Soft Sensing - Advanced Control and Supervision of Mineral Processing Plants - page 192

sensor is not available, is not fulfilled. Therefore, in an industrial application the
soft sensor models must be imbedded in systems that ensure robustness of the soft
sensor performance [1]. As an example, a soft sensor management systems called
soft sensor shell may be seen in [42] and [66].
In order to perform its task the functions of a soft sensor shell may be grouped in
modules.
1. Signal analysis module. Data collected through measurements in the plant must
be conditioned prior to being used in the determination of soft sensor models
and soft sensor operation [1, 44]. This involves filtering, data reconciliation in
general and, in particular, elimination of outliers, e.g. [31]. It is also suitable to
carry out a statistical normalization of the data so that all variables have zero
mean and unit variance.
2. Soft sensor set module. It is highly convenient to have not just one soft sensor
model, but rather a set of candidate soft sensors models. For instance, this is
necessary to keep the the primary measurement available in case one or more
secondary measurements become unusable for any reason. This condition will be
determined by the signal analysis module. An example is given below in Section
4.4.2 for the case of a particle size soft sensor. More examples are found in [11,
16, 17].
3. Model selection module. If at time t F the signal analysis module detects that the
actual sensor is unusable, then it must be substituted by one of the candidate soft
sensors of the soft sensor set module. A decision must then be made as to which
of the soft sensors candidates is best suited to replace the actual sensor. This on-
going decision is based on the quality of the primary variable estimation by each
soft sensors in the soft sensor set. This implies: (i) checking if the secondary mea-
surements for the candidate soft sensor models of the soft sensor set are available
and suitable; (ii) Estimating the mean square error between the outputs of these
soft sensors and the primary measurement while the sensor is available before
time t F when the actual sensor becomes unavailable. For example, using a mov-
ing average in a time window

[ − , ]

where t is the present time (alternatively
using a forgetting factor [44]); (iii) using other model validation indexes [1] that
may be determined on-line in the period

t T 1 t

.
4. Monitoring module. As long as the actual sensor is available the performance
of each soft sensor in the soft sensor set must be monitored. If, e.g. ,foranyof
these soft sensors the mean square error exceeds a given estimated threshold or
fails any other test during a given time window

[ t − T 1 , t ]

, its parameters should
be updated through the parameter estimation module until, e.g. , the mean square
error between the soft sensor output and the sensor measurement falls below the
given threshold during a given time window.
5. Parameter estimation module. When it becomes necessary to update the parame-
ters of the soft sensors of the soft sensor set, the secondary measurements should
be made changed so that the parameters are properly estimated as a consequence
of persistent excitation [44]. Measured disturbances may contain important sys-
tem information so they should be included in the candidate set of variables used
in the soft sensor design. But measured disturbances just happen and cannot be

[ t − T 2 , t ]