2
M. Bauer et al. / Journal of Process Control 38 (2016) 1–10
withseveral hundreds if not thousandsof PID control loops digitally implemented. With the advance of automatic control came the need to assess the performance of the control loops, according to the business stratagem “If you can’t measure it you can’t manage it.” Increasing computational power enabled assessments which looked, for example,at thestandarddeviation of process trends. Standarddeviation or variance is an obvious candidate as a performance index because many engineers selling control solutions usea depiction of data “before” control improvement and “after” where the “before” case shows a heavily fluctuating process variable and the “after” a near constant smoothened out time trend. A key challenge in the assessment of control loops is to distinguish between a controller that was performing poorly (A) because there was something wrong with it and (B) because of an outside disturbance. This was first addressed by Harris in 1989 when he published “Assessment of closed loop performance” [2]. The measure, later to be named the “Harris index” compares the variance of a process variable to the minimum achievable variance, which is caused by outside disturbances. This paper attracted a significant amount of interest in the academic world of control engineering and brought the problem formulation closer to research institutions. A reason for the focused attention on the assessment of control loops was that despite the prevalence of PID controllers ensuing studies found that the controllers were not doing as well as everyone had assumed [3]. In fact, before a measure was introduced, controllers were just used. Now these controllers were “good” or “bad”, “acceptable” or “poorly performing”. And most studies found that there more “poor performers” than expected. A key requirement for the assessment of control loops is that data from routine operation and closed loop control should be used. The initial assessment of a single control loop has since then expanded to frameworks and procedures, and includes diagnosis, fault identification, isolation, and root cause and plant wide disturbance analysis. This research area today is often called control loop performance monitoring (CPM). CPM now forms a substantial body of research articles and industry applications. Over the 25 years since the publication of the original article, several review articles and tutorials have been published. The first reviews were written about a decade later [4,5] f ocusing the multivariate extensions,feedforward control and industrial aspects. Hägglund [6] as well as Jelali [7] made significant contributions by explaining working indices in plain terms and providing codes for industrial implementation. More recently, the Springer series “Advances in Industrial Control” includes four monographs on control performance assessment, including valve stiction detection [8–11]. Furthermore, the recent textbook [12] explains established methods in detail and gives frameworks, implementation guidelines, applications and tools. CPM was developed in close cooperation with industry. In [7], a list of research articles and their industrial applications in the chemical, petrochemical, pulp and paper and other industries in provided. The same article also lists the commercial packets that comprise control loop monitoring tools, either as stand-alone solutions or built into automation and control software. In this article, we are scrutinizing what has been achieved in industry in the last three decades. The focus is on the production companies in the chemical, oil and gas, pulp and paper and other industries that use CPM to manage and assess the control of their processes. The questions addressed here are: •
• •
What works in industry?Whichmethods aremostuseful? Which frameworks and processes are successful? Is CPMa standard or only used by leading production companies? What are the keychallenges? What are theopen research topics?
These questions were gathered into an online survey that was distributed to control engineers at production plants in various industries and around the world. Since a survey requires a comparatively large number of participants we have focused the content on single control loops and particularly on PID loops. The results of this survey amongst CPM practitioners are presented in this article. The methodology of the survey and the background information of the survey respondents are described in Section 2. Section 3 discusses the awareness of CPM and the scope: how well are control loops performing today? Is there still a need for CPM? In Section 4, the prevalence of the various methodsas well as problems that canbe addressedby CPMare discussed while Section 5 investigates procedures, frameworksand workflow that further thesuccess of CPMin industry.Section 6 looks at future research directions and the satisfaction of CPM users with current tools and methods.
