Manage SmartSignal Models

SmartSignal models are used to predict equipment behavior under various operating conditions. A model consists of a group of related tags and historical data that have been filtered to reflect healthy equipment behavior, residual threshold settings, model settings, and testing information. Runtime data for all tags in a model are compared to the state matrix for the model to generate estimates. Through various model settings and rules, users are alerted when abnormal conditions are encountered. Models only operate when operating mode criteria have been met.

Access Model Data

You can access the data in a model associated with a SmartSignal deployment.

Procedure

In SmartSignal Maintenance, select .
In the Deployments pane, select the name of the model whose details you want to access.
To close the Deployments pane, select or select an area of the page outside the menu.
The page for the selected model appears. By default, the table on this page displays only tags that are active in the model.

Optional: If you want to display alternate tags, in the Filter menu, then select one of the following options:

Filter Option	Description
Active in Deployment	If you select this filter, all tags in the deployment are displayed in the table, even if some of them are not active in the selected model.
All Active and Inactive	If you select this filter, all tags associated with the selected model are displayed, regardless of whether they are active or inactive.

Access Settings for a Model

In the SmartSignal Maintenance module, you can access the settings associated with a model.

Procedure

Access SmartSignal Maintenance, and then select a model.
In the Actions box, select Settings > View Model Settings.
The View Model Settings window appears, displaying the various settings related to the model.

Model Settings

The following settings are available for a model.


Field	Descriptions
Active	Specifies whether or not the application should make predictions and perform data modeling using the model’s tags and configuration. Note: The deployment containing a model must be active for an active model to be processed.
Estimation Parameters
Estimation Generator	The engine used to generate estimated values. SBM Similarity Based Model. VBM Variable Similarity Based Model.
Similarity Operator	The algorithm to be used for modeling. SSCOP2 A similarity operator used during estimate generation. It is most useful for auto associative modeling. It tends to be a more fault tolerant than SSCOP3. SSCOP3 An alternative similarity operator. It provides a smoother estimate of the response variables, and tends to perform better in inferential modeling cases. However, it is less fault tolerant than SSCOP2 which is why we often assume that the predictor (input) variables can be considered good. If it turns out not to be the case, spillover can occur in one or more of the response variables (inferred variables).
Enable VSG	Specifies whether virtual signal (estimate) generation. If NaNs and/or outliers are encountered and the number of tags containing bad data is below the Maximum % of Bad Tags threshold, the system will fill in the missing or bad data for modeled tags with virtual signal data.
Maximum % of Bad Tags	The percentage threshold above which, the system inhibits virtual signal generation. This option is available only when Enable VSG is selected.
Variance of Style Factor	A multiplier for the standard deviation of normal residual signals.
Residual Smoothing
Smoothing Algorithm	Moving Average The simplest of the three smoothing algorithms. It is a moving window technique where the system averages the data within each window to produce the filtered sample for each window. Although simple and computationally efficient, there is an inherent delay in the response of the filter that is proportional to the window size. Also, the presence of a spike in the input data will affect the output signal for the length of the window size. Spline A real-time moving window technique that employs a cubic-spline fit to each of the individual sliding windows. The spline filter can produce a filtered signal with very little delay, but this is more computationally expensive. Also, a spike in the input data results in a shift in the output, but this effect will be short-lived compared to the moving average filter. Olympic This filter is very similar to the Moving Average filter, but the system does not use the maximum and minimum data samples for each window in the calculation of the average value. The resulting filter is very computationally efficient and effective at removing spikes. However, it suffers from the same response delay as the Moving Average filter.
Window Size	The total window (the number of data points or persistence) to use for smoothing.
Spline Smoothing Factor	A number between 0 and 1, where the value 0 causes the greatest smoothing and a value of 1 results in no smoothing (the output values equal the input values). A typical default value is 0.1
State Matrix Creation
Downsample Algorithm	The algorithm that will be used when creating the state matrix. Min Max This algorithm finds where the maximum and minimum values for each tag occurs and then uses those samples as the basis for the state matrix. If the same vector is identified from two different tags, the redundancy is eliminated. This method provides the minimum number of vectors in the training matrix. Vector Ordering This algorithm is the recommended choice. This algorithm identifies samples to include in the state matrix by spacing the magnitudes of each training snapshot. A training snapshot is a data sample containing a reading from each tag. If the same vector is identified from two different tags the redundancy is eliminated. The appropriate target number of observations in the state matrix is entered in the Target Vectors box. This method is useful for data that contains many states and is non-stationary (i.e., the mean of the tags varies over time). Min Max Split This method applies the MinMax algorithm to equal-sized sections of data. This method returns more data than the MinMax method, so you can use this method if the MinMax method doesn't return enough observations for model training.
Reference Data Splits	The number of sections to which the MinMax algorithm should be applied. Used when Min Max Split is specified in the Downsample Algorithm box.
Maximum Vectors	The maximum number of vectors to include in the state matrix. This option works with models that have auto adaptation. If the maximum vectors are larger than the target vectors, it will increase the size of the state matrix. For example, if the target vectors is set to 25 and the maximum vectors is set to 50, the first 25 times auto adaptation occurs, the state matrix will become larger, then it will stop increasing at 50. If the maximum vectors is smaller than the target vectors, this field will have no effect.
Target Vectors	The number of vectors that will be targeted when the state matrix is created. Used when Vector Ordering is specified in the Downsample Algorithm box.
Redundancy Check Time	Specifies when redundancy checks are to be executed (e.g., during runtime or setup of the model). They will be executed when the H-Matrix created or modified, when the Local D-Matrix is created, or both. Since the H-Matrix is relatively large, a redundancy check will take comparatively longer to run. The goal is to keep redundancies out of the estimate generation algorithm. H-Matrix Modified Anytime the H-Matrix is modified by either the user or by adaptation, an algorithm is executed to filter similar vectors from the H-Matrix. Local D-Matrix Created During runtime, VBM creates a local state matrix at every observation. Local D is typically small in size (approximately 10 vectors).Selecting this option runs the algorithm that filters similar vectors when the local state matrix is created. Since it is a small matrix, it executes quickly. Always Executes redundancy filtering when the H-Matrix is modified and when the Local D-Matrix is created.
Dynamic State Matrix Vectors
Number of Vectors	Fixed The VBM state matrix will contain exactly this many vectors. Variable The VBM state matrix will select a number of vectors between the minimum and maximum (below).
Similarity Percentile	Threshold used to determine whether or not vectors are different enough from each other to both be considered for the VBM state matrix.
Minimum	Minimum number of vectors in the VBM state matrix. Only applicable when Number of Vectors is Variable.
Maximum	Maximum number of vectors in the VBM state matrix. Only applicable when Number of Vectors is Variable.

