Class-8 Heavy Truck Duty Cycle Project Final Report - Center for ...

the loss of information, and consequently again, gaps in the database. Although for almost any data
analysis these gaps are irrelevant (most of the time they were proportionally small), they introduce
cumulative errors in the calibration of the fuel consumption, and would have resulted in an
underestimation of the fuel consumed as computed by the databus information. This problem could
be corrected by simply comparing the fuel reported in the fuel tickets with the databus information for
all the trips that were completed between two consecutive “gaps” in the database.
The second and third assumption would have permitted the determination of the fuel consumption
and efficiency between two consecutive fuel stops. Fuel time information was necessary for cases in
which the vehicle was parked at a truck stop overnight. In these cases, whether the vehicle was
fueled in the evening or the next morning would have made a difference (due to the engine idling
overnight) in the fuel consumption computed from the databus information and as reported in that
particular fuel ticket. That is, if the tank was filled up at the previous gas stop, then the fuel
consumed since then would have been the fuel necessary to attain a full tank.
Because the time at which a particular truck was fuelled was not reported, those trips in which the
vehicle was fuelled at a truck stop where it was parked overnight had to be discarded. This created a
new set of artificial gaps in the database, further reducing the data intervals between two consecutive
“gaps” in the database. However, more problematic was the fact that, in an effort to minimize costs, a
vehicle may not have filled up its tank every time it stopped at a gas station.
The combination of all these factors resulted in short segments of data for which the amount of fuel in
the tank (as provided by the information collected from the fuel tickets) was unknown at the
beginning and at the end of the data segment. This uncertainty regarding starting and ending fuel
levels would have introduced errors in the “ground-truth” measure of the consumed fuel, with these
errors increasing as the data segment decreased12. On the other hand, the fuel consumption
computed using information obtained from the vehicle databus could be calculated precisely for these
data segments. However, because of the described “ground-truth” errors, it would have been
impossible to calibrate the fuel consumption measurements obtained from the databus13.
To illustrate this point, a Monte Carlo simulation model was developed to determine the maximum
expected error in fuel consumption when the amount of fuel that is in the tank at the beginning and
end of the analysis period is not known. A four thousand mile data segment was considered in which
four stops to fuel the truck were made. The same amount of fuel was added at these four stops;
however, the amount of fuel in the tank at the beginning and at the end was randomly assigned for
each one of the 100 replications of the experiment. For each one of these 100 replications, the
difference (in percentage) between the fuel consumption calculated using the fuel tickets and the
databus information was generated for three cases: 1) a databus that underestimates fuel consumption
by 5%, 2) a databus that measures fuel consumption with total accuracy, and 3) a databus that
overestimates fuel consumption by 5%.
These differences or errors, which were sorted in ascending order, are presented in Fig. 62. Notice
that because of errors in estimating fuel consumption using the fuel ticket information as described
previously, there could be cases in which it would impossible to even determine if the databus is
underestimating, overestimating, or accurately reporting fuel consumption. Consider, for example
12 As the data segment decreases, there is less fuel consumed and therefore less fuel stops. Suppose there are n fuel stops which resulted in
a total of g gallons of fuel added at these stops. Assume further that the tank was at b% and e% of its capacity c at the beginning and at the
end of the data segment, respectively. The error in the computation of fuel consumption by using the fuel tickets would have been (e - b) *
c / g, which increases as g decreases.
13 Nevertheless, some preliminary calibration was done which indicated that the databus fuel sensors underestimates fuel consumption by
about 7% to 9%.
74