guidelines for evaluating control system performance · mitters, or changes in the operation of the...

Guidelines for EvaluatingControl SystemPerformanceGeorge Buckbee, P.E.

Introduction

Most plants have a huge opportunity to improve control system performance. A typical plant may exhibit:• 30% of Control Loops in Manual• 30% of Loops with tuning problems.• 10% - 40% Loops oscillating

The control system has a direct impact on the operation of the plant. The above problems combine to drive loss of efficiency, reduced production rates, higher energy costs, quality problems, and a host of other problems that directly affect the bottom line.

The opportunity for improvement depends on the specific plant, but has been seen to range from 2% to 10% of operating expense.

Why Is Control System Performance Important?

Introduction

There are many ways to measure the performance of a control system. In fact, ExperTune’s PlantTriage software measures over 80 different aspects of control performance for every single control loop.

These assessments of performance range from simple statistical measures, to complex analysis, to diagnostics for specific issues.

In the following pages, you will see how to combine this information with some engineering knowledge, to gain powerful insights, and to pinpoint the biggest payback improvements in your plant.

What to Measure?

Min of PVMin of PV Max of PVMax of PV Avg. of PVAvg. of PV

Min of COMin of CO Max of COMax of CO Avg. of COAvg. of CO

Std DevStd Dev VarianceVariance Noise BandNoise Band

Time at LimitsTime at Limits Service FactorService Factor

StictionStiction Valve SizingValve Sizing Noise BandNoise Band

SP ChangesSP Changes OscillationOscillation

Harris IndexHarris Index

PeriodPeriod

Opportunity GapOpportunity Gap

Introduction

Get insight into your process by applying engineering knowledge to the performance metrics.

For example, most instrument readings move at least slightly due to process variation or noise. If a process variable never, ever moves, you can be fairly sure the instrument has failed.

Somewhat surprisingly, it is common to find problems like this in most process plants.

There are dozens of engineering rules that can be applied broadly, once the control performance assessments are completed.

Use Diagnostics to Gain Immediate Insight

Min of PVMin of PVMin of PVMin of PV Max of PVMax of PVMax of PVMax of PV

Introduction

This e-book focuses on only the most important parts of monitoring process control performance. In the following pages, you will see:

• How to interpret performance

• When to measure performance

• The role of diagnostics

• How to find root cause

What These Guidelines Will Tell You

Guidelines

Loops Not in Normal Mode

What is Normal Mode?For most controllers, “AUTO” is

the normal mode of operation.

In more advanced schemes, the control loop should be operated at the higher level mode, such as “CASCADE”.

Symptom or Problem?A loop in “MANUAL” is usually a

symptom of an underlying problem. The operator will put a loop in “MANUAL” when he suspects the instrument, the valve, or does not like the tuning response.

Why Does it Matter?EfficiencyWhen a controller operates in “MANUAL”, it is effectively doing no control. The loop will tend to drift away from setpoint. There is no response to process upsets unless the operator happens to be paying close attention to the loop.

SafetyAlso, many industrial accidents have occurred because control loops have been left in the wrong mode.

Guidelines1.Track the % of time in normal mode for every control loop.

2.World-class plants have less than 10% of loops running out-of-normal.

3.Prioritize: Target the economically-important loops that are always out of normal mode.

4.Fix the root cause. Remember, operators usually have a reason to put a loop in manual. Be sure to ask why!

Guidelines

High VariabilityWhat Causes It?Variability can come from many different sources, including:• Load Disturbances • Sensor Noise• Valve Stiction• Setpoint Changes

Variability can be cyclical or random in nature.

Why Does it Matter?EfficiencyExcessive variability almost always causes a loss of efficiency.

Stability & QualityThe process capability decreases, and making it easier to exceed operating limits. Operators become hesitant to adjust the process, operating in ‘safe’ areas.

ReliabilityExcessive variability leads to excessive control response, wearing out control valves, and stressing other equipment.

Guidelines1.Track Variability for every control loop.

2.World class: Stable loops have less than 3% Variance as % of span.

3.Prioritize: Target loops based on variance and economics.

4.Fix the root cause. Use tools such as pattern-recognition to identify the true source of the variability.

Guidelines

Eliminate Oscillations

Where Do They Start? Unfortunately, oscillations can start just about anywhere. They can start from:• Sticky Valves• Overly-Aggressive Tuning• Batch Operations• Process Design

Once an oscillation starts, it spreads throughout the plant. In fact, it is common to see dozens of controllers cycling together.

Why Does it Matter?EfficiencyOscillating loops create inefficiencies in your plant. This is especially true for energy-consuming processes.

Eliminating oscillations can lead to instant energy savings, sometimes as much as 5% of overall energy.

QualityOscillations create swings in quality measures. This can be especially troublesome for producers of solid products, such as paper, or steel.

Guidelines1. Identify Oscillations for every control loop. Power Spectrum finds:

• Period of Oscillation• Strength of each Period• Significance to the Process

2. Sort by Period of oscillation to determine root cause.

3.Use pattern recognition, looking at the shape of the PV and CO signals, to find the root cause.

4% Production Increase4% Production Increase

View Case StudyView Case StudyView Case StudyView Case Study

http://www.expertune.com/r2p.asp?f=ISAemail&d=Oct11&t=EBEvalControlPerf&l=articles/UG2007/PerformanceAssessmentCementPlant.pdf

Guidelines

Get to the Root Cause of InteractionsWhere do They Start? Most process plants are full of interactions due to:• Process Design• Recycle Streams• Heat Integration• Control Strategies

Many plants have found that the original source of process upsets come from very far upstream, often in the utilities area.

Why Does it Matter?StaffingFinding the original source of an interaction can be like finding a needle in a haystack. Most plants simply do not have the enough people to fully investigate using older methods.

Quality, Efficiency, ProductionFinding the original source of an interaction problem can stabilize large portions of the plant. It is quite common to save $1 Million or more, from a single improvement.

Guidelines1.Use Process Interaction Maps on critical quality loops, to define the root cause of interaction.

2.Interaction Hot Spots can boost process understanding, showing all related measurements.

Interaction Hot Spots$1MM Energy Savings$1MM Energy Savings

View Case StudyView Case StudyView Case StudyView Case Study

http://www.expertune.com/r2p.asp?f=ISAemail&d=Oct11&t=EBEvalControlPerf&l=articles/wparcsabic.pdf

Guidelines

When to Measure Performance?In a New Plant:Get through the start-up curve faster, by identifying process and control problems quickly. Start monitoring during “water runs”.

In Existing Plants:Process plants are dynamic environments. There are changes to:• Raw Materials• Operating Procedures• Seasons and Weather• Product Specifications• Equipment• Economic Factors

24x7 is a RequirementChanges can happen at any time. Fortunately, most plants are already logging all the process data that you need to solve all of the performance issues described in this book.

Immediate ImprovementUsing the tools and techniques described above, you can tap in to your data history, and begin making improvements immediately.

Guidelines1. Start Monitoring control system performance right away.

2. Cast a broad net. Keep in mind that the root cause may be far away in the plant.

3. Use proven tools and techniques to start getting value right away.

Conclusions

Control Performance Matters!

Immediate BenefitsAs you have seen throughout this e-book, controller performance issues have a direct impact on the bottom line.

What many people find surprising is that improving these control problems results in an almost immediate improvement to the bottom line.

A Proven ApproachThese techniques have been proven in hundreds of plants worldwide, in many different process industries, including chemicals/petrochem, metals, pulp & paper, oil & gas, utilities, pharmaceuticals, and more.

Return on InvestmentOf all the projects you can do with a control system, control performance monitoring has one of the fastest returns on investment. Expect a complete return on your investment in less than 12 months.

$/ton$/ton

ton/hrton/hr

Improved Control

Improved Control Business Benefits

InterviewBusiness Benefits

Interview

View VideoView VideoView VideoView Video

http://www.expertune.com/r2.asp?f=ISAemail&d=Oct11&t=EBEvalControlPerf&l=http://media.expertune.com/videos/columbianbenefits.wmv

What is PlantTriage?Control Performance MonitoringPlantTriage monitors your plant 24x7, automatically identifying opportunities to improve.

