A quality by design approach to process plant cleaning

A quality by design approach to process plant cleaning

chemical engineering research and design 9 1 ( 2 0 1 3 ) 1095–1105 Contents lists available at SciVerse ScienceDirect Chemical Engineering Research ...

2MB Sizes 2 Downloads 256 Views

chemical engineering research and design 9 1 ( 2 0 1 3 ) 1095–1105

Contents lists available at SciVerse ScienceDirect

Chemical Engineering Research and Design journal homepage: www.elsevier.com/locate/cherd

A quality by design approach to process plant cleaning Elaine Martin a , Gary Montague a,∗ , Phil Robbins b a b

School of Chemical Engineering and Advanced Materials, Newcastle University, Newcastle upon Tyne NE1 7RU, United Kingdom School of Chemical Engineering, University of Birmingham, Birmingham B15 2TT, United Kingdom

a b s t r a c t The cleaning of process plant has traditionally been an activity that has been carried out in open-loop mode, with confirmation of cleanliness achieved through off-line sample assessment. Such strategies have partly arisen as the depth of scientific understanding of the cleaning process has been limited. With deeper understanding through the tracking and prediction of cleaning progression, more sophisticated approaches can be adopted allowing the timely termination of cleaning operations. This paper discusses the component needs of the improved system. At its heart is the need to use appropriate measurement devices for the soil of interest to measure the current process condition and to derive predictive strategies to specify when to terminate cleaning. Results from a case study application on the cleaning of a toothpaste pilot plant demonstrate the concepts. The use of spectroscopic measurements is contrasted with more traditional measurements such as turbidity to track the cleaning profile. Improvement is not achieved simply through better measurement, algorithmic methods for measurement enhancement and forecasting to predict end point of cleaning are both necessary in order to achieve the termination of cleaning operations in a timely manner. The capability to perform both these tasks is considered using the experimental cleaning case study. © 2013 The Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved. Keywords: Process cleaning; Forecasting; Modelling; Process measurement

1.

Introduction

Considering the relative time involved in process cleaning compared with operation, it is surprising that process cleaning is a subject that has attracted limited academic interest, with the possible exception of studies in the dairy industry (Wilson, 2005). Batch process operation, the predominant production strategy in many process sectors, by its very nature requires the cleaning of plant between batches. The extent of soil removal required depends upon the form of operation and the consequences of batch-to-batch carry over, with multiproduct batch plant generally having higher soil removal requirements than plant that is within a single product line. Whatever the cleanliness level required, there is generally a need to implement a cleaning operation between batch cycles. For processes subject to validation requirements, the general issues relating to cleaning and its validation in regulated environments are described in FDA Guidance to Inspectors (1993). Included in this are outlines of what is required in Standard Operating Procedures, validation of process cleanliness and



measurement issues and considerations regarding the setting of cleanliness limits. The design of a cleaning system begins with the original process equipment design. This paper does not consider these aspects as the majority of industrial opportunities arise in improvement of current plant although clearly a plant designed at the outset with ‘cleanability’ as a critical design parameter is the preferred option. The typical approach to process cleaning involves carrying out a series of plant trials of the clean in place (CIP) system during which process cleaning parameters (e.g. flowrates, temperatures, cleaning agent type and concentration) are varied. Off-line analysis of process swabs is used to determine whether the process is clean and CIP parameters selected which lead to consistently clean plant chosen. Following this, the cleaning strategy becomes fixed with the same CIP parameters used time after time. Off-line confirmation of plant cleanliness is periodically tested using swabbing, with the frequency of these tests dependent upon the criticality of success the cleaning operation. The precise levels of concentration that form the clean/soiled threshold

Corresponding author. Tel.: +44 191 2227265; fax: +44 191 2225292. E-mail address: [email protected] (G. Montague). Received 10 July 2012; Received in revised form 15 January 2013; Accepted 22 January 2013 0263-8762/$ – see front matter © 2013 The Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.cherd.2013.01.010

