Users, previously apprehensive about applying an artificial neural network (ANN), will likely find Emerson Process Management's (Austin, Tex.) soon-to-be-released DeltaV controller-resident neural network tools make developing and deploying ANN solutions easier.

Artificial neural networks use dynamic inputs and mathematical models to predict outcomes but implementation has often required getting process and control engineers in the same spot at the same time to achieve a successful ANN project.

Because neural network model accuracy is so dependent on the knowledge and abilities of the person(s) conducting the front-end data analysis, most process- and/or control-engineers haven't invested the time to implement an ANN.

What the DeltaV NN (neural network) solution provides is a tool set that delivers the analysis methodology and knowledge of ANN experts on the user's desktop.

Instead of users digging through existing historical data, DeltaV's NN data collection is automatic and begins when a user defines a neural network block, assigns up to 30 related input variables from anywhere in the system, assigns a manual laboratory or online analyzer data collector, and downloads everything into the controller.

When sufficient data are collected—usually data collected over about six times the time it takes the process to reach steady state—the next step is to segregate good data from bad.

For example, there may have been a couple of process upsets and a product grade change during the data collection period. Using DeltaV's drag and drop data selection tool, bad data samples are marked for removal leaving only data that reflect normal process operating conditions.

Artificial neural networks perform best when encountering data that fall within the range of data used to train the model, making it necessary to establish data outlier limits.

DeltaV's NN development software automatically screens training data against user-defined limits or, in the absence of defined limits, those data outside the range of three standard deviation around the mean value of the variable are considered outliers.

With outliers established, the next step is to determine each input's time delay and influence on the output. Traditionally this has been the most time consuming part of developing an ANN and required considerable discipline to avoid prematurely jumping to conclusions.

DeltaV's NN software uses a two-stage technique. Stage-one reveals individual input strength-dependency by calculating the correlation coefficient for each input-to-output using a fixed size window, resulting in parsed data with determined delays.

Stage-two consists of submitting the data to multiple analysis iterations to determine the sensitivity of combined inputs, with time delays applied, to influence the output.

With inputs and delays identified, the next step is to select a suitable number of hidden neurons and train the ANN model to predict outputs and not memorize data patterns.

DeltaV NN software splits good data into training and testing (cross-validation) data sets and uses them in an iterative process that automatically increases the number of hidden neurons and selects the ANN configuration that results in the best combination of train and test errors.

The final step is to use the testing data set to verify the ANN's ability to accurately predict an output.

Essentially, DeltaV NN software tools provide an ANN developer's skills in a nearly one-button solution, but for those users wanting more hands-on involvement, an expert mode is available.

Numerous applications that could benefit from online ANN solutions haven't received those benefits because of complexities of implementing ANN solutions. Emerson Process' DeltaV NN goes a long way to minimize complexities and make implementation of artificial neural networks practical.

For more information, Circle 498 or visit www.easydeltav.com