Analyzing Driver Behavior Compliance under HoS Regulations

∗

Ignacio Vellido-Exp

osito, Juan Fern

andez-Olivares

, Ra

ul P

erez

and Luis Castillo

Department of Computer Science and AI, University of Granada, Andalusian Research Insitute in Data Science and

Computational Intelligence, Granada, Spain

Keywords:

Fleet Telematics, Regulation Compliance, Artiﬁcial Intelligence, Automated Planning.

Abstract:

World wide spreaded Hours of Service (HoS) regulations constraint drivers’ amount of working and driving

time without resting. Transport companies are extremely interested on interpreting what their drivers are

doing, based on the raw information of event logs generated by ﬂeets’ onboard devices, considering the terms

deﬁned by HoS regulations. This work addresses the problem of analyzing the compliance of a driver wrt

HoS, by using AI techniques to classify and label the information of a driver’s log according to HoS terms.

The ﬁnal result is a human-interpretable descriptive model of driver behaviour that leverages the company

situational awareness, empowering staff responsible of operations to make better informed decisions.

1 INTRODUCTION

A problem of paramount importance in transport

companies is to determine whether driver activity

conforms with Hours of Service (HoS) regulation to

forestall illegal behavior and avoid costs due to sanc-

tions. HoS regulation are imposed by world wide

transport authorities (Meyer, 2011; Goel and Vidal,

2013), which constraint the amount of drivers’ work-

ing, driving and resting time when delivering a trans-

port service. As a consequence, scheduled driving

plans have to be aligned with laws that deﬁne the legal

behavior of a driver.

The widespread adoption of onboard IoT devices

in vehicle ﬂeets enables recording the activity of

drivers in event logs. Moreover, transport companies

are encouraged (currently almost forced by law) to

endow every transport asset with electronic sensors

(i.e. tachographs) which record the different activ-

ities a driver performs when delivering a transport

service. This context enables driver activity to be

recorded in event logs and enhances the need for driv-

ing work plans to be monitored and analyzed either by

companies’ decision makers or authorities to analyze

drivers’ legal compliance. Since raw event logs gen-

https://orcid.org/0000-0002-7391-882X

https://orcid.org/0000-0002-1355-1122

https://orcid.org/0000-0002-4910-7752

∗

This work is being partially funded by the Andalusian

Regional Project B-TIC-668-UGR20 and the Spanish Na-

tional Project RTI2018-098460-B-I00 with FEDER funds.

erated by onboard ﬂeet devices are difﬁcult to be di-

rectly interpreted, an important technical challenge is

to come up with easily interpretable descriptive mod-

els that help understand the huge amount of informa-

tion stored in such event logs.

In response to this need, the paper presents a

method that starting from a raw, real event log ex-

tracted from a tachograph, 1) classiﬁes and tags activ-

ities, and 2) generates a human-interpretable descrip-

tive model that consists of an extension of the original

raw log with additional attributes that, after the clas-

siﬁcation and labeling process, characterize each and

every activity in the log.

The characterization provides a semantics to the

raw log by extending it with additional information

about aspects such as the type of activity according

to HoS regulation, the type of HoS sequence/subse-

quence the activity belongs to, as well as whether the

activity (and hence the sequence which it is involved

in) is compliant with the HoS regulation. The method

is validated with an experimentation over real data

and by a proof of concept.

This method provides several advantages. On the

one hand, operational staff can determine at a glance

information about the legality of driver’s behavior,

without the need to resort on a burdensome inspec-

tion of the raw log. On the other hand, the extended

log (descriptive model) incorporates information with

sufﬁcient granularity to interpret what each of the

drivers is doing considering HoS regulations. In sum-

mary, the labeled event log can be easily interpreted

Vellido-Expósito, I., Fernández-Olivares, J., Pérez, R. and Castillo, L.

Analyzing Driver Behavior Compliance under HoS Regulations.

DOI: 10.5220/0011116200003191

In Proceedings of the 8th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2022), pages 463-470

ISBN: 978-989-758-573-9; ISSN: 2184-495X

463

Figure 1: A summary of HoS concepts, duration and additional constraints. We make use of ”<” and ”<=” symbols to

indicate whether the allowed duration is represented by either an open or closed interval, respectively.

