Modeling

In Continuous Troubleshooter you are able to generate powerful non-linear and rule-based process models. For a non-linear process model, the user must select the appropriate input and target fields. The rules model automatically uses this specified selection of inputs and targets for the rule-based model construction. The only difference between the outputs of the two models is that the output of the rules model is a discrete value that could be Low, Normal or High, as determined by the inner limit settings of the output variable.

Constructing a Model

Using history brushing to construct a model

If regional brushing has been applied, the intended process model can be built inside regional brushed data, outside regional brushed data, or on all the data. This means that the data set can be visually separated with a scatter plot, or trend from which a model can be built on either the marked (brushed) area or the unmarked (unbrushed) brushed

To construct a process model:

Access the Modeling view from the Troubleshooter project bar.
Click the Construct Model button on the modeling view. (If a model exists, click the Reconstruct Model button to reconstruct your model)
From the Input Field selection, select the desired input fields and model target for the intended process model. Fields marked as correlated are automatically not selected as inputs to the model. Fields with zero variance are also excluded from the model generation, and a possible error message will be generated.
- See more information on Field Correlation (Correlation Matrix) >>
Select the data for training:
- Construct model using all data for training
- Construct model using only history brushed data for training
- Construct model using data outside history brush for training
Set the smoothing window size.

The smoothing window represents the data window for which causes and IO relationships are calculated using a single ruler on the knowledge extraction view. Set the window period to a suitable value.

NOTE: The smoothing window must be entered using the correct format hh:mm:ss.
Click OK to begin model construction.

See more information on analysis parameters >>

Possible error message: Field x has zero variance. This will prevent the model from being generated, as fields containing zero variance cannot be used for modelling. This error message could be generated from one of two situations, and to correct the situation, the input fields database as a whole needs to be investigated, not only the fields listed as having zero variation.

Situation 1: If field y has a value missing in a row or multiple rows, the whole row will be marked as bad quality and not included in the model. If these deleted rows cause the variation required in another field such as field x to be excluded, then the error will state that field x has no variance and cannot be included. Although analyzing the trend of field x shows contrary to this, when the fields are considered together and some rows are excluded, field x might then show zero variance. It is necessary to investigate the data as a whole in order to correct this situation.

Situation 2: If a section of field y's trend is brushed, and the brushed data is selected to be used to generate the models, the same section of all other fields is also used in generating the model. It might occur that for that particular brushed section in field x, there is no variation. The error message will then be shown. It is necessary to investigate the brushed data trends in order to correct this situation.

NOTE: Model Construction can be stopped at any time during the construction phase by clicking the Stop button.

Model Statistics

Once process models have been constructed, the Modeling view displays the following statistics about your models:

Number of construction cases:

Represents the data on which the model was trained.
Number of validation cases:

Represents the data used for validation.
Number of bad quality cases:

Shows the number of data records with bad quality.
Model fit on construction cases:

Shows the accuracy of the model using the construction cases.
Model fit on validation cases:

Shows the accuracy of the model using the construction cases.

The accuracy is shown as a percentage and is interpreted as follows:

Example:
- 94% of the cases were predicted correctly by the model using the construction cases.
- 93% of the cases were predicted correctly by the model using the validation cases.

Rules Statistics

The rules are divided into three classes namely Low, Normal and High. The "% Correct" columns shows how accurately the classification is, while the "Number of cases" column shows the total cases falling into each class. It is interpreted as follows:

Example:

- 2132 cases are classified as Low Anode Temperature where 81% of the cases was correctly classified by the rules model.

For an Existing Trained Model

If you already have an existing trained model, the following options are available:

Reconstruct model: This reconstructs an existing model with different criteria.
Remove model: This removes an existing model completely.
Export model: This exports an existing model as an XML file to a location which you specify. The model can now be opened in the Architect for further editing.