1096

chemical engineering research and design 9 1 ( 2 0 1 3 ) 1095–1105

are product specific but are defined in each case by standard operating policies. Previous cleaning related research has considered the nature of the soil and its impact on strategy, fundamental science of soil adhesion and removal, modelling of the cleaning of process plant and improvement in strategy. A broad view, considering all four aspects, was provided by Wilson (2005) who highlighted current limitations in mechanistic understanding that need to be addressed if improvements in cleaning are to be achieved. Challenges in identifying appropriate instrumentation and its use to determine in a reliable manner end point were identified as critical considerations. Along with this, the need to understand the nature of the fouling deposit was also highlighted as the first step in improving a cleaning strategy. To this end, Fryer and Asteriadou (2009) developed a cleaning map where the nature of the soil was related to the cleaning approach taken. In moving towards improved cleaning, they highlighted the need to develop mechanistic understanding and descriptions of soil removal as a first step in the long-term goal of predictive modelling. Although they do raise the concern that such descriptions may need to account for stochastic effects. Liu et al. (2006) undertook a series of experimental investigations using micromanipulation to assess cohesiveness of soils and from that developed simple models for removal at the surface. While mechanistic level descriptions of surface behaviour are important, from an industrial perspective, being able to track cleaning of process unit operations and their interlinking pipework are of prime importance. Prosek et al. (2005) considered cleaning of pipes and how to predict cleaning rates as pipe geometry and layout changes. The approach adopted is conceptually similar to that used to determine pressure drop in pipes with equivalent length used to account for bends, valves etc. Pipe cleaning was also considered by Lelièvre et al. (2002) but in this case the removal of bacterial contamination along with soil was assessed. They developed models describing the relationship between cleaning rate and process operational parameters such as flowrate and chemical composition. In their analysis they drew on the work of Dürr (2002) who developed models for cleaning of milk heat exchangers and quantified dynamic behaviour and its relation to soil type. Importantly Dürr extended previous models that had been formulated for specific case studies, to a more general model that accounts for alternative mechanisms for soil removal. However, the work described considered only the modelling of milk cleaning from heat exchangers, leaving open the question of applicability to other systems and importantly the use of the model to improve the cleaning operation. In analyzing cleaning systems, it is important to fully appreciate the costs associated with the operation. Alvarez et al. (2004) analyzed the operating costs of the cleaning operation and how water usage could be optimized. An assessment of costs of operation is vitally important in making decisions regarding improvement projects but the challenge of where to set the boundary of consideration is important. Increases in plant availability and reduction in cleaning costs is obvious, but at the other extreme accounting for carbon emissions and associated taxes can be more complex. This paper does not attempt to make a full financial analysis of the business case but it is worth recognizing the complexity of the improvement decision that needs to be made. In making the decision to invest in improvement, several established techniques can be drawn on. Ahmad and Benson (2000) sets out a comprehensive approach for comparative assessment of process performance

Fig. 1 – Benefit justifications for a closed loop cleaning strategy. against what is achieved in comparable cases. This provides a high level comparison of the operational capability. At the control loop level the methods of cost benefit analysis (Anderson, 1996) are applicable to financial decision making. Both the approaches tackle the fundamental issue of how improvement can be estimated before investing in the project to make the improvement. In operational terms, predictive models are an essential component of many optimization and control approaches. In cleaning, previous research has thus demonstrated that models of the behaviour of soils at the surface of process plant can be developed. A universally applicable model structure has yet to be constructed, progress towards which is hampered by the variety of removal mechanisms that exist. While such scientific insight is invaluable, from a process operational perspective predictive models that determine cleaning rates of whole process units and relate these to process parameters such as flowrates and temperatures are more directly exploitable than descriptions of localized surface behaviour. The question then remains is how can such models be exploited for industrial benefit and do the necessary process operational modifications they indicate justify the expenditure involved. It is these aspects that this paper considers, firstly the strategy for improvement and then the instrumentation and control approaches that allow the strategy to be materialized.

2.

The cleaning strategy

The traditional CIP strategy is essentially open loop as there are no changes in operational policy in response to process cleaning performance. Diagrammatically this is shown in Fig. 1 where the open loop policy is set so that virtually all process variation is accommodated in normal operation, i.e. the natural spread of CIP performance results in cleanliness levels that do not violate cleanliness targets. The broad width of distribution is as a consequence of natural variation. The mean final open loop target is set by this distribution and the CIP parameters set to achieve a high percentage of CIP runs satisfy the cleanliness targets. Achieving these levels has an associated cost in terms of cleaning time, energy