Expertise Built InExperTune has specialized in control performance software for over 25 years.

The expert systems in PlantTriage capture engineering knowledge, and put it to work for you.

What’s IncludedSoftwarePlantTriage’s server-based software connects directly to your existing control system. A web browser interface makes it easy to get answers to your plant’s problems.

Service & TrainingWith PlantTriage, you get all the training and service you need, to start getting value from PlantTriage right away.

For More InformationContact ExperTune:

Phone: +1.262.369.7711

[email protected]

Product Information

www.planttriage.com

Case Studies

Contact the Author:

[email protected]

Sponsored By:

119

4Evaluating System

Performance

Practice

Overview Target product quality and production levels can be maintained by contin-uously evaluating the performance of the process and the control system. The maximum throughput and operating efficiency of a plant are ulti-mately determined by the process design and equipment selection. How-ever, in many cases a plant's operation is far from achieving the ultimate capability inherent in the plant design and equipment. For example, numerous studies done in the pulp and paper industry show that loop uti-lization ranged from 55% to 76%, depending on production area.

The reasons for process variation and poor control utilization can be attrib-uted to one or more of the following:

• Field measurement or final control element: analyzer and process measurement failures, noise, and repeatability errors; incorrect ranges and scan times; improper sensor and valve locations; damper or valve stick-slip; nonlinear installed damper and valve characteristics; and incorrect valve sizes

• Process or control design: insufficient capacity, process changes, undersized units, and lack of lab feedback

• Process interaction and mechanical problems

• Process, lab, and raw material problems (more so for chemical than pulp and paper)

120 Advanced Control Unleashed

One of the main reasons such conditions exist is that the downsizing of support services in many plants has resulted in plants operating with a minimal staff for process control and instrumentation maintenance. Often there is only enough manpower available to fix the critical problems found today that are limiting production and affecting product quality. Under these circumstances, there is little time to study the process operation to determine if an abnormal condition exists that could soon be a major source of process disturbance if not addressed. A problem may not become visible until the situation has deteriorated to the point where it affects product quality or production. Operating in this “firefighting” mode may lead to variations in operation that result in less than maximum production in a sold-out market, or to operating at less than best efficiency, or to wider variation in product quality, regardless of production rate.

The benefits from early detection of control system performance decay have been well documented by Bill Bialkowski [4.1] and others [4.2], [4.3]. The early detection of control system performance decay, of changes in process dynamics and the detection of new sources of variability, is abso-lutely necessary. This is an ongoing process that should focus on detection through continuous analysis and correction through diagnostics to mini-mize impact on plant operation.

One of the key factors is that the performance of control loops decays with time as a result of the wearing of control valves, loss of calibration of trans-mitters, or changes in the operation of the process. Some plants have real-ized the impact that control and field devices are having on their operation.

Efforts to achieve best plant performance must address both the areas of analysis and diagnostics.

• Process Analysis is the examination of a process, its elements and their relationships.

• Process Diagnostics is the investigation of the cause or nature of a process condition or problem.

The relationship between process analysis and process diagnostics is illustrated in Figure 4-1.

In some cases, additional staff has been added and a performance moni-toring tool has been purchased to periodically evaluate control loop and field device performance. However, this solution is costly and often hard to justify in the short term. Tools layered on top of a traditional Distributed Control System (DCS) to detect abnormal operation have had limited suc-cess in the detection of problems in instrumentation and fast processes

Chapter 4 – Evaluating System Performance 121

because of limitations in the speed with which information may be accessed — limitations that are inherent in the design of most DCS and Programmable Logic Controllers (PLC) systems in use today.

However, some manufacturers have introduced a whole new generation of field-based, scalable process control systems and Fieldbus field devices. The core technology used in some of these systems provides a solid foun-dation that allows the continuous monitoring and detection of abnormal control and field device operation.

In this chapter, the basis for automatic detection of abnormal operating conditions using this new technology is examined. The chapter covers per-formance monitoring tools for online analysis and monitoring of control loops. It also introduces the user to the use of multivariate Principal Com-ponent Analysis (PCA) and Partial Least Squares (PLS) for the online anal-ysis and monitoring of a unit operation in batch or continuous processes. A multivariate regression technique, such as PCA and PLS, can be designed to handle cases with a large number of correlated inputs and can extract contributions of individual inputs and identify breakdown in input correlation structure

Opportunity AssessmentThe justification for investment in tools and people for process and control analysis and diagnostics is the reduction in process variability and the associated improvement in plant profitability. A process parameter’s effect on a particular plant’s production depends greatly on that plant’s limita-tions and operating conditions. If the answer is yes to some of following questions, there is enough of an economic incentive to implement a perfor-mance monitoring system in a plant.

1. Are loops that affect plant capacity or efficiency always in their highest mode?

Figure 4-1. The Reduction of Variability Is an Ongoing Process

Measure Process Variation, Control Utilization

Identify Areas for Improvement

Determine Root Cause of Variation

Eliminate Source ofVariation

Analysis DiagnosticsWork Order -Operation Problem


2. Do loops that affect plant capacity or efficiency have too much variability?

3. Is equipment and instrument maintenance more reactive than predictive?

4. Are there batch operations where the quality or yield is affected by the variability or limits in feeds, process variables, or manipulated variables?

5. Are there batch operations where the cycle time is affected by the variability or limits in feeds, process variables, or manipulated variables?

If there is a large variability in variables that adversely affect production or product quality then by identifying these areas of variation and address-ing the source of variation, it is often possible to reduce variability and thus increase production, increase operating efficiency, or improve prod-uct quality.

The built-in diagnostic and analysis tools of modern scalable control sys-tems provide the unprecedented capability to automatically identify pro-cess variation caused by under-performing loops. This is done by continuously calculating the improvement in control that is possible and comparing this to the baseline loop performance. This approach to perfor-mance monitoring automatically examines input-output and control blocks in the controller for abnormal operating conditions that may indi-cate maintenance is required.

By automating the task of monitoring measurements, actuators and con-trol performance, it is possible to virtually eliminate the need for plant audits with portable PC-based tools and thus allow plant maintenance resources to focus on resolving the problems that affect plant operation most. Since these systems utilize the function-block architecture and model established for Fieldbus, they can use the input and output parame-ter status and block mode to determine abnormal behavior. In particular, this information is used to automatically determine the percent time a block was not in its designed mode of operation, or an input status was “bad,” “uncertain” or “limited,” or a downstream condition limited con-trol action. Blocks having an abnormal status or mode condition may be listed based on the percent time the condition existed. Plant control is summarized by the Performance and Utilization values. Such metrics can be used to judge overall control performance and whether control is being fully utilized.

Once a problem is detected, finding the solution to the problem may be straightforward for certain conditions. For example, a high variability


indication may in some cases only require that the PID control block be re-tuned to correct a change in operating condition. An uncertain indication on a transmitter may sometimes be quickly resolved using the online diag-nostics provided by the manufacturer of the device.

Unfortunately, in some cases the source of the problem may not be appar-ent and further analysis is required. In particular, resolving the source of high variability may not be as easy as just re-tuning the loop. Even though the loop may be correctly tuned, valve stick slip may cause the loop to continually oscillate, as discussed in Chapter 2 and illustrated in Figure 4- 2.

Tools for high-speed trend, histograms, and power-spectrum and cross-correlation analysis allow quick diagnosis of such problems. In a modern control system, such tools are integrated into the system design and allow such analysis to be done automatically.

Manufacturing companies are judged by the quality, cost, and availability of the products they make and sell. To achieve world-class performance in these areas, the manufacturing operation must develop specific metrics to assess how well they are performing day to day and guide them towards improvement. In order to improve control, the data must be taken fre-quently enough to show inertial effects. Also, statistical metrics should be established to evaluate process and control system performance. With such online tools in place, it is relatively easy to detect the presence of a new point source of variability, as individual signals will start to exhibit markedly different behavior.

