Vous êtes sur la page 1sur 6


Paper No. 142a Anand N. Vennavelli and M. R. Resetarits, Fractionation Research, Inc. Ting Huang and R. Russell Rhinehart, Oklahoma State University

Identification of both steady state and transient state in noisy process signals is important. Steady state, for example, triggers the collection of data for process model adjustment, process analysis, fault detection, data reconciliation, or neural network training. A transient state, for example, triggers the collection of data for dynamic modeling, or recognition of points of change. Often, engineers run a sequence of experiments to collect data throughout a range of operating conditions, and process operators sequence the next stage of a process. Each sampling event is initiated when the operator observes that steady conditions are met, and then the operator implements the new set of operating conditions. However, this visual method of triggering requires continual human attention, and it is subject to human error in the recognition of steady state when measurements are noisy, or process changes are slow, or there are multiple dynamics. An automated, online, real-time, steady state identifier would be useful. The Steady State Identification (SSID) and Transient State Identification (TSID) approach here uses an easily implemented statistical method [1] with defined critical values [2, 3]. It was extended to multivariable processes and demonstrated on lab-scale and pilot-scale processes to automatically trigger an experimental sequence [4]. It was also demonstrated as a convergence criterion in nonlinear regression optimization [5]. This study reports on an application to trigger sampling and transitions in a commercial scale distillation unit. Fractionation Research Inc. (FRI) [6] is a non-profit research consortium supported by memberships, which include many petroleum, chemical, and engineering companies. FRI is the only independent, commercial-scale, distillation experimentation program operating with hydrocarbon systems at pressures ranging from deep vacuum to 500 psia. Fractionation Research, Inc. collects hydraulic and mass-transfer data on distillation column internals in its 4-foot diameter low-pressure and high-pressure distillation columns in Stillwater, OK using several hydrocarbon systems. Almost all of these data are collected at steady-state conditions. A typical FRI test involves several transitions between steady-state conditions ranging from the maximum operating capacity (flood point) to the minimum operating capacity. In addition, column dynamics are influenced by the operating pressure (deep vacuum to 500 psia), internals (trays or packing), and the system physical properties. The steady-state identifier described in this paper was implemented at the FRI unit for real-time steady-state detection. The data from the FRI distillation unit was used to demonstrate that this SS and TS identification method works in an industrial process. The steady-state identifier methodology and preliminary results are described in the paper.

This SSID and TSID method uses the R-statistic method, a ratio of two variances, as measured on the same set of data by two methods [1]. In order to reduce computational effort, exponentially weighted moving average and variances are calculated in place of the conventional average or variance. Figure 1 illustrates the concept. The dotted line represents the true trend of a process. The value starts at 15, ramps to a value of 10 at a time of 50, and then holds steady. The diamond markers about that trend represent the measured data. The dashed line, the true trend is unknowable, only the measurements can be known and they are infected with noise-like fluctuations, masking the truth.

16 15 14



13 12 11 10 9 8 0 20 40 60 80 100 120


Figure 1: Illustration of concepts The SSID/TSID approach first calculates a filtered value of the process measurement, indicated by the curved line that lags behind the data. Then the variance in the data is measured by two methods. The deviation indicated by d2 in the upper left of the figure is the difference between measurement and the filtered trend. The deviation indicated by d1 in the lower right is the difference between sequential data measurements. If the process is at SS, as illustrated in the 80 to 100 time period, Xf is almost the middle of the data. Then a process variance , estimated by will ideally be equal to estimated by . Then the ratio of the variances, , will be approximately equal to unity. . Alternately, if the process is in a TS, then is not the middle of data, the filtered value lags behind, and the variance as estimated by will be much larger than the variance as estimated by , and ratio will be much greater than unity. >> 1. Figure 2 reveals the flow chart to describe the R-statistic method calculation.

