WO2000020939A9

WO2000020939A9 - Method and system for monitoring and controlling a manufacturing system

Info

Publication number: WO2000020939A9
Application number: PCT/US1999/023379
Authority: WO
Inventors: James D Keeler; Edward S Plumer; Joshua Brennan Ellinger
Original assignee: Pavilion Tech Inc
Priority date: 1998-10-06
Filing date: 1999-10-06
Publication date: 2000-08-24
Also published as: EP1042706A1; CA2313134A1; WO2000020939A1; JP2002526852A; AU6294999A

Abstract

Neural network models interface with distributed control systems associated with a manufacturing facility for performing a manufacturing facility. The neural network models receive measured variables of the manufacturing process to predict process performance data, and provide the performance data on a real-time basis to a communications server. A graphical user interface communicates over a network, such as the Internet or a corporate Intranet, to receive the real-time performance data, including performance metrics such as key performance real-time analyzers, for presentation to aid managers in making decisions regarding the manufacturing process. The communications server also interfaces with an off-line model engine to transfer the neural network model and real-time performance data for analysis on the off-line engine. Object oriented box transformes enhance publication and subscription of the performance data from the neural network models.

Description

METHOD AND SYSTEM FOR MONITORING AND CONTROLLING A MANUFACTURING SYSTEM

TECHNICAL FIELD OF THE INVENTION

This invention relates in general to the field of control systems, and more particularly to an improved method and system for monitoring and controlling manufacturing process performance indicators .

BACKGROUND OF THE INVENTION

Complex manufacturing facilities perform manufacturing processes that manipulate a wide variety of resources to produce desired end products. For instance, a petrochemical plant co-generation facility produces steam for plant operations and electricity for use by the plant and for sale on the public utility grid. Resources used for such a co-generation facility include tangible resources such as petrochemical raw materials and fuel for steam generation, and intangible resources such as environmental constraints, equipment limits and electrical grid configuration limits.

Equipment associated with a typical co-generation facility has a wide range of operating conditions and is distributed throughout the manufacturing facility in interacting and interconnected unit operations. These unit operations generate massive amounts of performance information available for monitoring the manufacturing process, including measured variables of the process such as pressure, temperature, flow rate, and production rate performance measurements. Performance information is generally measured at unit operations and translated to an electrical signal for transmission to a distributed control system (DCS) and for storage in a process data historian (historian) . Similar performance measurements are generally taken for other types of manufacturing processes, including pulp and paper processes, refining processes, food processing, and electric power utility production . Process performance measurements provide manufacturing facility management with essential information for monitoring and optimizing the manufacturing process. One difficulty associated with the mass of performance data available to facility management is that there is too much information from a managerial point of view to synthesize for decisions that allow optimal facility production. Using the example of the co-generation facility, a manager can alter the allocation of electricity and petrochemical production based on electricity price changes and steam demand changes, which are a function of the process production rates for various unit operations in the plant. The manager's production decisions may also face limiting constraints such as environmental limits, equipment limits, and electrical grid configuration limits. A manager attempting to make production decisions under fluctuating production conditions faces a challenging task of information overflow. In order to focus decision makers on the most relevant possible information, many companies use a standard set of performance metrics to monitor and analyze the economic results of a manufacturing facility and process. For instance, petroleum refiners typically use indices provided by Solomon Associates, such as the

Solomon Utilization index, to gauge refinery performance. The Utilization index calculates utilization of refinery processing units compared to their theoretical capacity. The use of a standard set of performance metrics limits the need for managers to monitor detailed manufacturing performance data, which is helpful from a business monitoring standpoint. However, performance metrics are generally computed from accounting data generated after completion of an accounting period, thus reducing their usefulness for management decision making between accounting periods .

