Related Experiment Videos
Deep Learning Based Superconducting Radio-Frequency Cavity Fault Classification at Jefferson Laboratory
Lasitha Vidyaratne1, Adam Carpenter1, Tom Powers1
1Jefferson Laboratory, Newport News, VA, United States.
Frontiers in Artificial Intelligence
|January 20, 2022
Summary
Deep learning models show promise for classifying superconducting radio-frequency cavity faults, offering faster inference speeds for accelerator operations. While comparable in cavity identification, they slightly trail traditional methods in fault classification accuracy.
Area of Science:
- Accelerator Physics
- Machine Learning Applications
- Superconducting Radio-Frequency (SRF) Technology
Background:
- The Continuous Electron Beam Accelerator Facility (CEBAF) at Jefferson Lab uses 418 SRF cavities for high-energy electron acceleration.
- Automating the analysis of RF time-series data from cavity failures is crucial due to the large data volume and manual inspection limitations.
- Previous work utilized traditional machine learning (ML) for fault classification, motivating exploration of deep learning (DL) for enhanced performance.
Purpose of the Study:
- To investigate the efficacy of deep learning (DL) models, specifically recurrent neural networks (RNNs) and convolutional neural networks (CNNs), for classifying SRF cavity faults.
- To develop a DL system capable of fast inference for potential real-time fault prediction and intervention.
- To compare the performance of DL models against existing state-of-the-art ML fault classification models.
Main Methods:
- Utilized a dataset of RF waveform data from past operational runs of CEBAF.
- Implemented and analyzed deep recurrent neural networks (RNNs), including those with long short-term memory (LSTM) layers.
- Implemented and analyzed deep convolutional neural networks (CNNs).
- Compared the performance of RNN and CNN models against a traditional ML fault classification model.
Main Results:
- Deep learning architectures achieved performance comparable to state-of-the-art ML models for cavity identification.
- DL models demonstrated slightly lower accuracy in fault classification compared to traditional ML methods.
- A significant advantage of DL models was their faster inference speed, crucial for predictive applications.
Conclusions:
- Deep learning approaches, including RNNs and CNNs, are effective for classifying SRF cavity faults in accelerator facilities.
- The faster inference times of DL models offer potential for real-time fault detection and mitigation strategies.
- Further research may focus on improving DL fault classification accuracy while retaining inference speed benefits.