Target Driven Anomaly Detection¶

The "Anomaly Detection" category offers automated journey for identifying the variations in your time series dataset as part of Predict offering using advanced machine learning technologies.

Target driven anomaly detection engine is created based on the normal behaviors of your assigned target datapoint from the dataset. The normal behaviors come from the target datapoint and also from the influence of other datapoints from the same datasdet. The engine can identify the anomalous behavivors of the target datapoint from multiple perspectives as detection logic. In industrial scenarios, it could contribute to monitoring continuously the anomalous behaviors from the target sensor which is the most critical sensor based on the expertise from the domain experts and could give early alerts to avoid more serious malfunctions in the near future.

Target driven anomaly detection engine could leverage the target datapoint and other influencer datapoints from asset and perform easy configuration for building the normal behaviors and perspectives to identify anomalies.

It could visualizes the anomaly spots, indicators, which could help you to analyze the anomalous behaviors and the root cause.

User Interface¶

The following image shows the Anomaly Detection screen:

① Displays the Dataset information
② Displays the variable for which anomaly needs to be created
③ Data visualization area
④ Displays the configuration parameters

Building the Anomaly Detection engine¶

To build the anomaly detection engine, proceed with the following steps:

1. Open "Predict" by clicking on below icon on Insights Hub launchpad.

   

2. Select the required "asset" and click "Anomaly Detections" in dataset tile.

3. Select "Target Driven" in anomaly detection approach section.
4. Select "Target Variable" from dropdown.
5. In the "Basic Configuration" section, select the "From Data" and "To Date" as the time period required for the training dataset for the engine building.
6. Select the "Predicator Variables" and the "Timeseries Type".
7. Configure the "Perspective" settings:
   • Residual: Detects the residuals with significantly higher magnitude than those observed on the in-sample period as anomalous.
     
   • Fluctuation: The fluctuation perspective helps to detect anomalies if different fluctuations are observed than those which were present during the in-sample period.
     
   • Imbalance: The imbalance perspective helps to detect anomalies accompanied by deviation of the residuals output from zero for a longer period of time.
     
8. Configure the "Data Imputation" setting by selecting the type from the drop-down.
   
9. For "Seasonality", click "On" or "Off" to enable or disable the seasonality.
   
10. Select "Polynomial" or "Fourier" to configure the type of "Feature".
11. Click "Submit".
    

The generated anomaly detection engine can be used to visualizes the anomaly spots, indicators.

Change Threshold¶

1. Click on Change threshold to change the sensitivity. User can change the threshold for a specific perspective and preview the results.
2. Once satisfactory results are shown, aka user is happy with the anomalies based on sensitivity then click on save to persist updated detection logic configurations.
3. Now, updated thresholds will be used for inferences.

Schedule the Anomaly Detection engine for production¶

It is possible to schedule the production of the engine which can be executed and evaluated on the regular intervals as per the preferred time range.

To schedule the production run of the anomaly anomaly engine, proceed with the following steps:

1. Open "Predict" in the "Analyze" tab.
   
2. Select the "Use Case", and click "Anomaly Detections".
   
3. Select the preferred engine from the list of available engines.
   
4. Click Schedule for production.
   
5. Select the preferred frequency for executing the engines.
   The anomaly detection engine can be executed as per the below intervals:
   • Minutes
     
   • Hourly
     
   • Daily
     
   • Weekly
     
6. Select the date in the "Schedule Start Date" and "Schedule End Date".
   
7. Enable 'Get notified when anomalies occur' button for seamless rules engine integration and getting notified.
8. Click "Deploy".

Anomaly Detection Results¶

Once the anomaly detection is scheduled for production, the anomalies are generated as per the schedule. The generated anomalies can be used to analyze the data for the selected variables of the asset. To view the anomaly detection, click "Production Detections" and select the preferred anomaly detections from the "Detail result for anomaly execution" drop-down.

Detecting Anomalies¶

Once the engine for anomaly detection is built, configure the "Testing Dataset" setting, the time period to execute the anomaly detection. Click "Detect" to start detecting the anomaly spots for the selected dataset manually.

Once the Anomaly detection is completed, the anomaly spots are displayed as shown below:

① Displays the selected variables and anomalies
② Displays the anomaly indicators as per the perspectives
③ Displays the top predictors
④ Displays the anomaly detection configuration details

Delete anomaly detections¶

To delete the generated anomaly detections, select the anomaly detections from the list, click and select "Delete".