chemical engineering research and design 9 1 ( 2 0 1 3 ) 1095–1105

usage and chemical consumption and the lower the target cleanliness measure the higher the cost. It is likely that this idealized open loop strategy is not followed, with the likelihood being that over cleaning beyond this occurs by those configuring the CIP system adopting a conservative approach to ensure that ‘the process is guaranteed clean’. A closed loop cleaning system would be one where the CIP parameters are varied in response to observed cleaning performance. In this instance part of the variation in cleaning performance is transferred to variation in CIP parameters resulting in reduced cleaning performance variation. There is a certain minimum cleaning time that always results in the process being clean as shown to the left of the figure. Using the traditional open-loop cleaning approach a fixed cleaning time is specified and as a consequence of natural variation, a distribution of cleanliness results but it significantly exceeds the cleaning target. With a measurement indicating the cleaning effectiveness, a closed loop scheme can be implemented. A distribution of cleaning times results as shown in the figure. Importantly the increased knowledge of process condition allows a reduction in cleaning time without violating the cleaning constraints and safety margin. This has a clear business benefit; cleaning times can potentially reduced and/or fluid temperatures or concentrations decreased. An additional benefit arises from the measurement of cleaning progress in that the greater insight into behaviour means that the degree of conservatism can be reduced hence less temptation to over clean. The business drivers to improve cleaning strategies have always been present but the increasing emphasis on reducing environmental emissions is adding to the gains possible. The savings from improved cleaning arise for process specific reasons but commonly major benefits can be found when the process is capacity limited. Increased plant production from reduced cleaning times can offer significant overall plant output improvements. The precise gains depend upon reduction in cleaning time and the time involved in cleaning compared to process operation. The relative time between process operation and cleaning differs but can change considerably between products and sectors. For instance, in certain pharma applications the time taken for cleaning can be several times longer than production whereas this tends not to be the case in consumer goods manufacture. Even in the latter cases where time savings can be small, as margins tend to be tight, improvements in plant output can still be a valuable gain. In situations where the plant is not capacity limited there are still benefits to be gained in reduced water usage through shorter cleaning or reduced flowrates, reduced energy usage by cleaning at lower temperatures or reduced chemical usage. To verify plant cleanliness, one approach would be to make on-line measurements of the surface cleanliness at the locations that are most difficult to clean. However, almost by definition such areas are difficult to place measurements in. In this paper the approach taken has been to measure the concentration of the cleaning fluid leaving the plant and from that infer the extent of plant cleanliness. This approach raises demands in instrumentation capability as low levels of contaminant need to be detected and it makes assumptions regarding the link between a clean fluid leaving the process and the process itself being clean. The analysis of the relationship between cleaning times and cleaning flowrates and temperature considered in this paper provides information that can be used to inform an optimization strategy to indicate likely cleaning time. As cleaning

1097

progresses, a closed loop strategy can be used to modify the target. The closed loop control concept in the development phase is as follows: • Measure on-line the concentration in the cleaning fluid leaving the plant using single or multiple measurements. It is likely that the detection limit will be higher than that required to indicate cleanliness but this is soil specific. • Take off-line samples for analysis to indicate the removal profile below the on-line detection limit. • Collect a ‘library’ of the cleaning profiles to indicate the varying profile shapes and the end-points. This requires variation of the CIP policy such as temperature of fluids, flowrates and cleaning chemical compositions if required. The on-line implementation of the scheme involves: • Measure on-line the concentration in the cleaning fluid leaving the plant. • Compare the on-line profile with the library of cleaning profiles using a Case Based Reasoning (CBR) or functional fitting strategy with the appropriate CIP operating parameters. The prediction generated then indicates when to stop the clean based upon measurement of cleaning progression. The CBR and function fitting strategies are discussed further in Section 4. The essential components of the scheme are thus to find a reliable and accurate measurement strategy with low detection limits, build a library of cleaning profiles and verify that the CBR strategy is able to predict cleaning end-point. It is this that this paper sets out to demonstrate using a case study example with toothpaste. The case study soil was chosen as it is possible to measure on-line the concentration down to low levels that are indicative of clean plant. Varying the CIP conditions provides a comprehensive set of data for the CBR library and also offers valuable insight into the optimization surface when cleaning improvement is considered. The case study soil investigation was undertaken using a pilot scale rig.

2.1.

Case study

A process pilot plant at the University of Birmingham in the UK was used to develop improved cleaning strategies. The plant was specifically designed to be flexible to investigate cleaning in pipes and more complex unit operations and allows for both once through cleaning and recirculation of fluids. In the case study, the plant was fouled with toothpaste obtained from GlaxoSmithKline in the UK. The section fouled was a 1 m long section of straight pipe. While not complex in terms of process operations, the purpose was to gain insight of the fundamentals of cleaning at surfaces in process representative pipe configurations. Water was used as the process cleaning fluid, with the possible addition of cleaning agents for specific soils but not required for toothpaste. The process pumps were capable of delivering flows in the range of 0–3.5 m/s in 2 in. diameter pipe. Typically process cleaning is undertaken at around 2 m/s fluid velocity. Cleaning could be carried out at temperatures from ambient up to 80 ◦ C. The cleaning procedure involved pumping toothpaste in with a soiling pump to ensure the soiled pipe was full of toothpaste. This ensured that the pipe was consistently fouled from run to run. Process cleaning was carried out following a design of experiments

