Annals of the Academy of Romanian Scientists

Series on Engineering Sciences

ISSN 2066-6950

Volume 18, Number 1/2026

REAL-TIME EXTRACTION OF NEUROMORPHIC

FEATURES FOR NUCLEAR AND INDUSTRIAL SECURITY

Gabriel VASILESCU⁵, Maria DIMA⁶,

Denisa TUDOR⁷, Augustin SEMENESCU⁸

Rezumat. Monitorizarea instabilităților reactorilor nucleari în timpul operațiunilor de

pornire și de regim in sarcină reprezintă o provocare atât pentru securitatea nucleară

(controlul reactivității, marjele de oprire), cât și pentru securitatea industrială (detecția

anomaliilor în timp real, integritatea buclelor de control). Dintre tipurile de reactoare cele

mai vulnerabile la astfel de instabilități neprevăzute, reactoarele de cercetare pulsate și

reactoarele modulare mici (SMR) se află în prim-plan. Această lucrare stabilește un cadru

independent de tipul reactorului pentru extragerea în timp real a parametrilor exogeni

potriviți pentru predictorii neuromorfici Neliniari Autoregresivi Exogeni (NARX). Detaliez

in aceasta lucrare un filtru Savitzky-Golay (SG) recursiv în timp real care actualizează

coeficienții polinomiali de aproximativ 100 de ori mai rapid decât metoda tradițională.

Filtrul SG separă variațiile lente exogene (bare de control, lichid de răcire, presurizatoare)

de zgomotul neutronic stochastic. Datele sintetice de reactor SMR relevă ferestre de

variație a parametrilor, esențiale pentru agenții TAG (Tool-Augmented Generation) bazați

pe NARX.

Abstract. Monitoring nuclear reactor instabilities during start-up and load-following

operations challenges both nuclear security (reactivity control, shutdown margins) and

industrial security (real-time anomaly detection, control loop integrity). Of the types of

reactors most vulnerable to such unanticipated instabilities, pulsed research reactors and

Small Modular Reactors (SMR’s) are at the forefront. This paper establishes a reactor-

agnostic framework for real-time extraction of exogenous parameters suitable for

Nonlinear Autoregressive Exogenous (NARX) neuromorphic predictors. I detail a real-time

recursive Savitzky-Golay (SG) filter that updates polynomial coefficients ca. ×100 faster

than the traditional method. The SG filter decouples exogenous drifts (control rods,

coolant, pressurizer) from stochastic neutron noise. Synthetic SMR data shows parameter-

swing windows, essential for NARX-based Tool-Augmented Generation (TAG) agents.

Keywords: NARX, SMR, TAG, nuclear reactor instabilities, Savitzky-Golay filter

DOI 10.56082/annalsarscieng.2026.1.31

⁵Senior Researcher I, Habil. PhD, Eng., National Institute for Research and Development in Mine

Safety and Protection to Explosion, Chief Laboratory of Explosives Materials and Pyrotechnic

Articles

–

INCD

INSEMEX

Petrosani,

Romania

(e-mail:

dragos.vasilescu@insemex.ro ).

⁶Doctoral School - University of Petrosani, Petrosani, Romania (e-mail: mstsds@proton.me ).

⁷Doctoral School - University of Petrosani, Petrosani, Romania (e-mail: deni_t.2301@yahoo.com ).

⁸Professor, PhD, Eng. Mat. Ec., National Science and Technology University Politehnica Bucharest,

Bucharest,

Romania,

Member

Academy

Romanian

Scientists;

(e-mail:

augustin.semenescu@upb.ro ).

Gabriel Vasilescu, Maria Dima, Denisa Tudor, Augustin Semenescu

1. Introduction

Instability challenges across reactor classes - nuclear reactors, whether

pulsed research reactors, conventional power reactors, or Small Modular Reactors

(SMR’s), exhibit oscillational power fluctuations during start-up and ramp-up

phases, load-following operations, transition between natural and forced

circulation, reflector or control rod positioning transients.

Power production reactors typically exhibit small fluctuations (20 times

smaller than those of fast pulsed facilities), however their smaller class homologues,

SMR’s intentionally have tighter operational margins in order to maximize

economic competitivity. This renders SMR’s more susceptible to instability-

reactivities and control challenges.

Nuclear- and industrial-security context - from the instability perspective,

unanticipated power oscillations create nuclear security vulnerabilities:

─

reactivity margins: fluctuations reduce the effective margin for emer-

gency shutdown thresholds,

─

proactively,

─

control system stress: automatic regulators compensate reactively, not

redundancy degradation: repeated instability episodes stresses protec-

tion systems,

─

SMR-specific: reduced on-site staffing means automated predictive se-

curity becomes essential.

In turn, industrial security (focusing on cyber-physical security and control

system integrity) is susceptible to nuclear-instabilities, which feed into the its

industrial control systems (ICS) and SCADA platforms - as early warnings.

2. SMR particularities

SMR’s introduce unique challenges that make prediction of instabilities

particularly important:

─

tighter operational margins: designed for economic competitiveness →

less buffer against fluctuations,

─

reduced staffing: advanced automation → fewer human operators to

catch anomalies early,

─

load-following duty: many SMR designs anticipate load-following →

frequent power changes → more instabilities,

─

Passive safety reliance: while passive systems handle accidents, opera-

tional instabilities still challenge control systems

─

Digital I&C proliferation: SMRs rely heavily on digital instrumentation

and control → cyber-security and process security converge

Therefore, for AI control systems to capture the wealth of signals and

features, a unified method for quantifying those from raw data stream is needed —

to power future NARX-based TAG (Tool-Augmented Generation) agentic models.

