On-Line Process State Classification for Adaptive Monitoring

Abstract

In process monitoring that is based on statistical models, adaptive monitoring techniques have been developed to reflect frequent changes in the operating conditions. The key to adaptive monitoring of real industrial processes is to distinguish process operating condition changes from variations due to disturbances. This paper proposes a systematic method for detecting process state changes and classifying them as operating condition changes or variations as a result of disturbances. The key idea of the proposed method is to extract process knowledge that is based on if−then rules for detecting the operating condition changes. When a state change is accepted by a defined set of rules, it is classified as an operating mode change. Otherwise, it is classified as a disturbance. A signed digraph and statistical data analysis are used to generate the rules from the process knowledge. A robust cumulative-sum algorithm is used for change detection. The proposed method was validated using a dataset from an industrial process heater. The results showed a classification power of 97%. Adaptive monitoring that is based on the proposed method could significantly reduce the number of false alarms, compared to previous remodeling approaches.

"Data Clustering Approach to Industrial Process Monitoring, Fault Detection and Isolation"

“…The proposed method extracts process knowledge and classifies process state changes. The case study taken by the authors is a refinery fired heater [7].…”

Nearest-Neighbor Method for the Automatic Maintenance of Multivariate Statistical Soft Sensors in Batch Processing

“…17,34 However, there are several batch process systems for which the model performance degradation does not follow a temporal pattern, mostly because the process or the plant is subject to sudden changes. 19 As an example, Figure 4 shows the mean relative error (MRE) of estimation (see eq 10 in section 5) for a product quality property (acidity number) in the industrial batch polymerization process when a PLS soft sensor calibrated on ∆I ) 30 batches with no model adaptation is used.…”

“…16,17 The robustness to outliers can be improved, in such a way as to avoid that data in abnormal conditions distort the model. [18][19][20] Finally, recent issues in terms of model maintenance for batch processes relate to the selection of the optimal number of reference batches, the selection of an optimal structure for the monitoring models, and the decision of when the model should be updated for a rapid but parsimonious adaptability. [21][22][23] Recursive adaptation approaches assume that the sequential order in which the batches are run also determines a sort of similarity between batches; that is, the batches that are closest in time to the current batch are assumed to be the most similar to the current batch.…”

Multimode Monitoring of Oxy-Gas Combustion Through Flame Imaging, Principal Component Analysis, and Kernel Support Vector Machine

“…Adaptive models, which adjust the monitoring models according to the mode changes, update frequently as the condition changes Lee et al, 2006). For example, Jin et al (2006) proposed a robust recursive PCA modeling procedure to reflect the operating mode change whilst Lee et al (2006) employed ifthen rules for detecting the operation condition changes. However, updating the model parameters depends on experience and process knowledge.…”