1098

chemical engineering research and design 9 1 ( 2 0 1 3 ) 1095–1105

Fig. 2 – Cleaning laboratory rig configuration. protocol to obtain appropriate operational data spanning the range of parameters of interest. A diagram of the laboratory based cleaning rig is shown in Fig. 2. The two water storage tanks of volume 0.75 m3 tanks are filled with mains water (filling system not shown). For each tank in turn, through appropriate manipulation of the valves, water is circulated through the steam plate heat exchanger

until the desired cleaning temperature was reached. During this operation, the pipe to be cleaned was removed, soiled and refitted. Consistency of soiling was critical for comparison purposes so for each experiment the pipe was completely filled and weight recorded to ensure no air bubbles were present. In all experiments a 1 m long 2 in. diameter stainless steel pipe was soiled. When both tanks reached the required cleaning temperature, water was pumped through the test section until the tanks ran dry or the pipe was clean. A number of measurements were taken of the water prior to in entering the drain (AR1–AR3). Process conditions were measured via a National Instruments based logging system (flows, temperatures, conductivities, turbidity) with a sampling frequency of 1 s, with OPUS gathering Bruker Near-infrared and Midinfrared probe signals with a sampling frequency of around 4 s. Matlab running in real-time and communicating via the OPC toolbox coordinated all data gathering and supervised the process cleaning scheduling operations. The selection of soil for study was carried out with reference to the cleaning map discussed in Fryer and Asteriadou (2009) and the need to select a soil that was characteristic of a number of industrial situations. With reference to the cleaning map, toothpaste is a compromise and represents a soil that is not simply washed off (as say would be the case with shower gel) nor does it require hot cleaning chemicals for removal (such as burnt on sugars). Toothpaste can be cleaned with large volumes of cold water pumped at a reasonable velocity or small volumes of warm water in a shorter time. In terms of cost of cleaning, an optimal condition thus exists balancing, time, cost of water and cost of energy. The question of generalization of results is also pertinent. The mechanisms for removal are different for hot cleaning chemical treatment compared

Fig. 3 – (a) Cleaning water temperature; (b) cleaning water flowrate; (c) turbidity probe 1; (d) turbidity probe 2.

chemical engineering research and design 9 1 ( 2 0 1 3 ) 1095–1105

1099

Fig. 4 – (a) Raw NIR data for process clean; (b) first derivative of NIR data for process clean; (c) raw MIR data for process clean; (d) first derivative of MIR data for process clean. to those arising in toothpaste cleaning thus complete generalization of the results is not possible but the approach adopted in arriving at the plant cleaning improvement is at a strategic level transferable. However, a further reason for selection concerned the ability to measure soil concentrations down to low levels in the plant exit stream. Toothpaste was selected since off-line tests had demonstrated that its concentration could be tracked down to very low levels using on-line turbidity measurement. Such accuracy will not be possible with all soils so it provided an ideal case study soil to investigate the impact of detection limits on the prediction of end-point. To simulate soils with less favourable measurement capabilities low concentration points could be assumed to be unknown but predictions verified by the measured data.

probes are situated around 3 m downstream of the test section. Signal processing of spectral signatures is common to remove instrument artefacts. A common approach is to take first derivatives of the signal to remove baseline shifts with the resulting response seen in Fig. 4b and d. A Savitsky Golay algorithm with a window width of 15 and a second order polynomial were used for this purpose. Further details of spectral data processing are provided in Section 3.1. A Design of Experiments (DOE) procedure was adopted to generate the process data. A full factorial design/central composite design, two levels (temperature, flowrate), three (or more) replicates and additional centre points was adopted. Around 60 cleaning runs were undertaken in total.

3. 2.2.

Analysis of experimental results

Toothpaste cleaning data characteristics

Examples of process data logged at a frequency of 1 s in a typical cleaning run are shown in Fig. 3. In Fig. 3a the switch between cleaning water from tank1 and tank2 can be seen at around 150 s. A feedback control system regulates the flowrate of cleaning water to setpoint as shown in Fig. 3b. The turbidity probes are positioned at around 2 m downstream of the test section. Turbidity probe 1 is to greater sensitivity and its response is shown in Fig. 3c where saturation of the signal is observed until around 170 s. Turbidity probe 2 is ranged to observe the whole cleaning response as seen in Fig. 3d. The raw spectral process measurements from on-line near and mid infra-red probes are shown in Fig. 4a and c. The

The first task in analysis was to get a comparative indication of cleaning time as the process conditions vary. A fundamental measurement need therefore is the ability to track the progression of cleaning through analysis of water leaving the pilot plant. The assumption is that once the concentration in the wastewater falls to what can be considered close to zero then the process is clean. While the quantification of what is close to zero is product and process specific, it follows that there is a need to measure down to concentrations below the clean threshold or less desirably measure down to low concentrations and forecast the further fall in concentration until it reaches below the cleaning threshold. In either case sensitivity of measurement at low concentration is important.

