we present

an example of a data-driven service in section 3. Our

proposed methodology with its six main steps is then

introduced in section 4. In section 5, we provide con-

clusions and an outlook.

2 RELATED WORK

Recent studies illustrate new data-driven business

models in the car industry by means of a digital vehi-

cle ecosystem. In this context, Seiberth et al. present

a deﬁnition of data-driven business models: ”data [...]

as primary business resource to deliver value to cus-

tomers and to convert this value into revenue and/or

proﬁt” (Seiberth and Gruendinger, 2018, p. 8). They

declare that in 2050 car manufacturers will achieve

50 % of the revenue from data-driven services. The

growing digitalization with its disruption process de-

stroys many traditional business models (Weill and

Woerner, 2014). The authors also picture more gen-

eral business models and the possibilities of digital

ecosystems for different industries. Car manufactur-

ers have different approaches to deal with digital ser-

vices and many tech start-ups are already develop-

ing sustainable business models with digital services.

Furthermore, OEMs enter already strategic partner-

ships and invest into such connected vehicle start-ups

(Kaiser et al., 2017).

Seiberth et al. discuss the available car data,

e.g. from sensors for autonomous driving, and high-

light the possible revenues when creating services

based on it (Seiberth and Gruendinger, 2018). In addi-

tion, there is a growing need for building trust towards

the customers regarding the use of their data for ser-

vices and therefore for the transparency of the data,

its use, and privacy and security (Kilian et al., 2020).

Beyond that, they present a ﬁgure of the connectiv-

ity ecosystem, which describes roughly a connectiv-

ity platform: it communicates with the data source

(car) and receives external data like weather, trafﬁc

etc. The connectivity platform is connected to the

OEM’s back-end system, as well as to third party ser-

vices and apps.

Many studies present new business models en-

abled by data (Seiberth and Gruendinger, 2018; Weill

and Woerner, 2014; Kilian et al., 2020). In most

cases, new service ideas are superﬁcially mentioned

and it is only brieﬂy discussed how to really beneﬁt

from each individual service. Some of the studies dis-

cuss the design of (vehicle) ecosystems, e.g. (Immo-

nen et al., 2016; Immonen et al., 2018), but a method-

ology for creating data-driven (customer) vehicle ser-

vices has not been a major topic of scientiﬁc research

yet.

3 A DATA-DRIVEN SERVICE FOR

CRASH DAMAGE PREDICTION

The variety of possible data-driven services is large.

A data-driven car service often assists the customer

(like a car pooling service) or the car (like predic-

tive maintenance services). Based on historical data

and with methods of artiﬁcial intelligence, models are

trained to predict behavior, e.g. in car pooling to pre-

dict the best possible route to carry the most passen-

gers or if a certain part of the car needs to be main-

tained in the nearby future.

Another example of a data-driven service is

a crash damage prediction system. Based on a

machine-learning model, such a system predicts the

damaged parts of a vehicle in a low speed crash. An

accident with a velocity difference below approxi-

mately 16 km/h is usually considered as low speed

crash. The baseline of this service is to use only on-

board data. Therefore, data from serial car sensors are

used for the prediction (e.g. acceleration). To gener-

ate an initial data set, low speed crash tests are per-

formed and certain on-board data are recorded. These

recordings are used together with the occurred dam-

age on the vehicle for training ﬁrst initial models. The

beneﬁt of such a data-driven service is e.g. immediate

transparency of the damage, which allows initiating

a faster and more convenient repair for the customer

(Seiberth and Gruendinger, 2018; Koch et al., 2018;

VEHITS 2020 - 6th International Conference on Vehicle Technology and Intelligent Transport Systems

Koch and B

ack, 2018).

However, the machine-learning model itself is

only one unit of the car service. When striving to cre-

ate seamless customer experiences with a data-driven

service, it is essential to consider the whole customer

journey. Such a journey describes the way how a

customer experiences the whole service. The overall

objective should be creating something which is so-

called convenient to the customer at all levels. This

can be achieved by designing the end-to-end service

with its technical challenges like data transfer or in-

tuitiveness of its handling as a whole picture. Based

on this, in the following we propose a general path to-

wards data-driven services to tackle and consider the

challenges with the one and only goal to create cus-

tomer value.

