The Virtual Driving Coach - design and preliminary testing of a predictive eco-driving assistance system for heavy-duty vehicles

In the context of driving, Fastenmeier defines the term situation as the environment of the driver–vehicle’s human–machine system from the driver’s perspective whereby situations are delimitable units in space, time and their characteristics [17, p. 27f].

This concept is illustrated by an example: Exceeding 85 km/h permanently increases fuel consumption excessively while hardly saving time due to traffic effects [18, 19]. The situation where excessive velocity is not recommended can be described using the situational variables street type (highway), topography (even) and driver intention (not overtaking).

This ensures that a driving error is only triggered in obviously inefficient situations, whereas the driver is given the opportunity to fully exploit the properties of the vehicle in topographically challenging terrain or when overtaking. If all the situational variables meet the required properties, the situation “excessive speed” is active and driving advice can be given if the driver operates the vehicle in an inadequate way.

The situational variables are derived by a systematic approach [15, 16]. Starting point is a test drive over 400 km on a commonly used test track for commercial vehicles [20], which is divided by experts into situations relevant for eco-driving. By analyzing situations relevant for fuel consumption, situational variables are discovered and divided into discrete states. An algorithmic description of the situational variables is formulated and the measurement data of the test drive is analyzed. For every situation relevant for fuel consumption, all occurring combinations of the situational variables are obtained. These combinations of the situational variables eventually define the situations which are detected by the situation detection module.

For the analysis of the current driving situation, it suffices to determine the situational variables at the very moment in time. However, anticipatory situation detection has to predict the future deployment of the situational variables in order to ascertain when a situation starts.

The eco-driving assistance system supports the driver in finding the optimum point in time to decelerate the vehicle by releasing the accelerator pedal. For this purpose, the velocity of the coasting vehicle has to be predicted.

Huber [21] proposes to predict the vehicle speed for the road segments ahead cyclically with a model of the longitudinal dynamics and to use this to determine the coasting phases. Assuming the vehicle is operated in coasting mode, the velocity v of the vehicle can be calculated with a simplified longitudinal dynamics model [22, 23].

By means of the relationshipthe equation of the longitudinal dynamics (1) can be expressed asdepending on the distance x on the path in front of the vehicle.

If the predicted velocity v(x) exceeds a pre-calculated maximum velocity v _max(x) the vehicle will reach a critical velocity at which a brake intervention is likely even in coasting mode. Thus the maximum velocity v _max(x) is defined as the maximum velocity a driver will accept without braking, and is calculated as a function of infrastructure elements lying ahead, such as speed limits, curve radii, turning angles at crossings, and street types.

The mere reaction to the maximum velocity would indeed reduce the fuel consumption most effectively, but may slow the vehicle down unacceptably during the coasting phase. If, for example, the speed limitation is situated in a long downhill slope the vehicle passes the beginning of the downhill slope very slowly. This driving strategy would both increase the driving time and interfere with the traffic behind. Therefore a driving advice is only given if the predicted velocity v(x) always exceeds minimum velocity v _min(x). This minimum velocity profile mainly depends on the road type. Only a short distance before the vehicle is expected to pass a speed limit, the minimum velocity is adapted to the maximum velocity (Fig. 3).

Furthermore, an advisory message will only be given if the drop in velocity during the coasting phase is significant. This avoids driving advice which would lead to only small fuel savings while putting the driver’s acceptance at risk.

ViDCo’s driving errors and the corresponding advisory messages are categorized. Two categories of driving errors are specifically important with regard to processing in the algorithms: the type of driving error and the time-wise detectability.

The type of driving error draws a distinction between strategic and tactical driving errors. Strategic errors occur with regard to choices and decisions about the mode of operation (e.g., choice of travel speed, use of longitudinal control systems). They often persist for a long period of time. Even though their impact on fuel consumption is rather small in the short term, they reduce fuel efficiency markedly when persisting in the long term.

Tactical driving errors, on the other hand, are typically committed during transitions of driving state (e.g., acceleration, deceleration). In contrast to strategic errors, tactical driving errors are of a rather short-term nature. Since their influence on instantaneous fuel efficiency is high, tactical driving errors should be corrected immediately.

Tactical driving errors can further be classified into tactical-predictive and tactical-retrospective driving errors. While tactical-retrospective errors are not connected to any specific event, tactical-predictive errors are related to infrastructure objects (e.g., speed limits, downhill gradient). Therefore, retrospective advice can only be given after the occurrence of the error, whereas predictive advisory messages are capable of correcting the driver’s misbehavior in the process of conducting the error.

