Neuromorphic Continual Learning for Sequential Deployment of Nuclear Plant Monitoring Systems

Samrendra Roy; Sajedul Talukder; Syed Bahauddin Alam

cs.NEApr 13, 2026

Neuromorphic Continual Learning for Sequential Deployment of Nuclear Plant Monitoring Systems

Authors: Samrendra Roy, Sajedul Talukder, Syed Bahauddin Alam

Organizations: University of Illinois Urbana-Champaign Department of Nuclear, Plasma & Radiological Engineering Urbana, IL, USA · University of Texas at El Paso Department of Computer Science El Paso, TX, USA · National Center for Supercomputing Applications Urbana, IL, USA

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.

Figures & tables

Panel of Figure 1 — Figure 1. Architecture overview. Heterogeneous sensors from four nuclear ICS subsystems are converted to sparse spike trains via delta encoding at rates reflecting each sensor’s natural dynamics (annotated on each connection). The SNN processes sliding windows through two PLIF spiking layers with learnable time constants. During sequential task deployment, the continual learning module (dashed red arrows) protects weights important for previously learned subsystems. Block diagram of the SNN anomaly detection pipeline: sensors to spike encoding to two PLIF spiking layers with continual learning module below.

Subsystem	Sensors	Spike Rate	Dynamics
P1 Boiler	38	0.005	Slow thermal
P2 Turbine	22	0.220	Fast mechanical
P3 Water Treat.	7	0.002	Very slow chemical
P4 HIL Sim.	12	0.061	Mixed simulation
Overall	79	0.073	92.7% sparsity

Table 1. Asynchronous spike rates by subsystem, computed via delta encoding (

\alpha=0.5

) on training data. The 100

\times

variation between P2 (turbine) and P3 (water treatment) reflects fundamentally different sensor dynamics.

Method	P1	P2	P3	Avg F1	AF
Sequential	0.821	0.901	1.000	.916 $\pm$ .002	.113 $\pm$ .016
EWC	0.814	0.916	1.000	0.910	0.135
SI	0.987	0.936	0.795	0.906	0.005
Replay	0.982	0.991	0.897	0.957	0.009
EWC+Replay	0.998	0.977	0.962	.949 $\pm$ .028	.035 $\pm$ .039
Joint (oracle)	0.711	0.826	0.822	0.786	0.000

Table 2. F1 scores after sequential training on all three subsystems. Per-task values are from seed = 42; Avg F1 and AF for Sequential and EWC+Replay show mean

\pm

std across seeds

\{42,123,456\}

Figure 2. F1 score evolution during sequential task training. (a) P1 Boiler: Sequential and EWC degrade from 1.0 to

\sim

0.82, while Replay and EWC+Replay maintain F1

>

0.98. (b) P2 Turbine: moderate forgetting under Sequential; other methods retain performance. (c) P3 is the final task, so no forgetting occurs, but SI achieves the lowest F1 (0.795), indicating that excessive regularization from prior tasks constrains new-task plasticity. Three line plots showing F1 score vs training task number for five CL methods across P1 Boiler, P2 Turbine, and P3 Water Treatment subsystems. Sequential and EWC drop on P1 while Replay and EWC+Replay remain high.

Figure 3. Left: Average F1 across all three subsystems. EWC+Replay (0.979) approaches ANN per-task accuracy (0.999) while Joint training (0.786) is the worst. Right: Average forgetting. Only EWC+Replay achieves near-zero forgetting; EWC and Sequential are nearly identical, confirming that regularization alone is ineffective in the spiking regime. Two bar charts. Left shows average F1 scores for seven methods, with EWC+Replay at 0.979 and Joint lowest at 0.786. Right shows average forgetting, with Sequential and EWC highest and EWC+Replay at zero.

Metric	ANN	SNN	Ratio
Ops/inference	535,040	42,433	12.6 $\times$
	(MACs)	(SOPs)
Input sparsity	0%	92.7%	—
Avg F1 (per-task)	0.999	0.979	2.0% gap

Table 3. Energy comparison between ANN and SNN anomaly detectors. The SNN achieves a 12.6

\times

reduction in operations due to spike-driven sparse computation.

