Alarm Monitoring

There are many examples of large and complex systems that are in a stable or benign state for most of the time but can occasionally stray into a dangerous or alarm state. Examples include

  • manufacturing plant
  • nuclear power stations
  • ecosystems, such as rivers
  • the human body

It is usually not possible to determine beforehand all the possible ways in which the system might exhibit an alarm state. However plentiful examples of the system in its stable or benign state are usually available.

When we entrust a human to monitor these systems we have the benefit of their skill and experience but, for systems that are infrequently in an alarm state, fatigue and boredom can set in and the response time when a real alarm occurs can be longer than ideal.

What can help here is a monitoring device that can learn the stable or benign state of the system and sound an alarm if the system strays out of this state.

Neural Solutions has developed a neural network training algorithm for this purpose called default training which is the subject of patent GB2258311.

The default training algorithm requires no examples of the alarm class. Instead it infers these by default from the examples of the stable or benign state.

Neural Solutions has applied default trained neural networks and other alarm detectors to a number of application areas with considerable success. The alarm detector is initially given a short period of training after which its performance is adequate. If a false alarm is encountered the alarm detector is given a small amount of additional training to incorporate the data leading to the false alarm into its understanding of the stable or benign state.

Further information is available on the application of alarm detection to the monitoring of patients in intensive care.