Process monitoring can take many forms, from a purely mechanistic approach guided by first principles to a purely statistical approach. In

Figure 4-2. Limit Cycle from Valve Stick-Slip

Stem Position

Implied Valve Position (IVP)

Controlled Process Variable (PV)

Setpoint (SP)


today’s world, data drives the decision-making process and the ability to make use of the vast amounts of process data is essential. In many plants, the rate at which we are accumulating this data surpasses our ability to extract knowledge from data and use it for better decision making. Multi-variate statistical analysis is a technique that allows us to analyze plant data in order to extract underlying themes in the behavior of the data, and to then use these themes to monitor the state of the process.

Traditionally, the task of monitoring plant data has fallen into the statisti-cal process control (SPC) or the univariate statistical analysis world. The manufacturing industries have made great progress by focusing on key quality variables and monitoring these variables with univariate analysis techniques such as SPC. But are these techniques still relevant now that the volumes of data being extracted by our automation systems have increased by orders of magnitude? The answer is, Not always! In previous years we had 10 engineers per quality parameter being monitored; today a single engineer is asked to monitor 10 quality parameters. The drive to understand quality coupled with the vast amounts of data can present a formidable task, one well suited to the application of multivariate tech-niques.

The multivariate techniques that are showing success today are called Principal Component Analysis (PCA) and Partial Least Squares or Projec-tion to Latent Structures (PLS). Multivariate statistical analysis is not a new technology, but the application of PCA and PLS has gained attention due mainly to the data explosion. The manufacturing and process indus-tries have invested heavily in real-time data acquisition systems or Process Information Management Systems (PIMS) and there is now a strong desire by manufacturing directors to see this accumulation of data put to use to improve plant operation; hence the strong interest in multivariate moni-toring techniques. Other multivariate monitoring techniques include Fac-tor Analysis, Eigen-vector Analysis and Singular Value Decomposition.

The field that uses multivariate statistical analysis tools for real-time anal-ysis of process data is called Multivariate Statistical Process Control (MSPC). The attraction of MSPC is in the ability to rapidly develop behav-ioral models of your process, akin to fingerprinting normal process behav-ior, and then to continuously compare current plant behavior to the normal fingerprint. If a deviation from normal plant behavior is identified, MSPC will allow you to identify which plant variables are the major con-tributors to the cause of the deviation.

The ability to monitor, identify and diagnose the cause of process variabil-ity is a task every engineer will recognize as important and one that will improve the performance of our manufacturing facilities. The technologies


of today are focusing on these monitoring challenges, while the monitor-ing technologies of tomorrow will be targeted more at relating variability to the bottom-line profitability of our plants—an important link still not well defined in most manufacturing environments.

Examples

Control Utilization and Its Impact Control utilization is an indicator that can be used to quickly determine if control and measurement problems exist within a control system. It is rela-tively easy, using most control systems, to take a snapshot of control oper-ation and to then calculate and document the control utilization. For example, using the documentation features of most control systems, it is possible to print the operation screens and later to compare the mode of the control blocks to the design mode for the control. In some cases, a low value of control utilization can be attributed to insufficient operator train-ing on the control. However, surveys of many plants indicate that the pri-mary reasons for control not being fully utilized fall into two areas:

• Field measurement or control element

• Process or control design

An immediate improvement in control utilization will be achieved by addressing these problems. To achieve full utilization, improved communications between maintenance and operations is important.

The automatic collection of control-utilization statistics is a major benefit in identifying problems in measurement, control or process design. In plants that do not have a control system that supports this capability, problems may go unrecognized that have a direct impact on the plant operation and efficiency of production. For example, at one pulp and paper plant a snapshot of the control utilization was collected to allow the state of the process control to be quantified. This survey revealed the fol-lowing:

Plant Area Total Control LoopsLoops not in Normal Mode Percent Utilization

Bleach Plant 78 60 76%Power House 185 130 70%

Pulp Mill 174 116 66%Paper Mill 236 134 56%


When the utilization was presented to the head of maintenance, he was surprised to find it so low. To address this problem, a number of things were put in place at this plant.

Automatic calculation of control utilization was implemented using fea-tures of their historian.

An instrumentation team was formed to investigate loops that were not running in their normal design mode. This team was responsible for mak-ing sure measurement, control valve, and process problems were addressed in a timely fashion.

The reduction in variability provided by these two steps led to significant improvements in plant throughput and product quality. Long after the team was formed, it continues to provide significant, quantifiable results to the plant through constant evaluation and improvement of control per-formance.

Control VariabilityThe response of a control loop in its designed mode of operation may not be adequate to compensate for process disturbance and changes in set point. The root cause of poor control performance may be traced to the control design, tuning, measurement or actuator performance. Many mod-ern control systems provide a means to automatically quantify the varia-tion seen while on automatic control. Using this embedded feature, it is possible to quickly identify control loops that require maintenance. If such capability is not available in the control system, it may be possible to uti-lize the statistical calculations included in a plant historian or to add loop performance-monitoring packages that connect directly to the DCS or to the plant historian through OPC connectivity. Both methods will monitor control loop variability and provide insight into changes in control perfor-mance.

By comparing the current variation over a shift or day to that previously seen for the same operating condition and timeframe, it is possible to iden-tify loops whose performance is degraded. Through the investigation of problems that cause increased variability, it is often possible to find the root cause and significantly improve plant operation. As discussed in Chapter 3, this reduction in variation often makes it possible to shift oper-ating points closer to operating constraints or to maintain the best operat-ing set point, and thus provide greater throughput or improved quality.


Quality Variable Monitoring We can use an SPC univariate chart to easily detect where the process is going out of control. Consider the analysis of the two variables X1 and X2 in Figure 4-3. Although neither variable appears to demonstrate any sig-nificant deviation from the norm, a scatter plot of the data clearly points to an outlier. This irregularity in the operation of the process could only be detected by revealing how this point differs from the normal correlation structure of the process data. X1 and X2 could not independently pinpoint the fault, but the relationship (correlation) of X1 to X2 clearly identifies a change in the process. Thus, multivariate statistical analysis models the correlation structure of the process data in time and detects abnormal behavior by pinpointing deviations from the normal correlation structure. The simple example of monitoring the correlation’s structure in the two variables can easily be extended to monitoring hundreds of process vari-ables.

Caster MonitoringThe next sample industrial process we will consider for MSPC is a caster monitoring application in a steel mill. The process is a continuous, straight mold, Demag slab caster, shown in Figure 4-4. Liquid steel enters the caster process where it is partially solidified in a water-cooled copper mould. The steel exits the mould, where the steel is contained within a thin solidified shell or skin, and cools as it moves along the process line towards the run-out. Upon leaving the mould, there is potential for the loss of containment of the molten steel. This loss of containment and release of molten liquid steel is a called a breakout. A casting breakout is an extremely hazardous occurrence that causes equipment damage and loss of production. An MSPC monitoring system was installed to provide real-time monitoring of the casting process variables in order to predict potential breakout conditions and provide operators with information to help diagnose casting process problems.

Figure 4-3. Scatter Plot Reveals an Outlier

Outlier

X1

X2


The MSPC casting application has a few multivariate SPC control charts representing the stable mould operation and alarms on detection of any significant anomalies in the behavior of the process variables. At any time, the operator can view how the measured plant values are contributing to these alarms, to help diagnose process problems. The final PCA model of the process contained over 240 inputs, collected selectively over several months of operation in order to cover many different operating scenarios. The PCA models were further summarized into control charts, using Hotelling’s T2 statistic, to be monitored along with the prediction-error control chart. T2 statistics provide an indication of variability. The T2 is named for Harold Hotelling, a pioneer in multivariate statistical analysis.

These statistical control charts were again summarized in a composite set of alarms for operator notification. This multivariate SPC information is displayed on the Caster Main control screen, along with other pertinent mould data, as shown in Figure 4-5.

These three charts form the basis of the real-time MSPC system, providing continuous information to the operator on the stability of the mould oper-ation. This example of an MSPC system, implemented in a steel plant, delivered benefits in the areas of increased production, improved safety and much improved process understanding—the ultimate goal of any monitoring system [4.12].

