A novel framework for integrating data mining with control loop performance assessment

Abstract

Data driven control loop performance assessment techniques assume that the data being analyzed correspond to single plant-controller configuration. However, in an industrial setting where processes are affected due to the presence of feedstock variability and drifts, the plant-controller configuration changes with time. Also, user-defined benchmarking of control loops (common in industrial plants) requires that the data corresponding to optimal operation of the controller be known. However, such information might not be available beforehand in which case it is necessary to extract the same from routine plant operating data. A technique that addresses these fundamental requirements for ensuring reliable performance assessment is proposed. The proposed technique performs a recursive binary segmentation of the data and makes use of the fact that changes in controller settings translate to variations in plant output for identifying regions corresponding to single plant-controller configurations. The statistical properties of the data in each such window are then compared with the theoretically expected behavior to extract the data corresponding to optimal configuration. This approach has been applied on: (1) raw plant output, (2) Hurst exponent, and (3) minimum variance index of the process data. Simulation examples demonstrate the applicability of proposed approach in industrial settings. A comparison of the three routes is provided with regard to the amount of data needed and the efficacy achieved. Key results are emphasized and a framework for applying this technique is described. This tool is of significance to industries interested in an automated analysis of large scale control loop data for multiple process variables that is otherwise left unutilized.

Hierarchical Multilabel Segmentation for System Identification Using Historical Data

“…The interval-halving method has been used in various data mining techniques. This technique has been used extensively in various fields such as econometrics, process control, fault diagnosis, trend analysis, etc. ,,, It can be used to detect any statistical property change in data. Interval-halving works by recursively splitting the data into two halves based on some criterion.…”

“…This bulk data set is called the historical data of the process. Such data are used for evaluating the controller performance, analyzing the effectiveness of various operations, and fault diagnosis. − These techniques utilize only data corresponding to the region of interest in a desired granularity. Historical data can be used for model identification purposes with a few additional techniques to isolate informative data.…”

Online Performance Monitoring of Processes Subject to Non-stationary Stochastic Disturbance

“…Although FCOR algorithm seems to be effective for on-line estimate of MV benchmark [10,13], it should be noted that this technique involves applying time series analysis to estimate the source white noise under an assumption of stationarity for output data [10,12,13]. However, such assumption is usually problematic in industrial process due to the facts: either the periodic disturbances, such as sine wave and rectangular signal, are common in industrial environment [15]-[19]; or deterministic disturbances such as sudden step changes may occur randomly in time [21]- [23]; or changes of disturbances dynamics are often encountered in chemical processes [14], [19].…”

“…In the presence of non-stationary disturbances or output data, the time series analysis techniques becomes invalid in application to CPA problem [13,14]. It may yield an unmeaningful performance evaluation if simple stochastic disturbance form is taken into account for benchmark calculation [23].…”

Assessment of predictive control performance using fractal measures

“…In the recent works authors [11] perform diagnosis of MIMO control loops with Hurts exponent evaluated through detrended fluctuation analysis (DFA) algorithm using Mahalanobis distance. The same approach is applied also to the disturbed univariate and multivariate systems with disturbances [12]. Methodology investigated in the paper performs comprehensive approach to the fractal methodology analyzing different persistence measures (not only single Hurst exponent).…”