arxiv: 2604.18611 · v1 · submitted 2026-04-13 · 💻 cs.NE · cs.AI· cs.LG

Recognition: unknown

Neuromorphic Continual Learning for Sequential Deployment of Nuclear Plant Monitoring Systems

Samrendra Roy , Sajedul Talukder , Syed Bahauddin Alam

Authors on Pith no claims yet

Pith reviewed 2026-05-10 15:01 UTC · model grok-4.3

classification 💻 cs.NE cs.AIcs.LG

keywords spiking neural networkscontinual learninganomaly detectionnuclear industrial control systemssequential deploymentneuromorphic computingenergy efficiency

0 comments

The pith

A hybrid continual learning strategy in spiking neural networks allows sequential deployment of anomaly detectors across nuclear plant subsystems while maintaining high accuracy and using far less energy.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that spiking neural networks can perform anomaly detection in nuclear industrial control systems even when subsystems are added one after another over time. It introduces a way to turn diverse sensor readings into sparse spike trains that match each sensor's natural pace, cutting input data by over 90 percent. Five different continual learning methods are tested on a real nuclear dataset covering boiler, turbine, and water treatment units. The best combination keeps detection performance near perfect across all units and cuts the number of operations by more than an order of magnitude compared with ordinary neural networks.

Core claim

The hybrid EWC plus replay approach in a spiking network reaches an average F1 score of 0.979, shows near-zero average forgetting (0.000 in a single run and 0.035 across seeds), and requires 12.6 times fewer operations than an equivalent conventional network while detecting every tested attack in 0.6 seconds on average.

What carries the argument

Spike-encoded asynchronous sensor fusion, a delta-based method that turns heterogeneous sensor streams into sparse spike trains at rates set by each sensor's own dynamics, combined with the hybrid EWC+Replay continual learning rule that protects important weights while replaying past examples.

If this is right

New subsystems can be added without retraining the entire model from scratch or risking loss of prior anomaly knowledge.
The same monitoring hardware can stay active continuously because the operation count drops by a factor of 12.6.
Attack detection latency stays under one second across all tested scenarios.
The approach directly supports staged plant commissioning where monitoring must begin before every unit is finished.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same spike-encoding and hybrid learning pattern could be tried on other industrial control systems that add sensors gradually.
Measuring actual power draw on a physical neuromorphic chip rather than using published estimates would give a tighter energy comparison.
Adding more than three subsystems in sequence would test whether the near-zero forgetting holds at larger scale.

Load-bearing premise

That strong results on the HAI 21.03 dataset and energy numbers taken from published hardware specifications will carry over to actual nuclear plant operations without extra tuning or real-world validation.

What would settle it

Running the trained system on live sensor streams from an operating nuclear facility and checking whether detection accuracy stays above 0.97 and forgetting remains near zero as additional subsystems come online.

Figures

Figures reproduced from arXiv: 2604.18611 by Sajedul Talukder, Samrendra Roy, Syed Bahauddin Alam.

**Figure 1.** Figure 1: Architecture overview. Heterogeneous sensors from four nuclear ICS subsystems are converted to sparse spike trains [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Figure 2: F1 score evolution during sequential task training. (a) P1 Boiler: Sequential and EWC degrade from 1.0 to [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Left: Average F1 across all three subsystems. EWC+Replay (0.979) approaches ANN per-task accuracy (0.999) while Joint [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: Replay buffer size sensitivity. Left: Average F1 [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

