pith. sign in

arxiv: 2606.05274 · v1 · pith:FNVYYMM5new · submitted 2026-06-03 · 💻 cs.LG

Anomaly Detection for Electro-Hydrostatic Actuators using LSTM Autoencoder

Nehal Afifi , Abdelmonem Elhendawi , Felix Leitenberger , Nadine Piat , Sven Matthiesen This is my paper

Pith reviewed 2026-06-28 06:56 UTC · model grok-4.3

classification 💻 cs.LG
keywords anomaly detectionLSTM autoencoderelectro-hydrostatic actuatorssensor signalsreconstruction errorfault injectionaerospace systemstime-series data
0
0 comments X

The pith

An LSTM autoencoder detects anomalies in electro-hydrostatic actuator sensors at 99 percent average accuracy.

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

The paper introduces an offline anomaly-detection framework for univariate sensor signals from electro-hydrostatic actuators that uses a reconstruction-based LSTM autoencoder. Traditional statistical and clustering methods miss the temporal structure in high-frequency EHA data and produce high false-alarm rates. The autoencoder is trained on normal sequences, thresholds are set from validation reconstruction errors, and performance is measured on injected faults in temperature and pressure channels under varying conditions. Reported results show average accuracy of 99.0 percent, precision reaching 100 percent, and F1-scores between 93.1 and 99.8 percent. These metrics indicate that data-driven reconstruction can separate normal from faulty behavior with low false alarms in safety-critical aerospace equipment.

Core claim

The LSTM autoencoder, trained to reconstruct normal univariate EHA sensor sequences and calibrated with thresholds from validation-set reconstruction-error distributions, detects injected anomalies in temperature and pressure signals with an average accuracy of 99.0 percent, precision up to 100 percent, recall between 90.2 and 99.6 percent, and F1-scores from 93.1 to 99.8 percent across multiple fault scenarios.

What carries the argument

Reconstruction-based LSTM autoencoder that learns normal temporal patterns in EHA sensor data and flags anomalies by elevated reconstruction error after calibration on validation distributions.

If this is right

  • The LSTM approach captures temporal dependencies that cause conventional methods such as Z-score, IQR, Isolation Forest, and k-means to underperform on EHA signals.
  • High precision and F1-scores imply sufficiently low false-alarm rates for practical monitoring of aerospace and industrial EHAs.
  • Sensitivity analyses under varying operating conditions support the claim that the framework remains effective beyond the specific test-bench settings used for calibration.
  • The offline framework establishes feasibility for data-driven anomaly detection on EHA temperature and pressure data before extension to real-time use.

Where Pith is reading between the lines

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

  • Extending the same reconstruction-error logic to multivariate sensor streams could improve isolation of specific fault types within the actuator.
  • Retraining or fine-tuning on data from different EHA hardware variants would test whether the learned temporal representations transfer without new threshold recalibration.
  • Integration with physics-based models of actuator dynamics might reduce the volume of labeled fault data needed for threshold setting.

Load-bearing premise

That reconstruction-error distributions computed on the validation set produce thresholds that reliably separate normal from faulty behavior under the injected fault scenarios and varying operating conditions described in the abstract.

What would settle it

A new collection of normal EHA sensor traces whose reconstruction errors systematically exceed the validation-derived thresholds would falsify the claim that those thresholds reliably separate normal from faulty operation.

Figures

Figures reproduced from arXiv: 2606.05274 by Abdelmonem Elhendawi, Felix Leitenberger, Nadine Piat, Nehal Afifi, Sven Matthiesen.