2. Survey description An online survey was conducted to capture the prevalence of CPM worldwide. There is an abundance of publications on survey methodologies for all survey purposes and groups of respondents [13]. Many providers offer free platforms to easilyformat questionnairesand capture results. Forthissurvey,the authors chose Google forms, which is part of Google drive and does not require any software installation. In addition, the results are stored in spreadsheet form and reports are generated automatically. The design of the questionnaire is the most important aspect of the survey. When putting the questions together many pitfalls have to be avoided. For example, the questions have to be phrased objectively and clearly in a coherent order. Non-exhaustive listings must beavoided.To ensurea high response rate,the questionsmust be meaningful and interesting [14]. To ensure thevalidityof thequestions, interviews with industry experts were conducted and the survey questions were discussed. These experts were Florian Wolff at BASF, Germany, Duane Muller at AngloAmerican, South Africa. For this type of survey the group of respondents are limited. The respondents were identified as lead control engineers in production companies from various industries. In order to address the target audience, several approaches were started. First, all personal contacts of the authors were approached. Second, what is referred to as ‘snowball sampling’ was pursued, that is, known responding control experts were asked for referrals among their colleagues. Thirdly, published authors in the area of CPM that now work in industry were approached. The contact data was retrieved from the journal article or conference proceedings. In addition, the survey was distributed in a Honeywell user group meeting and the participants filled the results in during the meeting. All responses, electronic or on paper, were anonymous. In total, 69 control engineeringexperts in production companies answered the survey. Fig. 1 lists the respondents by continent and by industry. Roughly half the respondents were from Europe (33 out of 69) because the authors’ contact were used to send out the questionnaire. The majority of respondents (64%) work in chemicals or oil & gas. This may be partly explained because chemical and petrochemical companies are traditional strongholds of CPM. Itshould benotedthattheanswers donot alwaysaddup 69because not all respondents answered all questions. The respondents have various levels of experience in control engineering, as indicated in Fig. 2. The total is about 1000 years of control engineering experience on which the survey is based on. Respondents were also asked how many loops are allocated for each control expert and the results are displayed in Fig. 3. On average, a control engineer is responsible forabout450 loops.However,
M. Bauer et al. / Journal of Process Control 38 (2016) 1–10
3
Fig. 1. (a) Respondents by continent; (b) respondents by industry.
www.processcontrollerperformance.com overview of the responses.
together
with
an
3. CPM pervasiveness
Fig. 2. Number of years respondents have worked in control engineering.
Fig. 3. Numberof PID control loops per control engineer.
the number of loops ranges from 30 to 2000. The work environment and the attention that can be paid to each loop is therefore tremendously different depending on the plant and application. In general, high or low numbers of loops do not correspond to certain industries. For example, a control engineer working in a chemical plant may be responsible for 50 loops or for 1000, depending on the application, on company policy and on the organization. The focus of the survey is on CPM users. Significance testing of such a small samplegroup is somewhatlimited.The results arenevertheless interesting and well worth reporting because of the vast experience of the respondents. Similar surveys conducted among control experts had a comparable sample size [15–17]. The survey results will be discussed throughout this paper. The questionnaire is available on the following website
About two thirds of the survey respondents already use CPM tools or packages. This may not be representative of the industry since a control engineer will have been more likelyapproachedif he is known for in the CPM community. However, CPM tools are relativelyrecentlyinstalled, in most cases they have been in use forless than ten years, see Fig. 4. On average tools have been implemented a decade ago, that is, in 2006. This is especially noteworthy when comparing it to the experience of the respondents in Fig. 2 where the average experience of the respondents is 14 years. This means that CPM has been introduced during the working lifetime of the respondents. From the number of recent installations we can conclude that in general the installation of CPM tools has increased. This leads to the question f control loops are performing better today than beforetheonsetof CPMin theearly 1990.Since then,several studies documented that control loops are not doing their jobs [3,18]. Over the past three decades, tools have been developed by control and automation system companies, by third-party software companiesand, in thebeginning, byuser companiesthemselves.As software packages become moresophisticatedand get easier access to historical data, own developments fall away and are outsourced. Today, automation companies can offer integrated solutions and small companies or divisions provide tailored services and consulting. In addition, process data is more easily accessible. A consequence is that in-house developments fall by the wayside.
Fig. 4. Years CPMtools have been used in theorganization.
M. Bauer et al. / Journal of Process Control 38 (2016) 1–10
4
4. CPM methods, time trends and faults
Some conclusions that can be drawn from Fig. 5 are as follows:
The organization and categorizing of CPM methods depends significantly on the viewpoint of the CPM practitioner and expert. For a comprehensive list of CPM methods see for example [12]. Generally, CPM methods can be grouped according to three criteria: 1. Mathematical method or algorithm. 2. Nature of time trend analyzed. 3. Type of fault or malfunctioning addressed.