Figure 2: Hierarchical relations and structure of the differ-

ent types of sub-sequences, extracted from (Meyer, 2011).

by company experts which can make more informed

decisions considering either the historical or current

labeled situation of a driver.

The remainder of this paper shows, ﬁrstly, a de-

tailed description of the novel problem addressed and

the main contribution of the approach (Section 2).

Then, we outline the overall methodology (Section

3), and explain some background concepts needed to

ease the overall description of the approach (Section

4). The technical details are shown in Sections 5 and

6. We end with an experimentation conducted over a

proof of concept of the application (Section 7), and

discuss related and future work (Sections 8 and 9).

2 HoS REGULATION

Transport companies are interested on making deci-

sions to govern the behavior of their drivers by pre-

dicting whether a driver is close to committing a regu-

lation violation, but previous to this they have to face

the challenge of characterizing drivers according to

their driving style with respect to the HoS regulation.

With the incorporation of IoT onboard devices, com-

panies are forced to deal with large data volumes of

daily event logs, with thousands of events, that need to

be interpreted with respect to HoS regulations. There-

fore, there is a clear need for an expert to directly in-

terpret what a driver has been doing during the period

of time encompassed by a given event log.

Concretely, every event in an activity log is a tu-

ple (a, start, end, dur, d id), where each component

refers to: activity identiﬁer (a), event start (start)

and end (end) times, event duration (dur), and driver

identiﬁer (d id), respectively. A value for a is any

of the labels [Driving, Other, Pause, Idle] meaning

that the driver is either Driving, performing Another

Work, at Pause or Idle during dur minutes, between

start and end. Our goal is to provide semantics to

each event, by completing this raw information with

additional attributes whose values are the terms pro-

vided by the HoS regulation, which are detailed in the

following.

This regulation has been extensively analyzed in

(Goel and Vidal, 2013; Meyer, 2011). It is applied in

several countries, but in this work we focus on the Eu-

ropean Union regulation (EC) No 561/2006 (Meyer,

2011). The basic terms refer to four types of driver

activities as break (short period for recuperation), rest

(period at driver free disposal with enough time to

sleep), driving (time during which the driver is op-

erating a vehicle) and other work (time devoted to

any work except driving, like loading). These ac-

tivities do not exactly correspond to the event labels

above deﬁned, but they are deﬁned according to such

events and their attributes. That is to say, the regula-

tion deﬁnes different types of rest and break periods

(see Figure 1, left-hand table) according to the dura-

tion of a Pause event. Moreover, there are different

types of driving sequences. A basic driving sequence

is composed of a totally ordered set of the elements of

[Driving, Pause, Other, Idle] constrained to the dura-

tion of any Pause is less than 15 minutes (see Figure 1,

right-hand table for more types). More constraints are

deﬁned over the duration of the rests and breaks, and

over the accumulated duration of driving sequences

(see both tables in Figure 1). The regulation provides

a set of basic and optional rules, should the former

not be satisﬁed, thus allowing more ﬂexibility to gen-

erate and interpret driving schedules under such con-

straints. For example, either a break of 45 mins has to

VEHITS 2022 - 8th International Conference on Vehicle Technology and Intelligent Transport Systems

464

be taken after 4.5 hours of accumulated driving or the

break can be taken splitted in two parts of at least 15

mins and 30 mins respectively. This feature is good

for drivers since it provides ﬂexibility to their work,

but complicates the interpretability of what they are

doing. Regulations also deﬁne additional constraints

(for example, the maximum number of occurrences of

a reduced rest in a weekly driving period), or the hi-

erarchical relationship between the different types of

sub-sequences, as well as their internal structure (see

Figure 2).

2.1 The Problem

In summary, the problem consists of classifying and

labeling events of a given log by considering the legal

terms above described about HoS regulations. Con-

cretely, the observed behavior has to be tagged in

terms of driving periods, daily driving periods (and

their types), and weekly driving periods (and their

types) as shown in Figures 1 (right-hand table) and

2. Last but not least, identiﬁed subsequences have to

be tagged as legal or illegal according to their compli-

ance with HoS regulation.