Figure 1
Figure 1. Figure 1: : LSTM autoencoder architecture with reconstruction for nominal temporal behaviour. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: : Design of the EHA used in the experimen [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: : Dataset preparation and evaluation protocol. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 6
Figure 6. Figure 6: : Sustained spike event: raw sensor signal [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: : Spike event example: raw sensor signal [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
read the original abstract

Electro-Hydrostatic Actuators (EHAs) are widely used in aerospace and industrial systems, where timely detection of sensor anomalies is essential to ensure safe and reliable operation. However, the large volume and high sampling frequency of EHA sensor data pose challenges for accurate and efficient anomaly detection. Conventional statistical and classical machine-learning methods such as Z-score, Interquartile Range (IQR), Median Absolute Deviation (MAD), Isolation Forest, Gaussian Mixture, and k-means often fail to capture the temporal dependencies inherent in EHA signals, resulting in limited detection accuracy and elevated false-alarm rates. Furthermore, systematic evaluations of data-driven anomaly detection approaches for EHA systems remain scarce, particularly under varying operational conditions. This study presents an offline anomaly-detection framework for univariate EHA sensor signals, focusing on temperature and pressure data collected from a controlled test bench. The method employs a reconstruction-based Long Short-Term Memory (LSTM) autoencoder, calibrated and evaluated using validation-set reconstruction-error distributions. Performance is assessed across multiple fault-injection scenarios using accuracy, precision, recall, and F1-score, complemented by sensitivity analyses under varying operating conditions. The LSTM autoencoder achieved an average accuracy of 99.0\%, precision up to 100\%, recall between 90.2\% and 99.6\%, and F1-scores from 93.1\% to 99.8\%, demonstrating high detection sensitivity and a very low false-alarm rate across all evaluated sensors. These results highlight the feasibility of data-driven offline anomaly detection for EHAs. Future work will focus on adapting the developed framework for an online (real-time) environment.

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

2 major / 0 minor

Summary. The manuscript presents an offline anomaly-detection framework for univariate EHA sensor signals (temperature and pressure) collected on a controlled test bench. It employs a reconstruction-based LSTM autoencoder whose thresholds are derived from validation-set reconstruction-error distributions, and reports strong empirical performance (average accuracy 99.0 %, precision up to 100 %, recall 90.2–99.6 %, F1 93.1–99.8 %) across multiple fault-injection scenarios while claiming superiority over classical statistical and ML baselines.

Significance. If the performance numbers are shown to be reproducible and robust to operating-condition variation, the work would provide a concrete, data-driven baseline for anomaly detection in safety-critical aerospace actuators—an area where systematic evaluations remain scarce. The use of real test-bench data and sensitivity analyses under varying conditions would be a positive contribution.

major comments (2)
  1. [Abstract] Abstract: the central performance claims (99.0 % accuracy, near-100 % precision, etc.) are stated without any information on network architecture (layers, hidden units), training details (optimizer, loss, epochs, batch size), exact threshold derivation from the validation reconstruction-error distribution, data volume, or train/validation/test partitioning. This renders the empirical results unverifiable from the supplied text.
  2. [Abstract] Abstract / evaluation description: no information is given on whether the validation set spans the same range of operating conditions (load, speed, temperature) as the fault-injection test cases, whether temporal leakage was avoided in the split, or whether metrics are from a single fixed partition or averaged over multiple runs. Consequently the reported metrics may be artifacts of the particular data partition rather than evidence of robust detection.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments highlighting the need for greater detail in the abstract. We agree that the original abstract did not provide sufficient information to verify the performance claims and have revised it to address both major comments.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central performance claims (99.0 % accuracy, near-100 % precision, etc.) are stated without any information on network architecture (layers, hidden units), training details (optimizer, loss, epochs, batch size), exact threshold derivation from the validation reconstruction-error distribution, data volume, or train/validation/test partitioning. This renders the empirical results unverifiable from the supplied text.

    Authors: We acknowledge that these implementation and evaluation details are absent from the abstract, which limits immediate verifiability. The full manuscript describes the LSTM autoencoder architecture, training procedure, threshold derivation from validation reconstruction errors, data volume, and partitioning in the Methods and Experiments sections. To resolve the concern, we have revised the abstract to include concise statements on these elements so that the central claims can be assessed from the abstract alone. revision: yes

  2. Referee: [Abstract] Abstract / evaluation description: no information is given on whether the validation set spans the same range of operating conditions (load, speed, temperature) as the fault-injection test cases, whether temporal leakage was avoided in the split, or whether metrics are from a single fixed partition or averaged over multiple runs. Consequently the reported metrics may be artifacts of the particular data partition rather than evidence of robust detection.