•
•
•
All methods can be grouped according to each category, however, some are strongly associated with a particular method (e.g. PCA), time trend (e.g. oscillation) or fault (e.g. sluggish tuning such as the Idle index [19]). It is often convenient to group the indices according to their main functionality, which can be either method, time trend or fault. For example, some indices may be referred to as ‘oscillation index’ while others may be labeled ‘stiction index’. The following sections present the survey results on the usefulness of methods as well as the occurrence of typical time trends and faults. 4.1. Mathematical methods All control experts were asked to rank certain methods according to their usefulness. The CPM methods were grouped into seven categories to keep the time to answer the survey within a reasonable limit. The authors felt that it would neither be feasible nor beneficial to assess each method individually. The results are presented in Fig. 5 and the provided more detailed description of the categories are given in Table 1.
•
The simpler the method the higher the number of respondents who find them useful. Operating mode statistics are easy to compute – though not always easily accessible since they are not alwaysstored in thedatahistorian – andequallyeasyto interpret. Multivariate or advanced statistics are more difficult to compute and to interpret. All methods are very useful to some CPM users, each user probably has his or her favorite method to use. Algorithmic indices have still untapped potential and could be promoted further. Most respondents who know about these indices find them very useful, however, they are unknown to about a quarter of the respondents. The impact which minimum variance indices, in particular the Harris index, had in the academic control community is not matched in the industrial community. The reason may be that the Harris index works well to explain the idea of CPM and on certain data trends but is meaningless if the data contains non stationary signals.
4.2. Nature of time trends Ideally, the time trend in regulatory control is stable or constant and lies within a certain operating region. In reality, the trends are never smoothbut show certain degrees of variability.Thereare several features or disturbancesin thetime trendsthat an experienced control engineer will recognize by visual inspection.Generally, disturbances can travel through a process through physical or data connections. Thus, a disturbance can be picked up in a number of control loops. Most CPMtechniques assess thecontroller performance by analyzingthe time trend.Typicaltimetrendsare shown in Fig.6 andthe
Fig. 5. Methods and howuseful respondents perceive them to be. Table 1 Methodcategories and methods that are grouped under this category. Category
Methods
O per at ing m od e s ta ti st ic s
T ime i n m anu al ve rs us t ime i n a uto , n umb er of op era to r i nt er ve nt ion s, number of tuningchanges,loops in saturation Minimum and maximum threshold crossings, mean, standard deviation Idle i ndex, i ntegrated a bsolute e rror ( IAE), s catter p lots o f m anipulated variable versus controller output (MV/OP plots) AR (X ) m od el s, a rt ific ia l i nt el li ge nc e ( AI) , s ti ct ion mo del s, H amm er st ei n m od el s Correlation, principal component analysis (PCA), partial least squares (PLS) Ha rri s i nd ex a nd r el ate d i nd ic es Gaussianity testing, probability density function, surrogate data analysis
Basic statistics Algorithmic indices M od el -b as ed te chn ique s Multivariate s tatistics M ini mum va ri an ce te chn ique s Advanced stati stics
M. Bauer et al. / Journal of Process Control 38 (2016) 1–10
5
Fig. 6. Time trendsof process variable and controller outputfor commondisturbances.
Fig. 7. Sample time trendsof faultsand their rating, howfrequently they occur.