The main contribution of this work consists of ad-

dressing the above described labeling of activities and

classiﬁcation of illegal sequences as a parsing process

based on automated planning (AP) techniques, as well

as an experimentation that validates our approach. AP

is a widely known ﬁeld of Artiﬁcial Intelligence de-

voted to develop algorithms, called planners, that im-

plement search-based problem solving processes to

determine a plan (sequence of actions) for a situated

agent to achieve a goal. The inputs provided to a plan-

ner are a model of actions that describes the behavior

of the agent, called a planning domain, and, a plan-

ning problem, a description of the current state of the

agent and the goal to be achieved.

The main use of these techniques is the generation

of plans, but recent works have shown their usefulness

to recognize sequences of actions. This last use lays

on the fact that when a planner is provided with an

initial state that incorporates a sequence of observa-

tions representing the past activities of an agent, and

a repertory of possible goals the agent is intended to

achieve, the planning process can be used to deter-

mine which of the possible goals is intended to be

achieved by the agent executing the observed actions.

This view of automated planning is called plan recog-

nition and it is the one followed by our approach.

Concretely we are using in this work temporal hi-

erarchical task network planning techniques (tempo-

ral HTN), a variant of AP where the planning do-

main is represented as a hierarchical knowledge base

in which the behavior of the agent is deﬁned in terms

of a compositional hierarchy of tasks/subtasks where

primitive tasks represent temporally annotated actions

to be executed by the agent, and compound tasks rep-

resent temporal ordering strategies of actions. The

key aspect in HTN is that compound task can be de-

composed in alternative ways by means of methods.

A method is a pair (t, d) that describes one way to

achieve the task t is to perform the tasks speciﬁed

in the task network d, a temporally constrained set

of tasks/subtasks representing things that need to be

done. The hierarchical planning process is based on

a search process that starts at a given top-level task

network, and recursively reduces it by opportunisti-

cally applying methods. The current task network is

transformed step by step in each reduction by insert-

ing tasks according to the order and temporal con-

straints speciﬁed in the methods, until it consists of

a plan only composed of primitive actions.

The reason to use a temporal and hierarchical

planning approach is two-fold: on the one hand, the

event log observations are temporally annotated and

constrained, hence we need to resort on a technique

able to manage such temporal constraints. On the

other hand, the HoS regulation can be naturally rep-

resented as a hierarchy of temporal tasks.

In our case a given event log represents a se-

quence of temporally annotated observations and the

planning process 1) identiﬁes different temporal sub-

sequences of a driver’s daily and weekly driving ac-

tivity, 2) labels them according to the terms deﬁned by

HoS regulation, and 3) classiﬁes them as legal/illegal.

It is important to note that the identiﬁcation of tempo-

ral sub-sequences is a process analogous to grammar

parsing, and in our case is addressed as a knowledge-

driven search process with two main features: i) it is

implemented in a hierarchical planner, ii) the planner

is provided with a knowledge-based model represent-

ing the HoS regulation. The parsing task is performed

over the entire event log considering that the decom-

position methods provided in the HTN model can be

seen as a set of production rules of an attribute gram-

mar.

In the following sections we describe the gen-

eral steps of our method, provide a description of

the knowledge-based representation of HoS regula-

tion, then we brieﬂy describe the knowledge-driven

search process to parse the log, and the experimenta-

tion and proof of concept carried out to validate our

approach.

Analyzing Driver Behavior Compliance under HoS Regulations

465

Figure 3: A schema of the methodology of the approach.

3 OVERALL METHODOLOGY

The methodology we have followed to provide a so-

lution to the problem consists of the following steps

(see Figure 3):

1. Formalizing HoS rules as productions of an at-

tribute grammar.

2. Transforming the raw event log into a sequence of

temporal observations.

3. Parsing the sequence of temporal observations by

using the attribute grammar representing the HoS

regulation.

4. Generating a human-interpretable descriptive

model that consists of an extension of the origi-

nal event log with additional attributes

The following sections describe in detail these

steps.

