Monitoring and Anomaly Detection Approaches with AI and Data Analytics for Pipelines
For this month's newsletter, we thought we'd take the opportunity to share that we've had a paper published in Pipeline Technology Journal - ptj . It continues with the AI focus that featured in our last two editions.
In the paper, we take you through two different approaches for operational asset monitoring and anomaly detection. On the one hand, purely physics-based methods and – on the other hand – machine-learned approaches. While we see great value in machine learning and AI driven approaches, hence our focus on AI in the last two newsletters, we argue that:
“While machine-learning is an excellent tool for building experts out of data, it is no replacement for engineering and physics expertise.”
If this piques your interest, you can read the full article here. Otherwise, keep reading for a brief summary.
?
Anomaly Detection
Anomaly detection finds events that vary from what you would expect. By identifying unusual incidents, operators can find important clues that will help them address problems before they evolve into incidents causing environmental, economic or reputational damage. Increasing safety is one reason why Klarian is developing technology in this domain.
Physics-based models
The physics-based models we investigated include computing hydraulic efficiency expectations versus reality, with cross-correlation of pumping and valve statuses, and breaking complex networks into legs and routes for a granular analysis, combined with a fuel-tracking system.
领英推荐
Machine-learning techniques
The machine-learning techniques we use include multivariate statistical analysis using Principal Component Analysis (PCA) and the Mahalonobis distance, artificial neural networks (specifically autoencoder networks), and Support Vector Machines.
Conclusion
While each method has its own trade-offs, research proved that both physics-based and machine-learned methods are effective monitoring and anomaly detection approaches. Deployment is highly contextual, dependent on specific operational requirements, system complexity, and available data sources.
What about you? Where do you think machine learning and AI should fit in relation to physics? Drop a comment with your thoughts below.
Further reading
?