Figure 2: Flow chart

The experimental data were recorded through a Yokogawa Exaquantum historian at the FRI control room. The data collection was performed as follows. The steam to the reboiler of the 4-foot diameter FRI distillation unit was set at a constant pre-determined rate. No other changes to the unit were made until the operator determined that a steady state was achieved. The unit was held at this steady state for a sufficiently long time to study the performance of the internals. The steam rate was then changed to a new value, and the process was repeated several times over a wide operating range. Several process variables were used to establish the steady state by the operator. These variables were broadly classified as hydraulic (flows, levels), pneumatic (pressures, differential pressures), and thermal (temperatures). IMPLEMENTATION OF THE SSID PROGRAM The implementation of the SSID program was done in three phases. In the first phase, offline data from the historian were collected, and the SSID program was implemented on the historical data as a preliminary check on the working of the program. Figure 3 shows the result of the application of the SSID program on the offline data. An SSID of zero indicates transient state, and an SSID of one indicates steady state. The steady state and the transient state times predicted by the program agreed with visual observations.

The offline analysis also confirmed that the sampling interval chosen for the SSID program was appropriate. Based on the offline analysis results, it was observed that the steady state was detected by the SSID program with a 95% confidence approximately one hour after a change in the reboiler duty. The one-hour window could have easily been shortened at the expense of false positives, but was deemed acceptable and no changes were made. In the second phase, the SSID program was implemented online. The SSID program was run simultaneously in a Microsoft Excel VBA program monitoring a subset of variables tracked by the operator. When one or more variables correlated with each other, the number of variables monitored by the SSID program was limited to one of the correlating variables. For example, not all temperatures along the column were included for monitoring as the temperatures all tend to move together. Such a simplification was expected to improve the accuracy of the SSID program. Depending on the nature of the internals in the distillation column, typically 6-12 variables were monitored for a given test. During the second phase, online real-time implementation of the program was effected. The online implementation results were first monitored by the authors and were deemed accurate enough to make the program available to the operators for the next phase. Figure 4 shows the results from the online implementation of the SSID program. In the third phase, the real-time implementation results were made available to operators. The program consistently detected steady state and transient state and was in perfect agreement with the operator-determined times for steady state. Figure 5 shows the results of the third phase of implementation. The operator determined steady state times, as shown in Figure 5, clearly agree with the steady state predictions of the program. Based on several runs of the program on real-time FRI data, the authors concluded that the program was accurate enough to predict steady state and transient state every time, and at the same time, robust enough to deal with the industrial process data.

Figure 3: SSID implementation on offline data

Figure 4: SSID implementation on real-time data. Program results agree with visual observations

Figure 5: Online implementation on real time data. Corroboration with operator observed steady states.

The SSID method described in this paper was successfully used to detect steady state and transient states in offline and online modes at the FRI distillation unit. The steady state times detected by the method were consistent with the operator-determined steady state times. The authors contend that the successful implementation of the program at FRI provides proof that the SSID method described in this paper has the potential to be easily implemented in an industrial process.

1. Cao, S., and R. R. Rhinehart, "An Efficient Method for On-Line Identification of Steady-State," Journal of Process Control, Vol. 5, No 6, 1995, pp. 363-374. 2. Cao, S., and R. R. Rhinehart, Critical Values for a Steady-State Identifier, Journal of Process Control, Vol. 7, No. 2, 1997, pp. 149-152. 3. Shrowti, N., K. Vilankar, and R. R. Rhinehart, Type-II Critical Values for a Steady-State Identifier, Journal of Process Control, Vol. 20, No. 7, pp 885-890, 2010. 4. Brown, P. R.; and R. Russell Rhinehart, Demonstration of a Method for Automated Steady-State Identification in Multivariable Systems, Hydrocarbon Processing, Vol. 79, No. 9, 2000, pp 79-83. 5. Iyer, M. S., and R. R. Rhinehart, A Novel Method to Stop Neural Network Training, Proceedings of the 2000 American Control Conference, June 28-30, 2000, Chicago, IL, Paper WM17-3, pp 929-933 6. http://www.fri.org/