pith. sign in

arxiv: 2604.11631 · v1 · submitted 2026-04-13 · 📡 eess.SY · cs.SY

Detectability of Subtle Anomalies in Dynamical Systems via Log-Likelihood Ratio

Alejandro Penacho Riveiros , Matthieu Barreau , Nicola Bastianello This is my paper

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

classification 📡 eess.SY cs.SY
keywords anomaly detectionlog-likelihood ratiolinear systemsGaussian noiseerror rateobserver designreal-time monitoring
0
0 comments X

The pith

The log-likelihood ratio detector admits a theoretical error-rate characterization for linear Gaussian systems.

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

The paper develops a real-time anomaly detector using the log-likelihood ratio for systems with multiple known plant models corresponding to different anomalies. It provides a theoretical characterization of the detector's error rate specifically for linear Gaussian systems. This characterization allows prediction of how detectable different anomalies are and can be used to improve observer design in industrial control. A sympathetic reader would care because accurate and quick anomaly detection prevents costly downtime in applications like manufacturing or power systems. The work bridges the gap between the practical use of likelihood methods and their formal analysis.

Core claim

We investigate a real-time anomaly detector based on the log-likelihood ratio and provide a theoretical characterization of its error rate when it is applied to linear Gaussian systems. We showcase the performance of this algorithm and the characterization obtained, and demonstrate how the latter can be leveraged for observer design.

What carries the argument

The log-likelihood ratio test, which compares the likelihoods of observed data under different known anomaly models to decide which anomaly has occurred.

If this is right

  • The error rate of the detector can be calculated theoretically for any set of linear Gaussian models.
  • This allows designers to assess in advance whether subtle anomalies will be detectable quickly enough.
  • The characterization can guide the selection or design of observers to minimize detection errors.
  • Performance can be showcased through simulations on linear systems.

Where Pith is reading between the lines

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

  • The approach might be approximated for mildly nonlinear systems using linearization techniques.
  • It could inform the design of adaptive anomaly detection systems that update models online.
  • Connections to change-point detection in statistics could be explored for faster implementations.

Load-bearing premise

The anomaly models are known beforehand and the system dynamics are exactly linear with additive Gaussian noise.

What would settle it

Running the detector on simulated or real linear Gaussian data with known anomalies and comparing the observed false alarm or miss rates to the predicted characterization; significant mismatch would falsify the theory.

Figures

Figures reproduced from arXiv: 2604.11631 by Alejandro Penacho Riveiros, Matthieu Barreau, Nicola Bastianello.

Figure 1
Figure 1. Figure 1: Upper: 2000 trajectories of the cumulative log￾likelihood for the stabilized inverted double pendulums. Lower: empirical and theoretical error rates, with shaded confidence bounds. decreases from ∼ 50% (α and β are equally accurate), reaching below 5% as more measurements are collected. This indicates that the algorithm successfully selects β as the better approximation of γ. As discussed in section IV-B.2… view at source ↗
Figure 3
Figure 3. Figure 3: Experimental (solid line) and theoretical (dashed line) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: Contours of tr(Σy ∗) and Λ for different gains of the observer L. Red areas show unstable observer configuration. Observer design: The goal now is to leverage the log-likelihood ratio to select an observer that maximizes detectability. To this end, for a range of possible observer gains L = [L1, L2], we compute: 1) matrix Λ characterizing detectability, and 2) tr(Σy ), which represents the average squared … view at source ↗
read the original abstract

Industrial control applications require detecting system anomalies as accurately and quickly as possible to enable prompt maintenance. In this context, it is common to consider several possible plant models, each linked to a different anomaly. The log-likelihood ratio method can then be used to identify the most accurate model and thereby classify which anomaly, if any, has occurred. Although the method has been applied to a wide variety of systems, there is no formal analysis of what makes anomalies more or less prone to detection. In this paper, we investigate a real-time anomaly detector based on the log-likelihood ratio and provide a theoretical characterization of its error rate when it is applied to linear Gaussian systems. We showcase the performance of this algorithm and the characterization obtained, and demonstrate how the latter can be leveraged for observer design.

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.

Referee Report

0 major / 3 minor

Summary. The manuscript develops a real-time anomaly detector based on the log-likelihood ratio test for linear Gaussian dynamical systems in which the plant models associated with each anomaly hypothesis are known in advance. It derives a theoretical characterization of the detector's error rate under these conditions, presents numerical performance results, and shows how the characterization can be used to guide observer design.

Significance. If the error-rate characterization holds under the stated assumptions, the work supplies a concrete tool for quantifying detectability of subtle anomalies and for synthesizing observers that exploit this information. This is a focused, practically relevant contribution within the linear-Gaussian setting common to many industrial control applications.

minor comments (3)
  1. §3, Eq. (8): the transition from the continuous-time likelihood ratio to the discrete-time recursion is stated without an explicit discretization step or sampling-time dependence; adding one sentence clarifying the Euler or exact discretization used would improve reproducibility.
  2. Figure 4: the caption does not indicate the number of Monte-Carlo trials or the confidence intervals shown; please add this information so readers can assess the statistical significance of the reported error-rate curves.
  3. §5.2: the observer-design example uses a specific cost function derived from the error-rate bound, but the mapping from the bound to the LQR weights is only sketched; a short appendix deriving the weights explicitly would strengthen the claim that the characterization is directly leveraged for design.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of our work on log-likelihood ratio anomaly detection for linear Gaussian dynamical systems and for recommending minor revision. No specific major comments were provided in the report, so we have no individual points to address. We are prepared to incorporate any minor editorial or presentation improvements identified during the revision process.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper presents a theoretical error-rate characterization for the log-likelihood ratio detector on linear Gaussian systems under the assumption of known plant models per anomaly hypothesis. No equations, derivations, or self-citations are provided in the available text that reduce any claimed prediction or result to a fitted input, self-definition, or prior author work by construction. The central claim remains a direct analysis within the stated linear-Gaussian setting and is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no free parameters, axioms, or invented entities are described.

