Data-driven models used in soft sensor
applications are expected
to capture the dominant relationships between the different process
variables and the outputs, while accounting for their high-dimensional,
dynamic, and multiresolution character. While the first two characteristics
are often addressed, the multiresolution aspect is usually disregarded
and confused with a multirate scenario: multiresolution occurs when
variables have different levels of granularity, because of, for instance,
automatic averaging operations over certain time windows; on the other
hand, a multirate structure is caused by the existence of different
sampling rates, but the granularity of the recorded values is the
same. This has two major and immediate implications. First, current
methods are unable to handle variables with different resolutions
in a consistent and rigorous way, since they tacitly assume that data
represent instant observations and not averages over time windows.
Second, even if data is available at a single-resolution (i.e., all
variables with the same granularity), it is not guaranteed that the
native resolution of the predictors is the most appropriate for modeling.
Therefore, soft sensor development must address not only the selection
of the best set of predictors to be included in the model, but also
the optimum resolution to adopt for each predictor. In this work,
two novel multiresolution frameworks for soft sensor development are
proposed (MRSS-SC and MRSS-DC) that actively introduce multiresolution
into the data by searching for the best granularity for each variable.
The performance of these methodologies is comparatively assessed against
current single-resolution counterparts. The optimized multiresolution
soft sensors are bounded to be at least as good as their single-resolution
versions, and the results confirm that they almost always perform
substantially better.