Real-time extraction of neuromorphic features for nuclear and industrial security

For a comprehensive review of data handling in two-phase flow instabilities

in SMR steam generators see [6].

3. Real-time feature extraction

Aside from actuator equipment, the foremost signal of interest in nuclear

reactor security is neutron noise. To follow this parameter, it must first be

deconvoluted from the instantaneous power signal and its modulated rolling

baseline. For this I used a Savitzky-Golay (SG) filter [1] due to its robustness (no

iterative processes to depend on final initial guesses) and its speed (being an

analytical method).

To test the method, I generated a 25 S/s power ramp, with neutron noise

based on the frequency profile of the Fourier transform of the point reactor kinetics

equations. In this low-frequency region the reactor acts as a low-pass filter to

reactivity noise. This is because all designs are dominated in this infra-frequency

region by the decay constants of delayed neutron precursors. The longest-lived

precursor group typically has a decay constant which corresponds to this infra-

frequency range. I additionally added accidental instabilities.

For a window of 2K+1 bins (in this paper K=100 centered at the current bin),

the "zero-power" neutron noise will be the residual from a cubic polynomial fit -

where the coefficients model the impact of exogenous parameters on the rolling

baseline:

(

To establish the coefficients, I render their derivatives to nil:

(

Due to symmetric summation to the left and right of the current point, sums

of odd powers vanish, leading to the system of equations:

(

where:

Gabriel Vasilescu, Maria Dima, Denisa Tudor, Augustin Semenescu

(

Denoting λ4 = S4/S2, the solution to the system is:

Given the ramp-up feature, or in general unknown reactor power, I use the

dimensionless ratios B/D and C/D to decouple exogenous drifts from neutron noise.

Under stable operation these ratios have a negligible noise pattern. For an SMR-

like context however, they display detectable swings 30-70 s before peak power

excursion - a narrow but still quite actionable window for NARX prediction.

While for a single point the solution is fast, note that modern DAQ systems

acquire data from a variety of sensors at up to 100 kS/s. Therefore, the direct

solution becomes computationally intensive. For real-time monitoring, I designed

a recursive procedure that reuses calculations from the previous point. For a new

data point y_nextreplacing y_last, the updates are:

(

where Y_k' are the sums at the current point, expressed in terms of those from

the previous point. This translation is possible because moving one step forward

corresponds to updating y_i(i − 1)^qrelative to y_ii^q, which can be expressed using all

previously known power sums. Thus, for high-order Savitzky-Golay polynomials

and large 2K+1 context windows, rapid updates in real-time for pedestal monitoring

are possible for a large array of simultaneously tracked quantities. The goal of

exogenous influence extraction is to uncover the trends that move the baseline. As

such I perform a 2K+1 average of B/D and C/D to eliminate any local fluctuations.

Computational gain: direct SG is O_Kper sample, my recursive SG is just O_I.

Given a typical K=100, this translates into a factor ×100 faster — enabling real-

time deployment in a signal-rich SMR environment. Recent work [7] demonstrates

Real-time extraction of neuromorphic features for nuclear and industrial security

real-time Savitzky-Golay filtering on FPGA hardware at 192 Hz with only 260 μW

power consumption, confirming feasibility for embedded SG deployment.

Fig.1. Synthetic SMR power trace (25 S/s). Top: neutron noise (black - as kW power swing)

〈

〉

〈

〉

_±100averaged over

with SG trace (yellow - D coefficient). Middle two panels:

_±100and

201 samples. Note their swing shortly (20–70 s) prior to power excursion. Lower panel: "zero-

power" neutron noise - essentially no correlation to exogenous influences driving the

instabilities.

Conclusions

The recursive Savitzky-Golay filter extracts self-scaling parameters at an OI

computational complexity cost. The parameters used display swing windows 30-70

s before power excursion in synthetic SMR data. This is sufficient for a NARX

neuromorphic predictor to act as a perceptron for a TAG AI-security agent.

R E F E R E N C E S

[1]

A. Savitzky and M.J.E. Golay, "Smoothing and differentiation of data by simplified least

squares procedures," Anal. Chem., vol. 36, no. 8, pp. 1627–1639, 1964.

J.A. Nelder and R. Mead, "A simplex method for function minimization," Computer Journal,

vol. 7, pp. 308–313, 1965.

M.F. Thielbar and D.A. Dickey, "Neural networks for time series forecasting: Practical im-

plications of theoretical results," North Carolina State University, 2011.

IAEA, "Security of Small Modular Reactors," IAEA Nuclear Security Series, 2021.

NARXSim, Universitat Polytechnica de Catalunya, 2019.

[2]

[3]

[4]

[5]

[6]

H.H. Abdellatif, W. Ambrosini, D. Arcilesi, P.K. Bhowmik, and P. Sabharwall, "Flow insta-

bilities in boiling channels and their suppression methodologies - a review," Nuclear Engi-

neering and Design, vol. 421, 113114, May 2024.

[7]

O. Dutra, L.H.C. Ferreira, G.D. Colletta, and L.B. Zoccal, "A low power R-peak detector

clocked at signal sampling rate," Journal of Integrated Circuits and Systems, vol. 19, no. 1,

Mar. 2024.