The autonomous anomaly detector service is a REST API-based solution that supports the training and prediction of events/severity from a time series.

Its value proposition is to mitigate the hard work required to build and maintain an anomaly detection system manually for assets streaming data signals for any industry.

In just a few clicks a swagger definition and test interface to simplify integration will be created once subscribed. No resource will be created in your Tenant so you will not have extra costs, just the subscription fee.

For each metric to be tracked, the developer creates a set of models. The time bucket intervals are flexible to what process is under evaluation. The basic workflow includes:

Creation of User Account
Creation of Model
Model training with historical data

Signal data
Event data
Severity data

Prediction

The training endpoint takes a signal and event/severity time series as input with basic configuration parameters. Internally the training process performs the following steps:

Train anomaly model
Identify anomalies in signal data
Break up the anomalies into clustered groups
Map events/severity points to anomaly clusters
Build event and severity prediction models
Save trained models into blob storage

The prediction process takes a time series sequence and returns a set of predictions that include event type and severity for the input data.

The system will evaluate the signal time series and generate predictions about the presence of anomalies and their potential cause and severity. See Figures 1 and 2.

Example: Let's use a typical Inverter asset example with this incoming data:

Inverter Data

Timestamp
Average Temperature
POA
Wind Direction
Wind Speed
Revenue Power Meter
AC Power
DC Power
DC Current
DC Voltage
Latitude
Longitude

Event Data

Timestamp
Type
Severity

Wouldn't it be nice to have an automated approach to the monitoring of inverter hardware that would improve your ability to schedule physical inspection of the system when there was potential serious risk of system failure? Our Autonomous Anomaly Detection monitoring system evaluates the plant data and alerts monitoring personnel to the presence of anomalous behavior and provides an estimate of the type and potential severity of the event.

Figure 3 includes the inverter signal data in gray with events and their corresponding severity. Events are color-coded red and blue while the size of their bar is the severity of the event.

Figure 4 shows how Substrate AI Autonomous Anomaly detection uses this data to train classifiers to predict the event type and estimate the potential severity of the event.

The solution relies on Microsoft Azure capabilities:

Azure Storage

Azure SQL Database

Service Plan

.NET

In combination with our own technology to ensure a great performance.

Autonomous Anomaly Detection

by SUBSTRATE ARTIFICIAL INTELIGENCE SA