4 METHODOLOGY TOWARDS

DATA-DRIVEN SERVICES

In this section we propose a methodology for devel-

oping data-driven car services. This interdisciplinary

method is illustrated in Fig. 1. The horizontal axis

shows the time of the development while the vertical

one describes the level of development, i.e. the matu-

rity of the service. The origin presents the time of the

initial idea about the service and the beginning of its

development. The methodological approach consists

of six overlapping phases:

1. Idea.

2. Potentials.

3. Modeling.

4. Deployment.

5. Process.

6. Finalization.

All phases are linked to each other. In order to

allow short development times the phases are partially

executed in parallel. The phases are described in the

following sections.

4.1 Idea

The ﬁrst phase of Fig. 1 is referred as idea. This pic-

tures the timeline from the ﬁrst idea about the service

to very concrete solution concepts. Principally, there

are many motivations or ideas for thinkable services,

but for successful and seamless services the business

potential and customer beneﬁts have to be evaluated

continuously in the next phase, the evaluation of the

potential.

Figure 1: The methodological approach: From the idea to a

deployed data-driven service.

4.2 Potentials

Seiberth et al. state that new car services follow

mainly two objectives: improvement of the brand im-

age or increase of proﬁt (Seiberth and Gruendinger,

2018). This shows that the motivation to create those

is based on image or proﬁt reasons or a combina-

tion of both. Therefore, data-driven services can

have strong impacts on the brand and can be used

for strengthening images with creating so-called cus-

tomer experience by building positive experiences

followed by an emotional bond between user and

product (Glattes, 2016).

Next to retail customers other stakeholders like,

e.g., ﬂeet operators, insurance companies or other par-

ties can strongly support such services with their own

advantages (Seiberth and Gruendinger, 2018). Cre-

ating a service with many beneﬁting parties exploits

its potentials and is key for a successful and seam-

less service. Therefore, in case of promising ideas, in

phase 2 of Fig. 1, it is important to analyze and contin-

uously evaluate all aspects of the data-driven service

regarding the own objectives and the targets of part-

ners. However, it is mostly very ambitious to evaluate

the real potential of a new service in an ad-hoc man-

ner. Therefore, it is important to quickly develop pro-

totypes for experiments, get customer feedback and

constantly monitor the need for the service and decide

continuously to proceed or cancel the development.

In this context, after revealing an initial potential

of the idea, data scientists begin the phase of model-

ing with collecting data and designing ﬁrst models.

4.3 Modeling

In the beginning of the modeling phase, data scientists

have to prove the feasibility of describing the desired

relations by the available data with methods from the

ﬁeld of artiﬁcial intelligence. A feasibility study helps

to quickly assess the practicability of the idea.

To start the modeling phase, an initial data set

is crucial. In some cases, the data has been already

Towards Data-driven Services in Vehicles

collected and is available or can be gathered quickly.

However, in most cases the data has to be generated

manually. When considering the damage prediction

system, data from low speed crash tests are required.

Performing large numbers of such tests is very tedious

and expensive. Therefore, in such cases only small

initial data sets are generated in order to evaluate the

feasibility. Prediction models based on small data sets

are often of poor prediction quality. In order to in-

crease its quality and especially to have informative

results for the feasibility study, the use of optimiza-

tion techniques is key.

Shearer proposed an approach to run data mining

projects in industry, the cross industry standard pro-

cess for data mining (CRISP-DM) (Shearer, 2000).

Figure 2: The modiﬁed CRISP-DM approach with opti-

mization and online learning components. Note that the

ofﬂine learning part follows the methodology proposed in

(Shearer, 2000).

This approach has become a very known standard

process to perform industrial data science projects.

Roughly, it describes the process from the business

understanding to data understanding, data prepara-

tion, modeling, evaluation until its deployment. We

have modiﬁed parts of the CRISP-DM and added op-

timization between modeling and evaluation in order

to enhance the model performance. Furthermore, we

have separated the data into initial data and ﬁeld data,

as well as the process streams into ofﬂine learning

(black) and online learning (yellow) (see Fig. 2). Of-

ﬂine learning describes the process of learning mod-

els with an ofﬂine generated (initial-) data set (Ser-

vice Data) to create an (initial-) prediction model. In

addition, car information data like the car type, the

equipment of the vehicle and geometry information,

as well as external data like, e.g. weather or trafﬁc are

used as additional data resource, because such data of-

ten contain valuable information for the service. Af-

ter deploying the initial model in an ofﬂine learning

process, we are updating it by online learning (yellow

stream). This means that the initial model is stabilized

step by step after deployment with newly generated

ﬁeld data.