4 HoS AS AN ATTRIBUTE

GRAMMAR

We have used attribute grammars (Knuth, 1968) as an

intermediate representation to formalize the descrip-

tion of the HoS regulation in Europe. The main reason

is that attribute grammars, apart from being an excel-

lent mechanism to describe context-sensitive gram-

mars, allow to easily represent the required temporal

and numeric constraints with grammar attributes. An

attribute grammar extends productions with seman-

tic rules

of the form {hconditionihassignmentsi},

where hconditioni deﬁnes the applicability conditions

of the production and hassignmentsi deﬁne how syn-

thesized attributes of parent nonterminals (at left-hand

side of the production) are calculated from attributes

of the right-hand side of the production. In the fol-

lowing we describe the production rules to recognize

basic breaks, rests and basic sequences (as described

previously).

We are using an ad hoc syntax for attribute grammar

in order to ease the explanation.

4.1 Recognizing Breaks, Rests and

Basic Sequences

The terminals of the grammar are the labels

[D, O, P, I], since they are used to distinguish the types

of observations which may be found in an event log

(Driving, Other, Pause, Idle). Terminals have asso-

ciated attributes such as duration that are represented

following the ”dot” notation. For example P.dur rep-

resents the duration of a concrete, recognized Pause

event in the event log. On the other hand, every non-

terminal has associated two attributes that represent

the driving time (attribute dt) and pausing time (at-

tribute rt) of the current recognized token. For exam-

ple, b

t1.rt represents the pausing time of the token

b t1 which in turn represents a Break of Type 1 (see

BREAK T1 in Figure 1).

Breaks Recognition. The pausing time of any pause

event in the log is further used to discriminate be-

tween different types of Breaks or Rests, according

to rules in Figure 1 (left-hand table). For example, a

production that recognizes a Break of Type 1, is rep-

resented as follows

b_ t1 : P {( a nd ( P . dur >= 45 min )( P. dur < 3 h rs )

( b _t 1 . rt := P. du r )

describing that a Break of Type 1 corresponds to a

pause greater than 45 minutes and lesser than 3 hours,

and that the duration of such pause event P.dur is as-

signed as the duration of the break b t1.rt . Other

types of breaks are represented in a similar way with

grammar productions.

Rests Recognition. Daily and weekly rests, repre-

sented by the non-terminal symbols rd and rw, may

be normal or reduced, and are deﬁned according to

the duration of pauses described in Figure 1 (left-hand

table). For example, the following productions repre-

sent this aspect of the regulation

rd : rd _no r mal { r d . rt := rd _no r mal . rt };

rd : rd _ red u ced { rd . rt := rd _ red u c ed . rt };

rd_ n orm al : P {( and (P. du r >= 11 h ) (P . dur < 24 h )

( r d _no r mal . rt := P . dur ) } ;

VEHITS 2022 - 8th International Conference on Vehicle Technology and Intelligent Transport Systems

466

The ﬁrst production describes the legal types of

daily rests, while the second represents that a normal

daily rest is a pause with a duration between 11 and

24 hours.

Basic Sequences Recognition. Figure 1 (right-hand

table) shows that a basic sequence is composed of

any number of activities such that the duration of any

Pause is strictly less than 15 mins. Firstly, we need

to deﬁne the legal activities allowed, which is repre-

sented by the non-terminal symbol elt and the fol-

lowing productions:

el t : D {( el t . dt := D . dur ; e lt . rt := 0)} ;

el t : P {( P . dur < 15 mins ) ;

( elt . dt := 0; elt . rt : = P. dur ) };

el t : O {( el t . dt := elt . dt := 0)};

el t : I {( el t . dt := elt . dt := 0)};

Then we can make use of the following recursive

production to deﬁne what a correct basic sequence

is. Moreover, it deﬁnes that the driving time of a se-

quence is the accumulated driving time of its compo-

nents. Note that in attribute grammars occurrences of

the same symbol are differentiated by subscripts to be

correctly identiﬁed.

ba seq : el t {( ba seq . dt := elt . dt )};

ba seq : el t bas eq

{( bas eq . dt := e lt . dt + bas eq

. dt )} ;

Finally the production rules to recognize the re-

maining HoS concepts described in Figure 1 are rep-

resented following the same principles illustrated pre-

viously.

