Implement model
- Fernanda Walker
If no model has been implemented yet, you will see the following message on the screen:
If there are already models in implementations, you will see the list of these models, as well as the button that allows you to add new ones.
After clicking on either of the two buttons above, the creation wizard with the necessary steps to set up the implementation of your model. The wizard will guide you through 4 steps:
When you have completed the 4 steps you can train and publish the implementation of your model.
Once you start the creation wizard, your implementation configuration is saved as a draft, so you can leave it at any time and come back when you want to complete it.
Creation wizard
Basic information
For the first step you need to name your implementation of the anomaly detection model and select an asset.
You can also optionally add a description.
Once you have filled in at least the required fields (name and asset), you can click on "NEXT >" to proceed to the second step.
Data/time selection
The data used to train the model is obtained from the dynamic properties of your asset. If your asset does not have such properties, the list will be empty.
You can create dynamic properties using the data flow module.
In this second step you are going to select the properties and the period of time to train with in order to establish the normal behaviour of your asset.
Click on the property selector. You can choose "All properties" if you want all properties of your asset to be taken into account or you can individually choose the ones you want to include in the training.
You can also use the "Search..." field to help you find specific properties you need to include.
Select the time interval you want to use for the training:
You can choose a specific time range using the calendar fields.
Or click on 'Use all data' to automatically select the date range in which your asset has data (the last 6 months with data of the properties you have selected is loaded. If no data is found after clicking on 'All data', the last 6 months will be automatically selected in the date range filter).
Once you have completed the properties information and the time range, the "Apply" button will be enabled, click on it.
For the training of your implementation it's important to use a time interval with a small number of anomalies (at most 5% of anomalous moments). To help you check this, the following graphs will be loaded:
Number of available data. Here you can see how much data is being received from your asset in the selected time interval.
Verify that there are no moments where no data is being received or where there is a significant interruption or variation in the amount of data received. You can see more information by hovering your mouse over the bars:
Alarms related to your asset and its selected properties. In this graph you can see a summary of the status of your asset in the selected time range.
Green color: indicates that no alert has been triggered at that time.
Pink, yellow, orange or red color: indicates that at least one alarm has occurred at that time. If there are several alarms, the color of the highest severity reached at that time is displayed.
Gray color: indicates that the asset has no alarms configured or did not yet have them configured in the selected date range.
Verify that there are no alarms or as few alarms as possible in the selected time range. You can get more information about the alarms by positioning the mouse over each time block:
Trends of the selected properties. By default they are disabled, but if you click on the switch, the graphs with the trends of all the properties you have selected will be loaded. These graphs can help you verify that everything is correct in the data being received for each of the selected properties.
Use the graphs to verify that the selected properties and date range are appropriate for training the model.
If you are not satisfied you can edit what you need in the "Properties" and "Time range" fields and click again on the "Apply" button to update the graphics.
Once you are ready, click "Next >" to continue with the configuration.
Statistical analysis
In the third step of the wizard you can optionally perform an advanced analysis if you need more statistical information on the data you have selected to confirm that they are suitable for the training.
If you want the analysis to be performed, click on "Yes, perform statistical analysis". If you do not require it, click on "Skip and continue to the next step".
A message may appear indicating that the process to perform the analysis can take a few minutes. If this is the case and you don't want to wait on that screen, you can exit the wizard with the "Close" button and reopen it when you are notified that the analysis is ready or whenever you need it.
If you exit the wizard, you can check the status of the draft at any time in the list. If the analysis is still in progress, you will see the word "Analyzing…" in the "DRAFT" column:
When the analysis is ready, you will see the buttons to return to the wizard:
Anyway, if you are out of the wizard at the time the analysis is completed, you will receive a notification indicating that it’s ready and you can go back to the configuration.
You can view the analysys it in the wizard window itself. You can also download it or view it in full screen by clicking on the buttons in the upper right corner.
Once you have reviewed the analysis, if you are satisfied click on the "Next >" button or, if the analysis has helped you to detect that the selected data is not the right one for the training, you can click on "<Back" to change it.
Pretreatment
The fourth step of the wizard is the data pretreatment configuration. In this final step you can specify whether you want to use temporal granularity to train your model.
Temporal granularity is disabled by default, you can leave that way and continue or click on the switch to enable it. Once activated, you can select whether you want it to be autocalculated or manual.
In the case of autocalculation, the value used is shown just below the selector:
If you change it to manual, you will be able to edit the value and adjust it to the one you need. We show you the ranges that it's recommended not to exceed:
Once you are ready, click on "Train model":
Training
When you have completed the steps in the wizard and clicked on "Train model", a message may appear indicating that the process can take a few minutes. If this is the case and you don't want to wait on this screen, you can exit the wizard with the "Close" button and reopen it when you are notified that the training is ready or whenever you need it.
If you exit the wizard, you can check the status of the draft at any time in the list. If the trainig is still in progress, you will see the word "Training…" in the "DRAFT" column:
When it’s ready, you will see the buttons to return to the wizard:
Anyway, if you are out of the wizard at the time the training is completed, you will receive a notification indicating that it’s ready and you can go back to the configuration.
Publication
Once the training process has been completed, the "Publish" button in the configuration wizard will be enabled.
After clicking on it a small popup window will open.
It will show you 3 fields required for publication:
Output property name. By default it is given the same name you have chosen for your implementation. If you are not satisfied, you can change it.
Alarm for anomalies. By default, a high severity alert is created upon publishing, so that it is activated each time an anomaly is detected. You can disable it if you do not want it to be created or change the severity at which it is activated. You can also add a label.
Calculation frequency. By default it's the same as the one defined in the temporal aggregation, but you can change it to the one you prefer.
When you are satisfied with the data in the publication form, click on "Publish and run". This will exit the wizard and return you to the list, where you can confirm at any time whether your implementation is still in the publishing process:
If it has already been published, the information in the "DRAFT" column will disappear, the "LAST PUBLISHED" column will show the date on which the publication was made and in "STATUS" you can always see whether the implementation is running or stopped.