Wind Turbine Anomaly Detection

GitHub project

This sample app demonstrates a time series use case by detecting anomalous power generation patterns in wind turbines, relative to wind speed. By identifying deviations, it helps optimize maintenance schedules and prevent potential turbine failures, enhancing operational efficiency.

In this article, you can learn about the architecture of the sample and its data flow.

If you want to start working with it, instead, check out the Get Started Guide or How-to Guides for Time-series applications.

How It Works

As seen in the following architecture diagram, the sample app at a high-level comprises data simulators (which can act as data destinations, if configured) - in the real world these would be the physical devices, the generic Time Series AI stack based on TICK Stack comprising of Telegraf, InfluxDB, and the Time Series Analytics microservice using Kapacitor and Grafana.

Data flow explanation

Below is an explanation of how this architecture translates to data flow in the wind turbine anomaly detection use case, in which the data is ingested using the OPC-UA server/MQTT publisher simulators and anomaly alerts are published to a OPC-UA server/MQTT broker.

Data Sources

The demonstration uses the dataset contained in wind-turbine-anomaly-detection.csv, which is a normalized version of an open source wind turbine dataset (T1.csv), as retrieved from Kaggle.

This data is ingested into Telegraf through the OPC-UA protocol using the OPC-UA data simulator OR MQTT protocol using the MQTT publisher data simulator.

Data Ingestion

Telegraf through its input plugins (OPC-UA OR MQTT) gathers the data and sends this input data to both InfluxDB and Time Series Analytics Microservice.

Data Storage

InfluxDB stores the incoming data from Telegraf.

Data Processing

Time Series Analytics Microservice uses the User Defined Function (UDF) deployment package (TICK Scripts, UDFs, Models) from the sample apps. The UDF deployment package for the Wind Turbine Anomaly Detection sample app is available in this folder.

Directory details is as below:

config.json

The task section defines the settings for the Kapacitor task and User-Defined Functions (UDFs).

Key	Description	Example Value
udfs	Configuration for the User-Defined Functions (UDFs).	See below for details.

UDFs Configuration:

The udfs section specifies the details of the UDFs used in the task.

Key	Description	Example Value
name	The name of the UDF script.	"windturbine_anomaly_detector"
models	The name of the model file used by the UDF.	"windturbine_anomaly_detector.pkl"
device	Specifies the hardware CPU or GPU for executing the UDF model inference.Default is cpu	cpu

Note: The maximum allowed size for config.json is 5 KB.

---

Alerts Configuration:

The alerts section defines the settings for alert mechanisms, using the MQTT protocol by default. For publishing OPC-UA alerts in Docker, refer to Docker OPC-UA Alerts. For OPC-UA Alerts in Helm, refer to Helm OPC-UA Alerts

Note: Enable only one type of alerts: either MQTT or OPC-UA.

MQTT Configuration:

The mqtt section specifies the MQTT broker details for sending alerts.

Key	Description	Example Value
mqtt_broker_host	The hostname or IP address of the MQTT broker.	"ia-mqtt-broker"
mqtt_broker_port	The port number of the MQTT broker.	1883
name	The name of the MQTT broker configuration.	"my_mqtt_broker"

udfs/

Contains the Python script to process the incoming data. Uses Random Forest Regressor and Linear Regression machine learning algos accelerated with Intel® Extension for Scikit-learn* to run on CPU/GPU to detect the anomalous power generation data points relative to wind speed.

tick_scripts/

The TICKScript windturbine_anomaly_detector.tick determines processing of the input data coming in. The file contains the details on execution of the UDF file, storage of processed data and publishing of alerts. By default, it is configured to publish the alerts to MQTT.

models/

The windturbine_anomaly_detector.pkl is a model built using the Random Forest Regressor algorithm from the Scikit-learn library. For more details on how it is built refer to the README.