Separating situation detection and driving error detection enhances the robustness of the systems since driving errors can only be detected during a given situation. On the other hand, positive driver behavior can be identified if a situation occurs but no driving error is detected. Therefore driving performance can be assessed independently of the route by evaluating the ratio of driving errors with respect to the corresponding situations which have occurred.

It is generally accepted that humans’ cognitive capacities for processing information are limited, which is expressed in several theories (for an overview see e.g. [24, p. 12ff] and [25]). In the context of driving, the primary task itself may be highly demanding depending on aspects such as traffic, weather, and road conditions. In this regard, in-vehicle information systems (IVIS) and advanced driver assistance systems (ADAS) require additional resources from the drivers’ mental or cognitive workload [26].

Without any kind of filtering module, the Virtual Driving Coach would issue an advisory message via the display unit for each and every detected state of suboptimal driving. If a driver causes several such detections within a short period of time, she or he would be confronted with numerous messages. In order to keep the information flow from the EDAS to the driver within acceptable limits, a filtering and prioritization module (FPM) is developed. A limitation of additional mental workload is expected to have a positive impact on three aspects which therefore act as distinctive design objectives during the development process of the filtering and prioritization module (FPM): system safety, user acceptance, and system impact on learning.

Advanced driver assistance systems (ADAS) or in-vehicle information systems (IVIS) “should enhance or at least not reduce road safety” [27, p. 206]. System safety is therefore a fundamental design objective in the development process of the Virtual Driving Coach (see also [27, 28]). Furthermore, ViDCo is capable of exploiting its potential only if the driver is willing to use the system. Dillon & Morris define user acceptance as “the demonstrable willingness within a user group to employ information technology for the tasks it [the information technology] is designed to support” [29, p. 5]. Since the two objectives system safety and user acceptance are eminent for the safe and effective operation of the EDAS, they are classified as primary design objectives of the FPM.

ViDCo issues advisory messages to the driver only when his operating behavior deviates from optimum behavior. An optimum driver would not receive any system messages at all. This strategy of fostering learning is called negative reinforcement in the theory of operant conditioning [30]. As this concept of imparting knowledge does neither aim nor expect the driver to intentionally produce advisory messages and thus to learn actively, it may be concluded that ViDCo aims on inducing knowledge implicitly to the driver [31, p. 49].

In order to achieve the three design objectives system safety, user acceptance, and system impact on learning, the algorithms of the FPM decide on the following tasks: which advisory message should when be transmitted, for how long, how often, and by which mode (visual or auditory)? The FPM consists of the three submodules prioritization, filtering, and output and timing (Fig. 4). In addition, it features a data memory.

Due to the characteristics of the different categories of driving errors (Subsection 5.1.2), the output and timing module handles the minimum period of time as follows:

Additionally, the output and timing module will initiate the repeated transmission of the synthesized voice message if a single strategic driving error is detected continually. This is meant to regain the driver’s attention and to draw it back to the display unit. In order to retain a high level of user acceptance, intervals between repeated transmissions are expanded. In the context of learning this strategy is called expanded retrieval [33] and is thus also implemented to support the driver’s process of learning.

The filter module ensures only a single driving error is forwarded to the output and timing module. The filter blocks driving errors regardless of the priority assigned by the prioritization module. In general, blocked errors are saved in the data memory. The data memory thus enables the FPM to issue advisory messages even though the error was committed in the past and is not active anymore. However, a delayed submission makes only sense within a restricted time period after the detection. The filtering therefore also eliminates errors which are considered inactive for too long without writing these to the data memory.

Furthermore, the module detects and filters refusal by the driver. If a specific advisory message is ignored repeatedly, it must be assumed that the driver is not willing to follow this kind of driving advice. Therefore, this advisory message will be blocked and not be displayed on the current trip anymore.

The priority of the driving errors is determined dynamically by the prioritization module. It receives active errors from the driving error detection and reads saved ones from the data memory. An algorithm assigns a rating score to each error. The higher the score, the higher the priority of the error. The rating score consists of a constant term, which is specific for each driving error, and a dynamic term. While the constant term reflects the importance of the driving error with regard to its influence on fuel consumption as well as its criticality regarding a transmission in a timely manner, the dynamic term of the score will increase over time. The dynamic rating of a saved driving error is set to an even higher score, when its status changes from active to inactive (i.e., the very same error is not detected by the driving error detection anymore). By doing so, newly inactive, saved errors are handled with a higher priority in comparison to active driving errors and are thus more likely to be issued to the driver before expiration. Only after the advisory message of a corresponding driving error has been submitted to the display unit or the error has been considered inactive for too long, the driving error is removed from the data memory. As a result, the rating score is reset to its constant value.