**Figure 5.** Figure 5: Asynchronous spike rates by subsystem. The 100 [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

read the original abstract

Anomaly detection in nuclear industrial control systems (ICS) requires continuous, energy-efficient monitoring across multiple subsystems that are often deployed at different stages of plant commissioning. When a conventional neural network is sequentially trained to monitor new subsystems, it catastrophically forgets previously learned anomaly patterns, a safety-critical failure mode. We present the first spiking neural network (SNN)-based anomaly detection system with continual learning for nuclear ICS, addressing both challenges simultaneously. Our approach introduces spike-encoded asynchronous sensor fusion, a delta-based encoding that converts heterogeneous sensor streams into sparse spike trains at rates dictated by each sensor's natural dynamics, achieving 92.7% input sparsity. We evaluate five continual learning strategies, including sequential fine-tuning, Elastic Weight Consolidation (EWC), Synaptic Intelligence (SI), experience replay, and a hybrid EWC+Replay approach, on the HAI 21.03 nuclear ICS security dataset across three sequentially deployed subsystems (boiler, turbine, water treatment). The hybrid EWC+Replay method achieves an average F1 score of 0.979 with near-zero average forgetting (AF = 0.000 single seed; 0.035 +/- 0.039 across three seeds), while requiring 12.6x fewer operations (an estimated 2.5x in energy based on published hardware specifications) than an equivalent artificial neural network. The system detects all tested attacks with a mean latency of 0.6 seconds. These results demonstrate that neuromorphic computing offers a viable path toward always-on, energy-efficient, and adaptable safety monitoring for next-generation nuclear facilities.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gets solid F1 and forgetting numbers on the HAI dataset with a hybrid EWC+replay SNN for sequential nuclear subsystem monitoring, but the energy and deployment claims rest on operation counts scaled from published specs rather than measured runs on neuromorphic hardware.

read the letter

The main thing to take away is that the hybrid continual learning approach with spiking networks holds performance across three sequential deployments on the HAI 21.03 dataset, reaching an average F1 of 0.979 and near-zero forgetting while using 12.6 times fewer operations than a comparable ANN. They also report 0.6-second mean latency for attack detection and 92.7 percent input sparsity from their delta-based spike encoding of sensor streams. That combination addresses a practical issue in nuclear ICS where subsystems come online at different times and forgetting prior anomaly patterns would be unacceptable. The work evaluates five strategies head-to-head, including fine-tuning, EWC, SI, replay, and the hybrid, and shows the mix performs best here. Credit for grounding the evaluation in a public dataset and running multiple seeds on the forgetting metric. The spike encoding step is a reasonable fit for asynchronous, heterogeneous sensor data and helps justify the neuromorphic angle. The central limitation is the efficiency story. The operation savings and 2.5x energy estimate come from counting MACs or spikes and scaling by generic published hardware figures, without executing the model on actual SNN silicon such as Loihi or BrainScaleS. That leaves routing costs, memory access patterns, and sensor interface overhead unmeasured, so the always-on viability claim for nuclear plants stays projected rather than demonstrated. The paper stays with one dataset, so how well the results hold under different attack distributions or real plant noise is still open. This is for people working on neuromorphic methods for industrial control or critical-infrastructure anomaly detection. Readers already following SNN continual learning will find the concrete metrics and method comparison useful as a domain-specific case study. I would send it to peer review because the dataset results and method evaluation are concrete enough to merit detailed referee input, particularly on closing the hardware measurement gap.

Referee Report

1 major / 2 minor

Summary. The paper presents the first spiking neural network (SNN) approach for continual anomaly detection in nuclear industrial control systems (ICS). It introduces spike-encoded asynchronous sensor fusion achieving 92.7% input sparsity and evaluates five continual learning strategies (sequential fine-tuning, EWC, SI, experience replay, and hybrid EWC+Replay) on the HAI 21.03 dataset across three sequentially deployed subsystems (boiler, turbine, water treatment). The hybrid method is reported to achieve an average F1 score of 0.979 with near-zero average forgetting while requiring 12.6x fewer operations (estimated 2.5x energy savings based on published hardware specifications) than an equivalent ANN, with mean attack detection latency of 0.6 seconds.

