ML-Driven Anomaly Detection for EPICS PV Streams at the Edge: Implementation and Evaluation on Raspberry PI IOCs
DOI:
https://doi.org/10.63282/3050-9246.IJETCSIT-V6I4P121Keywords:
EPICS, Anomaly Detection, Raspberry PI, Edge Computing, Process Variables (PV), Tinyml, Autoencoder, Isolation Forest, Control SystemsAbstract
Distributed control systems based on the Experimental Physics and Industrial Control System (EPICS) framework generate high-rate process variable (PV) streams that require robust, low-latency monitoring to ensure safe and reliable operation [1], [2]. Traditional centralized monitoring architectures can suffer from bandwidth limitations, latency, and single points of failure. This paper presents a comprehensive design, implementation, and evaluation of an edge-deployed, machine-learning-driven anomaly detection system for EPICS PV streams running on low-cost Raspberry Pi hardware. The proposed architecture integrates EPICS IOCs with lightweight Autoencoder-based and Isolation Forest models quantized for TensorFlow Lite [3], enabling continuous inference on PV windows and local generation of anomaly PVs. We present engineering details for EPICS integration, data preprocessing, model training, quantization, and deployment, and we evaluate system performance on a laboratory testbed with synthetic and recorded PV traces. Results show sub-20 ms inference latency, modest CPU footprint, and a significant reduction in central network traffic when deploying anomaly filtering at the edge. The paper also discusses operational considerations, fault modes, and directions for adaptive, federated learning across distributed IOCs [12]
Downloads
References
[1] L. R. Dalesio and M. Kraimer, ”EPICS Architecture and Implementation,” in Proceedings of ICALEPCS, 1994.
[2] EPICS Community, ”EPICS Base Documentation,” 2024. [Online]. Available: https://epics-controls.org/
[3] TensorFlow Lite, ”TensorFlow Lite Guide,” 2024. [Online]. Available: https://www.tensorflow.org/lite
[4] F. T. Liu, K. M. Ting, and Z.-H. Zhou, ”Isolation Forest,” in Proceedings of ICDM, 2008.
[5] G. E. Hinton and R. R. Salakhutdinov, ”Reducing the Dimensionality of Data with Neural Networks,” Science, vol. 313, no. 5786, pp. 504–507, 2006.
[6] D. Lane et al., ”TinyML: Machine Learning on Ultra-Low Power Devices,” Communications of the ACM, 2020.
[7] J. W. Hill, ”pyepics: Python interface to EPICS Channel Access,” 2020.
[8] S. Shi et al., ”Edge Computing for AI Applications: A Survey,” IEEE IoT Journal, 2021.
[9] H. Liu and Z. Zhao, ”On Performance of Isolation Forest in Streaming Settings,” Journal of Data and Information Quality, 2020.
[10] F. Ahmed and M. Mahmood, ”Anomaly Detection Using Machine Learning: A Survey,” IEEE Access, 2020.
[11] A. Smith and B. Jones, ”Containerizing EPICS IOCs: Best Practices,” Workshop on Control Systems, 2019.
[12] P. Kairouz et al., ”Advances and Open Problems in Federated Learning,” Foundations and Trends in Machine Learning, 2021.
[13] Google Coral, ”Coral USB Accelerator,” 2023. [Online]. Available: https://coral.ai/
[14] S. Brown, ”Security Considerations for EPICS-based Control Systems,” Control Systems Magazine, 2018.
[15] nnA. Doshi-Velez and B. Kim, ”Towards A Rigorous Science of Interpretable Machine Learning,” arXiv preprint, 2017.
