Article - Issue 8, May 2001
Looking for a needle in a haystack: fault-finding in process engineering
Elaine Martin · Department of Engineering Mathematics, University of Newcastle Julian Morris · Department of Chemical and Process Engineering, University of Newcastle
Fault-finding in process engineering
We use sophisticated technology to monitor many engineering processes but it is still often very difficult to locate the source of a problem on a production line. We have so much information but do not know how to sift through it. This article looks at a technique that could make fault-detection a great deal easier
Problems in process engineering can sometimes be very obvious and at other times are subtle and difficult to detect. The majority of engineering processes are now monitored by sophisticated instrumentation: a large number of process variables can be checked and recorded on a regular basis. An enormous amount of measured data is usually available, so that a problem such as a blockage may be very simply detected because a sensor downstream shows that no feed is present.
On the other hand, a more complex change in the process, such as small simultaneous changes in temperature, pressure and humidity that may later lead to a blockage, will be more difficult to identify. In such cases searching for the subtle process changes that caused the blockage can be likened to ‘looking for a needle in a haystack’. Fortunately help is at hand that does not involve further heavy investment, but uses what is already available, that is, data.
This article focuses on a technique that could revolutionise fault-detection in process engineering. Multivariate statistical process control (MSPC) combines mathematical analysis and fast computing power with the data that is already being collected and enables problems to be analysed which were previously considered almost insoluble. MSPC software packages now enable engineers to use the technique without having to acquire detailed mathematical knowledge. The strength of MSPC is its recognition that process variables are seldom independent of each other; MSPC can model interrelated variables and provide real understanding of how faults arise and how to correct them.
Over the last decade the emphasis in process manufacturing has changed. Quality and product consistency have become major consumer decision factors and are key elements in determining business success, growth and competitive advantage. Making products that meet their quality specifications at the first attempt results in higher productivity and reduced manufacturing costs through less rework, give-away and waste. This all contributes to reducing the impact of the process on the environment by minimising raw materials and energy usage. The achievement of first-time-right production requires a reduction in process variability; thus the monitoring of process behaviour over time – to ensure that the key process/product variables remain close to their desired (target) values – is essential.
Statistical Process Control
One of the foundations of modern quality-control procedures is (univariate) statistical process control (SPC). Process variables are regularly monitored and statistical tests are carried out to check that each variable lies within an acceptable range. Originally developed in the 1920s by Walter Shewart, SPC has subsequently been repackaged. For example, in the 1970s it formed part of the ‘Total Quality Management’ movement, in the 1990s it was incorporated within ISO 9000 and most recently it has been integrated into the ‘Six-Sigma’ philosophy.
However, SPC is still struggling to gain widespread acceptance. Why should this be? There are two angles to this debate. The first is that SPC is not simply a methodology, it is a philosophy. To be successful, it needs to be integrated in to the company’s quality procedures and business objectives, and not simply be seen as a quick fix to on-going problems.
The second aspect of the debate over SPC relates to the question: ‘Is SPC, as it has traditionally been applied, appropriate in today’s process manufacturing environment?’
The answer to this is ‘No’. With the increasing availability of process and analytical instruments, more and more measurement devices are being installed on processes and the resulting data stored on data collection systems. In utilising this data, engineers and scientists have tended to focus on a few variables that they believe, or have found in their experience, to be key to controlling the process and hence determining product quality and variability. This results in most of the data being unused.
It is only when a more subtle processing problem occurs – one that cannot be solved through examining a few variables – that the plant engineer or scientist will trawl through the rest of the data to try to isolate the problem. Problem-solving suddenly becomes much more complex because of the large amount of data involved. The key issue to bear in mind here is that by looking at the values and trends of individual variables, as has traditionally been done, the chances of success are very low.
Limitation of univariate statistical process control
Consider a situation in which two variables are being monitored using univariate statistical process control (Figure 1). The seven test results (measuring each variable for seven products 1, 2, …, 7) lie within the action limits agreed for that variable. Plotting the two univariate charts at right angles shows how it might be assumed that the resulting product meets the customer’s specifications: all values 1–7 lie within the univariate in-specification zone, the shaded square at the top left. However the customer complains that problems have occurred in the subsequent processing of products 1 and 3. Why might this be?
The answer is that many subtle faults are not identifiable from a single measurement. By applying univariate SPC, no account was taken of the interaction between the two variables. When their relationship is taken into account, the action limits can be revised accordingly to produce the elliptical region shown in Figure 1. It is now clear that product items 1 and 3 lie out of specification. Thus the solution to the identification of subtle process changes and faults is multivariate statistical process control (MSPC).