1100

3.1.

chemical engineering research and design 9 1 ( 2 0 1 3 ) 1095–1105

Spectral data analysis

Considering firstly the spectral data measurements from the NIR and MIR probes, in determining the fact that the process is clean two approaches can be considered: build a calibration model and translate certain key regions of the complex spectral signature over time to indications of concentration of soil or consider the spectra as a whole as an indicator of condition and distinguish between a pattern in the spectra when the process is clean and when it is not. In these studies the pattern based approach was adopted as it provided indication of cleanliness while also not requiring off-line sample analysis for calibration data generation. The extent of off-line analysis was limited to end of clean confirmation of cleanliness undertaken by surface swabbing. The procedure adopted for spectral data interpretation was:

(1) Take the first derivative of the spectra to remove changing base-line effects. (2) Identify the region or regions of the spectra indicative of toothpaste concentration. (3) For those regions, calculate the area between the current spectral signature and the ‘clean’ signature determined after an extended clean. (4) Display the difference in area as an indication of the extent of clean.

This procedure was applied to both the NIR and the MIR data. In Fig. 5 the NIR spectra collected over the course of a process clean can be observed. Fig. 5a shows the overall spectral region, while Fig. 5b focuses in on the specific region of interest associated with toothpaste. In this case only one region was found to be discriminatory for toothpaste. The change in area between the current spectra and the ‘clean’ spectra for the NIR analysis is shown in Fig. 6a. Here the integral of area deviation from the water signal is plotted. The dashed line also shows the fit of a first order transfer function. The time constant gives a quantification of the rate of clean. In this case the time constant is 52 s thus indicating steady state conditions are achieved after around 200 s. This is not confirmed by the conductivity readings shown in Fig. 6b where it takes around 250 s to reach clean. Thus it can be observed that while early in the cleaning the signal provides indication of the fall in toothpaste concentration in the exit stream, in the latter stages of the clean the signal is noisy and fails to give the accuracy required for determining the end-point of clean. Analysis of the MIR signal adopting the same approach provided significantly less capability than NIR measurements. Given the extremely short path length measured in MIR analysis it provides indications of soluble components whereas path lengths of NIR probes are of the order of several mm and thus give indication of soluble components and particulate matter. Given the nature of toothpaste, NIR is better suited to its measurement but for other soils this may not be the case. However, Fig. 6 suggests that the accuracy of the NIR may not be sufficient to track the clean and it was for this reason that turbidity was considered as discussed in Section 3.2. This is not to say that NIR does not add value to a monitoring system. While sensitivity of the NIR measurement is not as high as may be required, the NIR signal does provide a richer source of information than single indicators such as

Fig. 5 – (a) Complete NIR spectra for a process clean and (b) NIR spectral region associated with toothpaste for a process clean. turbidity. Take for example the cleaning signals observed in Fig. 7. The multipurpose pilot plant had been used to investigate the cleaning of deodorant base prior to the toothpaste trials. On the first toothpaste run the spectral data shown in Fig. 7 was logged. In the figure the spectral signature of pure toothpaste and pure deodorant are also shown. It can be seen that there is a peak at a wavenumber of 1250 cm−1 that is present in the process cleaning data that is characteristic of deodorant and is not present in toothpaste. This is a clear indication that the previous users of the pilot plant had failed to clean it effectively on its last use. Such contamination indications are not possible using measurements that give a single indicator such as turbidity.

3.2.

Turbidity analysis

While spectral measurement can provide differentiation between chemical components and a rich source of information, given the nature of toothpaste a more effective measurement is to measure the turbidity. Two turbidity probes were installed on the cleaning rig; one ranged to give an indication of the overall clean and the other set to measure the concentration towards the end of the cleaning process. It was the latter probe that was used to determine the end of cleaning. The most simplistic approach

chemical engineering research and design 9 1 ( 2 0 1 3 ) 1095–1105

1101

Fig. 8 – Surrogate cleaning time as a function of water temperature and flowrate.

Fig. 6 – (a) Extent of clean as determined by NIR analysis and (b) conductivity measurement for the clean. taken was for each run the time that turbidity probe comes off saturation was taken as a surrogate to the final cleaning time. At this time the surface was predominantly clean with only small areas of toothpaste remaining. While further cleaning was required, the measurement serves as a comparative indicator of cleaning effectiveness.

3.3.