The eco-driving assistance system is integrated into the tractor (MAN TGX, P _max  = 324 kW) of a tractor-trailer combination (Fig. 5) as such a combination is commonly used for long-haul transport in Germany [35, p. 24]. The tractor is equipped with an adaptive cruise control system based on a radar sensor and an automatic gearshift system. Therefore no driving advice for manual gearbox is issued. In order to intensify the system response, the trailer is equipped with additional weight to achieve a gross vehicle weight of 40 t.

Information about infrastructure elements, slope gradient and curvature are obtained from a digital map using commercially available software (ADASRP: Advanced Driver Assistance System Research Platform). Via a GPS sensor, the vehicle position is matched on the map and the relevant map data in the vicinity of the vehicle is transmitted to a prototype control unit as Electronic Horizon [36, 37] (Fig. 6).

A high-accuracy fuel consumption measurement system is used for system evaluation. All data, including fuel flow rate, vehicle states, internal states of the prototype control unit and the GPS-position, are recorded synchronously and combined with a video image of the driving environment.

Using the vehicle data along with the vehicle’s states and the driver’s input, the prototype control unit performs the calculation of ViDCo’s modules and triggers the output of driving advice on the display unit (Fig. 7).

The Virtual Driving Coach is evaluated on a demanding test track consisting of almost equal parts of highway and overland driving with an overall length of about 95 km (Fig. 8). The test track is driven six times by expert drivers to validate the functions of ViDCo.

Altogether 205 advisory messages are issued on the six trips on the test track; on average 48 advisory messages per trip, which translates to one piece of driving advice per 2 km.

Driving style has been influenced by advising reduced travel speed and the usage of cruise control and speed limiter. This resulted in a sufficient headway distance to preceding vehicles (Table 2).

At operational level, each single piece of driving advice is issued (Table 3). Most advisory messages addressed insufficient usage of the wear-free braking system.

On average, 17 advisory messages are displayed for anticipatory driving, leading to a coasting length of 11.1 km (11.7 % of total test track length). Speed limits contribute a major share to the coasting length with 5.4 km (Fig. 9).

No driving advice for coasting before a slow preceding vehicle is issued. It could be assumed that improved driving style leads to sufficient headway distance and thus to less relevant situations. In addition, on curvy road sections the detection range of the radar sensor and driver are equally limited. On straight road sections, drivers are able to perceive preceding vehicles prior to the radar sensor.

The real-world field tests demonstrate the technical feasibility to detect consumption relevant situations in real-world scenarios and that the driving advice leads to a more efficient driving style. Feedback of expert drivers gives a first indication of drivability which is evaluated more detailed in the driving simulator experiment.

ViDCo’s efficacy and aspects of driving behavior were evaluated by making use of a dynamic driving simulator. Participants were assigned to one of two groups: comparison group or experimental group. While both groups performed four runs on four tracks, only subjects of the experimental group were supported by ViDCo on the third run. N = 40 datasets (20 per group) were available for the analysis of the results. The functionalities output interval adaptation and refusal detection of the filtering and prioritization module were deactivated during the experiment as they were expected to negatively affect comparability (Section 5.1.3).

The experiment proved that ViDCo’s concept is capable of reducing fuel consumption significantly. Furthermore, the results indicate that drivers—at least in the short term—adapt driving advice of the system and therefore drive more efficiently without the EDAS after having been exposed to it (learning effect due to system experience). Both, system efficacy as well as the positive learning effect, were shown by means of a regression analysis of a linear mixed model (LMM). In this regression model, EDAS Usage and System Experience served as predictors (among three other predictors) on the regressand Average Fuel Consumption.

Several surveys were conducted during the experiment. Some of the surveys’ questionnaires were designed to evaluate the design objectives system safety, user acceptance and system impact on learning. As an exposure to the system is indispensable in order to evaluate these aspects, the relevant questionnaires were only asked to subjects of the experimental group (n = 20).

System safety is a necessity before introducing an IVIS to the market. System safety is evaluated in the context of two explorative questionnaires (EQ 1 and EQ 2). These questionnaires mainly contain items evaluating aspects of user acceptance. EQ 1 consists of 12 items, EQ 2 of 15 items. Items are selected from multiple sources ([38, 39], and own design). Seven of the 27 items ask questions regarding distraction and disturbance due to the EDAS and therefore provide a first assessment of system safety. The first of the two questionnaires is deployed after the system run (run 3) following the after-measurement of the SAS. The second questionnaire (15 items) is used as part of a survey which is conducted after all four runs.