prevalence of these time trends according to the survey are given in Fig. 7. In the following, each time trend is briefly discussed. • Saturation:
Saturation manifests itself at the controller output which is intermittently at its maximum or minimum (usually 0% or 100%). It will deviate from its extreme for a short while only to returnback to it.Saturation canresult from a poorlydimensioned actuators, poor controller tuning or may be caused by integral wind-up of PID controllers after a large setpoint change occurred [12]. • Sinusoidal oscillations: Oscillations are introduced in process variables by a number of means such as poor tuning settings, actuator faults or external disturbances. Generally,as oscillations become sinusoidal as they travel through a process via physical connections since the process acts as a low pass filter. • Manual control: If the controller is not acting satisfactory the operator or control engineer often switches it off. This is seen in the control system as a switch from automatic to manual mode. • Sluggish behavior: An unwanted behavior seen in the time trend is labeled as sluggish or slow. This means that theprocessvariable only slowly follows the setpoint. There is no overshoot after a setpoint change. The reason for sluggish behavior lies usually in the tuning settings. • Nonlinear oscillations: Some control loops contain a nonlinear element such as a faulty valve and thus cause nonlinearity in the time trend of process variable and controller output. A nonlinear
•
disturbance affects the process variable regularlyand repetitively but contains higher frequency components and can therefore be distinguished from a sinusoidal oscillation. Quantisation of process variables: All signals are digital in amplitude. The resolution of the signal is determined by the sensor. The sensor may have a coarse resolution so that the time trend shows step-like features. This often applies to temperature sensors which could have, for example, a resolution of 0.5◦ while the temperature only fluctuates by ±1◦ .
Generally, theresults shown in Fig.7 arenot new – control engineers in the 1990s looked for similar features in the data then as today: Which controller outputs are at their limits, which loops are oscillating and which are in manual control. Somewhat surprising in Fig. 7 is that saturation is the most frequent problem. Detection of saturation is straight forward. More intricate is the identification of the cause of saturation since this is often an indication of actuator dimensioning but can also occur when tuning settings are incorrect or no anti-windup algorithm is implemented. A similar issue is manual control which should always be investigated as it shows that there is an underlying problem with the loop. It means that the operator, for various reasons, does not trust the controller to do its job automatically. The root causes range from poor tuning over interacting loops to training needs.
6
M. Bauer et al. / Journal of Process Control 38 (2016) 1–10
sensorfaultsshould notbe pickedup bythe control loop monitoring system but by the maintenance personnel only. The degradation of process equipment as well as of actuators and sensors is strongly linked to condition monitoring, asset management and maintenance management. Despite several efforts to integrated control loop performance monitoring (for example, ISA95) with these neighboring disciplines, control loop performance monitoring is often conducted independently. This appears to be morebecause of workflow,data exchange andorganizationalissues rather than methodologically founded. Fig. 8. Control loop elements and associated faults accordingto the survey results, see Fig. 9. The height of thecolumn reflects theprevalence of thefault.
Though infrequent, it is surprising that quantization still occurs in this decade. The cause should not be technological but rather poor calibration of the sensor. In some instances, the responsible control engineer does not have the right to change sensor settings. 4.3. Prevalence of faults When controlling a process, several things can go wrong. Fig. 8 shows a control loop with its main components: controller, actuator, process and sensor. All of these components can cause a disturbance and the degradation of the control loop performance. The most frequently occurring faults are depicted in Fig. 8 showing that the actuator followed by the controller are the dominant causes for problems in the loop. The size of the columns reflects the prevalence of the faults according to the survey as shown in Fig. 9. The complete survey results as partly indicated in Fig. 8 are shown in Fig.9. For thefourmost prevalentfaults actions to remedy the cause can be directly inferred: A poorly tuned controller can be re-tuned either manually or automatically, a sticky valve will have to be replaced either immediately or in the next scheduled shut-down, actuators and sensors are replaced. Other faults may require more discussion and action to be taken are not as straight forward. Interacting loops for example are difficult to detect and often require a complete rethinking of the control strategy. Arguably, the picture of the most prevalent loops is still the same as in 1989. This means that although the overall control loop performance may have improved the same faults still occur. This is particularly surprising for the most prevalent cause of ‘wrong tuning settings’ since most modern controllers are equipped with an auto-tuning feature that should eliminate the problem. With the advent of smart sensors faults of the sensors should become less important in the future because the sensor will detect and diagnose faults and send a message to the system. Thus,
5. CPM framework, workflow and implementation Many authors have noted that CPM can only be successful if it is embedded in the company’s existing operation and workflows. For example, a control engineer will attend to the Top 10 poorly performing loops if his salary or bonus is linked to it. Wolff and Kramer [20] notethatCPM is a cycle of improvement.Analysis must be followed bythe derivation of corrective actions,whichthen need to be implemented before the next analysis is carried out. A critical aspect of CPM is to involve all key personnel that are affected by the performance of the controllers and the collaboration between different roles. Jelali [7] proposes a framework for implementing CPM methods and lists the main corrective actions as re-tuning, re-design or maintenance. Training is possibly a further corrective action. In some instances, the controller may do a good enough job but since the operator does not trust the algorithm, she may put the controller into manual. Further education on the process, control designor tuningmay be required.The training on CPMitself is criticalfor success. However, it is near to impossible totrainall involved personnel on interpreting the CPM results of individual methods. Instead, it is paramount that the CPM gives guidance toward the nature of the root cause of a poorly performing loop. The workflow is closely related to the implementation of the CPM solution, that is, as a stand-alone solution or even on the process control system. Survey results addressing both workflow and implementation are presented in this section. 5.1. CPM workflow Monitoring the performance of controllers is deeply embedded in theroutine operationof theplant. Plant personnelhave to ensure the continuous, profitable andsafe operation of the plant for which regulatory control is a key aspect. However, monitoring the performance of control loops is not necessarily part of the day-to-day
Fig. 9. Problems addressed by CPM.