5 GENERATING TEMPORAL

OBSERVATIONS

Every record of the event log is translated into

an observation (a, i, type, start, end, dur, d id)

where i is an index representing the order in the

sequence of events, a is a unique identiﬁer for

the action, type indicates the type of action (one

of [D, O, P, I] corresponding to the actions labels

[Driving, Other, Pause, Idle]), d id is the identiﬁer of

the driver, and start, end, dur represent the start and

end times of the action and its duration, respectively.

For example the event representing a Pause of 68 min-

utes from 15:46 to 16:54 on 05/01/2017 would be

represented as follows:

Event:

( P aus e " 0 5/0 1 /20 1 7 15 :46 "

"05 / 01/ 2 017 16 : 54 " 68 driv e r1 )

Observation:

( P15 39 P "0 5 / 01/ 2 017 15 : 46 "

"05 / 01/ 2 017 16 : 54 " 68 d r ive r1 )

6 PARSING PROCESS

Algorithm 1: Algorithm for classifying and labeling

based on event log parsing.

Input: G (productions), S (observations)

1 A ← [top level symbol] /*the agenda*/;

2 Π ← [] /*empty plan*/;

3 i ← 1 /*index of current event*/;

4 current event ← S [i];

5 Initialize contexts (Week, Day, DayType,

SeqOrder, SeqType, Token, Legal);

6 while A not empty do

7 extract the ﬁrst element t from A;

8 if t is TERMINAL then

9 if type of action(current event) is in [D O

P I]) then

10 if not Check temporal constraints()

then

11 Insert event in Π as ILEGAL

temporal action

12 else

13 Insert event in Π as temporal

action;

14 Label the action with contexts;

15 current event ← S [i++];

16 else

17 Insert in Π event as ILEGAL

18 if t is NONTERMINAL then

19 LEGAL ← Y ES;

20 Update contexts considering t;

21 if t is UPDATE

then

22 update t.att according to t.successor

23 else

24 Choose from G one production t:rh

not yet processed (rh

={t

.. t

});

25 if All productions already processed

then

26 LEGAL ← NO

27 else

28 t.successor ← rh

;

29 A ← A∪ {t

.. t

UPDATE

}

30 return Π with the labeled and classiﬁed actions

The algorithm of the parsing process is shown in Al-

gorithm 1. The algorithm receives as input the set

of productions representing the HoS regulation (the

grammar G) and the set of temporal observations S

(generated as above explained), and returns a plan Π

that contains the temporal observations transformed

in actions which are labeled with appropriate values

for the set of extended attributes [week day DayType

SeqOrder SeqType Token Legal] whose semantics is

described in subsections below. The ﬁrst steps of the

algorithm consist of initializing the agenda A, that is a

list containing the pending productions (or compound

tasks) to be processed, the plan Π that is incremen-

Analyzing Driver Behavior Compliance under HoS Regulations

467

tally generated as events are processed, the index i

that refers to the current event to be processed in the

log, and the extended attributes.

The main loop of the algorithm explores the set of

productions of the grammar G trying to match gram-

mar productions with the current event of the log (in-

dexed by the variable i). If for the current event to be

processed a matching with a terminal of a legal pro-

duction is found, the event is transformed in an action,

added to the plan Π (if the temporal constraints of

the temporal observation are consistent with the cur-

rent plan), and labeled accordingly. This parsing pro-

cess is carried out over the entire set of observations

in S. The main steps are described in the following

sections.

6.1 Event Recognition

Each type of event [Driving, Other, Pause, Idle] has

an associated primitive temporal action htypei p in the

domain, that is instantiated according to event infor-

mation and added to the plan Π when the parsing has

processed an event of that type. Therefore, recogniz-

ing a concrete event that is at any given position i with

label type consists of adding its corresponding primi-

tive action to Π such that (i) the temporal points of the

action are consistent with the temporal information of

the event, and (ii) guaranteeing that the temporal con-

straints of a are consistent with the rest of temporal

constraints of the actions already added to the plan.

These conditions are checked by the planning process

itself.

