Geometric process control moves into batch processing
12 Mar 2009
Geometric Process Control (GPC) is a new technology for monitoring and exploring process data through the application of n-dimensional geometry. Because it is based on simple geometric operations, it is easier to use with large multi-dimensional data sets than conventional techniques, such as principal components analysis or statistical process control, and requires minimal mathematics background.
Key to the application of GPC lies in the Inselberg parallel coordinate transformation, which allows visual representation of high dimensional data. This allows, for the first time, the leap from 2- or 3-dimensional representations to an arbitrary number of dimensions.
GPC delivers optimum results with continuous application over long periods aimed at keeping process performance at its best all of the time, not just occasionally, as is the case in many plants today.
Big gains have been reported from using GPC in data mining by Huntsman Petrochemicals and in rationalising alarm limits and alerts by INEOS Chlor, while Mallinckrodt Chemicals has detailed major benefits of the technology in operating envelope models for continuous process control. Other successful applications include continuous improvement methodologies (Six Sigma), quality by design, response surface analysis of design of experiments, and real-time optimisation.
Building on these applications, Curvaceous has been working to extend GPC to analyse and control multi-phase, multi-stage batch processes. Batch processes are common in pharmaceutical and speciality chemicals, and can represent some processes common in high tonnage continuous plants, such as PVC, polyethylene and coking in oil refineries. Process start-up of continuous processes is best viewed as a batch process. Semiconductor manufacturing, for example, is really a multi-stage batch process in which each stage has only one phase.
Different tools are used throughout the application of GPC. A visualisation tool is the basis of the technology, and allows the off-line exploration, analysis and review of process data. It is also used in the construction of process models that are used for other products. An online tool then provides real-time control and optimisation through OPC, using models provided through analysis to compute dynamic alarm and alert limits, guiding the operator through process optimisation.
Batch processes are inherently more difficult to analyse and control than their continuous counterparts. Rather than instantaneous operating conditions, the full time-varying history of the batch becomes important. There are separate phases of operation, transfers from equipment, and intermediate products, one batch of which may be used in multiple subsequent batch steps. These all add complexity to the analysis and control problems, adding more variation. Minimising run-to-run variability as well as variability across stages is a common goal as much higher dimensionality of batch processes means variability can appear in many more places.
In a typical three-stage manufacturing process where each stage consists of reaction and separation steps, the steps may be broken down time-wise into individual phases of operation that correspond to different activities taking place sequentially during the stage.
With a continuous process, the problem is often that of finding the Best Operating Zone (BOZ) or 'cloud' of operating points that produce product within specifications, obey process safety considerations, and may have some element of plant optimisation. Excursions from the BOZ correspond to unacceptable or untested conditions that are sub-optimal, poor quality or dangerous. In the simplest case for batch operations, this cloud becomes a 'tube' or tunnel through time containing the desired values of all parameters. The tube changes size and shape dramatically from phase-to-phase with the time of processing through the batch, and represents the interaction of the remaining parameters.
In online applications the time-dependence is addressed by the modification of the Curvaceous process model to include time as an explicit variable. The current acceptable cloud of variables and their interactions is restricted by the time, allowing the control not only of instantaneous process variables, but their trajectory through the entire batch. Modifying the model is simple from the user's perspective, requiring only the addition of time as an additional variable. Realtime control limits can then be calculated with no further input required, and can be fed back to the plant operator.
Multi-phase processes introduce an additional consideration. At different points in the operation, some variables may have no bearing on the process, for instance agitation rate before sufficient material is charged to meet the stirrer, or the temperature of a piece of equipment not currently being used. These result in a change of dimensionality of the operating cloud at discrete boundaries in the operating process. These are dealt with in GPC by the introduction of discrete phases and corresponding phase numbers that correspond to a set of variables that has meaning in each given phase. The number of these phases is arbitrary, and may vary from batch to batch.
