In general, the measurement of key process variables at a rate suitable for real-time control is a problem to many industrial processes. Due to the high installation and maintenance costs of equipment, measurements are often unavailable. When appropriate instrumentation is available, on-line measurement is often restricted by long analysis cycle times. This often leads to the monitoring of key process variables through the use of off-line laboratory analysis and result in the process deviating from desired operating conditions with long disturbance recovery times, leading to unsatisfactory variability and a reduction in profit. These adverse effects can not be acceptably overcome by the use of existing advanced control algorithms. Efforts towards alleviating this problem have included the development of inferential estimators. There are many variables measured on-line and are sampled relatively fast. These variables are indirectly related to the key 'difficult to measure' variables. Consequently, when an accurate inferential model is used on-line as a soft-sensor improved control performance can result simply because of higher sampling rate.

There are two general methodologies that may be adopted when building an inferential model. First, a 'bottom up', first-principle approach can be taken. This is generally the preferred approach for engineering modeling tasks. However, it requires an understanding of the basic physical and chemical properties of the process. In the chemical industries an incomplete understanding of the process often precludes this first-principle approach. A common alternative is to approach the problem from 'top down' and generate an input-output model of the process from plant data, also known as black-box or gray-box modeling approach.

To develop an optimal inferential model and to compare the performance of different "black box" algorithms in non-linear system identification, we apply several techniques to industrial data collected by Syncrude Canada. We exploit all available process information using the first principles of the chemical system to search for an approximate model structure. Using this model structure, the inferential modeling approaches such as partial least square technique, group method of data handling (GMDH) and genetic algorithms (GAs) are applied to construct nonlinear inferential models. These results are also compared with three-layer feed-forward Neural Network. To evaluate the effect of experiment design on nonlinear modeling, these approaches are also examined on the data obtained from a pilot scale CSTR reactor for different designed input signals. The results reveal that self organizing of the model structure in GMDH and GAs techniques have been very successful in building nonlinear inferential models.