Investigation of Geometric Process Control

Investigation on Curvaceous and VisuLab Software for Process Analysis, Supervisory and Control Tian Lin Institute of Particle Science & Engineering University of Leeds Leeds, UK School of Process, Environmental and Materials Engineering 04/08/2010

Purpose of the investigation Focus on the Curvaceous and VisuLab software Interests in multivariant visualisation methods, especially parallel coordinates Aims at process applications for analysis, supervisory and control

Introduction of Curvaceous software History Developed in 1990s by Robin W. Brooks, commercialised as Curvaceous Software Ltd. in 1994, and now Process Plant Computing Ltd. ( www.ppcl.com ) Brought to the chemical industry in 2001, with the first customer Ineos Chlor, UK, and many applications later in chemicals, oil and gas, materials, pharmaceuticals, etc. Main technology Geometric Process Control (GPC) , which analyses process data in parallel coordinates by geometric methods Data-driven real-time technique, without complex mechanic or mathematical models Components C Visual Explorer (CVE): offline data mining, e.g. to find out the underlying reasons of unexpected process behaviours and problems C Process Modeller (CPM): online monitoring, control and optimisation to keep the process operation aligned to the best experience and seek further improvement C Response Surface Visualiser (CRSV): process and product development to achieve the desired quality and performance, with the help of experiment design

Introduction of VisuLab History Developed as free software by Hans H. Hinterberger of Department of Computer Science, Swiss Federal Institute of Technology (ETH) Zurich in 1989 For e-learning and comparative study of multivariant visualisation methods, with application examples of static statistical data Our group applied VisuLab to visulise principle components and independent components of process variables for fault detection Main technology Selection of existing visualisation methods Emphases on user interactivity, more data visualisation and analysis methods, not for process-oriented, real-time or automated applications Components 4 visualisation methods in coordinated display: parallel coordinates, scatterplot matrix, permutation matrix and Andrew’s curves Data clustering tools Also many alternatives with similar functions as VisuLab

Parallel coordinates Invented by Alfred Inselberg in 1980s Multidimensional Points mapped as polygonal lines A real example of process data in parallel coordinates A point in traditional 3D coordinates and mapped in parallel coordinates A bunch of process data in parallel coordinates

Curvaceous’ view of alarm management Fig(a) - Ideal Hi/Lo limits at the exact boundary of the normal operation envelope Fig(b) - Actual Hi/Lo limits inside or outside the normal operation envelope Inside ones with frequent false alerts moved outwards Others start frequent false alerts due to variable interactions so moved as well Fig(c) - “Leap-frogging” until the Hi/Lo limits close to the HiHi/LoLo safety limits

Alarm problems shown in parallel coordinates Original alarm settings according to process data ranges and applied to all the modes (as coloured) Alarm settings changed after a period of operation and many lost effects

Geometric process control The idea Keep the operation consistent with the best historical runs and reproduce the best performance Process variable interactions are considered for tighter operation zones / alarm limits The alarm limits are warning for diversions of process operation and product quality Methodology Successful operation records are selected to form an operation envelope Operation zones / alarm limits are updated dynamically with real-time measurements If the operation diverts, GPC searches for optimal control actions in geometric sense With actual process measurements, potential quality scope can be predicted With desired quality indices, feasible control space can be determined Though without mechanic or mathematical models, GPC results are physically sensible

Operation envelope Points on nonlinear boundaries in traditional coordinates form sectional linear and convex polygonal lines in parallel coordinates An irregular operation envelope, formed by process data within constraints, also maps to sectional linear and convex polygonal lines; The envelope can be used as HiHi/LoLo limits Projection of an operation envelope in 2D

Best operation zone Tighter operation zones are derived with variable interactions With the fixed value V 1 of variable X 1 , the feasible range [L 2 ,H 2 ] of variable X 2 can be determined by the envelope E H and E L The feasible range of each variable can be determined by interactions from all other variables, and all these ranges form the best operation zone – the green lines

Suggestion on alarm clearance If any process variable (black dots) runs outside the best operation zone (green lines), it’s an alarm – the purple caret Then GPC will give suggested adjustments of control variables (blues dots), and the predicted best operation zone (blue lines)

Calculation of suggestions Calculate the effects of control variable changes on the offended limits of the alarmed variables The effect of change in control variable Xm on upper limit ULn of the alarmed Xn is dULn/dQm = -(L-x4)/x4 The effects of change in control variable Xm on all the offended limits Li are summed as dL/dQm = Σ i dLi/dQm The control variable with the maximum effects are adjusted for the time This is repeated until normal operation recovered But the step length is limited by VUm and VLm , and a precondition that ULn is determined by Xm but not other variables

Process setup and optimisation The methods for deriving best operation zones and alarm clearance suggestions can be applied to process setup (left figure) and optimisation (left figure) respectively Assign values to the previous (left) variables, then the range of the next (right) variable can be determined by the interactions from all the previous variables assigned with a value If the control variables are assigned with values first, the potential quality scope can be predicted; or vice verse, the feasible control space can be determined Maximise/minimise the lower/upper limits of the concerned variables, until a narrow operation zone is reached as working point The control variable with the maximum effect on the limits is adjusted These can also be implemented while online