In addition to the vast quantity of information available for monitoring and controlling a manufacturing process, a manager also faces critical decisions regarding the quality of the information on which he or she must rely. For instance, historical data relating to a manufacturing process provides insight for production variations that have occurred and the likely effect of such variations on future production goals. However, the use of historical data tends to rely on subjective determinations and has limited utility in situations where production fluctuates according to unpredictable parameters . One technique available for aiding management production decisions is the generation of a model of the manufacturing process and using the model to test varying production parameters. Pavilion Technologies, Inc. has developed a number of tools that use neural network techniques to develop models of manufacturing processes. For instance, Pavilion's Soft Sensor" technology can use historical data to train a neural network to model a given manufacturing process, enabling predictions of process behaviors such as quality measurements, production measurements or other relevant process information. Once a high-fidelity model of a manufacturing process is developed, the model can use measured variables of the process to predict process behavior and to optimize and control unit operations to achieve a desired result. For instance, a distributed control system can provide measured variables to a run time application engine (RAE), allowing manipulation and control of the variables to optimize production in view of the economics, production capabilities, and constraints of each unit operation. Pavilion's Process Insights, Boost, and Process Perfecter products enhance the ability of a run time application engine to optimize and control unit operations . Neural network models, such as those enabled by

Pavilion's technological advances, significantly enhance manufacturing process efficiency and optimization through direct interaction with distributed control systems. For instance, a neural network model can use real-time measured variables of a process controlled by a distributed control system to predict process results, and can command the distributed control system to vary predetermined parameters to obtain desired process results. Although the use of neural network models in communication with control systems generally results in more effective use of process resources, the neural network models have had limited utility in supporting real-time managerial decision making. One difficulty is establishing an interface between the neural network model and the management decision makers. The neural network models generally perform critical control functions which, if interrupted by requests for information from other systems, can disrupt the manufacturing process. Further, the neural network models use and generate a large quantity of data, making it difficult to sort out and present the most pertinent information. Another difficulty associated with the use of neural network models is the format used by the models to store performance data. Generally, neural network models maintain a flat table driven database from which it is awkward and difficult to extract desired data points .

SUMMARY OF THE INVENTION

Therefore a need has arisen for a method and system which enables the display and control of real-time and predictive manufacturing process performance information to enhance and optimize process control.

A further need exists for a method and system which enables access to neural network or other (first principles) model information without disrupting a manufacturing process. A further need exists for a method and system which allows individual users to customize the presentation of key performance metrics on an individual basis to provide information presentation relevant to individual user interests and responsibilities. A further need exists for a method and system which provides facilities for publishing and subscribing performance data of a manufacturing process from a neural network model associated with the manufacturing process. A further need exists for a method and system which enhances accessibility of predetermined manufacturing process performance data for use to manage, optimize and control the manufacturing process.

In accordance with the present invention, a method and system for presenting performance data of a manufacturing process enhances the ability of decision makers to access and use process performance data predicted by a neural network model. A uniform flexible architecture creates a distributed virtual plant to allow real-time display and manipulation of performance data, enabling modeling, simulation, model predictive control and optimization of a manufacturing process.

More specifically, the present invention provides a method for presenting performance data of a manufacturing process that is performed at a manufacturing facility. A neural network model of the manufacturing process is generated using measured variables of the process. The neural network model is able to predict process behavior such as quality measurements, production measurements or other relevant process information. The neural network model predictions provide real-time and future predictions of measured variables of the process as well as predetermined performance variables of the process. For instance, a key performance indicator (KPI) normally calculated with historical and accounting data is instead calculated as a key performance real-time analyzer (KPRA) by using real-time neural network models of performance measurements. The neural network model interfaces with the manufacturing facility, which provides the neural network model with real-time performance data, such as measured variables of the process . The neural network model predicts process performance data and transfers the predicted performance data over a network for display on a graphical user interface. In one embodiment the neural network model interfaces with a distributed control system and provides predictive performance data for enhancing control of the manufacturing process by the distributed control system. The neural network model also interfaces with a communications server. The manufacturing process can use plural neural network models, with one or more neural network or other models interfaced with distributed interconnected unit operations associated with the manufacturing process. Each such neural network model can interface with the communications server over a network, allowing the communications server to request and receive real-time performance data from the neural networks, including measured variables of the process, variables of the process predicted by the neural network, and performance variables predicted by the neural network, including KPRAs .

The communications server interfaces through a network, such as a corporate intranet or the Internet, with a graphical user interface. The graphical user interface displays performance data of the manufacturing process, including real-time and predictive performance data, in a user friendly format to enhance managerial decisions regarding the manufacturing process. The graphical user interface allows selection of predetermined performance data for display so that managerial decisions can be based on selective relevant information. For instance, the graphical user interface can display a KPRA for the manager=s standard set of performance metrics based upon real time data received on-line from the communications server.

