Function tolerances without component contributions

In a first step the tolerance calculation scheme is operated in a "golden-component­mode" i. e. the involved hardware components are assumed to have no tolerances at all. However even in that fictional scenario AMC function tolerances occur due to a limited parameter space of the learning function. Each learning information comes with a set of boundary conditions that are part of the picture. However AMC does not store the whole set of environment information with the learning point. The main reasons for this are that the time to achieve a learned system increases dramatically for each new parameter dimension and that the frequency for learning information to be used goes down as well. For a 5-dimensional learning problem with 10 supporting points for each considered parameter 10A5 matrix elements of learning information would have to be filled. The reduced dimensionality of AMC
significantly reduces the time interval that is required to fill in learning information while offering correction values for rail pressure control from a very early stage on. The price to be paid can be found in additional learning tolerances.

Engine speed [1/min] engine speed [1/min]

Fig. 8: Randomly distributed operation points with highlighted learning value thresholds (not considering component tolerances)

MC value within thresholds • MC value < lower threshold ■ MC value > upper threshold

MC value within thresholds • MC value < lower threshold ■ MC value > upper threshold

The calculation scheme essentially works as a point generator for AMC learning values that form a multidimensional cloud. Sectional cuts through this space of operation points are given in Fig. 8. The learning values are coded by size and shape. Circle and square shaped dots mark learning values below or above a virtual corridor. The largest learning values are to be found in the area of large injection quantities combined with small engine speeds and extreme temperatures. The fuel delivery is reduced by AMC for low temperatures (squares) while it is increased for high ones (filled circles).

Adding contour lines for the system demand (Fig. 9) it becomes obvious that the whole range of operation is covered. Following certain contour lines one can see that high as well as low learning values lie nearby.

The spread of generated learning points which can be seen in Fig. 10 shows a first order of magnitude of the AMC function tolerance. For a single given characteristic parameter several learning values exist due to the reduced set of input parameters of the learning function.

Function tolerances without component contributions
Putting the tolerance calculation scheme to work a first step is to generate random operation points defined by the vector [engine speed, desired injection quantity, fuel temperature]. This represents a worst case with respect to the fact that the boundary conditions are most probably not completely independent. Further input is generated out of the calibration data set. The choice of the "engine operation mode" (EOM) influences the rail pressure set point as well as the injection pattern. As described above the physical tolerances of the rail pressure sensor, the ECU current detection, the pump curve, the injection quantity and involved software functions are neglected in this first step.

Function tolerances without component contributions

Fig. 9: Contour lines of system demand

R— 1

Function tolerances without component contributions

MonteCarlo value

AMC values, upper boundary

AMC values, lower boundary

подпись: montecarlo value
amc values, upper boundary
amc values, lower boundary

characteristic parameter Fig. 10: AMC learning point spread w/o component tolerances

Добавить комментарий

Ваш e-mail не будет опубликован. Обязательные поля помечены *