Figure 4-4. Slab Caster Process

Ladle - 300 tonne

Tundish - 60 tonne

Mould

Bending

Runout

Unbending


Application

General Procedure1. Decide on the objectives of performance monitoring in terms of the

relative importance of diagnosing and predicting the following:

a. Measurement problems

b. Rotating equipment problems

c. Hydraulic resonance and shock waves

d. Pressure regulator and steam trap problems

e. Interlock and alarm system sequence of events

f. Control valve problems

g. Improper controller tuning

h. Improper controller modes

i. Interaction between control loops

j. Feedforward and disturbance analysis

k. Weeping and flooding of column trays

l. Distribution and mixing of feeds, components, phases, polymers, and particles in fluidized beds, columns, fermentors, reactors, and crystallizers

m. Analyzer and sample system problems

Figure 4-5. Display Interface for Slab Caster Example (Courtesy of Dofasco, Inc.) [4.12]


n. Lab sample and analysis process problems

o. Ambient temperature and humidity effects

p. Raw material problems

q. Fouling and coating (heat-transfer surfaces, polymer pipelines and reactors, crystallizers, sensors, electrodes, valve trim, column trays, particle filters, ion exchange beds, catalyst)

r. Activity evaluation (beds, catalyst, initiators)

s. Batch sequence and profile problems

2. Select the proper scale, exception reporting, filter time, and scan time so that information is not lost due to poor resolution or missing data points.

3. Activate process monitoring embedded in your control system or, if not available, install a performance monitoring application that meets your objectives.

4. Establish a best period of operation for each plant area or piece of process equipment and use this as the baseline for relative performance indices.

5. Turn off and flag the performance analysis of loops that are intentionally out of service due to the shutdown or sequencing of a unit operation.

6. Compute the economic performance of your system on line (efficiency and capacity, preferably in terms of cost of goods and revenue).

7. Set the deviation alert and alarm limits from the baseline wide enough to avoid an initial barrage of abnormal conditions when the system is first turned on, but narrow enough to flag important issues for the more critical process variables and control valves.

8. Work on down the list of alerts or alarms, prioritized by economic importance and relative performance, to fix automation and process deficiencies.

9. Reset the baseline and narrow the deviation limits as you progress and the performance improves.

10. Establish procedures and best practices for continual improvement.

11. For constrained model predictive control systems:

a. Display the constraints that are currently governing, the direction of the change in the affected controlled variables demanded by these constraints, the percent of time each


constraint has been active, and the percent of time that each constraint has been violated.

b. Provide trending and statistics of the bias correction to each trajectory.

c. Provide trending and statistics of predicted values for each trajectory.

d. Provide trending and statistics of optimization set points (set points from rampers and pushers).

Application DetailsWith a modern control system, all control loops are automatically moni-tored on a continuous basis and any degradation in loop performance or the detection of an abnormal condition in a measurement, actuator, or con-trol block is automatically flagged. These systems may thus identify prob-lem areas sooner than can audits done with portable PC-based tools. By using automated system performance monitoring and built-in diagnostic tools, the typically limited resources of plant maintenance can be used to advantage to resolve measurement, actuator and control problems. As a result, maintenance costs may be reduced or a higher overall level of sys-tem performance may be maintained and process variability reduced. Any reduction in variability can lead directly to greater plant throughput, greater operating efficiency and/or improved product quality.

Continuous ControlIn evaluating the performance of a continuous process, control systems may calculate indices that quantify loop utilization, measurements with a “bad,” “uncertain,” or “limited” status, limitations in control action, and process variability. In addition, for control loops, the systems show the potential improvement possible in control loop performance. The follow-ing are a few practical scenarios that show how such information might be used to determine an operating problem.

Loop Not in Normal Mode A critical process input flow loop is designed to normally run in cascade mode with its set point determined by a composition loop. However, the midnight shift operator has to change this flow loop to Auto because of wide fluctuations in the set point. The next morning, the process engineer for the plant sees that the system flagged the actual mode as being incor-rect for more than 1% if the time. This information allows the engineer to quickly spot that there was a problem in the primary loop tuning and thus minimize any off-spec product that resulted from manual adjustment of this key flow.


Limited Output An instrument technician for the power plant notices that the system shows the oil-flow control loop is limited in operation. Having been alerted to this problem, the technician determines that the set point for the oil header pressure control has been lowered below the design pressure, forcing the oil valve to go fully open under heavy load conditions. When the pressure control is readjusted to its designed target, the oil valve can meet its set point without going fully open.

Bad I/O A key temperature measurement is flagged by the system as having been “bad” more than 1% of the time over the last day. Having been alerted to the fact, the instrument technician re-examines the transmitter calibration and finds that the device has been calibrated for a temperature range that is too low. Re-calibrating the transmitter restores the accuracy of the mea-surement and improves the operation of the process.

High Variability During normal operation of the plant, the plant engineer sets all variabil-ity limits to the current value plus 5 percent. After a few weeks, he notices that the system has flagged a critical flow loop as having excessive vari-ability. Upon further investigation, he discovers that the valve positioner connection to the valve stem is loose, causing the control loop to cycle severely. Fixing the valve positioner returns the variability index to its nor-mal value.

All these cases are quite typical. Individually they may seem to have mini-mal impact on plant operation. However, the net effect of these problems could be significant if they were not addressed in a timely manner. An automated analysis system to evaluate performance can detect abnormal situations as they happen and thus play a key role in preventing produc-tion losses and product quality problems.

Built-in diagnostic tools may be used to quickly determine the source of a problem. Typical diagnostic tool sets will provide support for the follow-ing:

1. Trends

A high-speed trend provides the ability to trace any parameter associated with measurement or control actions at the same rate at which these values change. A trend resolution of 100 msec is usu-ally sufficient to analyze most measurement and control problems.


A high-speed trend requires over-sampling at the input card and access at the function-block level.

Experienced control specialists can readily spot a control valve that has excessive slip-stick from a trend recording of the process vari-able and controller output. Until waveform recognition systems are developed to reliably provide this analysis online, there are some general patterns to look for to help track down the causes of oscillations.

For example, a rounded sine wave and sawtooth oscillation with a period much larger than the integral (reset) time are symptomatic of a reset-induced cycle for a loop dominated by a time constant and time delay, respectively. If the PV oscillation has a flatter top (not as rounded for the sine wave or as sharp for the sawtooth), it is and indication that the control valve has excessive dead band (backlash).

If the PV oscillation becomes more like a square wave, then the control valve has appreciable stick-slip. The CO oscillation will resemble a sawtooth. For stick-slip, the oscillation is a limit cycle rather than a reset cycle. Any reset action at all will initiate the cycle and the cycle period will become proportional to the integral time. A trend of a low-noise flow measurement is the best way to see these oscillations. For sliding stem valves, the actual valve position from a smart positioner can be trended. In both instances, the unfiltered measurements are fast and the patterns for a dead time–dominant response should be observed on the trend.

2. Histograms

A histogram allows the distribution of the variation in a measure-ment value or actuator position to be analyzed. A bell shaped dis-tribution indicates that the source of variation is random in nature.

3. Power Spectrums

A power spectrum is a frequency distribution of the components that make up a measurement or actuator signal over a selected period of time. Such information may be helpful in determining the magnitude and frequency of process noise or the frequency of disruptions caused by loop interaction or upstream processes.


4. Cross Correlations

The influence of other control loops and upstream conditions on the variability of a control loop may be determined through cross correlation

Diagnostic tools for process and control analysis are available as embed-ded features in some control systems. These tools may be used to analyze both fast and slow processes. If an integrated tool set is not available, diag-nostic applications can be layered on the control system for diagnostics, although they are often limited to analysis of slow processes. Figure 4-6 illustrates the diagnostic information that is presented to the user by a typ-ical performance monitoring system.

Batch ControlThe application of online performance monitoring is most often associated with continuous processes. However, the percentage improvement from its application to batch control could be larger because the state of the con-trol loops typically receives less attention than the sequence of the batch operation. The emphasis to date in batch control has been on event sched-uling rather than on process control. The benefits of monitoring can be especially significant when you consider the higher profit margin and value-added opportunities for specialty chemical and pharmaceutical products, which are almost always batch processes. In such cases, the answers to Opportunity Assessment questions (4) or (5) are frequently Yes.