Significance. If the reported performance metrics and efficiency estimates hold under actual hardware deployment, the work would provide a concrete demonstration of neuromorphic continual learning for safety-critical, always-on monitoring in nuclear facilities. The use of a public dataset, concrete metrics (F1 scores, forgetting values, operation counts), and evaluation of multiple standard continual learning methods on sequential subsystem deployment are strengths that ground the empirical claims.

major comments (1)

[Abstract and Results] Abstract and Results: The headline efficiency claims (12.6x fewer operations and estimated 2.5x energy savings) are derived from MAC/spike counts on the HAI 21.03 dataset scaled by generic published neuromorphic hardware specifications rather than direct measurements on physical SNN hardware (Loihi, BrainScaleS, etc.). This is load-bearing for the abstract's conclusion that the system offers a 'viable path' for always-on nuclear ICS monitoring, because unmeasured factors such as AER routing energy, synaptic memory access patterns, and sensor-interface overhead remain unaccounted for.

minor comments (2)

[Evaluation] Evaluation: The manuscript provides no details on statistical significance tests, full per-subsystem F1 scores and forgetting values, specific hyperparameter choices for each method, or the post-evaluation justification for selecting the hybrid EWC+Replay approach.
[Abstract] Abstract: The multi-seed average forgetting value (0.035 +/- 0.039) would benefit from explicit description of how the variance was computed across the three seeds and whether it aggregates all subsystems.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the detailed review and constructive feedback on our work. We address the major comment on the efficiency claims below, agreeing that greater transparency is warranted regarding the estimation methodology.

read point-by-point responses

Referee: [Abstract and Results] Abstract and Results: The headline efficiency claims (12.6x fewer operations and estimated 2.5x energy savings) are derived from MAC/spike counts on the HAI 21.03 dataset scaled by generic published neuromorphic hardware specifications rather than direct measurements on physical SNN hardware (Loihi, BrainScaleS, etc.). This is load-bearing for the abstract's conclusion that the system offers a 'viable path' for always-on nuclear ICS monitoring, because unmeasured factors such as AER routing energy, synaptic memory access patterns, and sensor-interface overhead remain unaccounted for.

Authors: We agree that the reported efficiency figures are estimates obtained by counting multiply-accumulate operations in the ANN baseline and spike events in the SNN, then scaling by per-operation energy figures drawn from published neuromorphic hardware characterizations (e.g., Loihi and similar platforms). Direct end-to-end measurements on physical neuromorphic chips were not performed, as the study focuses on algorithmic and simulation-level evaluation using the public HAI 21.03 dataset. Consequently, overheads such as AER packet routing, synaptic memory access patterns, and sensor-interface circuitry are not included in the 12.6x operation reduction or the derived 2.5x energy estimate. To address this concern, we will revise the abstract to foreground the estimated nature of the energy claim, expand the methods section to detail the exact counting procedure and source references for the scaling factors, and add a new limitations paragraph in the discussion that explicitly lists the unaccounted hardware factors and calls for future physical deployment validation. These changes will make the 'viable path' statement more precisely qualified while preserving the core empirical contribution. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical evaluation on public dataset with standard methods

full rationale

The paper reports experimental results from training and evaluating five continual learning strategies (including EWC, SI, replay, and hybrid) on the HAI 21.03 dataset for anomaly detection across sequential subsystems. Performance metrics (F1=0.979, AF≈0) and operation counts (12.6× fewer) are obtained directly from these runs and compared to an ANN baseline; the energy estimate is a post-hoc scaling from external published hardware specifications rather than a derived claim. No equations, predictions, or uniqueness theorems are presented that reduce to fitted parameters, self-definitions, or self-citations. The central claims rest on measured experimental outcomes and are therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard assumptions from machine learning and neuromorphic computing applied to a domain-specific dataset; no new free parameters or invented entities are introduced beyond the described encoding and learning strategies.

axioms (2)

domain assumption The HAI 21.03 dataset is representative of real nuclear ICS anomaly patterns and attack scenarios
Used as the sole evaluation benchmark without additional validation mentioned.
domain assumption Energy savings can be estimated from published hardware specifications without direct measurement
The 2.5x energy claim is derived from operation counts and external specs.

pith-pipeline@v0.9.0 · 5597 in / 1459 out tokens · 57947 ms · 2026-05-10T15:01:36.857916+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

