Multivariable alarming using neural networks

The monitoring and alarming of processes is normally done in one dimension. Each measurable parameter is treated as independent of all others. The only time this is not true is when the dynamics of a multivariable process are "modeled" using equations that predict the chemical, physical, and/or thermodynamic properties of the process in terms of these parameters. Unfortunately, this requires that the relationship be known and accurately describable in mathematical terms. Using neural networks to map the "N"-space relationship among a set of "N"-variables, it is possible to create experience-based "models" of the multivariable relationships. This technique does not require that the relationship even be known, only that a "live" process be available to "teach" the neural network the various safe and unsafe states of the process.