M. Bauer et al. / Journal of Process Control 38 (2016) 1–10
7
Fig. 12. Prevalence of Top X lists of poorly performing loops. Fig. 10. Number of personnel involved in CPM.
work plan. It has to be integrated differently depending on the process and organization. Key questions when scrutinizing the CPM workflow comprise the following: • • •
Who assesses the control loops? How often are loops assessed? What are the actions following the assessment?
These aspects were investigated in the survey. First, a question concerningthe people involved wasasked: Who in the organization is involved in CPM? The results are presented in Fig. 10. The results show that these are mainlycontrol engineers, followed by APC engineers and process engineers. Maintenance staff and operators are not as often involved in the discussion of control loops though they often are affected by the results. Other staff involved in CPM includes KPImanagers, systemtechnicians or chemical engineers. Most of the time (70%) control engineers alone or one additional department are involved in CPM. When developing methods it is important to keep the collaborative aspect in mind. This can specifically mean that the results are distributed to all involved parties in an easy to understand format. The second question concerning the workflow relates to the frequency with which control loops are assessed. The results are shown in Fig. 11. Most plants conduct a CPM assessment on a weeklybasis often inconjunction with a fixed meeting.Others carry out the assessment less frequent while others do daily monitoring or assess the performance continuously and address issues as they arise. A key aspect of the CPM workflow that was captured in the surveyconcernsactions that aretriggered by theCPM results. This was asked as an open ended question. Most answers related to one of the following actions resulting from the CPM assessment: • Tuning
initiatives Maintenance, in particular referring to valves • Simple operating changes • Control configuration changes •
Fig. 11. Frequency of assessment of control loops.
Several respondents stated that CPM is used before the installation of advanced control solutions to ensure that the baseline is functioning sufficiently before installing high level control. In some organizations CPM is a formalized procedure with weekly meetings and follow-ups. More importantly, performance appraisal may be linked to CPM results giving monetary incentives to staff. On the other hand, CPM is carried out on an ad-hoc basis and adds to the already high workload of control engineers and other plant personnel. Very often, the loops that have been flagged as poorly performing are grouped into Top X lists, that is, a list of the five (Top 5) or ten (Top 10) malfunctioning loops extracted. These are usually the worst performing loops. The Top 5 loops for a process unit may be identified and discussed in a weekly meeting between the operators and thecontrol personnel. In thesurvey, the respondents were asked which Top X list they find most useful. The results are displayed in Fig. 12. This shows that most respondents prefer a Top 10 list which appears to be a balance between a manageable work load and a significant number of relevant control loops to be addressed. An important aspect of Top X lists is to incorporate the significance of the control loop. That is, in addition to a performance index such as the Harris index, the loops have to be assessed by their importance in the process. For example, a loop controlling the outside slurry feed might be performing very poorly. This will be a known fact to the operator and control engineer. However, no actions aretakento fix theloopsince theslurry feed is notcriticalto the production. Thus, loops have to be ranked by their significance to the overall production objectives. This requires a combination of detailed process knowledge and understanding of the control configurations. 