Cleaning end-point results

The results gathered indicate that while on-line spectral measurement using NIR or MIR provides a rich source of information and can discriminate between soils, the information from turbidity provides indications of toothpaste concentration down to lower levels as seen in Fig. 3c where visual

Fig. 7 – Batch cleaning in the presence of contamination.

inspection confirms the turbidity indication of the extent of cleanliness. Spectral measurement is valuable in that it provides contamination information but in this paper the main objective is to gain indications of soil concentration and therefore turbidity is used for end-point indication. Fig. 8 shows the surrogate cleaning time as a function of cleaning water flowrate measured in m3 /h for the turbidity measurement. Given the pipe diameter is 2 in. a flowrate of 14 m3 /h corresponds to a fluid velocity of 1.9 m/s. A number of important observations arise from Fig. 8. It is evident that the process will clean most rapidly with high temperature water at a high flowrate but this is most energy intensive. The increase in cleaning time from reducing the cleaning temperature to near ambient can be compensated for by increasing the cleaning flowrate. The information available from these experiments thus forms the foundation for optimization of the cleaning strategy given the relationship between cleaning time and flowrate and temperature is evident. It is also essential in designing the cleaning strategy that there is understanding of the variability in operation. Fig. 8 clearly indicates a degree of variation but this ideally needs to be quantified. Performing a statistical analysis of the data calculating the standard error of the mean indicated that the variability observed between cleans is not a function of flowrate or temperature. This is an important observation when the fact that temperature and flow could be varied in response to operational factors resulting in different operating conditions. From this analysis it follows that rather than obtaining a prediction that say at 40 ◦ C and 14 m3 /h water flowrate the cleaning time will be 95 s, the variation analysis provides additional information in that the cleaning time will be 95 ± 3.5 s with 95% confidence. The above results have been obtained using information from the turbidity probe ranged to give more precise indications of concentration towards the end of the clean. An alternative approach was to use information from the second turbidity probe and the flow measurement. As the second turbidity probe does not saturate, integrating over time turbidity multiplied by flowrate gives an indication of total soil removed. The method was found to work reasonably well and a first order + dead time response could be fitted, with the time constant indicating the rate of cleaning using the same approach employed for NIR analysis in Fig. 6a. However, visual observation of the removal shows that toothpaste is removed in lumps in the mid stages of the clean. As such turbidity measurements

1102

chemical engineering research and design 9 1 ( 2 0 1 3 ) 1095–1105

are not a true reflection of the amount of soil removed thus raising questions regarding the validity of this approach.

4.

Forecasting of cleaning progression

The results shown in Section 3 above indicate that in the case of toothpaste it is possible to track the concentration in the effluent down to levels that are indicative of a clean process. This situation will not arise for all soils and in many cases measurement limitations will dictate that it will be necessary to gain an on-line indication of cleaning end-point from a forecast of the progression of clean in the measurable region. The toothpaste case study provides an ideal test-bed of this concept as of all the soils considered in the study, toothpaste has the lowest concentration on-line detection limits. It can therefore be used to simulate the effect of less effective detection limits on forecasting ability. Forecasting end-point for the current cleaning operation can be achieved through a number of possible strategies; in this study mathematical function fitting has been contrasted with a case based reasoning strategy. A number of possible strategies exist taking either a model based approach or a pattern recognition technique. In this study the two that were chosen were contrasting approaches for forecasting.

4.1.

Function based forecasting

In function based forecasting, a mathematical description of the process is determined from past process batches and is used with data from the existing batch to identify likely future behaviour. The model can be either empirical with the parameter values determined using process data or it can have a mechanistic structure where the parameters are related to physical/chemical attributes. Trelea et al. (2001) compared the alternative approaches for the predictive modelling of a brewing fermentation and concluded that if mechanistic understanding is available it should be employed. Unfortunately in this case developing a mechanistic model of cleaning is beyond current capabilities for process systems of any complexity. The alternative approach is to adopt a pattern recognition strategy. For example, if the process trends follow a particular functional form, but with varying parameters, then the parameters can be determined from early cleaning behaviour and forecasts made using the functional patterns identified. The challenge in adopting this approach is to find a functional form that is characteristic of the pattern observed in cleaning a particular soil. Indeed as removal mechanisms are soil specific it is likely that different soils will require the specification of different functional forms. The Curve fitting Toolbox in Matlab provides the opportunity to search for a number of functional forms and returns the one that most closely fits the data. For the toothpaste case study it was found that:

Fig. 9 – Forecasting cleaning end point with function fitting. information is available to determine the parameters and the fitted function then used to forecast the remainder of the cleaning operation. Results from the application to a process clean are shown in Fig. 9 where the predicted cleaning time is shown along with the final cleaning time. It can be seen that from around 100 s before the end of the clean an accurate prediction of end-point is obtained.

4.2.