    Authors: The manuscript already reports sensitivity analyses under varying operating conditions and performance across multiple fault-injection scenarios. We agree the abstract should explicitly address coverage of operating conditions, avoidance of temporal leakage, and whether results reflect a single partition or multiple runs. We have revised the abstract to clarify that the validation set covers the same operating-condition range, time-based splits were employed, and metrics are aggregated across multiple scenarios and conditions. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical evaluation with no derivation chain

full rationale

The paper reports an empirical evaluation of an LSTM autoencoder on test-bench sensor data for anomaly detection. No equations, derivations, or first-principles predictions are present. Performance metrics (accuracy, precision, recall, F1) are computed directly from reconstruction-error thresholds on validation data and applied to fault-injection test cases. No self-citation load-bearing steps, fitted inputs renamed as predictions, or ansatzes smuggled via citation appear in the provided text. The central claims rest on experimental results rather than any reduction to inputs by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 0 invented entities

Only the abstract is available, so the ledger is necessarily incomplete. The framework depends on standard LSTM components and a validation-derived error threshold whose specific value and selection procedure are not stated.

free parameters (1)
  • reconstruction error threshold
    Calibrated from validation-set error distribution; exact value and selection rule not provided in abstract.

pith-pipeline@v0.9.1-grok · 5835 in / 1147 out tokens · 22316 ms · 2026-06-28T06:56:48.984185+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

24 extracted references

  1. [1]

    Chen, Z., C. K. Yeo, B. S. Lee, and C. T. Lau (2018). Autoencoder-based network anomaly detection. InWireless Telecommunications Symposium. D¨orr, M., F. Leitenberger, K. Wolter, S. Matthiesen, and T. Gwosch (2022). Model based control design of an eha position control based on multicriteria optimization.Machines

    2018

  2. [2]

    Ndubuaku, M

    Erhan, L., M. Ndubuaku, M. Di Mauro, W. Song, M. Chen, G. Fortino, O. Bagdasar, and A. Liotta (2021). Smart anomaly detection in sensor sys- tems: A multi-perspective review.Information Fusion 67, 64–79

    2021

  3. [3]

    Kitashov, and A

    Filonov, P., F. Kitashov, and A. Lavrentyev (2017). Rnn-based early cyber-attack detection for the tennessee eastman process

    2017

  4. [4]

    Schmidhuber, and F

    Gers, F., J. Schmidhuber, and F. Cummins (2000, 10). Learning to forget: Continual prediction with lstm.Neural Computation 12, 2451–2471. G¨unnemann, S., N. G¨unnemann, and C. Faloutsos (2014). Robust multivariate autoregression for anomaly detection in dynamic product ratings. InProceedings of the 23rd International World Wide Web Conference, pp. 361–372

    2000

  5. [5]

    Guo, T., X. Han, T. Minav, and Y . Fu (2022). A preliminary design method of high-power electro-hydrostatic actuators considering de- sign robustness.Actuators 11(11)

    2022

  6. [6]

    Jain, P., S. Jain, O. R. Za ¨ıane, and A. Srivas- tava (2022). Anomaly detection in resource constrained environments with streaming data. IEEE Transactions on Emerging Topics in Computational Intelligence 6(3), 649–659

    2022

  7. [7]

    A., J.-H

    Khan, M. A., J.-H. Jang, N. Iqbal, H. Jamil, S. S. A. Naqvi, S. Khan, J.-C. Kim, and D.-H. Kim (2025). Enhancing patient rehabilitation predictions with a hybrid anomaly detection model: Density-based clustering and interquar- tile range methods.CAAI Transactions on In- telligence Technology

    2025

  8. [8]

    Stolfo, P

    Lee, W., S. Stolfo, P. Chan, E. Eskin, W. Fan, M. Miller, S. Hershkop, and J. Zhang (2001). Real time data mining-based intrusion detec- tion. InProceedings DARPA Information Sur- vivability Conference and Exposition II

    2001

  9. [9]

    Leitenberger, F. and S. Matthiesen (2021). Archi- tecture of the digital twin in product validation for the application in virtual-physical testing to investigate system reliability. InDS 111: June 5, 2026 0:5 RPS ESREL2026 Proceedings/Edited Book: Trim Size: 221mm x 173mm esrel2026-paper 8Nehal Afifi and Abdelmonem Elhendawi Proceedings of the 32nd Sympo...

