WindSmart
Technology
Data
-
We use Supervisory Control And Data Acquisition (SCADA) signals, log remarks and past failure data in our product.
-
These data are already available in most wind turbine systems, meaning no new sensors need to be added.
-
Typical signal frequency is 10 mins.
Infrastructure
-
Product is hosted using an EC2 instance on Amazon web services (AWS).
-
Python scripts are used to preprocess the data sets and train machine learning models.
-
A Flask based web app is used to provide the model predictions through an intuitive user interface.
-
App is containerized using docker for ease of standardization and portability.
Modeling
-
Unsupervised learning technique used to detect anomalous turbine operation - Gaussian Mixture Model.
-
More than 65% of detected anomalies result in failures within 4-7 weeks.
-
Variables causing the anomalous behavior are identified.