Metric	Value
Attacks tested	25
Attacks detected	25/25 (100%)
Mean delay	0.6 s
Median delay	0.0 s
Max delay	16 s
Detected $\leq$ 10 s	24/25 (96%)

Table 4. Detection latency analysis using the EWC+Replay model. All 25 tested attacks were detected, with 96% identified within 10 seconds of attack onset.

Input	P1 Boiler	P2 Turbine	P3 Water
Standard	0.999	1.000	1.000
Spike-encoded	0.995	0.993	1.000
$\Delta$	$-$ 0.005	$-$ 0.007	0.000

Table 5. Per-task F1 with standard vs. spike-encoded input. Spike encoding incurs less than 1% accuracy cost while achieving 92.7% input sparsity.

$\alpha$	Sparsity	Rate	Avg F1	P1	P2	P3
0.10	85.4%	0.146	0.991	0.982	0.994	0.996
0.25	89.0%	0.110	0.991	0.995	0.984	0.993
0.50	92.7%	0.073	0.983	0.971	0.988	0.991
0.75	95.1%	0.049	0.965	0.931	0.985	0.979
1.00	96.9%	0.031	0.971	0.947	0.984	0.984

Table 6. Spike encoding threshold sensitivity. The threshold factor

\alpha

controls sparsity: higher

\alpha

produces sparser inputs but risks discarding attack-relevant sensor changes. F1 values differ from Table 5 because this ablation uses polarity-weighted spike input rather than the two-channel encoding of Table 5 .

Figure 4. Replay buffer size sensitivity. Left: Average F1 across all tasks remains above 0.92 for all tested buffer sizes. Right: Per-subsystem F1 shows that P1 Boiler (the earliest task, most susceptible to forgetting) benefits most from larger buffers. Two line plots showing average F1 score and per-subsystem F1 versus replay buffer size from 50 to 1000. Performance is robust, staying above 0.92 across all sizes.

Figure 5. Asynchronous spike rates by subsystem. The 100

\times

variation between P2 (turbine) and P3 (water treatment) reflects physical sensor dynamics and demonstrates natural alignment with neuromorphic event-driven processing. Bar chart showing average spike rates for four subsystems: P1 Boiler at 0.005, P2 Turbine at 0.220, P3 Water Treatment at 0.002, and P4 HIL Simulation at 0.061.

Industrial Control System Security

See all 17 papers

Apr 14, 2026

Gradient-Free Continual Learning in Spiking Neural Networks via Inter-Spike Interval Regularization

Continual learning, the ability to acquire new tasks sequentially without forgetting prior knowledge, is essential for deploying neural networks in dynamic real-world environments, from nuclear digital twin monitoring to grid-edge fault detection. Existing synaptic importance methods, such as Elastic Weight Consolidation (EWC) and Synaptic Intelligence (SI), rely on gradient computation, making them incompatible with neuromorphic hardware that lacks backpropagation support. We propose ISI-CV, the first gradient-free synaptic importance metric for SNN continual learning, derived from the Coefficient of Variation (CV) of Inter-Spike Intervals (ISIs). Neurons that fire regularly (low CV) encode stable, task-relevant features and are protected from overwriting; neurons with irregular firing are permitted to adapt freely. ISI-CV requires only spike time counters and integer arithmetic, all of which are native to every neuromorphic chip. We evaluate on four benchmarks of increasing difficulty: Split-MNIST, Permuted-MNIST, Split-FashionMNIST, and Split-N-MNIST using real Dynamic Vision Sensor (DVS) event data. Across three seeds, ISI-CV achieves zero forgetting (AF = 0.000 +/- 0.000) on Split-MNIST and Split-FashionMNIST, near-zero forgetting on Permuted-MNIST (AF = 0.001 +/- 0.000), and the highest accuracy with the lowest forgetting on real neuromorphic DVS data (AA = 0.820 +/- 0.012, AF = 0.221 +/- 0.014). On N-MNIST, gradient-based methods produce unreliable importance estimates and perform worse than no regularization; ISI-CV avoids this failure by design.