In addition, the graphical user interface accepts inputs to allow "what-if" and setpomt analysis. For instance, the graphical user interface accepts a variation in the value for one or more predetermined performance data points and transfers the variation to the communications server over the network. The variation can be, for instance, a variation in a setpoint to determine KPRAs associated with the new setpoint, or can be a variation in a KPRA to determine setpoints for achieving the KPRA in an optimal manner. The communications server can either request a neural network model to run a what-if or setpoint analysis based upon the input variations to the performance data, or can transfer the model in its real-time state to an off-line model engine so that the off-line model engine can run the analysis. The communications server provides the results of the what-if or setpoint analysis to the graphical user interface. Thus, for instance, a decision maker can input hypothetical conditions for the manufacturing facility and view projected KPRA values for the hypothetical conditions based upon real-time facility parameters . The graphical user interface can then accept production setpoints from the user for transfer through the network to the communications server and for implementation by the control and optimization models associated with the manufacturing process.

The present invention provides a number of important technical advantages. One important technical advantage is the presentation of real-time process performance data, such as real-time KPRA values, on a user friendly graphical user interface. By making predetermined and selective real-time performance data available from the neural network through the communications server, the present invention enables managers to study real-time a production process at a remote location and to implement production decisions through the graphical user interface.

Another important technical advantage of the present invention is that it makes real-time performance data available with minimal risk of disruption of the control process. For instance, the communications server can mediate requests for performance information to minimize the risk of overloading neural network models involved in process control. The communications server supports future state views of a production process by making real-time performance data available to an off-line model engine for what-if or setpoint analysis, thereby minimizing any impact on computing platforms that support the neural network models.

Another important technical advantage of the present invention is the ability of neural network models to accurately predict future process state conditions for presentation to decision makers at the graphical user interface. The performance data provided to the graphical user interface can include real-time measured variables of the process, predicted real-time measured variables of the process, and performance indicators of the process, such as KPRAs, that monitor and analyze economic results . Real time predictive and future performance indicator calculations simplify qualitative and quantitative economic decisions faced by managers of a manufacturing process.

Another important technical advantage of the present invention is that it presents a uniform flexible architecture for modeling a manufacturing process, for simulation of a manufacturing process, for model predictive control of the manufacturing process, and for optimization of the manufacturing process. By combining these four functions in a single system to make performance data available to a remote graphical user interface, the present invention simplifies and enhances control of a manufacturing process. Managerial-level production decisions can be made with reference to actual real-time production conditions, with simulation of variations to those production conditions, and with predictions for optimal setpoints based on anticipated future production conditions. With this knowledge, boiled down to a user friendly presentation of KPRAs, management can make optimization decisions and provide setpoints for those decisions directly to the control system.

Another important technical advantage of the present invention is the "pick-up and copy" of a neural or other network model and real-time performance data by the communications server for use on an off-line model engine. This allows a present state of the neural network model and manufacturing process configuration to be detached and frozen on a real-time basis for what-if and/or setpoint analysis. The present invention uses active transform building blocks to provide a hierarchical tree, allowing a variable name relationship for the transform blocks for simplified publication and subscription of desired data to the graphical user interface .

BRIEF DESCRIPTION OF THE DRAWINGS

A more complete understanding of the present invention and advantages thereof may be acquired by referring to the following description, taken in conjunction with the accompanying drawings in which like reference numbers indicate like features and wherein:

FIGURE 1 depicts a block diagram of a manufacturing process neural network control system interfaced with a graphical user interface; FIGURE 2 depicts a graphical user interface presentation of key performance real-time analyzers;

FIGURE 3 depicts a graphical user interface for selecting key performance real-time analyzers;

FIGURE 4 depicts a chart view of a graphical user interface presenting a temporal view of a key performance real-time analyzer;

FIGURE 5 depicts a graphical user interface presentation of selected manufacturing facility information; FIGURE 6 depicts a graphical user interface plan view of manufacturing facility;

FIGURE 7 depicts a block diagram of a neural network model;

FIGURE 8 depicts a block diagram of a neural network building block and data storage table;

FIGURE 9 depicts a generic object-oriented wrapper for processing element modeling;

FIGURE 10 depicts a block diagram for neural network processing of algebraic and iterative simulations; and FIGURE 11 depicts a block diagram of a neural network model for setpoint and what-if analysis.