5.2. CPM implementation The focus of this study is on plants with modern process control systems (PCS) in place. There are still many plants worldwide with old PCSs and a low level of automation. CPM requires a modern PCS and a connected data historian since it heavily relies on the logging of process data for most analysis methods. The survey deliberately did not include any mentioning of particular solutionsoffered by CPMcompanies.The reasons forthis are as follows. The number of respondents is not high enough to allow a meaningful ranking of anysort, regarding the most popular or the best solution.Also,the solutionsare notcomparableenough so that respondents can answer thequestions in a meaningful way. A third reason is that two of the authors are associated with a vendorcompany and may be accused of a biased view. A list of commercially available CPM solutions and their key features can be found in [7]. From an industrial application perspective, there are two different ways of supplying the product of CPM namely as either a product or a service. CPM as a product is when the solution is delivered as a software that is integrated with the data historian. The software solution is used by the end user, not the CPM supplier. CPM as a service means that regular reports are computed by a service provider and sent to the end user. The main difference is
8
M. Bauer et al. / Journal of Process Control 38 (2016) 1–10
Fig. 14. Satisfaction with CPM tools: Howsatisfied are youwith the current tools? 0 – notsatisfied, 9 – completely satisfied.
they see themselves as a service provider, while respondents who have purchased a commercial packet regard CPM as a product. Fig. 13(c) shows that CPM is predominantly implemented as a stand-alone solution and not on the control system (PCS or DCS). This is a very important question concerning the ease of use and whether the report can be updated easily. An advantage of running CPM on the PCS is the interfacing the data repository. However, in many instances the PCS might not be available to the user of CPM. Also, PCS and DCS are often difficult to access and do not provide a platform onto which additional applications can be easily implemented. Fig. 13. (a)CPM as product versus CPMas service,(b) CPMas in-housedevelopment versus commercial solution,(c) CPM used on PCS, as stand-alone solution.
that CPM as a service has to be more automated and standardized because it has to be interpreted by non-expert users. The report should highlight critical loops and underlying causes and suggest corrective actions. This means that eitherthe methods areclear, the results have to be definite, and that there is little room for interpretation and discussion. Fig. 13(a) shows the result of the survey: the split between product and service is nearly equal. This means that there is a need for both automated methods and methods that leave room for discussion. In the survey, the respondents indicated whether they use a commercial packet, their owndevelopmentor a combination of the two (Fig. 13 (b)). It is noteworthy that respondents using their own development predominantly regard the solution as a service, i.e.
6. Open research topics After more than 25 years of active research in control performance monitoring, it is of great interest to ask whether there are any relevant research questions left open. Here, we address this question from the industrial application perspective. Fig. 14 shows this level of satisfaction which appears message. Since most of the respondentsuse CPMthe message from this resultis mainly: People who use it see an advantage of the tools. Most respondents see value in applying CPM to their industrial processes and are reasonably satisfied with the current tools. At the same time, there appears to be improvement potential because some respondents arenot satisfied atall. This maybe because of the specific unsatisfactory tool these respondents are using but more likely because there is a research opportunity. In discussions with end users BASF and AnlgoAmerican areas of open research were identified and clustered. The resulting list of these topics together with the survey findings is shown in Fig. 15.
Fig. 15. List of topics for further research and development (split between what vendors and users think).
