A Complete Streaming Pipeline for Real-time Monitoring and Predictive Maintenance

Railway maintenance is changing as predictive maintenance (PdM) comes to prominence. In particular, the rapid progress of learning algorithms, commonly known as Machine learning (ML), strongly motivates data-driven PdM applications. However, traditional ML struggles with the large amount of data that arrive at high velocity in realtime streams. Facing big data-related issues, Stream learning (SL) is a new learning paradigm that adapts ML to the handling of fast, unbounded, and dynamic data streams. We deem SL suitable for online monitoring and relevant to our need of having incremental, drift-aware algorithms that quickly detect and predict anomalies. Aiming to enhance railway PdM with SL, we propose a complete streaming pipeline for real-time monitoring, anomaly detection, and anomaly prediction. A partial implementation of this pipeline has resulted in an interactive application named InterCE. Preliminary results on two real-world datasets supplied by a French railway company show that InterCE helps to improve the accuracy of the learning process.