Philosophy of MSPC
In practice in a modern processing plant, many more than two measurements are routinely made. The above approach readily generalises to situations where there are 10s, 100s or even 1000s of measurements made on a process. However, the overriding philosophy of MSPC is that, under normal operating conditions, there are not 10s or 100s of events determining the behaviour of the process. In fact there are typically only a few key operations that determine overall process operation. The resulting ‘key event’ variables provide a more readily understandable view of the overall workings of the process.
Benefits of MSPC
Gradually MSPC is being accepted in the process industries. Commercial products are now becoming available that enable the MSPC methodologies to be applied on-line in real time. For example, an MSPC software package called ANEX has been developed for the fibre manufacturing industry and InControl Technologies Inc have produced a software package, QualStat, for the petrochemical industry. Through various industrial applications, MSPC has been proved to:
translate data into information and hence knowledge;
provide a deeper understanding of the process;
allow early warning of process changes;
enable the identification of potential plant faults, process malfunctions and process disturbances.
A number of projects associated with the development of multivariate statistical process performance monitoring have been supported by the UK Research Council (EPSRC, IMI), the DTI and the EU. Applications resulting from these projects (involving the Foresight Centre for Process Analytics and Control Technology, University of Newcastle, in conjunction with a range of industries) have demonstrated that MSPC can contribute to assured process performance monitoring of very challenging process manufacturing problems.
Applications have included those described in the rest of this article and the at-line application of MSPC on a films line (DuPont) at Hopewell, USA. Other applications have been in food processing (Quest International), fine chemicals manufacturing, wastewater treatment (SmithKline Beecham), pharmaceuticals (SmithKline Beecham), materials processing (Corus), resin manufacturing and petrochemical processing (BP).
Multi-product multi-recipe process monitoring
Early applications of MSPC focused on the production of a single manufactured product, i.e. one grade or one recipe. Separate models would be constructed to monitor different products. Now that process manufacturing is increasingly driven by market forces and customer needs, there has been a shift of emphasis to multi-product flexible manufacturing.
The widespread acceptance of MSPC as an integral part of process performance monitoring is dependent on the technology being implementation-friendly. With many companies now producing a wide variety of products, some of which are only manufactured in small quantities, the number of MSPC models required can soon become excessive. This is impractical and non-user friendly; thus there is a need for a multi-group modelling methodology that allows a number of grades, products or recipes to be encapsulated within one process model.
As part of a European ESPRIT Project, Process Diagnostics for Plant Performance Enhancement (PROGNOSIS), involving Unilever (Netherlands), BASF (UK), BICC CEAT Cavi (Italy), Solvay (Belgium) and MDC Technology (UK), an alternative MSPC methodology has been developed. It has been applied on-line to an industrial semi-discrete batch process with Unilever. This industrial application demonstrated, for the first time, that it is possible to extend the MSPC methodology to situations where a number of different production recipes or grades are manufactured using a single multi-group model. The potential of such a development is far-reaching and ensures that MSPC is widely applicable within the process manufacturing industries.
An industrial application
The industrial manufacturing process of Unilever forming the basis of this example produces a range of liquid detergent products to meet the requirements of a rapidly changing market. For our specific brand, the company manufactures two specific formulations; each of these formulations comprises a number of different recipes. Furthermore, each recipe can then be sub-divided into different varieties.
We will look at two recipe families to demonstrate the different approaches that can be used in monitoring multiple recipes. Each recipe comes in two distinct varieties, let us say ‘lemon’ and ‘pine’, resulting in four data sets being included in the process model.
Figure 2 shows the result of merging the data from each of the individual data sets and applying principal component analysis to the resulting data. Often by adopting this approach, distinct clusters are observed, as here.
From Figure 2 we can see that the separation between the two recipes is primarily associated with principal component 1: the batches relating to Recipe 1, ‘o’ and ‘+’, lie on the left whilst those relating to Recipe 2, ‘x’ and ‘*’, lie on the right. Principal component 2 defines the variability resulting from the different varieties within each recipe.
The presence of the between-recipe variability in Figure 2 has a detrimental impact on the detection and diagnostic capabilities of the process model – it is not sufficiently sensitive to pick up process faults. To improve detection and diagnosis, the between-recipe variability needs to be eliminated. This can be accomplished by applying a novel approach developed at Newcastle within the EU project, PROGNOSIS, and the result is shown in Figure 3. Now there is no clear separation between the batches manufactured from the two recipes. Adopting this process model, we can focus our interest on within-recipe(s) variation. This is the key to providing assured production.