More functions Change of envelope when Serious diversion outside the current envelope Too low alarm frequency, i.e. margin for tighter control Different phase of processes Pre-defined events Real-time implementation Automatically send the GPC suggestions to controllers Production control Visualise the relationships between the economic performances and process variables

Functionalities of Curvaceous software Set up process operation envelopes in parallel coordinates based on clustered successful historical data; Calculate dynamic limits for process variables according to the envelope and real-time measurements, and estimate scopes of quality variables; Judge alarms for both process and quality variables, and derive adjustment of controllable variables to clear the alarms; Change to a tightened operation envelope appropriate for better performance; Change to corresponding envelopes for different process phases, modes and events; Optimise controllable variables by maximising lower limits or minimising upper limits; Optimise control spaces to achieve desired quality with aid of experiment design.

Advantages of Curvaceous software Gain new insights into the processes, and resolve hidden inefficiencies and mysterious irregularities; Reproducible good quality and performance, and QbD and PAT compatible methods; Increase production efficiency and throughput by accelerating unit start-up and shutdown; Scientific alarm management with fewer alerts, lower false ratio, promoted confidence and improved safety; Achieve business objectives by optimising the economic and legal factors such as yield, consumption, cost, profit, waste and emissions; Capture process essentials with easy and fast methods but without complex maths or expensive modelling or rule base construction.

Other issues of Curvace software Different with MSPC MSPC defines statistical concept of normal operation without exact limits GPC decides explicit operation envelopes and normal operation zones as alarm limits, with dynamic updates and suggested actions Dependence on the historical data GPC also utilises verified working points, in the format of clustered historical data sets, and keeps the process operation in a stable neighbourhood of them. In its algorithm to clear alarms (Section III.5), the limits imposed on the adjustment amount ( VUm and VLm in Figure 12) mark the boundary of the linearisation neighbourhood. The algorithm is kind of iterative search in gradient directions with full step length. The later changes of other variables due to the adjustment are not considered either. Hence the algorithm may not converge under some complex situations depending on the selected historical data sets. moves of the control-variables to bring the process back into the BOZ cannot be found ”, which was proposed to be dealt with an outer BOZ, i.e. a wider historical data space The selection should be well classified around a stable and robust operating point, and ensure ample historic data of satisfactory performance to broaden the feasible range of the algorithm. the process dynamics of each stage are expected to be handled by the control systems. That means, for complicated time-dependent dynamics that the process states of a certain instant depend on those of previous instants, GPC may not be able to deduce the correct results. Automated implementation CPM works as a layer above the process control system such as DCS, with little or no change to it. Any operation suggestions by the GPC algorithm can be automatically sent as set points to the control system for implementation. It’s said that Curvaceous is thinking of integrating CPM with DCS products. practical trials hadn’t been done up to 2007 most advices were accepted by the operators, but not all of them

Visualisation methods of VisuLab scatterplot matrix relationship between only two variables , pairs in a matrix layout permutation matrix Multivariant data can also be presented as a row of columns , stacked and exchange for patterns Andrew’s curves fx(t) = x1/20.5 + x2sin(t) + x3cos(t) + x4sin(2t) + x5cos(2t) + ... -π < t < π mean preservation, variance preservation, distance preservation, linearity preservation and proportional one-dimensional projections sequence-sensitive, as the low-frequency terms, i.e. the first few dimensions, are more readily recognized. clustering tools including nearest neighbour hierarchical clustering, furthest neighbour hierarchical clustering, average neighbour hierarchical clustering and K-means clustering, with options of Euclidean and Manhattan distance VisuLab also provides offline multivariant data visualisation in parallel coordinates, thus can also be applied to process analysis as CVE, though it may not be as powerful as CRSV.

Random walk model and DWS Random walk (Rogers 2008 ) Each photon is assumed to execute a random walk through the sample Light by all paths independently contribute to the detection Described by transport mean free path (distance between two scattering, related to individual particles scattering properties and particle concentration) , and possibility of certain photon travel length To solve a diffusion equation Diffusing wave spectroscopy (DWS) Measurements as DLS but interpreted by random walk (Scheffold 2002). DWS ResearchLab (LS Instruments, Switzerland)

Summary The Curvaceous software adopts geometric methods within parallel coordinates, and aims at keeping the process operation aligned to the best experience. It is a novel real-time solution for process supervisory, control and optimisation. It can also be applied to offline process analysis and setup, and has extraordinary performance in resolving hidden inefficiencies and mysterious irregularities. Though lack of news on its automated implementation, it has been admitted by many customers from industry. VisuLab is not designed for industrial processes. However, it can still be applied to process analysis with offline data.

Thank you ! School of Process, Environmental and Materials Engineering

Investigation of Geometric Process Control

More Related Content

What's hot (17)

Viewers also liked (14)

Similar to Investigation of Geometric Process Control (20)

Recently uploaded (20)

Investigation of Geometric Process Control