Nowadays, scientific and industrial interest on the development of nonlinear control systems increases day after day. However, before these models become reliable, they must pass through a hard and expensive implementation process. In this way, studies involving decision support methods try to develop low cost intelligent applications to build up advanced industrial control systems with excellent results, as in the petrochemical industry. In the distillation of oil derivatives, for example, it is very common the use of laboratorial sample analysis to identify if a substance has its physical- chemistry characteristics in accordance to international production rules. Besides, the analyses results allow the adjustment of production plant instruments, so that the process reaches a thorough control, and, consequently, a final product with higher quality. However, although laboratory analyses are more accurate to evaluate final product quality, sometimes it demands many hours of analysis, delaying the adjustments in the production equipment. In this manner, the process efficiency is reduced and some products have its production period increased because they should have its composition corrected with other reagents. Another disadvantage is the equipments´ maintenance costs and calibration, since these instruments are installed in hostile environments that may cause unaccurate field measurements, affecting also operator´s action. On the other hand, among the most applied intelligent systems in chemical industry process are the artificial neural networks. Their structure is based on biological neurons and in the parallel processing of the human brain. Thus, they are capable of storing and employing experimental knowledge presented to it earlier. Despite good results presented by neural network structures, there is a disadvantage related to the need for retraining whenever the process changes its operational point, for example, when the raw material suffers any change on its physical-chemistry characteristics. The proposed solution for this problem is a hybrid method that joins the advantages of a neural network structure with the ability of a stochastic filter, known as extended Kalman filter. This filter acts in the synaptic weights, updating them online and allowing the system to constantly adapt itself to process changes. It also uses specific pre-processing methods to eliminate scale mistakes, noises in instruments readings and incompatibilities between system input and output, which are measured with different acquisition frequencies; the first one in minutes and the second one in hours. Besides, variable selection techniques were used to enhance neural network performance in terms of inference error and processing time. The method´s performance was evaluated in each process step through different test groups used to verify what each step contributes to the final result. The most important test, executed to analyse the system answer in relation to a simple neural network, was the one which simulated process changes. For that end, the network was submitted to a test group with output samples added to a ramp signal. Experiments demonstrated that a system using simple neural networks presented results with MAPE error of about 1,66%. On the other hand, when using neural networks associated to an extended Kalman filter, the error decreases to 0,8%. In this way, it´s confirmed that Kalman filter does not destroy the original neural network quality and also adapts it to process changes, allowing the output inference without the necessity of network retraining.