Send for an EXPERT

There's no substitute for experience, as the cliche goes, and it's nowhere more true than in the control room of a chemical plant. But even the most grizzled operator needs a hand sometimes. With their plethora of different variables affecting performance, chemical plants are prime candidates for the use of expert systems.

Expert systems can act as an on-line database to help operators make sense of what their readouts are telling them about what is happening inside their plant, can act as an automatic fault detection, diagnostic and correction system, and can even combine the design and operation of a process. All of these were discussed at a recent meeting of TCCL (the Cogsys Club), a group of process and IT companies with interests in the expert systems field, at BNFL's nuclear fuel manufacturing site at Springfields, near Preston.

The control rooms of all chemical plants are equipped with arrays of alarms. However, pointed out Chris Hotblack of BP Chemicals' Sunbury research centre, `saturating an operator with floods of alarms' can be counterproductive. `It can cause a complete overload of reasoning ability,' he commented. This, he explained, is why BP, along with a consortium including Honeywell, Texaco and Gensym, are developing an expert system for abnormal situation management (ASM) - to `detect problems before they reach the alarm state.'

Unlike a control system, which can detect and cope with normal fluctuations in process parameters, but whose main task is maintaining the plant's throughput and reliability, ASM maintains the reliability of the plant and reduces its impact on safety, health and the environment. It is, however, integrated with the control system, monitoring the condition of the plant, detecting possible accidents and taking over if need be.

The ASM consortium aims to reduce `preventable' process upsets by 90 per cent. US-based and backed by the American government, the consortium has $20million-worth of funding to develop an expert system, which it has dubbed AEGIS (Abnormal Event Guidance and Information System).

AEGIS is not an artificial intelligence device, but an `embedded element' in the plant's automation system. It uses data from incident investigation reports, the plant's own operation profiles, and process hazard analysis as guidelines to analyse how the plant is working.It can then decide whether a slide away from normal operation is being handled by the control system, or whether it is the start of an accident, Hotblack explained.

Another fault-detecting system is operating at Springfields, again developed by a consortium. The ALSACA (automation for large-scale assembly) project, part of the European community's EUREKA scheme, was originated by Italian washing-machine manufacturers, but its techniques are applicable across a wide range of industries. BNFL uses ALASCA in a facility which assembles nuclear fuel bundles.

ALASCA uses a `virtual factory' simulation at the design stage to optimise the manufacturing process, setting factors like the duration of a production cycle and the maintenance schedule. Once the process is running, the virtual factory model is linked to the operating and control systems, where it generates shop floor data for future investment decisions, while tracking the running of the plant.

Part of BNFL's contribution to the project is an error recovery system, which uses case-based reasoning (CBR) to detect and diagnose errors within the plant.

Chris Bugby of Salford University Business Services, a member of the development team, explains that CBR compares data coming out of the PLCs which control the plant with the contents of a database of known faults and their remedies. `We had to design the system before we knew what the plant was going to look like or what sort of things might go wrong, so we needed something that required no predefined solutions,' he said. `CBR systems mature as the experience of operating the plant grows.'

The system collects data from PLCs, and compares this information with model data. If there is any discrepancy, it tries to compare the data with its set of abnormal situations. Matches and close matches are passed on to the operator, ranked in order of probability. This tells what's gone wrong, why it's happened, and how to put it right. The operator can then choose a response from this list. If nothing matches closely, the operator can enter the data, the cause of the problem, and the way it was put right into the database for future comparisons. Not so much `If at first you don't succeed, try, try again'; more `If at first you don't succeed, you will next time.