Detecting anomalies

You can detect anomalies, or outliers, in your time series data by using the built-in functions in the Analytics Service catalog.

In Analytics Service, anomaly detection is used to find patterns in your time series data that do not conform to expected behavior. Anomaly detection helps you to identify problems with your devices or assets early. For example, you might use an anomaly detector to identify that a critical device in a mechanical chain is failing before the device impacts the entire chain.

Analytics Service uses several anomaly models to detect and alert on anomaly conditions, such as the following examples:

• A faulty sensor is not sending data. The condition might manifest as a flat line on a graph.
• A device is malfunctioning or requires a battery replacement.
  You might see a vertical line on a line graph, or you might see that a peak value is exceeded or that the lowest minimum value is exceeded.
• A piece of equipment is out of calibration. A steady or flat line on a graph might become noisy.
• A piece of equipment is not performing as expected. A target variable that is correlated with dependent variables is not within its predicted range.

Anomaly detectors are classified as follows:

• Unsupervised anomaly detectors: Trains on unlabeled data, assumes that the data is normal, and looks for data that does not fit this assumption.
• Supervised anomaly detectors: Trains on labeled data and predicts the values of unseen data.

Before you run anomaly detectors on your input data, you can run data quality checks on the input data from your sensors using a built-in function. With the quality checks, you can detect and correct the faulty data before you perform anomaly detection. For example, faulty data might relate to the quality of sensors that are used to collect data rather than to usual behavior in the environment that is being measured. For more information, see the DataQualityChecks built-in function.