Explainable machine learning for incipient anomaly detection in compact molten salt heat exchanger with overlapping...

Area of Science:

• Nuclear Engineering
• Materials Science
• Computational Modeling

Background:

• High-temperature molten salt-cooled reactors (MSCRs) offer advanced nuclear power generation but rely on reliable heat exchangers (HXs).
• HX failures, caused by temperature gradients and channel plugging, pose risks to MSCR safety and efficiency.
• Current monitoring methods lack the spatial resolution for early detection of incipient HX faults.

Purpose of the Study:

• To propose and evaluate a novel compact salt-to-salt matrix-type HX design with integrated distributed temperature sensing (DTS).
• To enable localized, early-stage fault detection in MSCR heat exchangers.
• To benchmark machine learning models for anomaly detection in challenging, non-separable datasets.

Main Methods:

• Developed a novel matrix-type HX design with interleaved parallel tubes and integrated synthetic fiber optic DTS.
• Generated high-fidelity synthetic data via computational modeling, simulating channel plugging and incorporating sensor noise.
• Benchmarked eight supervised machine learning models, including XGBoost, and developed an explainability framework using Shapley values and POSETs.

Main Results:

• The novel HX design with DTS enables localized detection of incipient faults.
• XGBoost achieved the highest performance in classifying anomalies within a non-separable dataset.
• The explainability framework successfully quantified feature importance, enhancing model transparency and trust.

Conclusions:

• The integrated DTS and explainable ML approach significantly improves predictive maintenance capabilities for MSCRs.
• This technology enhances operational resilience and safety in advanced nuclear reactor systems.
• Intelligent feature selection combined with DTS and explainable ML is crucial for robust fault detection.