Auto-associative vs. Inferential Modeling Methods

At each sample, the system generates a residual vector–one element for each tag. A residual is the difference between an estimated value and a real-time value. If the system or process is behaving normally and the model is accurate, the residuals should be small with a random Gaussian distribution around 0

The system uses an Auto-Associative or Inferential modeling method when calculating residuals. When the model is Auto-Associative, the model uses every tag input to calculate the estimated state of every tag in the model. When the model is Inferential, the modeling component uses only a portion of the tag inputs to generate the estimated values for all the tags in the model.

In contrast to auto-associative modeling where all input tags contribute to the estimates for all outputs, inferential modeling uses a subset of the input tags to generate estimates for all outputs. An inferentially modeled (dependent) tag is not a contributor to the model’s output. This type of tag is useful when you want to insure that the inferred tags do not affect the estimates of the remaining tags in the model.

The modeling method is controlled by the "Is Independent" switches in the active model tags. (If the "Is Independent" column is not visible on the model tags page, enable it by clicking on the "Columns" button to view the column selector, and check the "Is Independent" checkbox.) The "Is Independent" flag on a tag indicates (if checked) that the tag's values are independent of the values of other tags or (if unchecked) that the tag's values can be inferred from the values of other tags. There are three cases that determine the modeling method:

If all of the model's tags have the "Is Independent" checkbox checked, then auto-associative modeling is used.
If any (but not all) of the "Is Independent" checkboxes are unchecked, then inferential modeling is used.
If all of the "Is Independent" checkboxes are unchecked, then inferential modeling is impossible (since it requires at least one independent tag), so modeling reverts to auto-associative.

About Virtual Signal Generation

SmartSignal provides you the option to generate a virtual signal so the modeling engine can generate estimates for tags that contain outlier or missing data values. The benefit of virtual signal generation is the ability to preserve good data that exists along with bad data in the same observation vector.

By default, when bad or missing tag data is encountered, the data feed replaces bad values with a NaN (Not a Number) and passes the data to the runtime engine. If the data is from a non-modeled tag, the runtime engine replaces the NaN with a Null and displays the affected data as a blank in graphs. If the data is from a modeled tag, the entire vector is ignored during modeling. Also, if the outlier filter algorithm is enabled and a vector contains a value that is considered an outlier, the entire vector is ignored during modeling.

If you enable the virtual signal generation function and the number of tags containing bad data (Outlier or NaN) is less than or equal to a threshold you set, the modeling engine generates estimates for tags that contain missing data values. To avoid creating NaN values in the residuals and other modeling engine outputs, the modeling engine sets the residual and SSCADI decision (for a tag that corresponds to a virtual estimate) to zero. The system also modifies all other modeling algorithms (those performing operations on input data) so they can handle the NaNs and outliers.