The primitive tasks represent the ”token” (for ex-

ample, a Pause) that has been read, and the algorithm

describes in essence in Line 13 that a primitive task

(for example of type Pause) has to be added to the

plan Π when there is an observation of the same type

at the reading pointer position. Interestingly, the abil-

ity to represent temporal constraints on tasks at any

level of the hierarchy makes possible to represent that

the start and end points of the primitive action are

constrained by the temporal information of the ob-

servation (described in Section 5). This is the point

where temporal planning capabilities play a central

role since the algorithm will fail to classify the ac-

tion as ILEGAL (Line 11) if the temporal constraints

cannot be met. If the primitive task is successfully

inserted, the action is classiﬁed as LEGAL, and the

index of current event to be processed moves to the

next position. The labeling of actions performed in

Line 14 is explained in the following.

6.2 Labeling of Events

The labeling process accounts for the different time

resolution contexts, deﬁned by HoS regulation, which

an event may belong to. Because of that, every event

in the event log is annotated with six labels

cor-

responding with six additional attributes with pro-

vide semantics to every event according to the fol-

lowing contexts: Week (integer values), Day (inte-

ger values) DayType (with possible values ndd or

edd, corresponding to the concepts of normal daily

driving and extended daily driving shown Figure 1),

SeqOrder (representing the order of appearance of a

sub-sequence as first, second or third), SeqType

(representing if an action belongs to a sub-sequence

of type splitted or continuous see Figure 1), token

( representing if the action is a break or a driving

action with possible values according to the types

of breaks in Figure 1 with the possible values b t1,

b t2, b t3, etc.). Finally, the last attribute Legal

describes if the activity has been classiﬁed by the

event recognition process as legal or illegal.

We have made extensive use of the features of the

attributes of every production in order to label records

of the event log. On the one hand, in Line 28 the set

of symbols for every non-terminal t is extended with

the special symbol {UPDATE

} for the algorithm to

be able to appropriately update the attributes of non-

terminals (Line 20). Moreover, a sequence is identi-

ﬁed as ILEGAL (Line 26) when all the productions

for its non-terminal symbol have been processed and

no UPDATE symbol has been reached. On the other

hand, the set of attributes of every non-terminal is ex-

tended with as many attributes as described above.

7 EXPERIMENTATION

We have validated our methodology with an exper-

imentation using real tachograph logs provided by

an industrial collaborator. We were provided with a

dataset formed by two-weeks-long sequences of ac-

tivities from 290 different drivers.

Our experimentation consisted of splitting the

dataset by driver, transforming each associated log

into a set of observations. Additionally the algorithm

has been implemented using an off-the-shelf HTN

planner. The planner is provided with an HTN do-

main that has been written by translating the produc-

tions of the grammar representing the HoS regulation,

and setting the goal as recognising each action while

outputting the activities with the appropriate contexts.

We are assuming six labels and a reduced set of possi-

ble values for each label to ease the explanation.

VEHITS 2022 - 8th International Conference on Vehicle Technology and Intelligent Transport Systems

468

Table 1: Labelling output after processing a daily sequence. This example shows an extended driving period where the ﬁrst

and second sequences contains split rests and the third is taken uninterrupted. The value ”A” in the Token column represents

any activity other than a Pause.

Original Log Annotated Labels

Driver Start End Duration Activity Week Day DayType Sequence BreakType Token Legal