We have developed this modiﬁed CRISP-DM ap-

proach, when we were dealing with the modeling

of the crash damage prediction system. Its required

crash data is extremely difﬁcult to generate at large

volume, because either crash tests or simulations have

to be performed. Therefore, we only created a small

test data set containing just enough observations to

verify whether it is feasible to use on-board data to

predict the damaged parts. We obtained a data set

with 100 observations. The number of damages of

some parts is less than 5 among the 100 tests. This

indicates a very small and class-imbalanced data set.

Due to the character of our data set, ﬁrst results with,

e.g., multi-label classiﬁcation methods were not lead-

ing to promising results. More and more we have tai-

lored our approach: we developed a part-wise classi-

ﬁcation, i.e., we generated individual prediction mod-

els for each part of the vehicle. This was very promis-

ing, because each model has its own set of charac-

terizing features and its own set of hyper-parameters.

However, creating hand-crafted predicting models for

each vehicle part is a very time consuming process.

As a result, we developed our own automatic ap-

proach for time series classiﬁcation, a so-called ma-

chine learning pipeline (Koch et al., 2018). The input

of our pipeline are time series with the correspond-

ing label. The outputs are predictive model perfor-

mance measures such as accuracy or F1-score, which

describe the quality of the prediction.

Our automated machine learning pipeline for time

series classiﬁcation consists of four steps:

1. Feature Extraction from Time Series,

2. Feature Selection,

3. Modelling and

4. Hyperparameter Optimization of the Classiﬁer.

This pipeline describes the modeling and optimiza-

tion part of our modiﬁed CRISP-DM approach more

in detail. It consists of the following components: fea-

ture extraction, feature selection, training of a classi-

ﬁer and hyperparameter optimization. When dealing

with time series instead of non-temporal data, repre-

sentations (features) from time series need to be ex-

tracted. Traditionally, time series features are identi-

ﬁed manually by a very time consuming process. To

VEHITS 2020 - 6th International Conference on Vehicle Technology and Intelligent Transport Systems

generalize and automate this process, we take the ad-

vantage of tsfresh package (Christ et al., 2018) that

automatically extracts a potentially large number of

features. In tsfresh, there are 63 built-in feature func-

tions and in our scenario 794 features are generated

by parameterizing those functions differently.

In a next step, a feature selection algorithm

searches in the generated features space for the ones

that contain signiﬁcant information about the damage:

some of those 794 features per time series describe the

relation better than others. Common methods for fea-

ture selection are, e.g., forward selection or RFE (re-

cursive feature elimination) (Guyon et al., 2008). In a

forward selection process, models are computed iter-

atively, starting with no features. Subsequently, fea-

tures are added in each iteration and those improving

the model performance are kept until no enhancement

can be achieved anymore (Guyon et al., 2008). Alter-

natively, in RFE, a predictive model is ﬁrstly built us-

ing all features and the weakest feature is removed ac-

cording to some well-deﬁned importance metric. The

model is re-trained after the removal and such a pro-

cess is repeated until the speciﬁed number of features

is reached (Guyon et al., 2006).

In the next step, after selecting the most important

features, these are used to train a random forest classi-

ﬁer. Due to our small number of observations, model

optimization strategies are indispensable. Therefore,

the hyper-parameters of this random forest classiﬁer

are optimized. Common hyper-parameter optimiza-

tion techniques are, e.g., grid search or randomized

search (Geron, 2017).

We developed this pipeline to efﬁciently gener-

ate individual models predicting the damage for each

part and, more importantly, the pipeline can be used

for automatically enhancing and stabilizing the ini-

tial model performance after deployment by online

learning following the methodology of our modiﬁed

CRISP-DM approach.