A ‘non-conforming’ batch, subsequently identified as resulting from a raw material overdose, is shown in Figure 4. The observations initially lie in the centre of the ellipse – the acceptable range – but after the overdosing they jump outside the control region to cluster in the top left of the graph. In many applications, the major sources of variability are identified in principal component 1 (which defines the direction of greatest variability) and principal component 2 (which defines the direction of next greatest variation), whilst the more subtle changes in the process are detected in the lower-order principal components. In this case the problem was identified from principal component 4.
A strength of this approach is that by identifying a trouble-spot using principal components, it is possible to examine the variables that have contributed to the unwanted change. A contribution plot (Figure 5) can be made of all the variables in principal component 4 at the point where the batch went out of specification.
Variables 69 and 70 (operating and dosing times) and 71 and 72 (flow meter and load cell dose weights) show up clearly as having changed dramatically when the process went out of control. Using this information, together with their process knowledge, the operational staff quickly traced the problem to a dosing valve malfunction.
Location of a dosing valve problem using normal fault tracing procedures is a particularly difficult and time-consuming task. Such a problem would not have been located by looking at the 83 individual variable trends. It was only through the examination of the specific combination of variables that caused the process to move out of statistical control that enabled the maintenance personnel to locate the source of the problem more quickly. If this problem had not been detected, a number of off-specification batches would have been produced. The cost of re-working an off-specification product is not only time consuming but expensive.
The benefits of the MSPC on-line performance monitoring, as cited by the plant manager, included: increased knowledge of the batch-manufacturing process; the identification of the most critical parameters relating to product quality and consistency; an awareness and understanding of process variability and how it can be reduced; an increase in the percentage of first-time-right products; an increase in process efficiency and output; and consistent operating procedures.
Increased instrumentation and greater emphasis on process consistency and quality, in conjunction with the drive towards flexible manufacturing, has raised industry’s awareness of the importance of MSPC. The methodology has been around in the statistical literature for the past century but it is only through recent developments in computer technology that the real potential of MSPC has been realised. The success of the technology depends on the willingness of a company to integrate the MSPC philosophy into their business culture. The technology can then contribute to increasing the proportion of first-time-right production and enhancing process consistency and efficiency. In achieving this, the company will also obtain an enhanced understanding of their engineering processes, and ongoing early warning and location of process deviations.
it is possible to extend the MSPC methodology to situations where a number of different production recipes or grades are manufactured
A strength of this approach is that it is possible to examine the variables that have contributed to the unwanted change
Case study 1: Limitations of univariate SPC
The power of MSPC is illustrated through the application of the technology to the manufacture of a polymer compounding process that comprises both batch and continuous type units. The product is subsequently used in the manufacture of high- and medium-voltage electrical cables.
Applying univariate SPC tools to the data (Figures A and B) no alarm was raised. Production lay within the control limits, although product quality was reported to be out of specification.
Analysing the data using MSPC, the onset of a production problem was clearly identified (Figure C) as observations started to progress outside the outer ellipse (99% action limit) on the operator’s screen. Good operation is indicated by points lying within the inner ellipse (the 95% warning limit). The problem was later traced to a drop in the power supplied to one of the mixers during a processing step, resulting in the next operating step being of longer duration.
This example highlights the inadequacy of univariate monitoring for complex processes.
Case study 2: MSPC for process understanding
Here MSPC is applied to a fluidised bed reactor in order to illustrate how it can help in understanding process behaviour. A total of 37 variables, including chemical quality and process variables, are monitored on the process. The process information in these variables can be presented in graphical form (Figure D).
The figure shows that the process operates in 21 clusters, all of which lie within the action limits (the outer ellipse). This particular reactor was operating in a number of different regions over the period of time during which the nominal data set was collected and, from an analysis of operator logs, the different clusters were found to be associated with different operating conditions. Based on this information the operators can undertake (for example) economic optimisation of the process. Through the application of MSPC technology, the company involved achieved annual savings of the order of £1 million by reducing the amount of catalyst used.
Elaine Martin* and Julian Morris‡
Foresight Centre for Process Analytics and Control Technology
*Department of Engineering Mathematics
‡Department of Chemical and Process Engineering University of Newcastle
Julian Morris is Professor of Process Control in the Department of Chemical and Process Engineering and Elaine Martin is a Reader in Industrial Statistics in the Department of Engineering Mathematics at the University of Newcastle. They co-direct the Foresight Centre for Process Analytics and Control Technology (CPACT). They have published widely in the area of process performance monitoring and empirical modelling. Email: email@example.com; firstname.lastname@example.org