CBR based forecasting

The overarching CBR philosophy involves comparing the measured concentration change during a cleaning operation and using that along with historical information to predict the remained of the profile. In forecasting through function fitting, the parameters of a mathematical relationship that is capable of ‘describing’ the shape of the cleaning behaviour are determined for the early stage of the process clean and used to forecast the subsequent trajectory. In the case based reasoning strategy, the data from previous runs are stored in a library and the measured data from the current run compared to the library and the most similar from the library used for forecasting. Rather than using current CIP behaviour and matching it or fitting some functional form, the alternative is to omit this step and match the pattern of the current CIP to the most similar previous profile and use the past CIP profiles as a predictor

Cout = aebt + cedt where Cout is the soil concentration in the outlet, t is cleaning time and a, b, c and d are the function coefficients that are determined using non-linear least squares. This functional form was found not to be applicable for other soils considered in the research study. The applicability of this function was found using historic data from complete cleans. In forecasting only part of the

Fig. 10 – Prediction of cleaning end time.

chemical engineering research and design 9 1 ( 2 0 1 3 ) 1095–1105

1103

decisions taken in those instances, the ‘best’ action to take for the current instance is determined. A critical metric in CBR is how to determine similarity between cases. In this paper, the time progression of the soil concentration in the outlet cleaning fluid is considered. To compare the current profile with the library of examples the following procedure is used: (1) For the first example from the case library, compute the time difference between the case example and the current cleaning profile coming off saturation. Use this to align the start points of the profiles. (2) Determine the value of ‘n’ the stretch coefficient in the following expression that minimizes:

  t2  fit =  (n × ylibi−tcorr − ymeasi ) i=t1

Fig. 11 – The stretch coefficient variation.

of current performance. Watson and Marir (1994) presented a review of CBR and discussed potential areas of application. Among the characteristics of problems suited for the application of CBR are those where a model of the system is not available but numerous historical examples of behaviour exist. The fundamental principle behind CBR is that the historical records are examined and the most similar situations to the current condition are used to make decisions. The historical records come from normal process operation and capture natural deviation as a result and are not generated by experimental design principles with defined deviations. The approach used in this paper is described in detail in Montague et al. (2008). In CBR, a library of previous process behaviour is established. Current behaviour is then compared with previous experience and based on past observations and

where tcorr is the time correction which aligns the start time of the two profiles where start time is when the probe comes off saturation at time t1. ylib is the soil concentration profile from the case library and ymeas is the soil concentration from the current clean. t2 is the time window over which the comparison between the profiles is made. (3) Repeat the procedure for all examples in the case library. (4) Determine the case with the minimum value of fit. (5) Using the library example with the modified start time and the calculate value of the stretch coefficient ‘n’ predict the remainder of the cleaning profile. Essentially, the procedure aligns the profiles and allows the case library profiles to squash or stretch and thus seeks the profile with the most similar underlying pattern that is not anchored in time but has the same characteristic shape.

Fig. 12 – Example of fit to using the closest fit in library.

1104

chemical engineering research and design 9 1 ( 2 0 1 3 ) 1095–1105

To investigate the functionality of the procedure, a small case library was built comprising of five cases. The above procedure was followed and the results shown in Figs. 10–12 obtained. Fig. 10 shows end time prediction at a range of times prior to the end of cleaning as well as the final cleaning time. It can be seen that the end point prediction lies within about 10 s of the final value using information from only half of the cleaning profile. The stretch parameter values over these predictions are shown in Fig. 11. The fit that these variations achieve is shown in Fig. 12 where the original and modified library profiles are shown alongside the profile of the current clean. The time shift and stretching effect are clear from these profiles.

5.

Conclusions

This paper has raised the prospect of a new paradigm in process cleaning in the move from open loop control that is not responsive to process behaviour to an end-point detection philosophy where natural variation is taken into consideration. The traditional open-loop operation also fails to allow an appreciation for the response to variation of operational parameters to be gained. Through variation of CIP parameter settings an appreciation of the impact on cleaning times has been acquired. Important in this is the need to gain an understanding of the inherent variability in cleaning as this impacts on control scheme design and operation. Furthermore, such information is crucial as cleaning strategy improvement moves from closed loop end-point detection to optimization of the cleaning procedures. The benefits to be gained from improvement of industrial process cleaning are complex to determine and depend on many factors. Among the key factors are is the manufacturing output capacity limited and how long the cleaning operation takes relative to the manufacturing operation. It is indeed the case that for short, easy to clean products, an increase in sophistication of cleaning procedures is not financially justifiable. At the other extreme, for multi-purpose fine chemical batch manufacturing plant it can be the case that the cleaning operation is significantly longer than the reaction. In such cases the benefits from improved cleaning operations could bring significant financial benefit, especially if the plant is production limited and time saved on cleaning can be transferred to time spent on additional production. As discussed previously, part of the challenge is to assess the benefits ahead of making changes and here the cost/benefit analysis strategy comes into its own. Improvements to cleaning strategies could by many just be considered from a purely financial perspective and justified as such but when change of environmental impact arises public perception issues serve to complicate and ‘fuzzify’ the case for change. In our research programme both extremes were witnessed: the justification based on financial case irrespective of environmental consideration and decisions solely based on the environmental case. The studies reported in this paper were undertaken as part of programme aimed at increasing the understanding of process cleaning operations and underpinning the design and operational decisions with basic scientific understanding. This paper has reported on the cleaning of toothpaste from pipework as a case study system. This has the advantage that toothpaste can be measured down to low concentrations

