Configuring the Nonlinear Model block

You are able to configure the Nonlinear Model block both preceding and following training of the model.

Preceding training allows you to change the output field(s) name(s) to something more meaningful. This changed output name automatically populates the Offline Trainer block's property page.

Following training, you can enable Input Range Validation on the Input Min/Max tab, ensuring that the inputs to the model are always within the range on which the model was trained. We also explain how to enable analysis on the trained model.You are able to view the statistics and model audit information of the model, shown on the Statistics and Model Audit Properties tabs.

Models can be imported from XML if the block is untrained. Trained models can be exported.

Return to Overview of the Nonlinear Model block.

Configuring the model block:

Block configuration is specified on the block property page. To open this configuration page, either double click on the block, or right-click on the block and select Block Properties.

Configuration prior to training

Prior to training, there are no configuration options on the Input Min/Max, Statistics or Model Audit Properties tabs prior to training.

Configuration tab

Change the name of the output fields in the Nonlinear Model block by selecting the field in the list, and pressing the F2 key, or double-clicking it.
- Note: The output field name(s) specified in the Nonlinear Model block must be the same as the target data field name(s) in the target data source.
Add additional outputs to the block by selecting Add output.
Remove outputs from the block by selecting the output and then Remove output.
Select to Import model if you want to import a trained model from an XML file in a location that you specify.
- After import, the property page reflects the information imported from the XML file.

Return to top

Configuration following training

After training, the only configuration that can be altered is:

Exporting a trained model.
Adding, removing, and renaming output fields.
Enabling input range validation.

All other information is gleaned from the model, and cannot be altered.

Configuration tab

Once a model is trained, you can do the following on the Configuration tab.

Select to Export model. This produces an XML file, in the location you specify, containing the trained model.
Change the output name of the Nonlinear Model block by selecting the name in the list, and pressing the F2 key.
- Note: The output field name(s) specified in the Nonlinear Model block must be the same as the target data field name(s) in the target data source.
Add additional outputs to the block by selecting Add output.
Remove outputs from the block by selecting the output and then Remove output.

Warning: If you add or remove outputs from the model, the model will no longer be trained, and will have to be retrained.

Input Min/Max tab

Once the model is trained, the input field names will be listed with their minimum and maximum values alongside. These are from the database, and cannot be altered.
Enabled Input Range Validation: by enabling this, the model will check the values of the input fields to validate that they are within the range of the training data set when the model is run.

The model output is marked bad if the model inputs are outside of the range of the training data. The model is, in effect, extrapolating results when using input fields that are outside the range of the training data set and the resulting value of the output field could be grossly incorrect.

Return to top

StatisticsStatistics

On the statistics tab, the following calculation results are displayed for both the training and validation data sets (Training Set and Validation Set respectively) as well as the entire training data set (Overall includes the training and validation data sets):

R2: Correlation coefficient: how closely the value of the model target output matches the process output. A R2 value of 1 indicates a direct correlation of the model to the process, and the model can therefore be used to accurately predict the process.
Explained Variance: a measure of the proportion to which the model accounts for the variance. The model is effective when the explained variance is high relative to the total variance.
RMS: Root Mean Square: An indication of the quantity of error in the model. This value is totally dependant on the normalized input data, there is no specified good/bad ranges.

The RMS value is calculated from the square root of the mean of the square values. Specifically, it is calculated from the difference between the model and target outputs, which are then squared, and all the squared values summed, and divided by number of data points used (to calculate the mean square value of the differences). The square root of this value is then determined.

If the model has more than one output, the statistics for each output fields of the model can be viewed by changing the field name in the [Output box] at the top of the Statistics tab.

Return to top

Model Audit Properties

There are no configuration options on this page, it simply lists information regarding the training data set, including:

number of records used for training (NumTrainingPatterns),
number of records used for validation (NumValidationPatterns),
number of bad quality records (NumBadPatterns),
the start and end dates for the training data set (StartDataSet and EndDataSet), and
the start and end dates over which training took place (StartedTraining and EndedTraining).

Return to top

Enabling analysis on the block

Right-click on the model block and toggle Analysis on. Analysis can only be enabled on models with a single output field.

Right-click on the block again (as it closes automatically after the analysis is toggled on or off) and click on Analysis properties to specify certain features of the analysis.

Features to be specified in the Analysis Properties pop-up box include:

- Display Title: this will be the name on the analysis view tab; it will also be the name of the monitored variable on the analysis view y-axis.
- Smoothing Window Size: a moving average filter is applied to the input tags before they are used in the various analysis calculations. Specify the size of the time window over which the input tags are averaged.
- Setpoint Value: the value of the fixed setpoint where applicable. The Analysis Properties box obtains this value (if it is available) from the Alarming and Analysis block - it may be altered in either of the two blocks and the changes will reflect in both.
- History Buffer Size: this is the number of analysis records Architect will keep in memory at any one time. If you are running 3 days worth of data where the records are 1 minute apart, you will have 4320 records; thus, if you wish to be able to view all of the data once the simulation is complete, you will have to set the history buffer size to 4320.

Return to top