driver3 12/01/2017 09:55 12/01/2017 09:57 2 Driving

3 6 edd

ﬁrst

split 1

A yes

driver3 12/01/2017 09:57 12/01/2017 10:41 44 Break B T2 yes

driver3 12/01/2017 10:41 12/01/2017 10:43 2 Driving

split 2

A yes

driver3 12/01/2017 10:43 12/01/2017 11:57 74 Break B T3 yes

driver3 12/01/2017 11:57 12/01/2017 12:00 3 Driving

second

split 1

A yes

driver3 12/01/2017 12:00 12/01/2017 12:06 6 Break B T0 yes

driver3 12/01/2017 12:06 12/01/2017 12:08 2 Driving A yes

driver3 12/01/2017 12:08 12/01/2017 12:36 28 Break B T2 yes

driver3 12/01/2017 12:36 12/01/2017 13:07 31 Driving

split 2

A yes

driver3 12/01/2017 13:07 12/01/2017 13:13 6 Break B T0 yes

driver3 12/01/2017 13:13 12/01/2017 15:31 138 Driving A yes

driver3 12/01/2017 15:31 12/01/2017 15:39 8 Break B T0 yes

driver3 12/01/2017 15:39 12/01/2017 15:59 20 Driving A yes

driver3 12/01/2017 15:59 12/01/2017 18:00 121 Break B T3 yes

driver3 12/01/2017 18:00 12/01/2017 18:54 54 Driving

third uninterrupted

A yes

driver3 12/01/2017 18:54 12/01/2017 19:08 14 Break B T0 yes

driver3 12/01/2017 19:08 12/01/2017 19:09 1 Driving A yes

driver3 12/01/2017 19:09 12/01/2017 19:14 5 Other A yes

driver3 12/01/2017 19:14 13/01/2017 05:16 602 Break DR T2 yes

Once the system returned the labelled log, we selected

multiple sequences at random, both legal and illegal,

and manually veriﬁed that the output was the appro-

priate under the HoS regulation.

The additional labels provide an exact summari-

sation of the driver behaviour according to the HoS

regulation, resulting in a human interpretable descrip-

tive model, much more readable by experts and con-

sequently enabling business decisions based of tacho-

graph data, like assigning drivers to routes according

to their driving styles or identifying problematic ten-

dencies. Furthermore, because the system is capable

of identifying illegal sequences, a driver could obtain

immediate feedback after committing an infraction,

and therefore forestall future action that could result

in sanctions for the company.

Table 1 shows an output example for a daily driv-

ing sequence. As we can see, the original data is not

easily interpretable, since they are formed by differ-

ent patterns of Driving, Other and Break actions of

distinct duration. However, thanks to the annotated la-

bels the output becomes more informative and under-

standable, giving an overview of the driver behaviour

in multiple levels of granularity. In our example, it be-

comes more explicit that the driver carried out an ex-

tended daily driving with three sequences, where the

break of the ﬁrst and second are taken in two parts,

and the break of the third sequence is taken uninter-

ruptedly at the end. Additionally, the last two columns

provide further information of how each action is con-

sidered under the regulation, and that none of them is

illegal.

Another example is shown in Table 2, display-

ing the response of our system against non legal se-

quences. As they do not ﬁt correctly under the HoS

regulation, some contexts cannot be recognised and

they are labelled accordingly. Nevertheless, because

there could be legal subsequences belonging to a big-

ger illegal sequence, the architecture tries to label as

many contexts as possible to provide that additional

information. This functionality produces more under-

standable logs simplifying the process of pinpointing

the cause of the infraction.

8 RELATED WORK

The problem of generating driving plans compliant

with HoS regulations have already been addressed in

(Goel, 2018) and (Goel and Irnich, 2017), and HTN

planning can be considered an enabling technology

for this problem. However, we are using this planning

technique from a different perspective: recognizing

a preexisting temporal plan. As previously shown, a

key issue is to formalize the HoS regulation, and an al-

ternative formalization can be found in (Goel, 2018),

which is aimed to use classical scheduling techniques

to check the compliance of schedules (only plan ver-

iﬁcation) or to generate compliant schedules (plan

generation), but not suitable for the problem of plan-

goal recognition here addressed.

This application can be, in part, viewed as an in-

stance of Runtime Veriﬁcation with Planning, where

Analyzing Driver Behavior Compliance under HoS Regulations

469

Table 2: Labelling output after encountering an illegal sequence. The system adds as many contexts as possible to ease the

interpretation of the sequence.

Original Log Annotated Labels

Driver Start End Duration Activity Week Day DayType Sequence BreakType Token Legal

