Data Quality Assessment for Closed-Loop System Identification and Forecasting with Application to Soft Sensors.

摘要

In many chemical plants, data historians store thousands of variables at fast sampling rates. Much of this collected data is routine operating data that could easily be used for system identification and forecasting, especially in the design of soft sensors. Currently, there is no framework for assessing the quality of these data sets. Therefore, this dissertation proposes a two-step method, consisting of data segmentation and data quality assessment, with application to soft sensor development.;The first step, data segmentation, seeks to partition the extracted data set into regions that can be described using the same model. This step is necessary to avoid making the data set seem more informative than it truly is. Data segmentation is performed using a signal-entropy based approach for which the statistical properties were developed. Based on these results, it was proposed that monitoring the difference in entropy between the input and output signals for routine operating data is sufficient to partition the data set into its constituent models. Furthermore, it was shown that this difference is independent of the input signal properties under the assumption that the process is running in closed-loop without external excitation.;The second step, data quality assessment, seeks to assess the actual data quality of each region. For data quality assessment for system identification, the condition number of the Fisher information matrix is shown to agree well with the developed and extended parameter- and order-based theoretical conditions for identification that included generalised pole-zero cancellations. For forecasting, data quality assessment, a Z-score based on the distribution of the measurement errors was proposed.;Finally, in the process of testing this framework on a soft sensor system, it was discovered that the configuration of a closed-loop soft sensor can determine the success of the soft sensor development. Based on a detailed theoretical analysis of the soft-sensor system, it was shown that the presence of an integrator in the bias update term improves the tracking performance of the system.;The individual steps as well as the complete framework were extensively tested using different simulation configurations and a pilot-scale, heated tank system.