pith-pipeline@v0.9.0 · 5437 in / 1195 out tokens · 60740 ms · 2026-05-10T15:05:57.824950+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages

  1. [1]

    Review of recent applications and future perspectives on process monitoring approaches in industrial processes,

    S. Li, B. Zhou, J. Shang, X. Chen, and J. Yu, “Review of recent applications and future perspectives on process monitoring approaches in industrial processes,”Journal of Manufacturing Systems, vol. 82, pp. 509–530, Oct. 2025

    work page 2025

  2. [2]

    Industrial Process Monitoring in the Big Data/Industry 4.0 Era: from Detection, to Diagnosis, to Prognosis,

    M. S. Reis and G. Gins, “Industrial Process Monitoring in the Big Data/Industry 4.0 Era: from Detection, to Diagnosis, to Prognosis,” Processes, vol. 5, no. 3, p. 35, Sep. 2017

    work page 2017

  3. [3]

    Validating control of extreme disturbance of an organic Rankine cycle using VMGsim,

    “Validating control of extreme disturbance of an organic Rankine cycle using VMGsim,” inComputer Aided Chemical Engineering. Elsevier, Jan. 2018, vol. 44, pp. 745–750

    work page 2018

  4. [4]

    S. X. Ding,Model-Based Fault Diagnosis Techniques: Design Schemes, Algorithms and Tools, ser. Advances in Industrial Control. London: Springer, 2013

    work page 2013

  5. [5]

    Anticipation, earliness, alarm cardinality: A new metric for industrial time-series anomaly detection,

    R. Dion, M. Alamir, and T. Le Magueresse, “Anticipation, earliness, alarm cardinality: A new metric for industrial time-series anomaly detection,”IFAC-PapersOnLine, vol. 58, no. 4, pp. 192–197, Jan. 2024

    work page 2024

  6. [6]

    An incipient fault diagnosis methodology using local Mahalanobis distance: Detection process based on empir- ical probability density estimation,

    J. Yang and C. Delpha, “An incipient fault diagnosis methodology using local Mahalanobis distance: Detection process based on empir- ical probability density estimation,”Signal Processing, vol. 190, p. 108308, Jan. 2022

    work page 2022

  7. [7]

    Grant and R

    E. Grant and R. Leavenworth,Statistical Quality Control. New York, N.Y London: McGraw-Hill Science/Engineering/Math, 1996

    work page 1996

  8. [8]

    Fault diagnosis in dynamic systems using analytical and knowledge-based redundancy: A survey and some new results,

    P. M. Frank, “Fault diagnosis in dynamic systems using analytical and knowledge-based redundancy: A survey and some new results,” Automatica, vol. 26, no. 3, pp. 459–474, May 1990

    work page 1990

  9. [9]

    Isermann,Fault-Diagnosis Applications: Model-Based Condition Monitoring: Actuators, Drives, Machinery, Plants, Sensors, and Fault- tolerant Systems

    R. Isermann,Fault-Diagnosis Applications: Model-Based Condition Monitoring: Actuators, Drives, Machinery, Plants, Sensors, and Fault- tolerant Systems. Berlin, Heidelberg: Springer, 2011

    work page 2011

  10. [10]

    Fault Detection and Isolation in Industrial Processes Using Deep Learning Approaches,

    R. Iqbal, T. Maniak, F. Doctor, and C. Karyotis, “Fault Detection and Isolation in Industrial Processes Using Deep Learning Approaches,” IEEE Transactions on Industrial Informatics, vol. 15, no. 5, pp. 3077– 3084, May 2019

    work page 2019

  11. [11]

    An inte- grated monitoring system based on deep learning tools for industrial process*,

    A. Rodr ´ıguez-Ramos, C. Verde, and O. Llanes-Santiago, “An inte- grated monitoring system based on deep learning tools for industrial process*,”IFAC-PapersOnLine, vol. 58, no. 4, pp. 168–173, Jan. 2024

    work page 2024

  12. [12]

    Shalev-Shwartz,Understanding Machine Learning, 1st ed

    S. Shalev-Shwartz,Understanding Machine Learning, 1st ed. New York: Cambridge University Press, Jul. 2014

    work page 2014

  13. [13]

    Basseville and I

    M. Basseville and I. V . Nikiforov,Detection of Abrupt Changes: Theory and Application. Englewood Cliffs, N.J: Prentice Hall, 1993

    work page 1993

  14. [14]

    Gustafsson,Adaptive Filtering and Change Detection, 1st ed

    F. Gustafsson,Adaptive Filtering and Change Detection, 1st ed. Chichester New York: Wiley, Sep. 2000

    work page 2000

  15. [15]

    Sequential Tests of Statistical Hypotheses,

    A. Wald, “Sequential Tests of Statistical Hypotheses,”The Annals of Mathematical Statistics, vol. 16, no. 2, pp. 117–186, Jun. 1945

    work page 1945

  16. [16]

    Sequential Generalized Likelihood Ratio Tests for Online Item Monitoring,

    H.-A. Kang, “Sequential Generalized Likelihood Ratio Tests for Online Item Monitoring,”Psychometrika, vol. 88, no. 2, pp. 672–696, Jun. 2023

    work page 2023

  17. [17]

    T. M. Cover and J. A. Thomas,Elements of Information Theory 2nd Edition, 2nd ed. Hoboken, N.J: Wiley-Interscience, Jul. 2006

    work page 2006