Process monitoring, using methods from system identification and chemometrics
Event details
| Date | 26.02.2010 |
| Hour | 10:15 |
| Speaker | Prof. R.Ergon - Telemark University College, Norway |
| Location |
ME C2 405
|
| Category | Conferences - Seminars |
On-line measurements of product qualities (primary outputs y) from industrial processes
(chemical plants, food industry, etc.) are often not feasible. Instead, samples are taken at
more or less regular intervals and brought to the laboratory for costly and time-consuming
analyses. There is thus a need for primary output estimation at a high sampling rate, based
on known inputs u and secondary process measurements z (flows, temperatures, etc.). In my
talk I will discuss several related aspects of this:
For dynamical systems, we may use Kalman filtering and system identification methods, also
when the primary samples are obtained at a very low and irregular sampling rate. Spectral
secondary measurements (NIR, acoustics, etc.) must then be compressed into principal
components, using principal component or least squares regression (PCR/PLSR). The
estimates of y may also be used in, e.g., Smidt controller feedback structures.
PCR and PLSR may also be used directly for primary output estimation, using statistical limits
for normal process operation. The Hotelling's T2 statistic is then used to see if new z samples
are acceptably close to the normal operating point within the projection space, while the
squared prediction error SPE or Q statistic is used to detect abnormal deviations outside of
the projection space. For this purpose new samples are split into z = zmodel + e, and for
PCR this is unproblematic. The corresponding splitting in PLSR is much discussed at the time,
and the issue is also complicated by sampling errors in y, which are often larger than errors
in z.
The splitting of z in PLSR, i.e. the definition of zmodel and e, also affects the score-loading
correspondence, and is thus of interest for fault diagnosis methods.
Links
Practical information
- General public
- Free