Our initial pipeline models have achieved F1-

scores between 0% and 94%. This indicates, that

based on the small number of data points the pre-

dictability depends strongly on the part, i.e., the dam-

age of some parts can be predicted more precisely

than of others. Additional methods like frequent

pattern mining could help analyzing which parts are

likely to be damaged together within one crash. This

is especially important for parts with a low prediction

quality. Furthermore, by considering, e.g., the learn-

ing curves from our results we were able to foresee

an improvement of the performance with increasing

data (Koch et al., 2018). Among other things, this

helped us to evaluate the feasibility for practical use.

The solution space for such modeling problems

is usually large and often the combination of differ-

ent methods is leading to usable results in practice.

In our opinion automatic machine learning methods

(AutoML) like our pipeline approach are very promis-

ing, because of its practical use and due to its perfor-

mance and its efﬁciency (short computation times).

Modern vehicles employ many different sensors

and can produce large amounts of data. Sometimes

it is not obvious what data would be promising for

modeling. Therefore, from a possibly large number

of sensors, the most promising ones for the task at

hand have to be discovered.

As mentioned, after the data generation a feasibil-

ity study shows the practicability of the data-driven

service. When receiving results matching the expec-

tations, the models for the serial application can be

developed.

In conclusion, the key of the modeling phase is to

generate efﬁcient predictive models and to check the

validity of the service model approach by assessing

the quality of the models that can be learned from the

data. After the feasibility is identiﬁed and conﬁrmed,

the deployment of the service should be prepared.

4.4 Ecosystem/Deployment

The deployment of a data-driven service in an auto-

motive environment depends mainly on whether it re-

quires on-board or off-board running services. On-

board services are deployed on embedded systems in

the car. This needs data storage and computing power

on control units of the vehicle. Off-board services

are running in back-end systems. In this case, data

is transferred via the internet interface from the car

to the back-end, provided a sufﬁcient bandwidth is

available. In both off-board and on-board running

services the car needs to provide the required data.

In this regard, in the deployment phase the electronic

components of the car have to be enabled to deliver

the needed data. Fig. 3 shows a typical vehicle net-

work of a passenger car. Such architecture consists

of different communication systems like Ethernet and

more traditional bus systems like FlexRay, CAN or

LIN. Ethernet is a local area network (LAN). It is

designed to transmit data between computers (Spur-

geon, 2000). BES are bridged end stations (switches),

which can send and receive transmissions. Bridges

communicate to other bridges, to the gateway (router)

and to end stations (ES), which is in an automotive

environment, e.g., the head unit (Spurgeon, 2000;

Spurgeon and Zimmerman, 2013). Bus systems like

FlexRay, CAN or LIN differ in various bandwidths

and each system transfers data between components,

called electronic control units (ECU). ECUs are em-

Towards Data-driven Services in Vehicles

bedded systems, which control electrical systems in

the vehicle. A car contains many ECUs like the en-

gine control unit, the airbag control unit, the bat-

tery management system or the telematic control unit,

which sends and receives data via the mobile network.

All bus systems are connected via gateways (Robert

Bosch GmbH, 2014; Matheus and K

onigseder, 2015).

Sensors are connected to the ECUs, which process the

sensor raw data and route it partly to the bus system.

This bus data can be used from other ECUs within

the connected bus or by the gateway. In some cases,

data from one bus is required on another bus. Then,

the gateway routes this data from one bus to another.

However, mostly data is only available in the ECUs

or on the initial bus system.

Figure 3: A typical schematic in-vehicle network.

Deploying a service on-board in an ECU de-

mands a high effort regarding receiving/delivering the

needed data, matching the data quality requirements

and the general deployment of the software within the

automotive system. This additional software needs to

harmonize with all car systems and thus implies very

costly technical security. In addition, the capacity

of storage within the ECU and its computing power

is limited for additional services due to the fact that

ECUs serve most likely other essential vehicle func-

tions. Furthermore, it is very challenging to deploy

on-board services in the lifetime of cars due to com-

patibility issues and the additional technical security

required.