DETAILED DESCRIPTION OF THE INVENTION

Preferred embodiments of the present invention are illustrated in the FIGURES, like numerals being used to refer to like and corresponding parts of the various drawings .

Neural network technology has greatly enhanced manufacturing facility operations by providing models that simplify and optimize process control decisions at a facility level. However, the tremendous power of neural networks has not yet been harnessed in terms of supplying high-level information for facility management with facility-wide view towards optimizing manufacturing process economics. This is partly due to the large quantity of information made available by neural network models, and partly due to the complicated data structures used by neural networks . The present invention provides a link between performance information available through neural networks and performance metrics used by managers for making facility-wide economic production decisions. Referring now to FIGURE 1, a simplified block diagram of a manufacturing facility for producing electricity is depicted. A boiler 12 produces steam for rotating a turbine 14 so that the turbine can generate electricity. Boiler 12 provides measured variables, such as steam temperature and pressure, to a distributed control system 16. Distributed control system 16 controls the measured variables provided by boiler 12 by manipulation of facility operations. For instance, distributed control system 16 can maintain a predetermined temperature or pressure setpoint for boiler 12 by manipulating electromechanical controls for fuel flow to boiler 12. Similarly, distributed control system 18 interfaces with turbine 14 to receive measured variables of the turbine operation such as turbine RPM, power output, and turbine steam pressure and temperature. Boiler 12, turbine 14 and their associated distributed control systems form interacting and interconnected unit operations for the production of electricity. Thus, for instance, distributed control system 16 and turbine distributed control system 18 might manipulate electromechanical controls to maintain predetermined setpoints for production of a predetermined amount of electricity. A historical database 20 tracks and stores the measured variables for analysis and accounting purposes .

The block diagram of the electricity manufacturing facility 10 depicted by FIGURE 1 is simplistic compared to typical co-generation facilities. For instance, a typical co-generation facility has a large number of boilers and several turbines. In addition, steam from the boilers also supports petrochemical production. Thus, a small change by a single distributed control system to achieve an existing or new setpoint can have ramifications along the production process that are difficult to predict. In addition, some process variables are difficult to measure. Neural networks aid a production process at a manufacturing facility by accurately modeling measured variables and by predicting unmeasured variables for use in estimating impacts of setpoint changes to the manufacturing process. For instance, a neural network model 22 for boiler 12 can predict temperature and pressure changes associated with changing electric production setpoints. Similarly, neural network model engine 24 for turbine 14 can predict turbine RPM changes associated with changing steam pressure setpoints. By communicating within manufacturing facility 10 in cooperation with distributed control systems, neural network model 22 and neural network model 24 can optimize setpoints over time to achieve desired production results. The neural network models can also predict unmeasured variables, such as fuel usage, to provide estimates of production costs. The first step in establishing a neural network model for controlling a manufacturing process is to build the model. Soft Sensor^", sold by Pavilion Technologies, Inc., enables development of neural network models of manufacturing processes by a number of techniques. For instance, a neural network model can be trained using historical data gathered from the manufacturing process. Alternatively, the neural network model can be trained by linking in a first principles model or by training the neural network from data generated by a first principles model. Once the neural network model is trained, it is saved as a file and transferred to a computing platform associated with the manufacturing facility.

The neural network model interfaces with distributed control systems through an application programming interface 26, such as the Pavilion Data Interface (PDI) . The Pavilion Data Interface ties the neural network model into the distributed control system to provide measured variables of the manufacturing process to the neural network. Pavilion's runtime application engine (RAE) accepts measured variables of the process online from the distributed control system, and processes the data to provide the distributed control system with predictive values and desired setpoints. The runtime application engine can use Pavilion's dataflow architecture 28 to provide predictive values. Pavilion's dataflow architecture 28 is described in greater detail in U.S. Patent No. 5,768,475, entitled Method and Apparatus for Automatically Constructing a Data Flow Architecture, by Godbale, et al . In summary, the runtime application engine accepts measured variables of the process in a temporal flat table format and uses the neural network models to provide desired predictive values. For instance, the runtime application engine calculates desired values used from a column of a flat table and repopulates the table with the calculated values in a new column. The runtime application engine can interface with other facility runtime application engines to generate predictive values for populating the table. Thus, for instance, the runtime application engine of neural network model 22 for boiler 12 accepts measured variables from boiler distributed control system 16 and neural network model 24 for turbine 14 to predict electricity production levels. The runtime application engine also provides data back to the distributed control system, such as setpoint data for achieving desired process results. This process repeats at predetermined time intervals, allowing the neural network models to exert control over the distributed control systems. Once a high fidelity network models of manufacuturing process operation units are developed, other Pavilion tools are available for using the model to optimize and control the unit operations. For instance, Pavilion's Process Insights, BOOST and Process Perfecter products enable optimization and control of unit operations using economic, process and constraint information. In this way, neural network models provide a powerful tool for enhancing control at a facility level.

