Autogenerate a neural network model

After you have selected a range of data to use for training the neural network, you can have the Neural application generate the network. To do so:

  1. Click the Autogenerate button on the toolbar or select FileAutogenerate to start.

    You can also create a model from data stored in a valid NN data file format.

    To use a data file, from the File menu select Autogenerate from file, then browse to the desired data file.

    The application performs a sensitivity analysis on the data selected in the trend view, trains the network, and creates the network model.

    Training stops when one of the following four criteria is met:

    • The error limit has been reached
    • There is no significant improvement in error reduction
    • The maximum number of epochs has been reached
    • The maximum number of hidden nodes has been reached
    • It has completed convergence to the minimum possible error for the data set
    • The Stop button on the Training Progress dialog is clicked

    During training the Status field is updated with messages, including any error messages. Upon completion of training the application updates the Status field of the dialog.

    If you are an Expert user, after a few moments a message stating the condition which caused training to complete.

    If you are an Expert user, click OK and the application updates the DeltaV database.

    The name of the created model appears in the status bar of the application window next to Assigned Model field.

    The Hierarchy View is updated with the new Training Model node, which appears as the selected node of the view. The Assigned Model icon also appears next to the model name.

    A message appears saying that the configuration has changed and that you should download the module and reconnect the application to the block.

    The blue triangle also appears next to the NN module in DeltaV Explorer.

  2. Click OK to continue.

    The Configuration View changes to the for the trained network.

    If design and training complete successfully, the application generates a neural network model and saves it in the DvData\Neural folder of the ProfessionalPLUS workstation. The signature (path name, unique database id) of the specific block is saved in the file as well. This model file keeps all the information needed for creating an intelligent sensor, including input parameters, weights, mean and variance, input, hidden, and output limits, and training and testing errors.

    The DVData\Neural\Data folder has the training data used to create the NN. The format of this file is the same as that used for creating an NN from file.

    Note

    Do not modify the names, contents, and formats of these files. These files should be left as is except when you use a saved data file to create another model.

    Note that there is another folder at this level, DVData\Weights, that contains the weights used in the networks. Do not modify or delete the files in this folder.

  3. Download the model to the controller

The name of the selected model file appears in the status bar of the application window next to the Control Model field. The Downloaded Model icon appears next to the model name in the Hierarchy View.