    2021

  10. [10]

    Li, S. and W. He (2019). Vidanomaly: Lstm- autoencoder-based adversarial learning for one- class video classification with multiple dynamic images. InProceedings of the IEEE Interna- tional Conference on Big Data (Big Data)

    2019

  11. [11]

    Matthews, and I

    Memarzadeh, M., B. Matthews, and I. Avrekh (2020). Unsupervised anomaly detection in flight data using convolutional variational auto- encoder.Aerospace 7(8)

    2020

  12. [12]

    Munir, M., S. A. Siddiqui, M. A. Chattha, A. Den- gel, and S. Ahmed (2019). Fusead: Unsu- pervised anomaly detection in streaming sen- sors data by fusing statistical and deep learning models.Sensors 19(11)

    2019

  13. [13]

    Nguyen, H., K. Tran, S. Thomassey, and M. Hamad (2021). Forecasting and anomaly detection approaches using lstm and lstm au- toencoder techniques with applications in sup- ply chain management.International Journal of Information Management 57, 102282

    2021

  14. [14]

    Zafar, Z

    Nizam, H., S. Zafar, Z. Lv, F. Wang, and X. Hu (2022). Real-time deep anomaly detection framework for multivariate time-series data in industrial iot.IEEE Sensors Journal

    2022

  15. [15]

    Kay, and L

    Paffenroth, R., K. Kay, and L. Servi (2018). Ro- bust pca for anomaly detection in cyber net- works.arXiv

    2018

  16. [16]

    Qiao, G., G. Liu, Z. Shi, Y . Wang, S. Ma, and T. Lim (2017, 12). A review of electromechan- ical actuators for more/all electric aircraft sys- tems.Proceedings of the Institution of Mechan- ical Engineers, Part C: Journal of Mechanical Engineering Science 232, 095440621774986

    2017

  17. [17]

    Leckie, M

    Rajasegarar, S., C. Leckie, M. Palaniswami, and J. C. Bezdek (2006). Distributed anomaly de- tection in wireless sensor networks. In2006 10th IEEE Singapore International Conference on Communication Systems, pp. 1–5

    2006

  18. [18]

    Pokharel, and B

    Sharma, B., P. Pokharel, and B. Joshi (2020). User behavior analytics for anomaly detection using lstm autoencoder – insider threat detection. In Proceedings of the 11th International Confer- ence on Advances in Information Technology (IAIT ’20), Bangkok, Thailand. Association for Computing Machinery

    2020

  19. [19]

    Siegel, B. (2020). Industrial anomaly detection: A comparison of unsupervised neural network architectures.IEEE Sensors Letters 4(8)

    2020

  20. [20]

    Tax, D. M. J. and R. P. W. Duin (2004). Support vector data description.Machine Learning

    2004

  21. [21]

    Tran, P. H., C. Heuchenne, and S. Thomassey (2020). An anomaly detection approach based on the combination of lstm autoencoder and isolation forest for multivariate time series data. InDevelopments of Artificial Intelligence Tech- nologies in Computation and Robotics

    2020

  22. [22]

    Wang, X. R., J. T. Lizier, O. Obst, M. Prokopenko, and P. Wang (2008). Spatiotemporal anomaly detection in gas monitoring sensor networks. In R. Verdone (Ed.),Wireless Sensor Networks,

    2008

  23. [23]

    Zhang, J

    Yin, C., S. Zhang, J. Wang, and N. N. Xiong (2022). Anomaly detection based on convolu- tional recurrent autoencoder for iot time series. IEEE Transactions on Systems, Man, and Cy- bernetics: Systems

    2022

  24. [24]

    Zong, B., Q. Song, M. R. Min, W. Cheng, C. Lumezanu, D.-k. Cho, and H. Chen (2018). Deep autoencoding gaussian mixture model for unsupervised anomaly detection.International Conference on Learning Representations

    2018