Neural Networks for Property Estimation

One of the key objectives of a process control system is to maintain product quality parameters within specifications. An on-line analyzer may not be available to make a sampled or continuous measurement of a physical property or the composition of a liquid or gas stream, or other conditions may prevent their use. In such a case, the measurements must be obtained by lab analysis of a process grab sample.

When a product quality measurement is available only from the lab, the delay associated with processing the sample and the lack of the measurement’s continuous availability impacts the way this information can be used to control the process. In many cases, these lab measurements are used by the operator to manually make corrections in the process, such as changing feed flows or temperature that impact the parameter(s) reflected in the lab test. To obtain an immediate indication of the results of process changes that impact a quality-related parameter, it is often possible to use upstream measurements such as flow rate, temperature, and feedstock composition to calculate an estimated value of the quality parameter as illustrated below.

As discussed in Chapter 7 of Advanced Control Foundation – Tools, Techniques and Applications, some newer control systems used in the process industry include the capability to estimate product quality parameters. Control system manufacturers have adopted different technologies to implement property estimation. To address the non-linear response of product quality parameters to changes in process inputs, the estimator can be based on a neural network model. My next few blogs will address different aspects of using neural networks for property estimators.