In the Curvaceous application of GPC, phases are explicitly accounted for through the use of a table indicating the changing role of each process measurement and variable throughout the batch. Thus a temperature may be a manipulated variable (through a local controller) during one phase, a measured process variable in another, and the value may then be carried to later phases as an external variable.
Many batch processes involve multiple stages in producing a final product. These previous stages may be explicitly recognised as producing intermediates that may be used in many different subsequent processes, or may be conceptual in that the entire batch is moved on to the next stage of processing, as with semiconductor manufacture. GPC deals with this by allowing the construction of an explicit model for each stage, while retaining the ability to analyse across stages to determine the effect on the end-product of operations at an early stage. The top image in the figure above shows such an investigation, examining the effects of processing in a particular phase of the first stage of a three-stage process on the final product qualities.
The models for these stages are more powerful by allowing the propagation of quality or process measurements from previous stages to subsequent ones. Use of external variables (those not reflecting current operating conditions, but which are expected to affect them) allows the propagation of batch data for precursor or intermediates through each step of the process to final product. This allows models and operation to explicitly account for, and adjust to, the conditions and qualities of intermediates previously made and used in the current batch.
GPC has benefits throughout the process. The n-dimensional projection allows representation of data, which would normally require a collection of control charts, to be displayed in one simple graph. Interactions between the variables are specifically accounted for so enabling the simultaneous consideration of cross-variable effects, and identifying excursions from good operation that would not be detected by traditional means. The 'no-maths' and visual features of GPC allow models to be quickly understood and modified.
In batch processing, time-dependent dimensionality requires new extensions. The static graphs used to show compliance for a continuous process are now replaced with a time-series, or animation that captures the entire timeline of the batch. With this one animation, compliance with both static and dynamic variable limits can be shown, and readily understood.
A real-time image during processing is shown in the diagram above. The process variables are highlighted in grey, the quality variables are in cyan. As these will not be measurable until the end of the batch, the values displayed are currently predictions of the final qualities. The green limits have been automatically displayed from the prepared model, showing acceptable ranges of the variables. This allows the operator to quickly and easily see the potential operating range for each variable, and which ones require close control.
The application of GPC to batch processes is still being developed but early successes include reduction in run-to-run variation, reducing start-up times, and improving product yield. New opportunities continue to emerge.
Key to the application of GPC lies in the Inselberg parallel coordinate transformation, which allows visual representation of high dimensional data. This allows, for the first time, the leap from 2- or 3-dimensional representations to an arbitrary number of dimensions.
GPC delivers optimum results with continuous application over long periods aimed at keeping process performance at its best all of the time, not just occasionally, as is the case in many plants today.
Big gains have been reported from using GPC in data mining by Huntsman Petrochemicals and in rationalising alarm limits and alerts by INEOS Chlor, while Mallinckrodt Chemicals has detailed major benefits of the technology in operating envelope models for continuous process control. Other successful applications include continuous improvement methodologies (Six Sigma), quality by design, response surface analysis of design of experiments, and real-time optimisation.
Building on these applications, Curvaceous has been working to extend GPC to analyse and control multi-phase, multi-stage batch processes. Batch processes are common in pharmaceutical and speciality chemicals, and can represent some processes common in high tonnage continuous plants, such as PVC, polyethylene and coking in oil refineries. Process start-up of continuous processes is best viewed as a batch process. Semiconductor manufacturing, for example, is really a multi-stage batch process in which each stage has only one phase.
Different tools are used throughout the application of GPC. A visualisation tool is the basis of the technology, and allows the off-line exploration, analysis and review of process data. It is also used in the construction of process models that are used for other products. An online tool then provides real-time control and optimisation through OPC, using models provided through analysis to compute dynamic alarm and alert limits, guiding the operator through process optimisation.
Batch processes are inherently more difficult to analyse and control than their continuous counterparts. Rather than instantaneous operating conditions, the full time-varying history of the batch becomes important. There are separate phases of operation, transfers from equipment, and intermediate products, one batch of which may be used in multiple subsequent batch steps. These all add complexity to the analysis and control problems, adding more variation. Minimising run-to-run variability as well as variability across stages is a common goal as much higher dimensionality of batch processes means variability can appear in many more places.