About the State Matrix

In a model, the state matrix is a matrix of vectors in which each vector represents one of the normal operating states (for healthy behavior) of the equipment or system. Historical data from all tags in the model are included in the matrix. Using the state matrix, the runtime engine can, in real-time, compare each observation of live data to the set of normal operating states represented in the state matrix. With this comparison, the runtime engine generates an estimate for each tag in the observation and then compares the actual value to the estimate value.

Modify Data in the Page for a Model

Use these steps to modify the values associated with a model.

Procedure

Lock for editing the analytic that you want to modify.
Select any cell to modify the value.

What To Do Next

When you have finished making changes, deploy the analytic instance.

Create or Rebuild a Model

In the SmartSignal Maintenance module, you can create or rebuild a model for a deployment.

Procedure

In the Applications menu, navigate to the RELIABILITY section, and then select SmartSignal Maintenance.
Select an analytic deployment.
Depending on whether the analytic is locked for editing by another user, select one of the following options:
- If the analytic is not locked for editing by another user, select Edit.
- If the analytic is locked for editing by another user, confirm that the user who has the analytic locked for editing is no longer editing the analytic, and then select Take Edit Control. When prompted, select Take Edit Control again.
The table changes so that the values can be modified. If you took edit control from another user, any changes that user had in progress are transferred to you.
Note: If you took edit control from another user and you do not want to keep the changes that user had in progress, you can select Actions > Discard Changes to remove the changes that were in progress.
In the Actions box, select Model Training > Update Models.
The Update Models window appears.

If you want to import training data, select the CHANGE TRAINING DATA check box, and then complete the following steps:

Select one of the following options:


Option	Description
APPEND	Adds new training data to the existing training data.
REPLACE	Replaces existing training data with new training data.
SYNC	Synchronizes new tag training data with existing models using time stamps of existing training data. All impacted output tags are re-evaluated.

Select the source of the training data from one of the following options:


Option	Description
PREDIX TIME SERIES	Uses raw time series data.
ANALYTIC OUTPUT	Uses only the data that the SmartSignal Cloud Runtime has processed in this analytic. This source data is available for the Append and Replace options, but not for the Sync option.
FILE (.CSV, .ZIP, .TAB)	Uses the data from the specified file.

Select a date range.
If you selected PREDIX TIME SERIES, specify a date range and interval.

Note: If you selected ANALYTIC OUTPUT, the data is taken at the interval at which it was collected and processed for this deployment, unless you select the Downsample by selected interval checkbox. This checkbox allows you to reduce the data by increasing the sampling interval.

If you want to include imported data in reference data, switch the Include imported data in reference data toggle, and then select one of the following options:

Option	Description
Conservative	This is the default option, and it is recommended in most cases. This option presumes that your training data set includes a high proportion of undesirable operating behavior on which to train the model. Therefore, this option uses a small subset of the available training data as reference data, which results in a model with a narrow training range. This narrow range reduces the possibility of including undesirable data, but it may omit some relevant data.
Moderate	This option uses a balanced approach to include a larger set of available training data as data patterns while reducing the potential of training in undesirable data.
Aggressive	This option is only recommended when asset operation is generally controlled, and available training data is considered reliable. This option uses a wider set of the available training data as reference data. While this may result in a model with a wider training range, it also increases the potential to train the model using undesirable patterns.
All	This option uses all available or imported training data as reference data and is only recommended for especially advanced users or when the amount of available training data is limited.

If you want to change which model are being rebuilt, change the selection of the models in the grid.
To enable display of recommended model residual thresholds, select the Recommend Model Residual Thresholds checkbox and the application will calculate values for the positive and negative thresholds for the models selected. Only the active model tags that have their corresponding analytic template positive and negative thresholds set to zero will have a recommendation generated. Previously customized settings will not be modified.
If you want to change the default behavior and not apply any configured High/Low, Flat-Line, or Spike filters when rebuilding the model, clear the Apply Filters checkbox for the corresponding model.
If you want to change the User Data Selection value for any model, use the check box in the gird to set its value. For more information, refer to the Modify Training Data section of this documentation.
Select Continue.
A progress indicator appears in the Edit Deployment workspace. When the process is complete, the Training Result Data window appears, summarizing the added data.

Activate or Deactivate a Model

You can activate or deactivate a model to specify whether the SmartSignal software should make predictions and perform data modeling using the tags and configuration associated with the model.

About This Task

Note: The deployment containing a model must be active for an active model to be processed.

Procedure

Access SmartSignal Maintenance, and then select a model.
Select Edit.
In the Actions box, select Settings > Edit Model Settings.
The Edit Settings window appears.
Select the Active toggle to activate or deactivate the model.
Your changes are saved automatically.
Select Close.
When you have finished making changes, select Deploy, and then, when prompted to confirm the deployment, select Deploy again.

Note: If you want to undo your changes, you can select Discard Changes in the Actions menu.