Samrendra Roy, Kazuma Kobayashi, Souvik Chakraborty +2

University of Illinois Urbana-Champaign, Department of Nuclear, Plasma & Radiological Engineering, Urbana, IL, USA · Indian Institute of Technology Delhi, Department of Applied Mechanics, New Delhi, India · University of Texas at El Paso, Department of Computer Science, El Paso, TX, USA +1

Spiking Neural Networks Catastrophic Forgetting Importance Weighting Neuromorphic Computing

Ask PDF

Mar 11, 2026

Spatio-Temporal Attention Graph Neural Network: Explaining Causalities With Attention

Industrial Control Systems (ICS) underpin critical infrastructure and face growing cyber-physical threats due to the convergence of operational technology and networked environments. While machine learning-based anomaly detection approaches in ICS shows strong theoretical performance, deployment is often limited by poor explainability, high false-positive rates, and sensitivity to evolving system behavior, i.e., baseline drifting. We propose a Spatio-Temporal Attention Graph Neural Network (STA-GNN) for unsupervised and explainable anomaly detection in ICS that models both temporal dynamics and relational structure of the system. Sensors, controllers, and network entities are represented as nodes in a dynamically learned graph, enabling the model to capture inter-dependencies across physical processes and communication patterns. Attention mechanisms provide influential relationships, supporting inspection of correlations and potential causal pathways behind detected events. The approach supports multiple data modalities, including SCADA point measurements, network flow features, and payload features, and thus enables unified cyber-physical analysis. To address operational requirements, we incorporate a conformal prediction strategy to control false alarm rates and monitor performance degradation under drifting of the environment. Our findings highlight the possibilities and limitations of model evaluation and common pitfalls in anomaly detection in ICS. Our findings emphasise the importance of explainable, drift-aware evaluation for reliable deployment of learning-based security monitoring systems.

Kosti Koistinen, Kirsi Hellsten, Joni Herttuainen +1

Aalto University School of Science Computer Science Department P.O.Box 11000, 00076 AALTO, Finland

Industrial Anomaly Detection Industrial Control System Security Interpretable Anomaly Detection Unsupervised Anomaly Detection Cyber-Physical Systems Spatiotemporal GNNs

Ask PDF

Apr 29, 2026

Neuromorphic Graph Anomaly Detection via Adaptive STDP and Spiking Graph Neural Networks

Anomaly detection in dynamic networks is critical for applications from cybersecurity to industrial monitoring, yet existing methods face challenges in energy efficiency, temporal precision, and adaptability. This paper introduces ASTDP-GAD, a novel Adaptive Spiking Temporal Dynamics Plasticity framework for Graph Anomaly Detection that integrates spiking graph neural networks with STDP learning for energy-efficient neuromorphic detection in dynamic networks. Our framework unifies spiking neural computation, STDP learning, and graph-based anomaly detection through the following key innovations: temporal spike graph encoding with adaptive Leaky Integrate-and-Fire (LIF) dynamics; LIF-based graph attention with lateral inhibition; event-driven hypergraph memory with STDP-inspired prototype updates; spike rate contrast pooling based on spiking irregularity; adaptive STDP layers capturing causal temporal relationships; and multi-scale temporal convolution with multi-factor anomaly fusion. Theoretical analysis provides rigorous guarantees: spike encoding preserves input information with resolution scaling linearly in simulation steps and hidden dimension; LIFGAT approximates any continuous attention function; hypergraph memory converges to optimal prototypes; contrast pooling achieves provable anomaly selection bounds; STDP learning converges stably; and multi-factor fusion produces calibrated scores with up to

5\times

variance reduction. Extensive experiments on nine datasets on both dynamic and static graphs demonstrate superior anomaly detection accuracy while maintaining biological plausibility and energy efficiency for neuromorphic deployment.

Abdul Joseph Fofanah, Lian Wen, David Chen +2

School of ICT, Griffith University, Australia

Graph Neural Networks Spiking Neural Networks Spike-Timing-Dependent Plasticity Neuromorphic Computing Anomaly Detection

Ask PDF

Abstract

Industrial Control System Security

Explore similar work

Gradient-Free Continual Learning in Spiking Neural Networks via Inter-Spike Interval Regularization

Spatio-Temporal Attention Graph Neural Network: Explaining Causalities With Attention

Neuromorphic Graph Anomaly Detection via Adaptive STDP and Spiking Graph Neural Networks