The project demonstrates how sensor data can be used in a real-time detection system for hydro power plants. A vast majority of the historical data from the process control systems is stored in data historians. This data is utilized for the development of predictive maintenance models. One of the approaches is to use a real-time anomaly detection system based on AI-algorithms to predict machine failures before they actually happen and give insights into the current state of hydro plant components.
The technology stack for this project includes an OPC-UA-gateway, StreamSets, Kafka, InfluxDB and Grafana on the infrastructure side. For the AI-modelling “R” is used with a Python-Wrapper for execution. The AI-models were developed in-house by Data Science and AI-experts, the infrastructure setup was designed with external partners.
“Empowering AI is one of our core values on the digital journey.”
The first image below shows the monitoring of one of the machine sensors. The red curve shows the actual sensor measurement, the green curve is the predicted measurement from the AI model. The green area is calculated based on the predictions and shows the normal operation mode. If the actual value (red curve) exceeds the green area, the anomaly score will go up significantly. These calculations are executed for any sensor of the AI-model as shown in the two images below.