M. Bauer et al. / Journal of Process Control 38 (2016) 1–10
Theclearmainresult is that therespondentsask forbetter guidance on corrective actions to be taken. This does not come as a surprise because proper corrective action is a pre-requisite for economic payback of the investment in CPM. The second most important research need relates to the mentioned prioritization of control loops. A corrective action is only relevant if the asset relating to it has a certain importance for the overall performance or criticality of the whole plant. The topics of training,guidance on parametertuning,automated data selection and graphical display relates to the implementation of CPM methods in industrial solutions. This result of the survey fits the current tendency to not program own CPM tools but to try to buy commercial products or services. Commercial tools often offer more advanced functionality for long-term monitoring, data handling, graphical display and training material. An interesting point is thesecond last statementthatCPM methods relates to condition monitoring of rotating equipment. As a matter of fact, there is a certain disconnectbetweenCPM on theone side and traditional condition monitoring Condition monitoring refers to fault detection for mainly mechanical process equipment. If applied to the same process at the same time, both are often dealt with in different software tools that are managed by different people. Nowadays, there is much more information available for analysis, such as production data, condition monitoring data, operating conditions, plant parameters, etc. Such data will increasingly be accessible also for CPM1 . Hence, methods should be developed that make use of that data fusion. The benefit of integrating different data sources should be an active field of research. This is often the result of recent trends referred to as ‘data analytics’ and ‘big data’. Such gaps can be bridged by research, which brings both technologies closer together, or by organizational structures. The latter can be realized by specific staff that share the responsibility of both aspects in the same plant. Chemical company BASF in Germany which were interviewed initially for the purpose of the survey, has deployed so-call Asset Managers for each of their plants. These asset managers connect plant data from the automation system, the maintenance management system and other ERP data. Other areas such as automatic, data-driven detection of stiction behavior in process control valves has been extensivelyresearched. There are more than 15 different methods published already. Also, efforts have been made to integrate them into an “index-fusion” in order to increase the reliability of detection [11]. While it may be worthwhile to further improve the reliability of existing methods, it does not appear to be fruitful to develop yet another detection algorithm that will outperform the existing ones. A conclusion of the answers about further research direction from an industrial perspective is that ‘simple’ methods work best in most cases. Simplicity is here referring to three aspects:
1. Simplicity of parameterization: Methods that need to individually be tuned, parameterized, trained etc. will most likely not survive in the industrial CPM practice. 2. Simplicity in interpretation of the results: Methods that require significant training, experience of interpretation, do not resolve ambiguities and interpretation guidelines will most likely not survive in industrial CPM practice. 3. Simplicity in computational complexity. Even though computational power still increases, CPM is typically applied often and
to a large amount of control loop data. Hence long computation time still can be a drawback.2
In order to address therelevant research questionsin thefuture, it is recommended that academic institutions collaborate with the process industry, that is, with end users. This will guarantee that therelevant problems aresolved, theresults arefeasibleto be engineered, interpreted and used. Also, new CPM algorithms should be testedon real-lifedata ratherthan only in simulation because often root-causes interact which can hardly be simulated realistically.
7. Conclusions This paper presented results from a survey on industrial application of CPM in the process industries. It has become evident that there is considerable knowledge about control performance present in many process industries. The topic of monitoring and assessing control performance has been important in the past and will remain to be important also in the future. We conclude that after 25 years of intensive research there are still relevant research questions to be solved in the area of CPM. We will here disregard all indirect aspects such as software implementation, usability, presentation of results, data access, and integration into automation systems. They may lead to research effort, however this is hardly specific for CPM application. Staff responsible for control performance in process plants are increasingly given more and more complex tasks so that extensive tool support for CPM has become a standard. The application of such tools, however is still quite heterogeneous. One reason is the different experiences companies have gathered over the years. A second reason is the different maturity of commercial tools, they differ in philosophy (product versus experttool versus service) and in price. There is not yet any standard which methods to use and how to present the results to the users, though attempts have been made [21]. CPM tools are mainly needed to prioritize maintenance actions related to deteriorated control performance. Guidance for corrective action is a vital requirement that is not yet solved sufficiently well.For this,veryoftensimplemethodsseemto besufficient. Rootcauses diagnosing is still an open problem, but, at the present state, one can detect loop malfunctions with a good level of confidence in many situations. One key finding is that CPM is more that the application of numerical algorithms to plant data. Companies that successfully apply CPM have undertaken thorough efforts to integrate CPM into their daily plant operation and asset management. This also includes different work procedures. A final question of interest is if the control performance in the process industry hasbeen improved duringthe last 25 years (orif it would have deterioratedmore without CPM). This question is difficult toanswer.The companiesthatsee valuein CPMobviouslyseem to apply it successfully. In those companies, it can be expected that control is performing on a better level than earlier. There are, however, still many companies neglecting the advances in CPM such that the industry average of non-optimal control performance can be expected to be much lower than desirable. There is hence still significant potential to improve control in the process industry.