Figure 4-6. Process and Control Diagnostic

FT101/PID1Analysis

+ANAL_FT101PID020602ANAL_FT101PID020802

PowerSpectrumHistogramAutocorrelationTrend

Cross CorrelationCORR_FT101PID020802

PI103/PID1/OUTAI321/AI/PV

+CORR_FT101PID020902+CORR_FT101PID021002

Trend Auto correlation

Power Spectrum Histogram


Figure 4-7 shows a batch reactor for a sold-out high profit–margin spe-cialty chemical plant. One of the feeds and a byproduct can be recovered from the vent system if there is sufficient condensing to reflux unreacted feed back to the reactor and there is no liquid carryover from high level or foaming. Periodically there are interruptions due to high reactor pressure, level, and exchanger temperature. The cooling water pressure is some-times too low and its temperature too high. The feed profile is scheduled to minimize the activation of interlocks and excessive riding of the con-denser, vent system, and exchanger limits. The efficiency and safety of the operation depends upon a uniform quality, flow rate, and distribution of the feeds within the batch. There is no online composition measurement of the product. The yield varies significantly with each batch.

A performance monitoring system is installed that provides indices on variability, the detection of limits, power spectrums, and cross correla-tions. The results are as follows:

The standard deviation and variability indices are high and the control output blocks are limited at times during the batch for the reactant B and antifoam feed loops, the condenser and exchanger temperature loops, and the reactor pressure loop.

Figure 4-7. Batch Reactor Control System

TC 1-1

TT 1-1

LT 1-2

LC 1-2

TT 1-2

TC 1-2

RSP

Batch Reactor 1

Coolant

Discharge

Coolant

TT 1-3

PT 1-3

PC 1-3

PT 1-1

PC 1-1

TC 1-3

Eductor

Vent System

Feed B

FC 1-2

FT 1-2

Feed A

FC 1-1

FT 1-1

Anti-Foam

FC 1-3

FT 1-3


The low-limit flag for the output function block of the reactant B flow con-troller is periodically activated throughout the batch. After some investi-gation, it is discovered that the original sliding stem control valve was replaced with a ball valve that has too much capacity and too much fric-tion near the closed position. The reactant B flow controller ramps up its output but there is no appreciable increase in flow above the leakage flow until its output reaches 6 per cent. Then there is a burst of flow so high above the set point that the controller must close the valve. This square wave of flow from the stick-slip of the oversized control valve continues through most of the batch.

The antifoam controller output during some batches ramps up to its high-output limit. A correlation analysis shows that this coincides with the simultaneous demand for antifoam by other batch operations and a dip in supply pressure. An undersized antifoam pump is pinpointed as the cul-prit. A further look at other variables shows there is a significant cross cor-relation between high antifoam flow controller output and high reactor pressure controller output and high condenser temperature controller out-put. This indicates the possibility of carryover into the vent system that can result in contamination of the byproduct and a corresponding loss in yield of the main product.

A trend of the condenser temperature variability indices and limit flags show that the condenser performance is often worse during hot summer afternoons. There appears to be no correlation with the sequence of other batch operations. A review of the cooling-water system shows that the cooling-water valves for other unit operations are left wide open all the time and that the reactor condenser is at the highest elevation in the plant. Furthermore, the cooling tower water temperature on hot days reduces the Log Mean temperature difference for the exchanger by more than 20%. This, combined with the inability to increase cooling water flow due to the low pressure at the condenser elevation, causes the condenser control valve to go wide open and the condenser temperature to float above the set point. The high condenser temperature decreases the reactor yield and increases the variability in the composition of the byproduct and recycle recovered by the vent system. The installation of a higher head cooling tower water pump and the commissioning of temperature control loops for the other batch processes is a much lower-cost solution than the addi-tion of a new cooling tower as favored by operations.

A power spectrum analysis of the eductor pressure loop in the vacuum system and the reactor pressure controller shows that there are large peaks at the same frequency and that these peaks get larger during hot summer months. Upon inspection, the integral time of the eductor pressure loop is much less than the period of oscillation that corresponds to the common


frequency with significant power. This is a clue that there is a reset cycle from excessive integral action from the eductor pressure loop. During hot summer days the process gain and time constant are less, making the reset cycle more severe.

The above analysis reveals the importance of extending the scope of per-formance monitoring to other unit operations, utility systems, vent sys-tems, and ambient conditions. It also shows the value of being able to trend performance indices, status flags, events, and variables to look for coincident events. It is important that the trend be fast enough to deter-mine what happened first and that the oscillation period or waveform not be distorted by aliasing. Finally, there is obviously a need for some basic understanding of the system to determine the actual cause-and-effect rela-tionships.

The knowledge gained from performance monitoring systems leads not only to immediate improvements in valve and pump sizes and controller tuning, but also to the implementation of advanced control techniques. In the chemical plant example, the identification of the limits can be used to provide constraints for model predictive control. In this case, the con-straints are condenser, exchanger, and vent valve position; the variable to be optimized is reactor feed. A model predictive controller could be set up to maximize reactor feed without violating utility and vent system limits.

Once the basic system has been improved, the use of performance moni-toring can be extended to include online multivariate principal component analysis (PCA). A PCA worm plot can provide rapid recognition of batches that start to trend away from normal operation to warn of abnor-mal conditions before they cause significant problems of reduced yield or capacity. It can also lead to Partial Least Squares (PLS), Neural Network, and dynamic linear estimators of batch end time and product concentra-tion based on various peaks during the batch cycle. For the above reactor, it turned out that the amount of fresh feed added to reach the first peak in temperature was an indication of the composition of the recycled feed and a sustained low condenser and vent valve position could be used to pre-dict the batch end point.

While the above analysis was for a specialty chemical, the opportunities for fermentors could be even greater due to the need for tight dissolved-oxygen and pH control despite interactions and interferences and the advantages of having online estimators to detect deviations of batch con-ditions and cell and product concentration. Similarly, the analysis should be extended to substrate and nutrient feed systems, antifoam systems, vent systems, and utility systems. Figure 4-8 shows the main control loops for a batch fermentor.


Multivariate AnalysisAlthough the methods for monitoring plant data have traditionally been limited to statistical process control (SPC) or univariate statistical analysis, manufacturing industries have made good use of these methods to iden-tify and monitor key quality variables. Today, the volume of data being extracted by automation systems has increased by orders of magnitude. The drive to understand quality coupled with the vast amounts of data can present a formidable task.

What are the factors that make process data analysis a challenge?

• Correlation: Although we may be monitoring hundreds of variables, there are not hundreds of independent events occurring. The significant correlation in the variables, with respect to each other, points to only a few underlying themes.

• Dimensionality: The average chemical plant now records and stores thousands of variables every minute in a plant historian. This volume of data is difficult to handle, forcing people to look at only a small subset of variables, and thus only a fraction of the overall information.

Figure 4-8. Batch Fermentor Control System

TT 2-2

TC 2-2

TT 2-1

TC 2-1

coolant

tempered water

Batch Fermentor

RSP

AT 2-2AT

2-1

VSD

Anti-Foam

FC 1-3

FT 1-3

FC 1-2

FT 1-2

Substrate

AC 2-1

Reagent

steam

FC 1-1

FT 1-1

Air

pH

AC 2-2

DissolvedOxygen

Vent

C S

PT 1-1

PC 1-1

RSP

RSP


And what are the elements that make PCA and PLS good multivariate techniques for modeling plant behavior from real-time operating data?

• Handle large volumes of data

• Deal with highly correlated variables by reducing the space to a set of independent factors

• Robust to missing data and noisy data

• Ability to drill down and understand contribution of plant variables

• Standard set of graphical tools to represent results

The flow of the MSPC analysis path is quite simple.

1. Build a correlation model of your process

Although a process may contain hundreds of variables that are being monitored by plant information systems, there are never hundreds of independent occurrences. Most of the data moves in a correlated fashion, which is the basis of our model. MSPC will form new variables, called latent variables, which are simply linear combinations of the existing plant variables, but which are selected to model the greatest directions of variability in the data. These latent variables are generated to not only extract the maximum degree of variability from the data, but to do so such that the new latent variables are independent of (orthogonal to) each other. Often, a very large portion of the variability in plant variables can be summarized and compressed into a relatively small number of latent variables.

2. Monitor the reduced set of latent variables

Reducing the order of the problem is the key to MSPC. If the vari-ability in hundreds of plant variables can be captured and modeled with several new latent variables, our task is much more manage-able.