A more ﬂexible way is established by transmit-

ting the data from the car to a back-end system and

running the data-driven system off-board. As soon

as the required data is available on a bus system, this

data is routed by the gateway to the bus where the

sending unit (telematic) is located and it transfers the

data from the car to a back-end system. The back-end

computes the results and transmits it to the involved

systems of the stakeholders. In this regard, the inter-

net interface (telematic) of the car needs to be enabled

to transmit different data package sizes in order to

provide efﬁciently the required data (Arena and Pau,

2019). One crucial baseline is, that the bandwidth of

the mobile network allows such data transfers.

As mentioned before, the deployment of a service

on-board (embedded) is complex, time consuming

and not as ﬂexible as a data-driven service is meant

to be. Some services need to run on embedded sys-

tems like autonomous driving (Liu et al., 2017). The

functionality of most of the other services allows run-

ning outside of the car like in case of the damage pre-

diction system. Off-board car services provide, next

to their simpler deployment, more ﬂexibility regard-

ing faster model updates and adaptations. The key for

off-board services is the availability of data: The car

has to be able to send the required data on demand,

according to GDPR (General Data Protection Regu-

lation) and any applicable regional law and after the

conﬁrmation of the customer, i.e., the electronic ar-

chitecture of the vehicle must be enabled to provide

the requested data. This is the foundation of a digital

ecosystem, which can collect demanded data and al-

lows deploying new services and interactions with the

car and other involved parties quickly.

Fig. 4 shows the basic principle of a vehicle

ecosystem for data-driven services. When develop-

ing a new service, the requested data is sent via mo-

bile network from the car to a data layer, called ser-

vice data, of a protected service cloud (security layer).

This service cloud is a part of the whole vehicle back-

end system. Next to the service data, the back-end

system receives external data from third parties like

trafﬁc, weather and other service important informa-

tion. Furthermore, the back-end contains car infor-

mation like, e.g., the car type, the equipment of the

vehicle, the drive technology, geometry information,

service information. Such information are often very

valuable for a data-driven service. Hence, using these

additional data can increase the prediction perfor-

mances. The data-driven model (AI-unit) is deployed

to the back-end system, as well as other units (ﬂex-

ible units) like data processing. This back-end sys-

tem communicates on demand with the car. Beyond

that, the architecture of the ecosystem allows train-

ing the models with newly collected data from time to

time or automatic (see Fig. 2). Such a digital ecosys-

tem gives even the possibility to provide partly ac-

cesses to third parties to create new valuable services,

e.g. (BMW GROUP, 2018). Generated customer in-

formation from the service cloud are provided to cus-

tomer devices in the car (control panel) or outside the

car, e.g. mobile applications.

When considering an example like the damage

prediction system, this would in concrete terms imply

VEHITS 2020 - 6th International Conference on Vehicle Technology and Intelligent Transport Systems

Figure 4: A digital ecosystem to run data-driven service.

Note that this illustration follows partly the methodology of

the BCG analysis (Kilian et al., 2020).

the following sequence of events: after a low speed

crash event and a conﬁrmation of the customer, a

small data package is transmitted to the data-driven

prediction model in the back-end. With the data

package as input the model predicts, based on his-

torical data, the damaged parts and the repair costs.

This information can be provided to participants like,

e.g., the customer, the insurance companies for seam-

less claim settlements or the workshops for faster re-

pair (Koch et al., 2018).

In all cases, before any data is transmitted, the

customer has to conﬁrm the certain service with an

overview of the transmitted data. Furthermore, a gen-

eral transparency of the service and its intuitiveness in

understanding and handling must be provided within

seconds to the customer. A conﬁrmation can be can-

celed anytime. In this context, recent studies show

that 94 % of connected car owners are interested in

apps and services. Out of those 94 %, 84 % are will-

ing to share personal automotive data for new services

(Otonomo, 2018).

An ecosystem with seamlessly operating data-

driven services requires data exchange from the

ecosystem not only to the car but also to other stake-

holders/participants. These processes need to be de-

signed in regards to the business processes. This is

described below.

4.5 Process

Running a service requires an interconnection of all

stakeholders/participants. Without data transfer to all

involved parties the potential of the service can not

be exploited. Therefore, shortly after having a rough

idea about the deployment, the business process needs

to be designed with taking all necessary stakeholders

into account.

When looking at the example of the damage pre-

diction system, the data of the damaged parts and the

cost for repair are computed in the back-end system.

