Industrial use of safety-related artificial neural networks

[Lisboa Paulo]

A contract research report on industrial use of safety-related artificial
neural networks has been published on the web by the contractors, the UK's
Health and Safety Executive.  A link address and abstract are appended to
this email.

This is in the nature of a consultation paper, so feedback regarding any
aspect of the paper is very welcome.

Paulo Lisboa.

 http://www.hse.gov.uk/research/crr_pdf/2001/crr01327.pdf

Abstract

The overall objective of this study is to investigate to what extent neural
networks are used, and are likely to be used in the near future, in
safety-related applications.

Neural network products are actively being marketed and some are routinely
used in safety-related areas, including cancer screening and fire detection
in office blocks.  Some are medical devices already certified by the FDA.
The commercial potential for this technology is evident from the extent of
industry-led research, and safety benefits will arise.  In the process
industries, for instance, there is real potential for closer plant
surveillance and consequently productive maintenance, including plant life
extension.

It is clear from the applications reviewed that the key to successful
transfer of neural networks to the marketplace is successful integration
with routine practice, rather than optimisation for the idealised
environments where much of the current development effort takes place.  This
requires the ability to evaluate their empirically derived response using
structured domain knowledge, as well as performance testing.  In controller
design, the scalability of solutions to production models, and the need to
maintain safe and efficient operation under plant wear, have led to the
integration of linear design methods with neural network architectures.

Further research is necessary in two directions, first to systematise
current best practice in the design of a wide range of quite different
neural computing software models and hardware systems, then to formulate a
unified perspective of high-complexity computation in safety-related
applications.

There is a need to develop guidelines for good practice, to educate
non-specialist users and inform what is already a wide base of
practitioners. Combined with a safety awareness initiative, this would be of
as much of benefit to the development of this commercially important new
technology, as to its safe use in safety-related applications.