3. Identify deviation from normal plant behavior

As with univariate SPC, MSPC relies heavily on visualization and graphical tools to monitor the reduced set of latent variables. Although the new latent variables do not have any physical mean-ing, tracking these variables in a control chart fashion can clearly identify deviation from normal behavior.


4. Identify which and how the original plant variables contribute to the deviation

One of the principal strengths of the MSPC approach is the ability to drill down and identify how the original plant variables contrib-ute to the abnormal behavior in plant performance. Knowing how the original plant variables contribute allows action to be taken to bring the plant back within control.

In order to better understand the process described above we can graphi-cally interpret MSPC applied to a simple process monitoring 3 process variables: temperature, pressure and flow. Time-series plots of the vari-ables are shown in Figure 4-9 with no apparent abnormal behavior. The data can also be plotted in a 3-D space, as shown in this figure.

The first step in MSPC is to draw a line through the data in the direction of maximum variability. This line is called the first principal component and projecting the data points onto this component defines our first latent vari-able. A second line can now be drawn through the data, orthogonal to the first (linearly independent), and in the direction defining the second great-est variability as shown in Figure 4-10. This is known as the second princi-pal component, and again the data points can be projected onto this principal component to generate latent variable 2.

These 2 new principal components now form a new plane in our data space, which is commonly referred to as a scores plot. The scores plot for our temperature, pressure and flow data is shown in Figure 4-11; it cap-tures over 92% of the variability in the dataset. The scores plot has reduced the order of our data from 3 dimensions to 2 dimensions. Monitoring how the data points move on this plot is one of the graphical tools used in iden-tifying abnormal behavior.

Figure 4-9. Plots of Process Variables

Temperature (X1)

Pressure (X2)

Flow (X3)


For example, the temperature, pressure and flow data show significant deviation from model norm, behavior not easily detected from the time series trends. Although we have limited our analysis to reducing 3 vari-ables down to 2 principal components, we can often reduce hundreds of variables into only several principal components and still capture a signif-icant amount of the variability in the dataset.

In any statistical online monitoring application, graphical output and visualization of the data are important tools for identifying abnormal behavior. In MSPC the most common plots include the scores plot, the model residual plot and contribution plots.

Scores PlotThe scores plot allows the observations to be viewed in the new co-ordi-nate system of principal components. The new principal components will

Figure 4-10. First and Second Principal Components

Figure 4-11. Scores Plot Shows Contribution of Principal Components and a Strong Outlier

X3

X1

X2

PC2

PC1

PC-2 Scores

PC-1 Scores-20 -15 -10 -5 0 5 10

543210

-1-2-3-4-5


form a sub-space, and projecting the observations down onto this sub-space will allow us to visualize how our observations relate to our PCA model of the process. A common technique for real-time monitoring is to connect the observations in the scores plot, forming a worm of a preset length. Every new observation adds a value to the head of the worm, while an older observation is dropped from the tail. Following the direc-tion of the worm can indicate when the process is trending towards abnor-mal behavior.

Hotelling T2 The Hotelling T2 is a statistical parameter indicating variability of the mul-tivariable process. It can be used to classify when an observation in the data is a strong outlier. A strong outlier in the data indicates abnormal behavior. The Hotelling T2 statistic is most often represented as an ellipse on the scores plot with 95% or 99% confidence intervals. Thus, any obser-vation outside of these limits would indicate a strong outlier that does not conform to the normal correlation model of the process data. Figure 4-11 illustrates the result of the Hotelling T2 classification for a process with two principal components. Prediction Error or Model Residual Plot

The prediction error is an indication of a moderate outlier and is com-puted as the residual for each observation. In MSPC, the prediction error is usually plotted in time with a critical tolerance level.

Contribution Plot One of the many advantages of the PCA approach is that the information in the model is open for inspection. If a situation is detected the system can present how the original variables are contributing to the detected fault. These are most often presented as a contribution difference from model center or contribution difference between two observations.

Although this treatment of the MSPC topic has been limited to the moni-toring and analysis continuous systems, the techniques can also be applied to batch modeling and monitoring. In the analysis of batch data, many batch variables can be modeled using the techniques of PCA to gen-erate a set of new latent variables. Analysis of known good batches can then generate a series of latent variables that define the golden batch tra-jectory and target path for the batch process. This is known as intra-batch modeling and helps provide early detection of abnormal batch operation, within a specific phase of the batch. Moreover, MSPC has also been effec-tively applied to the monitoring of overall batch performance, with the ability to condense the overall batch information such that it can be com-pared to several other batches.


Rules of ThumbRule 4.1. — Make sure the scan time of the I/O, module, and performance moni-toring system is faster than 1/5 of the dead time for fault diagnostics. This will prevent misalignment of events and aliasing of loop oscillations because it will insure at least 10 samples per oscillation for dead time–dominant loops and 20 samples per oscillation for self-regulating processes with a large process time constant.

Rule 4.2. — To analyze items 1a through 1e in the General Procedure, the scan time should be as fast as the input card or field device to avoid improper identifica-tion of periods of oscillation (aliasing) and the sequence of events. The speed and richness of diagnostics from Fieldbus devices can be an important advan-tage.

Rule 4.3. — To track down the cause of oscillating control loops, locate the loops with significant power at the same frequency on a process flow diagram (PFD) and trace the direction of manipulated flows between the loops. The loop from which the manipulated flow originates is the source of the oscillations. It is generally, but not necessarily, the furthest loop upstream.

Rule 4.4. — Watch out for periodic disturbances. Aggressive tuning of level loops, valve stick-slip, too small a reset time on loops dominated by a time lag (reactors, fermentors, evaporators, crystallizers, and columns), and too high a controller gain for loops dominated by a time delay (webs, sheets, and pipelines) are the most common causes of a periodic disturbance. If the loop manipulates an inlet or outlet flow to the volume, the oscillating loops will be upstream or downstream, respectively.

Level loops on surge and feed tanks tend to be tuned too tightly. Level measurement noise is a frequent problem, particularly if a high gain or any rate action is mistakenly used. Valve stick-slip often appears as a square wave limit cycle. If the reset time is less than the oscillation period for a loop dominated by a large time constant, the reset time is probably too small. These loops need to overdrive the manipulated variable and require more gain than reset action (see Chapter 2). Conversely, the pulp and paper industry tends to have loops that are dominated by large time delay, and consequently the use of a gain or reset time that is set too high is more of an issue.

Rule 4.5. — If the oscillation period is less than 4 seconds, it probably originates from rotating equipment, pressure regulators, actuator instability, burner insta-bility, resonance, or measurement noise. Incipient surge and buzzing due to dips in the characteristic curves of compressors and high-pressure pumps are a principal cause of fast oscillations in gas and boiler feed water sys-tems, respectively. Oversized field pressure regulators and inadequately


sized actuators and springs can cause large rapid fluctuations in flow. Burner instability can cause rapid oscillations in fuel flow and furnace pressure.

Rule 4.6. — Items 1a through 1i in the General Procedure can usually be handled by performance monitoring systems that provide relative statistical measures of variability, capability, utilization, sustained cycling, oscillation frequency and power (Power Spectrums), saturated controller outputs, and limited process vari-ables.

Rule 4.7. — The diagnosis and prediction of process equipment, operation, sequence, allocation, and raw material problems (items 1j through 1s in the Gen-eral Procedure) generally require cross correlation analysis and multivariate prin-cipal component analysis (PCA).

Rule 4.8. — Online property estimators should be created to enhance the repeat-ability and reliability of online and lab analysis systems by the addition of Partial Least Squares (PLS) analysis, dynamic linear estimators, or neural networks. Time delays and time constants must be used to align the inputs with out-puts. Some of the need for data alignment can be reduced by the use of long scan times but this slows down the calculation for estimation and fault detection to a point where it may be too late or unable to resolve what occurred first. This is a more important issue for continuous opera-tions than batch operations since batch outputs can be tied to a batch, step, and phase identification number and time (see Chapter 8).