and therefore it can be used to develop strategies where such concentration measures are not possible. Fundamental to the move to closed loop strategies is the measurement of soil concentration to levels of detection that are sufficiently low that tracking of natural variation is possible. In this case with toothpaste this was possible and with other soils considered in project in which this work was undertaken similar capabilities were present but this will not be the case for all soils. In this sense, it must be recognized that the mechanisms of removal are soil specific so further work is required to assess the applicability of approach. Also in this case cleaning was restricted to pipework rather than more complex plant. Given the approaches adopted in this paper involve end of pipe measurement, they are not invalid when applied to more complex process configurations but require further study to gain an insight into the profiles and variation of the profile for more complex process geometries. So in summary, this paper by no means answers all the questions that arise with regard to process cleaning. From a scientific perspective there is a need for a fundamental understanding of the mechanisms of cleaning and their relation to soil type. Through such understanding the control engineer can implement a more robust and capable control system. This philosophy is at the heart of the FDA Quality by Design methodology. This paper has demonstrated that improvements in process cleaning are possible but further work is required to build this level of understanding that will allow for greater generalization of the concepts. From an industrial viewpoint, there is a need to optimize the CIP procedures against many objectives and arrive at a responsive and environmentally more considered solution. In both cases though the need to understand and be able to predict the behaviour of process cleaning is first and significant step in making changes. The methods described represent a significant step in this direction.

Acknowledgements The authors would like to acknowledge the financial support of the UK Technology Strategy Board. The contribution of the members of the ZEAL project consortium is also acknowledged for their development of the pilot plant at the University of Birmingham and discussions regarding process plant cleaning and best practice assessment. In particular, thanks go to Dr Konstantia Asteriadou for her considerable contribution to ensuring the functionality of the pilot plant.

References Ahmad, M., Benson, R., 2000. Benchmarking in the Process Industries. IChemE, ISBN 0852954115. Alvarez, D., Garrido, N., Sans, R., Carreras, I., 2004. Minimization–optimization of water use in the process of cleaning reactors and containers in a chemical industry. J. Clean. Prod. 12, 781–787. Anderson, J.S., 1996. Control for profit. Trans. Inst. Meas. Control 18 (1). Dürr, H., 2002. Milk heat exchanger cleaning: modeling of deposit removal II. Trans. IChemE 80 (Pt C), 253–259. Fryer, P.J., Asteriadou, K., 2009. A prototype cleaning map: a classification of industrial cleaning processes. Trends Food Sci. Technol. 20, 255–262. Lelièvre, C., Antonini, G., Faille, C., Bénézech, T., 2002. Modelling of cleaning kinetics of pipes soiled by Bacillus spores assuming

chemical engineering research and design 9 1 ( 2 0 1 3 ) 1095–1105

a process combining removal and deposition. Trans. IChemE 80 (Pt C), 305–311. Liu, W., Zhang, Z., Fryer, P.J., 2006. Identification and modeling of different removal modes in the cleaning of a model food deposit. Chem. Eng. Sci. 61, 7528–7534. Montague, G.A., Martin, E.B., O‘Malley, C.J., 2008. Forecasting for fermentation operational decision making. Biotechnol. Progr. 24 (5), 1033–1042. Prosek, M., Krizman, M., Kovac, M., 2005. Evaluation of rinsing-based cleaning process for pipes. J. Pharm. Biomed. Anal. 38, 508–513.

1105

Trelea, I., Titica, M., Landaud, S., Latrille, E., Corrieu, G., Cheruy, A., 2001. Predictive modelling of brewing fermentation: from knowledge-based to black-box models. Math. Comput. Simul. 56, 405–424. US Food and Drug Administration, 1993. Technical Guides, Validation of Cleaning Processes (7/93). Watson, I., Marir, F., 1994. Case-based reasoning: a review. Knowl. Eng. Rev. 9 (4), 327–354. Wilson, D.I., 2005. Challenges in cleaning: recent developments and future prospects. Heat Transfer Eng. 26 (1), 51–59.