As the number of actuators and sensors increases in modern
combustion engines, the task of optimizing engine performance
becomes increasingly complex. Efficient information processing
techniques are therefore important, both for off-line calibration
of engine maps, and on-line adjustments based on sensor data.
In-cylinder pressure sensors are slowly spreading from
laboratory use to production engines, thus making data with high
temporal resolution of the combustion process available. The
standard way of using the cylinder pressure data for control and
diagnostics is to focus on a few important physical features
extracted from the pressure trace, such as the combustion phasing
CA50, the indicated mean effective pressure IMEP, and the ignition
delay. These features give important information on the combustion
process, but much information is lost as the information from the
high-resolution pressure trace is condensed into a few key
parameters.
The final objective of engine calibration and control is to
achieve low fuel consumption and emissions, and high reliability
and durability of the engine. In light of this objective, it is
proposed to approach the problem of extracting key features from
the cylinder pressure data in a more systematic way. Here, a method
is suggested to extract low-dimensional features from the
high-dimensional pressure data such that the information retained
is maximized. This leads to a principal component analysis
approach. It is shown that only a few components are required to
accurately describe the pressure trace.
The low-dimensional principal component coefficients are then
used as input to a neural network that can be trained to predict
engine outputs of interest. It is shown that NOx and λ can be
accurately predicted using the principal components coefficients
and the neural network. Benefits of the scheme during transients
are illustrated, where conventional sensors are too slow to provide
cycle-to-cycle measurements.