Rule 4.9. — Control loops that are intentionally not in service due to batch oper-ations, product or grade produced, or trains of equipment that are shut down to be maintained or cleaned, must be flagged as such and other diagnostics automati-cally suppressed. Otherwise, true problems are camouflaged by false alerts and alarms. Also, the “Normal Mode” of loops used in such batch opera-tions should be automatically updated to reflect the needs of the batch sequence.

Guided TourThis tour illustrates the potential ease of use and convenience of an inte-grated interface that is possible for a performance monitor embedded in an industrial control system. The following areas are addressed:

• Summary of measurement, actuator and control performance

• Examination of detail state

• Information provided to an operator


This performance monitoring capability is accessed from any operator and engineer stations within the control system. The primary interface is illustrated in Figure 4-12.

From the explorer view at the left of the interface, the user may select the entire plant or an individual process area to examine. Based on this selec-tion, a summary is provided that shows the number of modules that are being utilized for control, monitoring, and calculation within the selected area. Also, a summary is provided of the modules that have an I/O or con-trol block with abnormal conditions. The individual modules that have an abnormal condition are listed in the right portion of the view in the Sum-mary tab. The types of problems that have been detected are shown using a frown face.

When a module is selected in the summary list, the individual I/O and control blocks in that module are listed in the bottom right portion of the interface. The percent time that an abnormal condition existed is shown for each block. If this time exceeds the defined limit, then this condition is flagged as an abnormal condition in the interface; that is, a frown face is shown for the module. Through the Filter selections, it is possible to view information for the previous or current hour, shift or day. Also, the user may select the type of blocks and whether all modules or just those with abnormal conditions are displayed.

The following types of abnormal conditions are automatically detected:

• The actual block mode is not in normal (designed) mode of operation

• Control action is limited by a downstream condition

Figure 4-12. System Performance Monitoring Interface


• Input status indicates the value is limited or has a Bad or Uncertain status

• The calculated variability index and standard deviation exceed their limits

In many plants, it is common for a piece of equipment to be shut down at the end of a batch or to take it out of service for maintenance. During these periods of down time, the measurements on that equipment may indicate a Bad status. To prevent this information from being included in the performance monitoring view, the user may select an area and then indicate that the associated equipment if off-line. Such areas are clearly indicated in the view of the equipment and are moved from monitoring. An example where the RSU unit has been removed from monitoring is shown in Figure 4-13.

The user may find out more information on an abnormal condition by using the detail tabs in the interface. For example, by selecting the Mode tab, further information is provided on the actual and design mode of the blocks within the module, as shown in Figure 4-14.

Figure 4-13. Disabling a Selected Area of Monitoring

Figure 4-14. Detail Information for an Abnormal Condition


By selecting the Print icon, the user may choose to print a module sum-mary report or a detail report. Also, to allow the operator to view the per-formance and utilization information from his displays, a standard function block and dynamo are used to include this information in the operator interface. Through this dynamo an operator may also enable and disable the monitoring in the associated area. This dynamo appears in an operator display as shown in Figure 4-15.

TheoryThe ability to quickly inspect control and measurement loops has a pri-mary importance in industrial applications. Both poorly tuned loops and malfunctioning field devices jeopardize product quality and production. In the last decade this problem has received significant attention from both academia and practitioners. Much of this work has focused on an assess-ment of control loop performance using minimum variance controller as a reference; see Harris [4.1] and Desborough and Harris [4.2] and [4.3]. Bea-verstock et al. used heuristic definitions of unit production performance tailored to specific applications, rather than loop performance [4.4]. Numerous other researchers have advanced performance monitoring. In particular, Rhinehard [4.5] explored simple ways of computing standard deviation from measurement-to-measurement deviations and filtering. Harris et al. [4.6] and Huang et al. [4.7] provided guidelines for perfor-mance assessment of multivariable controllers. Qin reviewed recent works on performance monitoring in [4.8]. Shunta gave an excellent practical summary of the performance monitoring in the monograph [4.9].

Automatic detection of abnormal operating conditions is possible with Fieldbus architecture and Function Blocks (FB), features available in mod-ern control systems. In older control systems, a monitoring application

Figure 4-15. Operator Access to Performance Information


may often be added as an OPC client. Based on these features, it is possible to quantify loop utilization, “bad” or “limited” measurements, limitations in control, and process variability. In addition, for control loops the poten-tial improvement in loop performance is calculated. This functionality is possible due to the Function Block features, defined by the Fieldbus Foun-dation. Every function block has specific parameters for current block input or output parameter status, block mode and total and capability standard deviation. These block parameters are used by process monitor-ing applications and for creating summary reports for hour, shift and day.

Major manufacturers of process instrumentation in the process industry have introduced a new generation of field devices based on the standards established by the Fieldbus Foundation. To allow interoperability of devices from different manufacturers, a common implementation is sup-ported using a standard set of function blocks for measurement, control and calculations. A key concept in allowing control and calculations to be distributed between devices or between a field device and a controller is a common implementation of block mode supported by each function block. Also, each input and output of a function block is designed to pro-vide a status attribute along with a value, as illustrated in Figure 4-16.

The status associated with each function-block output gives a direct indi-cation of the quality to the measurement of control signal. Quality is defined by describing the measurement as suitable for control (Good), questionable for use in control (Uncertain) or not suitable for use in con-trol (Bad). In addition, the status provides an indication of whether a mea-surement or control signal is high or low limited. For example, if a measurement is operating above its calibration range, the quality of the measurement may be shown as “Uncertain.” Downstream blocks may

Figure 4-16. Block Parameter Status and Mode are Utilized in Performance Monitoring

Function Blocks Support Mode

Function Block Inputs and Outputs Provide Engineering Unit Value and Status


provide a status of “Good Limited” to the upstream control block to indi-cate that a valve is at its upper or lower end of range.

The modes of Fieldbus blocks consist of four attributes, which may be very valuable in examining the operation of a control block. When a con-trol block is initially configured, the “Normal” attribute of the mode parameter is defined based on the design mode of the block. During nor-mal operation, the operator sets the “target” mode attribute to enable or disable automatic control and to indicate whether the operator or another block or application has control of the block set point and block output. The block reflects the mode of operation it can achieve in its “actual” mode attribute based on the target mode and the status of inputs to the control block. By comparing the actual mode to the normal mode attribute it is possible to determine if a function block is operating in its designed mode of operation.

By continuously monitoring the status and mode supported by Fieldbus function blocks, it is possible to automatically determine the following abnormal conditions in control and I/O function blocks:

• Bad I/O — the status of the block process variable (PV parameter) is “bad,” “uncertain” or “limited.” A sensor failure, inaccurate calibration, or measurement diagnostics have detected a condition that requires attention by maintenance.

• Limited (control action) — a downstream condition exists that limits the control action taken by the block. Such limitations may prevent the loop from achieving or maintaining set point.

• Mode not Normal — The actual mode of the block does not match the normal mode configured for the block. An operator may change the target mode from normal because of equipment malfunction.

The percent time that these conditions exist over an hour, a shift, and a day is computed for every block and compared to a configured global limit for each condition. When one of these limits is exceeded, the associated module is displayed in the main summary display.

It is possible that the status of all inputs to a control block is normal and the mode of the block is correct and yet the control provided is poor. No indication is included in the standard block defined by Fieldbus that directly indicates this problem. However, statistical techniques exist that allow the quality of control to be determined in a reliable manner. Leading companies in the process industry use such techniques to evaluate the per-formance of control loops[4.9]. Based on a knowledge of total and capabil-ity standard deviation of the control measure, illustrated in Figure 4-17, it is possible to compute a variability index for control measurements that


compares current control performance to the best achievable for the pro-cess dynamics.

However, to accurately determine total and capability standard deviation, it is necessary that the measurement of fast processes be sampled at a rate that can not be achieved by the communication networks of most DCS sys-tems used in plants today. Thus, in the past, the application of this technol-ogy has been limited to the analysis of slower control loops or to special dedicated tools that support the sample rates required for faster processes. It is now possible for the manufacturer of the Fieldbus device and modern control systems to build these calculations into the control and I/O func-tion blocks to support control analysis. By taking this approach, the con-trol of even the fastest process may be accurately evaluated.

