Speaker
Description
In modern high-energy physics experiments, effective data quality monitoring is essential. Histograms serve as a primary tool, ensuring experimental efficiency and data integrity. However, the sheer volume of histograms produced in large-scale experiments makes manual monitoring impractical. Although traditional statistical methods handle large datasets, they often struggle with high false-negative rates and lack adaptability to complex patterns. To address this, this work presents a machine learning framework for real-time anomaly detection, aimed at enhancing the automation, accuracy and scalability of monitoring extensive histogram data. Our method utilizes an autoencoder to learn normal patterns from histogram data, identifying anomalies as instances with high reconstruction errors. We established a complete, end-to-end pipeline spanning feature extraction, model training, evaluation, and online deployment, delivering a highly automated and scalable solution. The framework was developed and systematically validated using a dataset of approximately 200,000 SPMT histogram samples from the Jiangmen Underground Neutrino Observatory (JUNO). Experimental results demonstrate that our approach can effectively monitor tens of thousands of histograms simultaneously, reducing processing latency to mere seconds while significantly improving detection precision and recall compared to conventional methods. The successful deployment of this system in a live experimental environment proves its engineering utility and potential for broader adoption, presenting a new paradigm for intelligent experiment operations. This talk will detail the system's design, implementation, and application.
| Minioral | Yes |
|---|---|
| IEEE Member | No |
| Are you a student? | No |