The data available from neural network models also provides important process information for managerial level decisions relating to the facility, assuming that relevant data is available at needed management levels in a timely way. The present invention provides a two-tier system to get neural network data away from the facility for use by decision makers in a manner that avoids burdening operations at the facility. A communications server 30 interfaces with neural network model 22 and 24 through a network 32, such as an intranet or the

Internet. Communications server 30 can interface with other facility data sources 33, such as information generated in Excel, SAP and ODBC formats, as needed. Communications server 30 mediates requests for information from associated neural network models and, depending upon the availability of a predetermined neural network model or models to provide data, performs a "pick up and copy" operation to transfer a present state of a predetermined neural network model or models from facility 10. The neural network model or models received by communications server 30 maintain their configuration in their actual real-time state, effectively detached from facility 10 and frozen at the present time. Communication server 30 interfaces with an off-line model engine 34 and can transfer a neural network model or models received from facility 10 to off-line model engine 34. Once communication server 30 has received a neural network model, communication server 30 provides a unified, virtual plant tool to analyze the manufacturing process of facility 10. Communication server 30 acts as a central base to perform modeling, simulation, model predictive control and optimization. Communication server 30 can mediate requests for these functions to facility neural network models on an as-available basis, or alternatively, can use off-line model engine 34 to perform the functions with real-time data received from facility 10.

Once the models for predicting the facility manufacturing process are completed and transferred to the client-server architecture supported by communications server 30, KPRAs generated with the models can be made available so that executives, plant managers or operators can have real-time indicators of how the facility s operating. For instance, real-time KPRAs can provide key information regarding production rate, quality and limiting constraints. Furthermore, once this information is available for display, it can also be used to run facility operations m a more profitable manner.

For instance, Pavilion multi-unit optimization technology can optimize overall facility operations using real-time pricing information, contract information, operational targets and constraints, and environmental costs or constraints. Multi-unit optimization is available online through communications server 30 so as to optimize each unit and the combined unit operations. A multi-unit optimizer calculates optimal profit setpoints and sends the setpoints to the open or closed-loop controller/optimizer of each individual unit to implement setpoint changes .

Off-line model engine 34 supports optimization analysis by accepting variations in performance setpoints and goals to run Awhat-ιf@ and setpoint analysis for a variety of changing constraints, priorities and unit configurations and status. Based upon this analysis, a user can modify setpoints provided by multi-unit optimization for implementation at the facility. As an example, an analysis of total fuel cost includes not only the price of fuel, but also the costs associated with fuel transportation, fuel efficiency based on the heat rate for the unit burning the fuel, emissions from burning the fuel, and αegradation of equipment caused by buring the fuel, such as slagging of burners and degrading catalysts. The off-line engine supports modeling with such fuel cost factors to aid in fuel buying and selling decisions. Communications server can receive inputs for these values from a user, or can use tools such Internet web browsers to automatically update needed information on a real-time basis from outside of the manufacturing facility.

Communications server 30 transfers performance data from manufacturing facility 10 over a network 36, such as a corporate intranet or the Internet, to a graphical user interface 38. Graphical user interface 38 provides a point and click environment that allows remote viewing of real-time performance data for manufacturing facility 10. For instance, by using TCP/IP and HTML pages transferred over the Internet, managers at a headquarters facility can view real-time manufacturing facility performance data for facilities located around the world.

By selecting the KPRA option on graphical user interface 38, a display such as the display depicted by FIGURE 2 provides KPRA performance data for viewing.

