Anomaly detection is a process in machine learning that identifies data points, events, and/or observations that deviate from a dataset’s normal behavior. To detect anomalies from unlabeled time series data is a pain point that is critical to address for industrial applications.

In manufacturing industries, IoT sensors are pervasive for monitoring assets. The data from these sensors can be analyzed with advanced statistical and machine learning algorithms to yield insight on pending failures, and optimal maintenance and operation strategies. A gas turbine, compressor, boat engine are examples of valuable assets that can benefit from such insights. These industrial assets fail due to many reasons, and often they behave in abnormal ways preceding the failures. In order to have an effective predictive maintenance program, it is crucial to be able to detect these abnormalities so that action can be taken to avoid downtime loss. However, identifying these tell-tale abnormalities before a failure event is difficult since, in general, failures are comparatively rare events so there is often a scarcity of tagged data for building predictive models.

This anomaly detection API service can help users detect anomalies from the entire time series or predict anomaly status of the last data input. The anomaly detection algorithms supported in the service provide an unsupervised way of learning a model from data that mostly contains “normal” operating data with a sparse number of actual failure events. Currently the service supports both univariate and multivariate time series.

To contact developers for any issues and feature suggestions, please check Getting Started page for more details.