Distributed embedded intelligence blazing processing trails

Progress is being made in diagnostics, production changes, abnormal situations, optimization, dynamic scheduling, and advanced control.

Intelligent systems can now be implemented in a distributed way at the sensor level, as well as at the vaguely defined manufacturing execution system (MES) level. As industry seeks to achieve the goal of "closing the loop" with enterprise business systems, it also will be necessary to embed intelligence throughout the distributed systems.

Embedded intelligence at the sensor level for calibration, failure detection, and data filtration is important for all of the higher levels of operation. Sensors can drift or malfunction in various ways. They can be noisy and may need correcting for interacting variables.

Correcting sensor data from a centralized distributed control system (DCS) or supervisory control and data acquisition (SCADA) system requires a lot of data communication, which results in performance limitations. Moving the corrections to the sensors themselves provides higher system performance. This is possible in a flexible and adaptive way if the sensors allow open communication and use intelligent agents that communicate to the higher-level systems.

Pilot trial planned

At the MES level, applications are becoming available using embedded intelligence to achieve higher productivity and better quality and to assist operators in complex or abnormal situations.

Distributed embedded intelligence at the MES level, and even at the enterprise management level, can offer new and exciting possibilities. For example, at the National Center for Manufacturing Sciences (NCMS), a pilot trial will start soon, which uses distributed intelligent agents for dynamic scheduling in an automotive production facility.

The general concept is that scheduling is often too complex and real-time for a centralized system to operate well. However, distributed intelligent agents could make intelligent local decisions.

An application is also being considered in which distributed agents could be in several different factories that supply parts for an automotive operation. However, the concept could be applied to many different types of manufacturing.

There are a number of other MES functions that can benefit from embedded intelligence, such as diagnostics, process optimization, production change control, and dynamic scheduling (see Figure 1). Several vendors already offer embedded intelligence in these areas, while others are developing embedded capabilities for abnormal situation management as well as advanced control dynamics that handle multiple variables in intelligent ways.

Improves process diagnostics One of the most obvious targets for embedded intelligence is process diagnostics. We are not talking about single-point alarms; these are handled by DCS systems or SCADA systems. The intelligent process diagnostics applications typically consider a number of measurement points around a process unit and the way that these measurement points are dynamically changing.

The observed behavior is compared with "knowledge" of expected behavior. That knowledge may be in various forms: it could be a rigorous mathematical model, a learned model provided by neural networks, or in the form of fuzzy logic rules defined by process experts. When the observed behavior departs from the expected behavior, the intelligent diagnostic system suggests possible causes and corrective actions.

Clearly, intelligence is required when operations are changed, whether this is during startup, shutdown, or major production changes. Right now most conventional control systems are designed to regulate only during steady operations. They are not used for control during production changes.

In fact, control systems are typically put on "manual" during such changes, with operators using direct manual control. However, plants are increasingly complex, and operator training and attention vary. Therefore, an intelligent monitor and adviser could add a margin of safety to these operations.

Several vendors, including YamatakeHoneywell, are offering intelligent MES products for process unit startup and shutdown that provide greater safety during these operations.

According to a paper presented by David W Beach at ISPE '95, BP America already has installed intelligent systems to assist in process startup and shutdown. Such systems may even become mandatory.

Managing abnormal situations

The Abnormal Situation Management (ASM) Consortium represents one of the most advanced initiatives involving both intelligent diagnostics and production change management. ASM is led by Honeywell and funded by NIST and a group of process manufacturers including Exxon, Mobil, Chevron, and others.

The ASM initiative seeks to improve the safety of process operations by guiding operators during abnormal situations, and a number of prototype applications have been developed by individual consortium vendors.

Estimates indicate the inability of automated control systems and personnel to control abnormal situations costs the U.S. economy at least $20 billion per year. This more than justifies the investment by the ASM, which seeks innovations in human-machine interaction, system architecture, and system configuration tools. Toward this end, more than 30 technology-development studies are reportedly being conducted.

Optimization products emerging

Optimization has been a topic of interest for years. However, the growing power of computers has made optimization increasingly more practical, and a number of applications and products are emerging. The greatest potential is for optimization that combines rigorous mathematical models, when available, with learned neural-network or statistical models and with rules learned from human experts.

Shell Oil Co. and Simulation Sciences (SimSci) recently implemented a rigorous online modeling (ROM) system for a large gas plant processing gas from the North Sea field and downstream NGL plants. It provides effective performance monitoring of the process equipment and a tool for open-loop optimization of plant operations that takes into account equipment capacities, product qualities, fuel/electricity costs, and downstream plant constraints, while providing set points for operating the plant.

Shell Oil and SimSci have recently agreed to combine their technology into a new generation of optimization system called ROMEO (rigorous online modeling and equation-based optimization), which is designed to solve larger-scale optimization problems.

Objects ease production tracking The oldest recognized MES application is production tracking. In fact, several vendors use the MES label simply based on their capabilities in this area. That's a distortion of the label because production tracking is just one of the areas of MES operations.