KPRA column 40 includes a number of performance metrics typically associated with a co-generation facility, but calculated as KPRAs using real-time data received from communications server 30. For instance, as described in description column 42, graphical user interface (GUI) 38 depicted by FIGURE 2, provides information such as total electricity production, which can be a measured variable of the process, and Utilization factor, a Solomon index normally computed using historical accounting data. The KPRA values depicted m column 44 can result from a direct download of real-time measured variables of the process, variables of the process calculated by neural network models, or performance metrics such as KPRAs calculated by neural network models. The performance data depicted in column 44 is requested by GUI 38 from communications server 30. Communications server 30 mediates with facility 10 neural network models to obtain real-time data. Once communications server 30 has obtained real-time data, it can provide the data directly to GUI 38 and off-line model engine 34. Subsequent requests for the same data by GUI 38 can result in updates directly from neural network models of facility 10 or predicted values generated by off-line model engine 34.

A given manufacturing facility can have a large number of unit operations, each unit operation having performance data including KPRAs. However, a decision maker may have a particular interest in only limited numbers of performance data. To enable a user to display only desired information, an add/remove index selection 46 is available. Selection of the add/remove index option 46 provides the user with the a display depicted by FIGURE 3. A user selects a KPRA from selection window 48 and then selects add or remove as appropriate. Once a user has identified KPRAs of interest, the user can save these KPRAs to allow the user to call up this information on a repeated basis. Thus, a number of users can identify KPRAs on an individual basis and store their individual configurations on communications server 30 or a computing platform associated with GUI 38 for future reference. The user can also customize the KPRAs by selecting the create new KRPA option 50 or by selecting the new data source option 52. New KPRA calculations are facilitated by Pavilion=s Data Flow architecture and transform calculation engine.

Graphical user interface 38 also cooperates with communications server 30 to provide detailed historical data. By selecting a particular KPRA, such as total electricity production, and then selecting chart option 54, a chart display of historical data for the selected KPRA is available. For instance, referring now to FIGURE 4, a chart display for total electricity production over the last 24 hours is depicted. Additional historical data is available by scrolling the window to the left. A status window 56 provides information for unit operations associated with electricity production at plant site XYZ . By selecting about option 58, a user obtains updated information for the unit operations shown in status window 56, as is depicted by FIGURE 5. Alternatively, the user can receive a plan view of unit operations as is depicted by FIGURE 6.

The client server architecture supported by communications server 30 enables presentation of real-time performance data at graphical user interface 38. However, the flat table data formats typically used by neural network models are generally not well suited for publishing and subscribing purposes. For large projects with large numbers of unit operations, such flat table formats are aw ward to configure. Further, the flat table formats can lead to flaws in neural network model outputs m some instances. Referring now to FIGURE 7, a hierarchal neural network model composition is depicted that enhances publication and subscription of data from the neural network model to other models or a communications server. A system neural network 60 encapsulates a unit A neural network model 62. Unit A model 62 includes a series of transforms, A, B, C and E connected together in a wiring context. Data flows from a distributed control system 64 to transform A for processing, with the processed data flowing from transform A to transform B. A feedback loop is depicted by data flow from transform B to transform E, transform C, and back to transform A. Transform C also receives input from a dependant variable 66. Altogether, the transforms included in unit A model 62 act as a composite transform providing data to transform F of system model 60. Transform F, in turn, provides processed data to distributed control system 64.

The object-oriented box approach depicted by FIGURE 7 provides significant advantages over the flat table approach of run time application engines in terms of data publication and subscription. For instance, the box approach provides a hierarchy structure with identifiable subsets of data that enhance the ability to browse the model for specific data. An example is illustrated by data name 68 "System. Uni . B . x" . Data name 68 identifies by name data X resulting from transform B within unit A 62. The availability of data names, as opposed to table positions, greatly simplifies publication and subscription of the data. Referring now to FIGURE 8, a neural network model 70 is depicted in communication with a distributed control system historian 72 through a Pavilion data interface 74. Interface 74 provides distributed control system tags 76 to model 70. Tags 76 are processed in turn by transforms C, A and B. Transform C outputs data Z, transform A outputs data Y and transform B outputs data X. Thus, model 70 encapsulates the exchange of preconfigured variables for the specified distributed control system. A data monitoring client 78 retrieves data X, Y and Z by either interactively subscribing to the variables within model 70, or directly querying the variables within model engine 70. The object-oriented box architecture enhances interactive browsing by placing variables in a browsable, hierarchical name space. Referring now to FIGURE 9, an air flow filter object-oriented wrapper 80 provides an example of how a transform models a processing element. Air flow filter transform 80 models filter air flow as an exponential moving average. Inputs flow 1, flow 2, flow 3 and CF to transform 80 can originate from a number of sources, including other neural network objects, empirical models, first principles models, signal processing, sensor validation and direct data access such as a Pavilion data interface, OPC or ODBC. Transform 80 processes the input to provide outputs of filtered flow 1, filtered flow 2 and filtered flow 3. Transform 80 provides a uniform model for an air flow filter which is reusable in a number of different neural network models for a given facility that uses a number of different air flow filters. Similarly, generic object-oriented wrappers can model other elements within a facility. The object-oriented box approach using generic object-oriented wrappers for processing elements provides a flexible architecture that allows rapid modeling of related unit operations.