28 extracted references · 2 canonical work pages

[1]

Abhishek Chaudhary, Junseo Han, Seongah Kim, Aram Kim, and Sunoh Choi
[2]

Anomaly Detection and Analysis in Nuclear Power Plants.Electronics13, 22 (2024), 4428

2024
[3]

Dokania, Thalaiyasingam Ajanthan, and Philip H

Arslan Chaudhry, Puneet K. Dokania, Thalaiyasingam Ajanthan, and Philip H. S. Torr. 2018. Riemannian Walk for Incremental Learning: Understanding Forgetting and Intransigence. InEuropean Conference on Computer Vision (ECCV). Springer, 532–547

2018
[4]

Mike Davies, Andreas Wild, Garrick Orchard, Yulia Yu, Tularaj Kuber, Shyam Risbud, et al. 2021. Advancing Neuromorphic Computing with Loihi: A Survey of Results and Outlook.Proc. IEEE109, 5 (2021), 911–934

2021
[5]

Konstantinos Demertzis and Lazaros Iliadis. 2017. A Spiking One-Class Anomaly Detection Framework for Cyber-Security on Industrial Control Systems. In Engineering Applications of Neural Networks (EANN) (CCIS, Vol. 744). Springer, 122–137. 7 Samrendra Roy, Sajedul Talukder, and Syed Bahauddin Alam

2017
[6]

Wei Fang, Zhaofei Yu, Yanqi Chen, Timothée Masquelier, Tiejun Huang, and Yonghong Tian. 2021. Incorporating Learnable Membrane Time Constant to Enhance Learning of Spiking Neural Networks. InIEEE/CVF International Con- ference on Computer Vision (ICCV). 2661–2671

2021
[7]

Young, Andrew D

James Ghawaly, Aaron R. Young, Andrew D. Nicholson, Brett Witherspoon, Nick Prins, Mathew Swinney, Cihangir Celik, Catherine D. Schuman, and Karan Patel
[8]

InInternational Conference on Neuromorphic Systems (ICONS)

Performance Optimization Study of the Neuromorphic Radiation Anomaly Detector. InInternational Conference on Neuromorphic Systems (ICONS). ACM
[9]

Mark Horowitz. 2014. 1.1 Computing’s Energy Problem (and what we can do about it). InIEEE International Solid-State Circuits Conference (ISSCC). 10–14

2014
[10]

Ismail Hossain, Jahangir Alam, Sai Puppala, Kazuma Kobayashi, Samrendra Roy, Syed Bahauddin Alam, and Sajedul Talukder. 2025. A Privacy-Aware Cyber Attack Detection Framework for Advanced Reactors Using Data Fusion and Quantum Deep Learning. InNuclear Plant Instrumentation and Control & Human- Machine Interface Technology (NPIC&HMIT 2025). American Nuclea...

2025
[11]

Raisa Hossain, Farid Ahmed, Kazuma Kobayashi, Seid Koric, Diab Abueidda, and Syed Bahauddin Alam. 2025. Virtual Sensing-Enabled Digital Twin Framework for Real-Time Monitoring of Nuclear Systems Leveraging Deep Neural Operators. npj Materials Degradation9 (2025), 1–12

2025
[12]

Innatera. 2026. Pulsar Neuromorphic Microcontroller. https://www.innatera.com. Accessed: April 2026

2026
[13]

Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al

James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Oriol Vinyals, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. 2017. Overcoming catastrophic forgetting in neural networks.Proceedings of the National Academy of Sciences114, 13 (2017), 3521– 3526

2017
[14]

Kazuma Kobayashi and Syed Bahauddin Alam. 2024. Deep Neural Operator- Driven Real-Time Inference to Enable Digital Twin Solutions for Nuclear Energy Systems.Scientific Reports14 (2024), 1–15

2024
[15]

A. Kosari. 2025. Real-time network traffic anomaly detection using spiking neural networks with adaptive learning.Contributions to Science, Technology and Engineering2, 2 (2025), 17–22