A notification appears when your changes have been deployed successfully.

Results

In the tree, the model appears gray if it is inactive or black if it is active.

Add Reference Data to a Model

In the SmartSignal Maintenance module, you can add reference data to an existing model.

About This Task

For best performance, add only the minimum amount of data required for a representative data set. It is recommended that you do not exceed 20,000 total vectors of training data.

Procedure

Access SmartSignal Maintenance, and then select a model.
Select Edit.
In the Actions box, select Model Training > Update Models.
The Update Models window appears.
Select Change Training Data to add, replace, or sync data into your models.
Specify a date range.

If you want to include imported data in reference data, switch the Include imported data in reference data toggle, and then select one of the following options:

Option	Description
Conservative	This is the default option, and it is recommended in most cases. This option presumes that your training data set includes a high proportion of undesirable operating behavior on which to train the model. Therefore, this option uses a small subset of the available training data as reference data, which results in a model with a narrow training range. This narrow range reduces the possibility of including undesirable data, but it may omit some relevant data.
Moderate	This option uses a balanced approach to include a larger set of available training data as data patterns while reducing the potential of training in undesirable data.
Aggressive	This option is only recommended when asset operation is generally controlled, and available training data is considered reliable. This option uses a wider set of the available training data as reference data. While this may result in a model with a wider training range, it also increases the potential to train the model using undesirable patterns.
All	This option uses all available or imported training data as reference data and is only recommended for especially advanced users or when the amount of available training data is limited.

Select Continue.
A progress indicator appears in the Edit Deployment workspace. When the process is complete, a notification appears, confirming that the model was created.
Select OK.
The reference data is added to the model.

Access Constants for a Model

In the SmartSignal Maintenance module, you can access the constants associated with a model.

Procedure

Access SmartSignal Maintenance, and then select a model.
In the Actions box, select Settings > View Model Constants.
The View Model Constants window appears.

Rebuild a Model

When you make changes to a model (for example, by modifying tag data or training data charts), you may need to rebuild the model in order for your changes to take effect. Complete the following steps to rebuild a model.

Procedure

Access the model that you want to rebuild.
Select Edit.
In the Actions menu, select Model Training > Rebuild Model.
When you have finished making changes, select Deploy, and then, when prompted to confirm the deployment, select Deploy again.

Note: If you want to undo your changes, you can select Discard Changes in the Actions menu.

A notification appears when your changes have been deployed successfully.

Copy Reference Data

You can copy the reference data from one model to other models in the same mode if the analytic is locked for editing.

About This Task

You can copy all reference values or just the new reference values that have been green stripped in the current locked session.

Reference data is the combination of green stripes and adaptation stripes. Adaptation stripes are maintained by the SmartSignal Runtime Engine and you cannot modify the adaption stripes. New Reference data refers to the stripes that have been added but not been deployed into the production instance. If you choose the option to copy only New reference data, observations in adaptation stripes are not copied to target models. Adaptation stripes in target models are not affected when copying from source model, for both options. If you choose the option to copy all reference data to target models, the adaptation stripes (observations) in source model are included. However, the adaption stripes in target models are not affected.

After the reference data is copied, if the reference data of target models are changed, the models become invalid for state matrix, and the reference data are set as user-selected.

Procedure

Access the model that you want to use as the source of the reference data.
Select Edit.
In the Actions box, select Model Training > Copy Reference Data.
After the changes are made, select Copy.

Note: If you want to undo your changes, select Discard Changes in the Actions box.

The copy operation starts. A result summary appears when the operation completes.

Modify Constants for a Model

In the SmartSignal Maintenance module, you can modify the constants for a model.

Procedure

Access SmartSignal Maintenance, and then select a model.
Select Edit.
In the Actions box, select Settings > Edit Model Constants.
The Edit Constants window appears.
Modify the values as necessary.
Select Close.
When you have finished making changes, select Deploy, and then, when prompted to confirm the deployment, select Deploy again.

Note: If you want to undo your changes, you can select Discard Changes in the Actions menu.

A notification appears when your changes have been deployed successfully.

Deploy a SmartSignal Analytic Instance

After you modify an analytic instance or model, you must deploy your changes for them to take effect.

Procedure

Access a SmartSignal analytic.
As needed, modify the model.
When you have finished modifying the model, select Deploy.
The Deploy window appears, displaying a box in which you can enter notes about your modifications to the model.
Optional: In the Note box in the Deploy window, enter notes about your modifications to the model.

Note: Any notes you enter appear when you view the analytic instance history.
When you are ready to deploy your changes, select Deploy.

Results

The analytic instance is deployed, and you can view details about the analytic instance in the Overview section in the SmartSignal Deployments page.
The analytic instance and any notes you added appear in the Analytic Instance History window for the analytic instance for which you deployed changes.