It can be beneﬁcial for the customer to send certain

information to other participants like the workshop to

order the parts immediately and to prepare the work-

shop visit. Beyond that, with detailed damage infor-

mation the insurance company could approve the re-

pair immediately, which simpliﬁes the whole insur-

ance claim settlement and would avoid an interaction

of customer and insurance company. Such connec-

tions are identiﬁed and designed in the process phase.

Furthermore, business architects establish customer

oriented processes for running the data-driven system

with all necessary parties connected. Often, the whole

potential can be only reached when all parties are con-

nected in a beneﬁcial way.

4.6 Finalization

Fig. 1 indicates that the ﬁnalization phase starts ap-

proximately half-way of the process phase. More pre-

cisely, when having ﬁrst working systems, the ﬁnal-

ization phase starts with testing, validating and im-

proving the service. In most cases, it is indispensable

to test the service with, e.g. deﬁned customer groups

to use this feedback for further improvements. In this

phase it is key to consider and connect the ﬁve pre-

vious phases seamlessly with each other in order to

create a customer experience.

5 CONCLUSIONS AND

OUTLOOK

Nowadays, the expectations regarding data-driven

business models in the car industry are massive. This

paper illustrates a track towards an efﬁcient develop-

ment and deployment approach for data-driven ser-

vices in vehicles. It presents the important steps, as

well as the main challenges. Through the explanation

of the methodology, examples are drawn to show pre-

cisely the key points. A ﬂexible and various service

generation can be reached with a full integration of

vehicles in a digital ecosystem, which means that the

car delivers data according to GDPR (General Data

Protection Regulation) and any applicable regional

law and in cooperation with the customer to a back-

end system. The main service runs on this back-end

system, processes the data and transfers the results to

the participants like the customer or other involved

parties like, e.g., ﬂeet operators. The shown method

enables generating fast data-driven services in order

to integrate cars more seamlessly into our lives.

Towards Data-driven Services in Vehicles

As an outlook, we mention that data enables much

more than creating data-driven services: data is trans-

forming car manufacturers from traditional engineer-

ing companies to data-driven companies. This indi-

cates that not only service creation but also car devel-

opment in general is progressively based on data.

REFERENCES

Aeberhard, M., K

uhbeck, T., and et al. (Retrieved May

6, 2019). Automated driving with ros at bmw.

https://roscon.ros.org/2015/presentations/ROSCon-

Automated-Driving.pdf.

Ardelt, M., Coester, C., and Kaempchen, N. (2012). Highly

automated driving on freeways in real trafﬁc using a

probabilistic framework. IEEE Transactions on Intel-

ligent Transportation Systems, 13(4):1576–1585.

Arena, F. and Pau, G. (2019). An overview of vehicular

communications. Future Internet, 11(2).

BMW GROUP (Retrieved November 19, 2018). BMW

group launches BMW cardata: new and innova-

tive services for customers, safely and transpar-

ently. https://www.press.bmwgroup.com/global/

article/detail/T0271366EN/bmw-group-launches-

bmw-cardata:-new-and-innovative-services-for-

customers-safely-and-transparently?language=en.

Christ, M., Braun, N., Neuffer, J., and Kempa-Liehr, A. W.

(2018). Time series feature extraction on basis of scal-

able hypothesis tests (tsfresh a python package). Neu-

rocomputing, 307:72 – 77.

Geron, A. (2017). ”Hands-on machine learning with

scikit-learn & tensorﬂow”. O’Reilly Media, Inc., Se-

bastopol, CA, USA.

Glattes, K. (2016). ”Der Konkurrenz ein Kundenerleb-

nis voraus: Customer Experience Management –

111 Tipps zu Touchpoints, die Kunden begeistern”.

Springer.

Guyon, I., Gunn, S., Nikravesh, M., and Zadeh, L. (2006).

”Feature Extraction: Foundations and Applications”.

Studies in Fuzziness and Soft Computing. Springer.

Guyon, I., Gunn, S., Nikravesh, M., and Zadeh, L. (2008).

”Feature Extraction: Foundations and Applications”.

Studies in Fuzziness and Soft Computing. Springer.