2025
[16]

Santhi, S

Srimanta Mandal, B. Santhi, S. Sridhar, K. Vinolia, and P. Swaminathan. 2014. Anomaly detection in nuclear power plant data using support vector data de- scription. InIEEE International Conference on Computational Intelligence and Computing Research

2014
[17]

Arash Mehrabi and Andre Van Schaik. 2024. Intrusion Detection on Indus- trial Networks Using Neuromorphic Computing. InInternational Conference on Electrical, Control and Instrumentation Engineering (ICECIE). IEEE

2024
[18]

Zaidao Mei, Boyu Wang, Daniel Patrick Rider, and Qinru Qiu. 2023. SEnsitivity Modulated Importance Networking and Rehearsal for Spike Domain Incremental Learning. InInternational Conference on Neuromorphic Systems (ICONS). ACM

2023
[19]

Jaekyoung Park, Jung-Ho Kim, and Dong-Jo Kim. 2018. A Deep Learning Ap- proach to Anomaly Detection in Nuclear Reactors. InIEEE Nuclear Science Sym- posium and Medical Imaging Conference

2018
[20]

Sai Puppala, Ismail Hossain, Jahangir Alam, Kazuma Kobayashi, Samrendra Roy, Syed Bahauddin Alam, and Sajedul Talukder. 2025. Secure Anomaly Detection in Advanced Reactor Through Quantum Federated Learning with Advanced Cryptography. InNuclear Plant Instrumentation and Control & Human-Machine Interface Technology (NPIC&HMIT 2025). American Nuclear Society...

2025
[21]

Kaushik Roy, Akhilesh Jaiswal, and Priyadarshini Panda. 2019. Towards spike- based machine intelligence with neuromorphic computing.Nature575, 7784 (2019), 607–617

2019
[22]

Samrendra Roy, Kazuma Kobayashi, Souvik Chakraborty, Rizwan-uddin, and Syed Bahauddin Alam. 2026. Adversarial Vulnerabilities in Neural Operator Digital Twins: Gradient-Free Attacks on Nuclear Thermal-Hydraulic Surrogates. arXiv preprint arXiv:2603.22525(2026)

work page arXiv 2026
[23]

Hyeok-Ki Shin, Woomyo Lee, Seungoh Choi, Won-Seok Hwang, Jeong-Han Yun, Byung-Gil Min, and HyoungChun Kim. 2021. Two ICS Security Datasets and Anomaly Detection Contest on the HIL-based Augmented ICS Testbed. InCyber Security Experimentation and Test Workshop (CSET). ACM, 36–40

2021
[24]

Hyeok-Ki Shin, Woomyo Lee, Jeong-Han Yun, and HyoungChun Kim. 2020. HAI 1.0: HIL-based Augmented ICS Security Dataset. In13th USENIX Workshop on Cyber Security Experimentation and Test (CSET). USENIX Association

2020
[25]

Afzal, and Ángel Miguel García-Vico

Xabier Vázquez, Javier Sedano, M. Afzal, and Ángel Miguel García-Vico. 2025. Vacuum Spiker: A Spiking Neural Network-Based Model for Efficient Anomaly Detection in Time Series.arXiv preprint arXiv:2510.06910(2025)

work page arXiv 2025
[26]

Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. 2024. A Comprehensive Survey of Continual Learning: Theory, Method and Application.IEEE Transac- tions on Pattern Analysis and Machine Intelligence46, 8 (2024), 5362–5383

2024
[27]

Shuang Yi, Sheng Zheng, Senquan Yang, Guangrong Zhou, and Jiajun Cai. 2024. Anomaly Detection for Asynchronous Multivariate Time Series of Nuclear Power Plants Using a Temporal-Spatial Transformer.Sensors24, 9 (2024), 2845

2024
[28]

Friedemann Zenke, Ben Poole, and Surya Ganguli. 2017. Continual Learning Through Synaptic Intelligence. InInternational Conference on Machine Learning (ICML). PMLR, 3987–3995. 8

2017