Referring now to FIGURE 10, an object oriented box transform 82 depicts an algebraic transform 84 and an exponential transform 86. Transforms 84 and 86 illustrate the flexibility of the box architecture for on-line and off-line solutions. In an on-line neural network model, these transforms provide uniform and self- contained computation of performance data without a need for referencing tabular historical data since the data for each object oriented wrapper within the transforms is readily available for processing. These transforms are easily transferred with present state data to an offline engine for further analysis without a need for calling up data from tabular formats. Referring now to FIGURE 11, a transform 88 is depicted for providing setpoint and what if analysis. Setpoint analysis is accomplished by selecting targets and/or range constraints and solving the transform to determine optimal setpoint values for achieving the targets. What-if analysis is accomplished by setting setpoint values and determining results provided by the setpoints at desired reference times. Tne what-if analysis involves mixed-mterger optimization in which on/off decisions are combined with optimization of continuous parameters .

Although the present invention has been described in detail, it should be understood that various changes, substitutions and alterations can be made hereto without departing from the spirit and scope of tne invention as defined by the appendeα claims.

Claims

WHAT IS CLAIMED IS:

1. A method for presenting performance data of a manufacturing process, the process performed at a manufacturing facility, the method comprising the steps of: generating at least one model of the manufacturing process; interfacing the model with the manufacturing facility to provide the model with process performance data; predicting process performance data using the model; and transferring the predicted performance data from the model over a computer network for display on a graphical user interface.

2. The method according to Claim 1 wherein the model comprises a neural network.

3. The method according to Claim 1 wherein the model tranfers real-time predicted performance data to the graphical user interface.

4. The method according to Claim 1 wherein the model transfers future predicted performance data to the graphical user interface.

5. The method according to Claim 1 wherein the model transfers real-time performance metrics data to the graphical user interface

6. The method according to Claim 5 wherein the performance metrics data comprises at least one key performance real-time analyzer.

7. The method according to Claim 6 wherein the key performance real-time analyzer comprises the Solomon Utilization index.

8. The method according to Claim 1 further comprising the steps of: transferring the model and performance data to an off-line engine; accepting variations of the performance data from the graphical user interface; transferring the variations of the performance data to the off-line engine to determine predicted performance data associated with the variations; and transferring the predicted performance data for display on the graphical user interface.

9. The method according to Claim 8 wherein the variation of the performance data comprises a variation of a process setpoint, and wherein the predicted performance data comprises predicted performance metrics associated with the variation of the process setpoint.

10. The method according to Claim 8 wherein the variation of the performance data comprises a variation of a performance metric, and wherein the predicted performance data comprises predicted process setpoints for achieving the variation of the performance metric.

11. The method according to Claim 10 further comprising the step of transferring the predicted process setpoints to the neural network model for implementation in the manufacturing process.

12. The method according to Claim 1 further comprising the step of selectively adding or deleting KPRAs for display on the graphical user interface on a individual user basis.

13. A system for presenting performance data of manufacturing facility unit operations, the system comprising : plural control systems, each control system associated with at least one manufacturing facility unit operation for maintaining measured variable setpoints of the unit operation; plural models, each model interfaced with at least one control system to receive measured variables of the process from the control system; a communications server interfaced with the models for receiving real-time performance data from the models; and a graphical user interface interfaced with the communications server through a network, the graphical user interface for displaying real-time performance data to a user.

1 . The system according to Claim 13 wherein at least one model comprises plural object-oriented box transforms .

15. The system according to Claim 14 further comprising an off-line model engine interfaced with the communications server, the off-line model engine for accepting a model and associated performance data from the communications server, the off-line model engine operational to analyze model responses to variations in performance data.