driver1 11/01/2017 20:49 11/01/2017 20:50 1 Driving

2 10

unknown unknown

uninterrupted

A no

driver1 11/01/2017 20:50 11/01/2017 20:52 2 Other A no

driver1 11/01/2017 20:52 11/01/2017 23:06 134 Driving A no

driver1 11/01/2017 23:06 12/01/2017 00:00 54 Break B T1 no

driver1 12/01/2017 00:00 12/01/2017 03:00 180 Driving

unknown

A no

driver1 12/01/2017 03:00 12/01/2017 03:02 2 Other A no

driver1 12/01/2017 03:02 12/01/2017 03:03 1 Driving A no

driver1 12/01/2017 03:03 12/01/2017 03:06 3 Other A no

driver1 12/01/2017 03:06 12/01/2017 03:13 7 Driving A no

driver1 12/01/2017 03:13 12/01/2017 03:17 4 Other A no

driver1 12/01/2017 03:17 12/01/2017 06:30 193 Break unknown no

driver1 12/01/2017 06:30 12/01/2017 06:33 3 Driving

ndd unique uninterrupted

A yes

driver1 12/01/2017 06:33 14/01/2017 12:02 3209 Break WR T1 yes

regulations are expressed in a suitable formal lan-

guage, most often based on temporal logic, as de-

scribed in (Bensalem et al., 2014). This paper is in-

terestingly different, in that it encodes regulations in-

volving metric temporal constraints, expressed in an

attribute grammar. Our work is also related with con-

formance checking, in the area of Process Mining,

that consists of determining whether an executed pro-

cess conforms with a process model (Aalst, 2016).

Nevertheless, we do not only provide an answer to

whether a plan is or is not generated by a regulatory

model, but additionally our approach classiﬁes sub-

sequences of the input plan and recognizes/identiﬁes

them with concepts of the model, providing an inter-

pretation of the plan in such terms. Therefore, this

is a clear instance of a Plan-Goal Recognition (PGR)

problem that is traditionally addressed either by pro-

viding a plan library to represent the set of possible

plans to be recognized, or by representing the behav-

ior of the agent as a planning domain (also known as

PGR as Planning) (Ramırez and Geffner, 2009).

9 CONCLUSIONS

We have presented a method intended to provide sup-

port to experts on the task of interpreting what drivers

are or have been doing by recognizing their activity

recorded in an event log according to HoS regulations.

The main contribution is the proposal of a temporal

HTN-based approach to classify and label the activ-

ities in a real log where the plans observed are tem-

poral sequences of events, using an off-the-self tem-

poral HTN planner, with a domain generated from an

attribute grammar that describes de HoS regulation,

conﬁgured to recognize and classify the events ac-

cording to the goals represented as HTN tasks. The

novelty here is that the planning domain requires to

represent temporal and numerical information, and up

to authors’ knowledge there is no approach to which

ﬁts that requirements. Another contribution is the for-

malization of concepts and constraints of the HoS reg-

ulation as an attribute grammar, a necessary step in

the knowledge engineering process proposed to rep-

resent the HTN domain.

REFERENCES

Aalst, W. v. d. (2016). Process Mining: Data Science in

Action. Springer-Verlag, Berlin Heidelberg, 2 edition.

Bensalem, S., Havelund, K., and Orlandini, A. (2014). Ver-

iﬁcation and validation meet planning and scheduling.

International Journal on Software Tools for Technol-

ogy Transfer, 16(1):1–12.

Goel, A. (2018). Legal aspects in road transport optimiza-

tion in europe. Transportation research part E: logis-

tics and transportation review, 114:144–162.

Goel, A. and Irnich, S. (2017). An exact method for vehicle

routing and truck driver scheduling problems. Trans-

portation Science, 51(2):737–754.

Goel, A. and Vidal, T. (2013). Hours of service regu-

lations in road freight transport: An optimization-

based international assessment. Transportation sci-

ence, 48(3):391–412.

Knuth, D. E. (1968). Semantics of context-free languages.

Mathematical systems theory, 2(2):127–145.

Meyer, C. M. (2011). European Legislation on Driving and

Working Hours in Road Transportation. In Meyer,

C. M., editor, Vehicle Routing under Consideration of

Driving and Working Hours: A Distributed Decision

Making Perspective, pages 9–24. Gabler, Wiesbaden.

Ramırez, M. and Geffner, H. (2009). Plan recognition

as planning. In Proceedings of the 21st interna-

tional joint conference on Artiﬁcal intelligence. Mor-

gan Kaufmann Publishers Inc, pages 1778–1783.

VEHITS 2022 - 8th International Conference on Vehicle Technology and Intelligent Transport Systems

470