Huber, W., Steinle, J., and Marquardt, M. (2008). Der

Fahrer steht im Mittelpunkt - Fahrerassistenz Aktive

Sicherheit bei der BMW Group. Integrierte Sicher-

heit und Fahrerassistenzsysteme, 2048:123–138.

Immonen, A., Ovaska, E., Kalaoja, J., and Pakkala, D.

(2016). A service requirements engineering method

for a digital service ecosystem. Service Oriented

Computing and Applications, 10(2):151–172.

Immonen, A., Ovaska, E., and Paaso, T. (2018). Towards

certiﬁed open data in digital service ecosystems. Soft-

ware Quality Journal, 26(4):1257–1297.

Kaiser, C., Stocker, A., and Viscusi, G. (2017). Digi-

tal vehicle ecosystems and new business models: An

overview of digitalization perspectives. In Platform

Economy & Business Models workshop at i-KNOW

2017.

Kilian, R., Gauer, C., Stein, J., and Scherer, M. (Retrieved

March 7, 2020). Connected vehicles and the road

to revenue. https://image-src.bcg.com/Images/BCG-

Connected-Vehicles-and-the-Road-to-Revenue-Dec-

2017 tcm9-179631.pdf.

Koch, M. and B

ack, T. (2018). Machine learning for

predicting the impact point of a low speed vehicle

crash. In 2018 17th IEEE International Conference on

Machine Learning and Applications (ICMLA), pages

1432–1437.

Koch, M., Wang, H., and B

ack, T. (2018). Machine learn-

ing for predicting the damaged parts of a low speed

vehicle crash. In Proceedings of Thirteenth Inter-

national Conference on Digital Information Manage-

ment (ICDIM 2018), pages 179–184.

Liu, S., Li, L., Tang, J., Wu, S., and Gaudiot, J. L. (2017).

”Creating Autonomous Vehicle Systems”. Synthesis

Lectures on Computer Science. Morgan & Claypool

Publishers.

Matheus, K. and K

onigseder, T. (2015). Automotive Ether-

net. Cambridge University Press.

Nilsson, N. J. (1969). A mobile automaton: An applica-

tion of artiﬁcial intelligence techniques. In 1st Inter-

national Conference on Template Production, pages

509–520.

Otonomo (Retrieved November 19, 2018). Are con-

sumers willing to share connected car data?

https://otonomo.io/are-consumers-willing-to-share-

connected-car-data/.

Robert Bosch GmbH (2014). ”Bosch Automotive Electrics

and Automotive Electronics: Systems and Compo-

nents, Networking and Hybrid Drive”. Springer,

Wiesbaden, 5th edition.

Schmidhuber, J. (Retrieved November 30, 2018). Prof.

Schmidhuber’s highlights of robot car history.

http://people.idsia.ch/∼juergen/robotcars.html.

Seiberth, G. and Gruendinger, W. (2018). Data-driven busi-

ness models in connected cars, mobility services &

beyond. In BVDW Research, volume 01.

Shearer, C. (2000). The crisp-dm model: the new blueprint

for data mining. J Data Warehouse, 5:13–22.

Spurgeon, C. (2000). ”Ethernet: The Deﬁnitive Guide”.

Deﬁnitive Guide Series. O’Reilly.

Spurgeon, C. and Zimmerman, J. (2013). ”Ethernet

Switches: An Introduction to Network Design with

Switches”. O’Reilly Media.

Stone, P., Brooks, R., and et al. (Retrieved Novem-

ber 5, 2018). Artiﬁcial intelligence and life in

2030. one hundred year study on artiﬁcial intel-

ligence: Report of the 2015-2016 study panel.

http://ai100.stanford.edu/2016-report.

Tsugawa, S., Yatabe, T., and et al. (1979). An automo-

bile with artiﬁcial intelligence. In Proceedings of the

6th International Joint Conference on Artiﬁcial Intel-

ligence - Volume 2, pages 893–895, San Francisco,

CA, USA. Morgan Kaufmann Publishers Inc.

Weill, P. and Woerner, S. L. (2014). Thriving in an increas-

ingly digital ecosystem. MIT sloan management re-

view, 56(4):27–34.

VEHITS 2020 - 6th International Conference on Vehicle Technology and Intelligent Transport Systems