This project is a successful demonstration
of how case-based learning (CBL), a type of artificial intelligence
(AI), can solve process control problems in the pulp and paper industry.
To develop the system, a British SME and a Belgian university, both
experts in AI, worked with a large German manufacturer of process
control systems. CBL takes a middle ground between AI techniques that
need extensive training on historical data and others that demand
an accurate mathematical model of the process.
Two other partners, a Portuguese pulp mill and a Dutch paper mill,
have both made significant cost savings while maintaining product
quality. The system will soon be available commercially and could
be used by other industries such as food processing.
Industrial processes that use natural raw
materials can be difficult to control. A mill producing wood pulp
for the paper industry must produce a consistent product, yet wood
varies in character throughout the year and often comes from different
sources as market prices fluctuate. Pulp quality varies as a result.
Paper mills using recycled paper rather than virgin pulp have the
To ensure that the pulp always meets customer specifications, the
mill must set its standards a little higher. If there is high process
variability this is expensive, for it means that for much of the
time the mill is making pulp that is better than its customers need
or will pay for. Paper manufacturers and food processors have the
Process control - the technology of maintaining product quality
in process industries such as paper and chemicals - has recently
started to use computer-based artificial intelligence (AI) to tackle
difficult problems. AI has a long academic pedigree, but its successful
application to real-world situations is problematic.
CLEAN is a highly successful attempt to bring AI to the pulp and
paper industries using a new approach known as case-based learning
(CBL). The project has linked two AI specialists, British Maritime
Technology in the UK and the Catholic University of Leuven in Belgium;
two user companies, and a control systems supplier, Siemens of Germany.
The users are a small Portuguese pulp mill, Companhia de Celulose
do Caima, and Roermond Papier, a Dutch paper mill that makes only
recycled paper. Both have been delighted with the results and have
experienced significant cash savings.
Problems with existing methods
Pulp and paper are competitive industries, and to gain a commercial
advantage mills are investing heavily in advanced control techniques
such as AI. Existing AI techniques, however, have disadvantages.
One example is neural nets, which can learn to control a process
by studying past production records. Neural nets need little prior
knowledge about the process but they do need plenty of 'training'
data. This is not always available: the Caima mill, for example,
includes processes for which full data may only be available once
a week, so collecting enough information to train a neural net could
take up to a year.
Another technique, 'fuzzy logic', is quicker to set up because its
reasoning is pre-programmed using information collected from experienced
human operators. Unfortunately this means that fuzzy logic systems
must be carefully tailored for each site, and they may need to be
re-tuned if the process or the raw materials change.
By the early 1990s Roermond engineers were looking for a control
system that would help the mill cope better with changes in the
quality of its recycled paper feedstock. They had already tried
neural nets without success, so they asked control systems manufacturer
Siemens to develop a better system.
CBL: the best of both worlds
Siemens, which has an estimated 80% market share for control systems
in the pulp and paper industries, approached British Maritime Technology
for help with more advanced AI. BMT specialises in marine engineering
but has a history of successful AI research and development in other
industries. Together with the Catholic University of Leuven and
the Caima pulp mill, the CLEAN project started in 1993.
BMT and Siemens decided to use case-based learning, a technique
that combines some of the flexibility of neural nets with the predictability
of fuzzy logic. CBL uses past experience rather than a rigid mathematical
model of the process, making it well suited to processes based on
natural materials, yet it can also incorporate existing engineering
knowledge where this is available.
The challenge for AI is to use the current state of the process
to predict how the product quality will change as variations in
raw material quality and process conditions travel through the plant
from input to output. CBL does this by comparing the current set
of process data with historical data and looking for the best match.
If the system finds a large number of past cases like the current
one, and all these cases led to product of acceptable quality and
cost, then it knows the process is on target. If all the past cases
led to poor-quality or expensive product, the system steers the
process towards conditions which it knows from past experience will
produce better results. It does this by using the plant's existing
low-level control system to change process temperatures, flowrates
and motor speeds.
If the CBL system finds that the matching cases gave variable product
quality, or that the current case has no precedent, it has to make
an informed guess about how to get the process back on track. Experienced
process operators are often wary of handing over control to a 'black
box', but a CBL system has two features that help to win their confidence.
First, it can explain how it arrived at its predictions, and if
it lacks enough information to come to a decision it can say so
- something other AI systems find hard to do. Second, it can be
set up so that initially it never strays too far from known conditions.
As operators gain confidence in its abilities, they can give the
system more responsibility.
Installations and results
By the time the project ended in January 1996 both the Roermond
and Caima mills had working CLEAN installations. The results were
impressive. At Roermond, the factory's energy consumption per tonne
of paper fell by nearly 9%, starch usage fell by more than 3% and
water consumption was down by 24%. Product degradation fell by around
20%. Production staff also find that the new system makes it easier
to switch between different grades of paper, so shorter production
runs have become profitable.
At Caima, the basic yield of the pulp process has increased by 3%,
with a simultaneous saving in steam of 1.4% and a reduction in the
amount of reject material produced of more than 50%. Together these
amount to a cost saving of more than 500,000 ECUs a year. Product
quality has also improved. Caima is now installing a new CLEAN system
in another process area.
In both plants the system has a modular architecture that simplifies
software testing and maintenance. The modules run on a personal
computer under Microsoft Windows, sharing data through a series
of simple text files. The heart of the system is the Site Control
Unit (SCU), a program that assembles, classifies and stores the
operating 'cases' using plant data supplied from other parts of
the system. Roermond has a single SCU; Caima has three, reflecting
the more complex nature of its pulp process, which includes batch
The other parts of the system are a module for obtaining raw data
from the plant's low-level control system, a display unit for the
process operators, and a modelling system. The latter, which was
designed by the Catholic University of Leuven, helps the CBL system
by predicting information, such as the results of laboratory analyses,
that is not yet available.
A toolkit for the future
At present, the highly-customised nature of the Roermond and Caima
installations means that CLEAN is not available as an off-the-shelf
commercial product. This will soon change, however, because Siemens
is working on a CBL 'toolkit' that operating companies can use to
add CBL to their processes quickly and simply. Siemens plans to
launch the toolkit in 1997 and BMT will share the revenues.
As well as pulp and paper, the companies will be targeting the fibreboard
and food industries, for which CBL is also a good choice. Siemens
and BMT have taken out two joint patents to protect their know-how.
The next generation of AI is already on its way BMT is working on
an even newer technique, known as conceptual clustering, that may
produce even better results.