In a typical three-stage manufacturing process where each stage consists of reaction and separation steps, the steps may be broken down time-wise into individual phases of operation that correspond to different activities taking place sequentially during the stage.
With a continuous process, the problem is often that of finding the Best Operating Zone (BOZ) or 'cloud' of operating points that produce product within specifications, obey process safety considerations, and may have some element of plant optimisation. Excursions from the BOZ correspond to unacceptable or untested conditions that are sub-optimal, poor quality or dangerous. In the simplest case for batch operations, this cloud becomes a 'tube' or tunnel through time containing the desired values of all parameters. The tube changes size and shape dramatically from phase-to-phase with the time of processing through the batch, and represents the interaction of the remaining parameters.
In online applications the time-dependence is addressed by the modification of the Curvaceous process model to include time as an explicit variable. The current acceptable cloud of variables and their interactions is restricted by the time, allowing the control not only of instantaneous process variables, but their trajectory through the entire batch. Modifying the model is simple from the user's perspective, requiring only the addition of time as an additional variable. Realtime control limits can then be calculated with no further input required, and can be fed back to the plant operator.
Multi-phase processes introduce an additional consideration. At different points in the operation, some variables may have no bearing on the process, for instance agitation rate before sufficient material is charged to meet the stirrer, or the temperature of a piece of equipment not currently being used. These result in a change of dimensionality of the operating cloud at discrete boundaries in the operating process. These are dealt with in GPC by the introduction of discrete phases and corresponding phase numbers that correspond to a set of variables that has meaning in each given phase. The number of these phases is arbitrary, and may vary from batch to batch.
In the Curvaceous application of GPC, phases are explicitly accounted for through the use of a table indicating the changing role of each process measurement and variable throughout the batch. Thus a temperature may be a manipulated variable (through a local controller) during one phase, a measured process variable in another, and the value may then be carried to later phases as an external variable.
Many batch processes involve multiple stages in producing a final product. These previous stages may be explicitly recognised as producing intermediates that may be used in many different subsequent processes, or may be conceptual in that the entire batch is moved on to the next stage of processing, as with semiconductor manufacture. GPC deals with this by allowing the construction of an explicit model for each stage, while retaining the ability to analyse across stages to determine the effect on the end-product of operations at an early stage. The top image in the figure above shows such an investigation, examining the effects of processing in a particular phase of the first stage of a three-stage process on the final product qualities.
The models for these stages are more powerful by allowing the propagation of quality or process measurements from previous stages to subsequent ones. Use of external variables (those not reflecting current operating conditions, but which are expected to affect them) allows the propagation of batch data for precursor or intermediates through each step of the process to final product. This allows models and operation to explicitly account for, and adjust to, the conditions and qualities of intermediates previously made and used in the current batch.
GPC has benefits throughout the process. The n-dimensional projection allows representation of data, which would normally require a collection of control charts, to be displayed in one simple graph. Interactions between the variables are specifically accounted for so enabling the simultaneous consideration of cross-variable effects, and identifying excursions from good operation that would not be detected by traditional means. The 'no-maths' and visual features of GPC allow models to be quickly understood and modified.
In batch processing, time-dependent dimensionality requires new extensions. The static graphs used to show compliance for a continuous process are now replaced with a time-series, or animation that captures the entire timeline of the batch. With this one animation, compliance with both static and dynamic variable limits can be shown, and readily understood.
A real-time image during processing is shown in the diagram above. The process variables are highlighted in grey, the quality variables are in cyan. As these will not be measurable until the end of the batch, the values displayed are currently predictions of the final qualities. The green limits have been automatically displayed from the prepared model, showing acceptable ranges of the variables. This allows the operator to quickly and easily see the potential operating range for each variable, and which ones require close control.
The application of GPC to batch processes is still being developed but early successes include reduction in run-to-run variation, reducing start-up times, and improving product yield. New opportunities continue to emerge.