Emphasis is put on the evaluation of user acceptance in this early stage of the development process. If ViDCo and its concept were not accepted, at worst it would not yield any improvements of fuel efficiency. In this case, the concept would need to be re-designed to a great extent.

In order to analyze user acceptance, the System Acceptance Scale (SAS) of van der Laan et al. is used [40]. The English version of this scale consisting of nine items was translated into German. The before-measurement was performed after describing the EDAS as follows:

“Imagine a new driver assistance system is installed in your truck. The system can help you to drive more fuel efficiently. The system provides you with instructions during certain situations on how to adjust your driving style. Please specify what you would feel about such a system?”

Afterwards the subjects perform the system run (run 3), before conducting the after-measurement of the SAS (System Acceptance Scale). The after-measurement represents the actual assessment of the Virtual Driving Coach with regard to system acceptance.

A straightforward yes-no question is asked to obtain a self-assessment of the subjects, whether they believe they learned anything from ViDCo. We call this the participants’ judgment of learning.

In order to analyze whether the results of the survey with regard to system acceptance and the self-reported judgment of learning help to explain parts of the variances for the average fuel consumption, the LMM is refitted with data of the survey. The refittings are done by replacing the original variable System Experience either by the SAS scores or Judgment of Learning. Other explanatory variables of the original model remain unchanged. Evaluation of the quality of the model is done by fitting the model with a maximum log-likelihood (ML) approach and analyzing the corrected Akaike information criterion (AICc) [41, 42]. For the estimation of the fixed-effect values regarding the specific predictors, their t- and p-values, the restricted maximum log-likelihood (REML) method is used.

As there is no control condition existing for the evaluation of EQ 1 and EQ 2, the results of the items regarding system safety are analyzed qualitatively. As illustrated in Fig. 10, the overall rating of the system is positive (M from 1.10 to 2.15 on a 5-point Likert scale, with 1 representing a positive and 5 a negative rating).

The results of the SAS questionnaire’s before- and after- measurement are presented in Table 4. For both measurements (before and after) as well as for both scales (Usefulness and Satisfying) the values of Cronbach’s α indicate a high internal consistency with α coefficients ranging from 0.89 to 0.93. Usefulness and Satisfying Scale improve both from the before to the after measurement. While a paired-samples t-test indicates a significant before–after difference for the Satisfying Scale (t(19) = 2.875, p = 0.01), the difference of the Usefulness Scale is statistically not significant (t(19) = 1.407, p = 0.18).

14 subjects out of the 20 participants of the system group answered with “yes” to the question whether they believe they have learned anything from the system while six truck drivers answered in the negative. Figure 11 shows a boxplot/beanplot comparison of the combined SAS over these judgments of learning. The combined SAS is the arithmetic mean of the Usefulness and Satisfying Scale. Each of the two scales alone, Usefulness Scale and Satisfying Scale, leads to similar results as the combined SAS of Fig. 11. However, the combined SAS is able to improve the LMM quality of the refitting better then Usefulness Scale and Satisfying Scale.

The higher overall SAS rating for the subjects answering with “yes” (M = 1.21, SD = 0.49) is statistically significant compared to the scores for “no” (M = −0.44, SD = 0.95). This can be shown by means of an independent-samples t-test (t(6.2) = 4.019, p = 0.007).

Table 5 contains the results of the original model fitting of the linear mixed model (LMM) [11] as well as two refittings regarding the predictor System Experience. When replacing System Experience by SAS scores, model quality is improved (as AICc becomes smaller). The original model has a probability of exp ((658.3 − 659.7)/2) = 0.50 compared to the refitted model to minimize the information loss [42, p. 74]. However, this improvement is less strong then the one using Judgment of Learning as predictor. In this case the likelihood of the original model is reduced to not even 0.12.

Analyzing the estimations of the fixed-effect β, the original model assesses the learning effect due to System Experience a fuel consumption reduction of −1.04 L/100 km or −2.53 % compared to a run without EDAS (run 2) [11]. Using the SAS scores as predictor suggest an average fuel consumption reduction of 1.21 · (−0.90) L/100 km = −1.09 L/100 km or −2.67 % for those subjects answering with “yes” on the judgment of learning and an increase of 0.4 L/100 km (0.97 %) for the subjects answering in the negative. When a positive learning effect is only considered for those participants stating they have learned something by the system (Judgment of Learning), fuel consumption is reduced by −1.64 L/100 km or −4.01 %.