Using Statistics for Control Performance EvaluationThe valuable role that statistics can play in evaluating control performance is well known. Standard deviation of the process measurement (Equation 4-1) is commonly used and is an easy-to-understand measure of process variability

(4-1)

where n is the number of sample, typically n = 120.

Figure 4-17. Capability and Total Standard Deviation for Process Variable

“Capability” Only Random or Short Term Variability

“Total”: All Data Including Short and Long Term Variablity

Process Value

Time

Stot

Xi X–( )2

i 1=

n

∑n 1–

--------------------------------=


(4-2)

When loop performance is of concern, then standard deviation alone may not provide sufficient information to evaluate loop tuning. To gain an objective judgement, a reference value for the process control performance is required. The best performing feedback control theoretically is mini-mum variance control, Sfbc. This value may be calculated [4.9] directly based on the knowledge of the total, Stot, and capability, Scap, standard deviation of the process measurement as shown in the Formulas 4-3 and 4-4.

(4-3)

Scap is calculated from process samples XI as :

(4-4)

Based on a the value of total standard deviation and standard deviation for minimum variance control, it is possible to define a variability index for the control loop that reflects how close control performance comes to minimum variance.

Variability Index (4-5)

where s is the sensitivity factor.

Finally, for control blocks, overall control performance is as follows:

Performance Index (PI) = (100-VI) (4-6)

The overall control utilization is an indication of what percent of time loops were in normal mode:

Utilization = Average Percent Time in Normal Mode

To support the VI and PI calculations, the I/O and control function blocks in the scalable system and Fieldbus devices calculate Scap and Stot and make them visible as block parameters. Intermediate calculations are done

X

Xii 1=

n

∑n

---------------=

Sfbc Scap 2ScapStot----------

2–=

Scap

Xi Xi 1––( )2

i 2=

n

∑2 n 1–( )

----------------------------------------=

100 1 fbctot

S sVI

S s+

= − +


each execution of the function blocks. The parameter values are then updated per every n executions of the function block. For a typical imple-mentation, the update is done after 120 executions of the function block.

To simplify calculations and memory requirements, the total standard deviation, Stot,, can be calculated by mean absolute error, MAE, computa-tion as follows:

(4-7)

where:

(4-8)

The measurement value is used in I/O blocks to calculate the mean value. In control blocks, either the working set point or the measurement value is used depending on block mode.

The relation between standard deviation and a mean absolute error can be verified by computing a mean absolute value for the normalized Gaussian distribution, p(x):

(4-9)

(4-10)

Since for the normalized Gaussian distribution standard deviation σ=1

(4-11)

The capability standard deviation Scap , is calculated as follows:

(4-12)

where:

(average moving range) (4-13)

Only the summing component associated with the MAE and MR is done each execution. The division of the sum by N or N-1 is done as part of the

Stot 1.25MAE≅

1|))((|1 t

N

ytyN

MAE −= ∑

∫∞

∞−

−= 2

2

21)(

x

expπ

7978845.221

21||

21||

0

2

0

2

0

2

0

22

22222

==

−=

+−==

∞−

∞−

−∞

−

∞−

−∞

∞−

−

∫∫∫ ππππxxxxx

eexexeexx

||2533141.1 or 7978845.|| xx == σσ

ScapMR

1.128-------------=

2|))1()((|

11

−−−

= ∑ tytyNMRN


Stot, and Scap calculations only once every N executions (default N=120). Capability standard deviation calculation requires that the sampling rate be fast enough. The requirements for the sampling rate are similar to the scan rate of a control loop. A practical estimate for selecting scan rate for control loops is sampling five or more times per time constant. [4.9].

Extending the Concept to the Multi-variable EnvironmentIn many products designed for control and measurement system perfor-mance evaluation, the scope of detection is limited to traditional I/O and traditional control capability. This capability has been extended to address I/O and control blocks in Fieldbus devices and to provide interfaces and analysis tools for high-speed trend collection, device diagnostics and tools for commissioning control. Since MPC control and Neural Network capa-bility are being embedded in modern control systems, it is a natural exten-sion of this work to include performance evaluation support that includes advanced control.

For the purpose of performance evaluation, advanced control can be defined as “monitoring or control applications that use multiple process measurements to produce one or more outputs that are used directly or indirectly to improve the operation of a plant.” As illustrated in Figure 4-18, this definition fits the advanced control applications that are currently used in industry today: model predictive control, fuzzy logic control, adaptive tuning, soft sensors based on neural networks, plant optimizers (such as RTO+), multi-stream blending.

In spite of the diverse technologies associated with each of these applica-tions, their requirements for correct operation are basically the same as those for base-level measurement and control.

Figure 4-18. Example of an Advanced Control Applications

TT 1-1

FT 1-2

AT 1-2

Neural Network

AT 2-1

FT 2-9

TT 2-5

FT 3-8

TT 3-8

FT 4-3

FT 4-6

LT 4-6

FT 6-2

FT 6-6

FT 6-8

FT 5-7

AT 5-8

FT 5-8

Model Predictive

Control

Adaptive Tuning

BlendingReal Time Optimization

MultivariableFuzzy Logic

Control

Control


• Valid measurements (good and not limited).

• Control action is not limited by downstream conditions.

• The operator and operating conditions allowed the algorithm to operate as designed; i.e., the application is in the correct mode of operation.

• Control or calculation objectives were met.

When compared to the measurement and control-blocks requirement addressed in single loop control, the primary differences are (1) multiple inputs (rather than one) are used for the control or calculation and to indicate downstream limitations, and (2) the measurement of control performance is based on multiple inputs and targets.

Thus, to extend system evaluation to address advanced control applica-tions, the techniques used for single loop control to determine abnormal input, control limited, incorrect mode, and control index exceeds limit must be generalized to include multiple inputs and targets.

Addressing Advanced ControlThe detection of abnormal conditions or poor performance of an advanced control application should take into account the multiple inputs and tar-gets used in the control or calculation. Also, in many cases multiple inputs must be examined to determine if control has been limited by downstream conditions. Thus, the four basic indications used by system performance evaluation to determine abnormal or under-performing control must be generalized to address the following:

1. Inputs used in the application are abnormal (“bad,” “uncertain,” or “limited”).

2. Inputs associated with downstream conditions indicate control action was limited.

3. The algorithm was not able to operate as designed; i.e., its mode of operation was not normal.

4. The process output did not meet the target-operating objective.

In this section, the manner in which each of the calculations may be generalized to include multiple inputs and targets is detailed.

Abnormal InputsFor a single input control block such as the PID or FLC control block, the status of the primary input (PV) of the block may be used to determine if the input is abnormal. For example, if the PV status is uncertain or “bad”


or is “limited,” then the block will be flagged as having an abnormal input. In advanced control applications in which multiple inputs may be required for correct operation, any of these inputs that have a status of uncertain, bad or limited may prevent the control from achieving its goal. Thus, for advanced control applications, the indication of abnormal input(s) will be based on all inputs that directly affect the control or calcu-lation. This may be calculated as follows:

Abnormal Input Condition Advanced Control =Abnormal Condition on Input I or block

where:

= a logical sum of I = 1, …, N Conditions

Control Limited Where advanced control is providing multiple process inputs, the control interface to the process will be done through I/O or control blocks. Both provide an output that will be used as a back-calculation input to the advanced control application. If the control action taken is limited down-stream, then this is reflected in the status of the associated back-calculation input. Since the control objective will not, in general, be met if any of the control outputs becomes limited, then the indication that control is limited must consider the back-calculation input status associated with all control outputs. This may be calculated as follows:

Control Limited = Control Action (I) is downstream limited

where:

= a logical sum of I = 1, …, N Conditions

Incorrect ModeFor single input/output blocks, the status of the primary input, cascade input, back-calculation input, and target mode must be used in determin-ing achievable mode of operation. This mode of operation is reflected in the mode parameter as the actual mode attribute. When a block is config-ured, the customer may indicate the normal mode that the block is designed to operate in. By comparing the actual mode attribute to the nor-mal mode attribute for the block, it is possible to determine whether the block is operating in its designed mode. Advanced control applications may be engineered to use the standard status and mode definition. To cal-culate mode, the status of all inputs used i

guidelines for evaluating control system performance · mitters, or changes in the operation of the...

Documents