In its simplest form, production tracking is just transactional, as product moves from one point to another. However, the new object technology opens new capabilities for advanced production tracking. Foxboro, for example, has introduced the Oil Movement Information System, which applies intelligence to the complex task of managing petroleum logistics operations.

Objects add capability to production-tracking applications. As products enter manufacturing, objects can be created that represent the product. These objects can be moved through a "virtual reality" graphical representation of the actual plant.

At each stage of manufacturing, the object can record the actual process conditions and quality tests affecting any particular product. This approach captures the actual manufacturing process as attributes of the object. Information in the objects can be coupled with neural networks to automatically produce better manufacturing models.

Such objects also provide better records for traceability and can show exactly how the product was manufactured. All this leads to a new capability: visualizing the production operation as objects in a virtual reality display. In advanced applications, "zoom" controls will allow the operation to be viewed at high levels, even as a supply chain, or in great detail. An example, shown in Figures 2 and 3, is the G2/ReThink-based PROSPER tool of the EDS Energy Division, with the supply chain of a petroleum company shown as a process from crude oil supply to final customer delivery.

Tools like this allow the enterprise manager to visualize the profitability and other key operating parameters at the top level as they are affected by decision alternatives. Each top-level object is an abstraction, driven by structures of objects at the more detailed levels, and so on. The supply chain is driven by customer demand and by crude supply.

Speeds dynamic scheduling

One of the most important areas for MES is dynamic scheduling, the minute-by-minute decisions of how to actually optimize production given the current resources, constraints, and priorities. Generally, the master schedule is determined by the business systems, but the real-time scheduling decisions are made by operating personnel.

The ability to build object models of production allows intelligent dynamic scheduling to be more naturally implemented. Mathematical optimization or simulation-testing approaches can use the object model to evaluate decisions and choose the best one.

For example, Aspen Technology, Inc. has the Pro-Sked intelligent product for scheduling, which has been implemented at several refineries. Also, Glaxo Wellcome has implemented intelligent dynamic scheduling for its pharmaceutical operations at a new production facility in Canada.

Expands advanced control choices

Advanced control is an MES application area that has been around for some years. However, the ability to embed intelligence into advanced control opens new possibilities.

For example, rules about process performance can be used to determine when one type of advanced control should be used instead of another. Neural networks could be set up to learn models that then can be used for control. These models can be multivariable, nonlinear, dynamic, and adaptive. This extends their potential beyond today's conventional notions of advanced control.

Amoco, for example, implemented neuralbased advanced control for high-quality diesel production. This system, which provides a realtime model of the product quality without the delay of lab tests, also provides tighter control and much faster on-spec performance after grade changes. Faster on-spec control has increased profits by an estimated $500,000 per year.

Some vendors are already using embedded intelligence in advanced control products. EIMCO, for instance, provides advanced control systems for mineral recovery in mining operations. The neural networks provide models for the advanced control.

In mineral recovery, the relationships are nonlinear and difficult for an operator to handle with a "mental model." Also, the raw material is variable, and so the model changes. But the neural network is able to handle the issues of adaptive models and nonlinearity, thereby improving mineral recovery. Improving recovery by even a few percentage points has a large economic benefit.

Embedded intelligence distributed

Putting embedded intelligence into sensors obviously requires a distributed approach because the individual programs or agents must be downloaded to the sensor level. This sort of approach also is needed at the MES level, which encompasses many functions.

Furthermore, as integration occurs over the supply chain, there is also considerable geographic distribution. The overall implementation of embedded intelligent systems, therefore, requires open communication between these intelligent MES applications and integration with intelligent remote applications and agents across a network (see Figure 4).

Anther requirement of intelligent MES applications is that they operate concurrently, even within a single application. For example, intelligent diagnostics requires diagnosing multiple simultaneous problems and prioritizing actions.

In turn, this requires the applications to be built on real-time intelligent system tools that provide concurrent multithreaded reasoning, actions based on priorities, and a basic understanding of dynamic behavior. Modern computers are fast enough for the applications, but the software base must allow for concurrent real-time reasoning.

Significant impact seen

What will be the impact of embedded intelligence? At the sensor level, it will produce better measurements that are available for all the higherlevel operations. At the MES level, embedded intelligence will provide safer operations, better productivity, more consistent quality, and more flexibility.

It is widely accepted that businesses should be viewed and managed as processes, but there has not yet been much application of dynamic modeling and intelligent process control at the enterprise level where embedded intelligence will allow entire supply chains and enterprises to take advantage of modern process operation technology.

Behind the byline

With a strong background in intelligent systems and process control, Dr. Robert L Moore has served as Gensym Corp.'s president since he helped found the company in 1986. In the early 1980s, he conceived the first real-time expert system for process control. His professional experience includes the Foxboro Company, Sentrol Systems Ltd., AI Systems, Inc. (a company he founded in 1983), and LMI. He holds a Ph.D. in electrical engineering from MIT.

Copyright Instrument Society of America Mar 1997
Provided by ProQuest Information and Learning Company. All rights Reserved