Utilisation of on-line machine learning for SCADA system alarms forecasting

This paper describes a prototype design and implementation of a real-time (on-line) knowledge generation component which can be utilised in industrial Supervisory Control and Data Acquisition (SCADA) systems. The overall architecture of our SCADA scenario, which utilise proposed knowledge generation is based on a multi-agent approach. This design is different from what we can see in conventional commercial SCADA solutions. Nowadays, there is a big pressure on operators to precisely analyse a huge amount of data coming from technological processes and make right decisions in the right time. This is where a real-time knowledge generation can highly improve decision making strategies in complex industrial processes. Nevertheless, the actual state of the art solutions are usually not using the knowledge generation directly, or there are often restricted so called off-line learning approaches. The recent development in the area of machine learning lead to the creation of distributed solutions which could process real-time data and dynamically adapt the generated knowledge. We applied this on-line machine learning approach in our proposed prototype. The experimental agent is focused on the specific scenario of the process alarm forecasting, which is considered to be a binary classification problem. We describe our solution for useful classifier feature vector construction. The classifier itself is based on Passive-Aggressive algorithm. Furthermore, in order to evaluate a performance of the classification the results from knowledge generation experiments were provided in form of Matthews Correlation Coefficient (MCC) together with Receiver Operating Characteristic (ROC). The proposed prototype shows how to design and implement an on-line knowledge generation component for novel SCADA solutions.