2
1 This unlimited access to data is called Industrie 4.0 in Europe and Industrial Internet in North-America.
9
This comment will in the future possibly be less important when CPM calculations largely are done in the cloud. Some companies already offer the experiment with CPM-as-a-Service where thealgorithms are performed in a powerful (local or private) cloud infrastructure.
10
M. Bauer et al. / Journal of Process Control 38 (2016) 1–10
Acknowledgements Theauthors wouldlike to thank Florian Wolff at BASF andDuane Muller at Anglo American for their insight as well as Richard Salliss at Honeywell for distributing the questionnaire at a user group meeting. This paper is a result of many discussions with senior control practitioners and academics. We would like to thank Russ Rhinehart,KevinBrooks, TomEdgar, Biao Huang,Mike Grimble and Gerrit van der Molen for their extremely valuable contributions. References [1] S. Bennett, A brief history of automatic control, IEEE Control Syst. Mag. 16 (3) (1996) 17–25. [2] T.J.Harris,Assessment ofcontrolloop performance,Can. J.Chem. Eng.67 (1989) 856–861. [3] D. Ender, Process control performance: not as good as you think, Control Eng. 40 (1993) 180–190. [4] S.J. Qin, Control performance monitoring – a review and assessment, Comput. Chem. Eng. 23 (1998) 173–186. [5] T.J. Harris, C.T. Seppala, L.D. Desborough,A review of performance monitoring and assessment techniques for univariate and multivariate control systems, J. Process Control 9 (1999) 1–17. [6] T. Hagglund, Industrial implementation of on-line performance monitoring tools,Control Eng. Pract. 13 (2005) 1383–1390. [7] M. Jelali, An overview of control performance assessment technology and industrial applications, Control Eng. Pract. 14 (2006) 441–466.
[8] B. Huang, S.L.Shah, Performance Assessment of Control Loops, Springer,Berlin, 1999. [9] A.W.Ordys,D. Uduehi, M.A.Johnson, ProcessControlPerformanceAssessment: From Theory to Implementation, Springer, Berlin, 2007. [10] M.A.A.S.Choudhury,S.L. Shah, N.F.Thornill,Diagnosis of ProcessNonlinearities and Valve Stiction – Data-Driven Approaches, Springer, Berlin, 2008. [11] M. Jelali, B. Huang, Detection and Diagnosis of Stiction in Control Loops: State of the Art andAdvanced Methods, Springer,Berlin, 2010. [12] M. Jelali, Control PerformanceManagement in Industrial Automation,Springer, Berlin, 2013. [13] P.V. Marsden, J.D. Wright, Handbook of Survey Research, 2nd ed., Emerald Publishing Group, Bingley, UK, 2010. [14] D.A. Dillman, Mail andInternetSurveys: TheTailored DesignMethod, 2nded., Wiley,New York, 1999. [15] M. Bauer, I.K. Craig, Economic assessment of advanced process control – a survey andframework, J. Process Control 18 (2008) 2–18. [16] M. Bauer, K.S.Brooks,C. Sandrock,Industry expectations andacademicpractice in control engineering education – a South African survey, in: IFAC World Congress, Cape Town, 24-20 August 2014, 2014. [17] D. Wei, I.K. Craig, Grinding mill circuits – a survey of control and economic concerns, Int. J. Miner. Process. 90 (2009) 56–66. [18] L. Desborough, R. Miller, Increasing customervalue of industrialcontrolperformance monitoring – Honeywell’s experience, AlChE Symp. Ser.98 (326)(2002) 153–186. [19] T. Hägglund, Automatic detection of sluggish control loops – the single-loop case, Control Eng. Pract. 7 (12) (1999) 1505–1511. [20] F. Wolff, S. Kramer, Regelguetemanagement.Atp, 3/2014 (2014) 56–64. [21] NAMURUser Association of Automation Technology in ProcessIndustries,Controller Performance Management: Monitoring and Optimisation of Regulatory Control in Production Plants. NAMUR Recommendation NE 152, 2014 www. namur.de.