Pith. sign in

REVIEW 3 major objections 6 minor 24 references

Federated Learning with Reservoir State Analysis for Time Series Anomaly Detection

T0 review · 3 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper establishes that federated time series anomaly detection can be made mathematically equivalent to centralized MD-RS by having each client share only the covariance matrix of its reservoir states.

desk verdict Useful and simple federated anomaly detection method, but the central exact-equivalence claim is overstated without stated initialization conditions. read the letter →

arxiv 2502.05679 v2 pith:CKNLS4RM submitted 2025-02-08 cs.LG

Keigo Nogami , Hiroto Tamura , Gouhei Tanaka This is my paper · ORCID
classification cs.LG
keywords federatedlearningreservoircomputingtimeseriesanomalydetectionMahalanobisdistanceechostatenetworkcovarianceaggregationsubsamplingprivacy-preservingmachine
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

IncFed MD-RS is a federated learning scheme for time series anomaly detection built on reservoir computing. Instead of training deep networks locally and averaging their weights, each client runs the same fixed random reservoir on its normal training data and sends only the covariance matrix of the resulting reservoir states to the server. The server sums these matrices and inverts the sum to obtain a global precision matrix, which the paper argues is mathematically equivalent to the precision matrix of a centralized MD-RS model trained on all data at once. Anomaly scores are then Mahalanobis distances from new reservoir states to the learned normal-state distribution. The paper reports that this scheme matches centralized MD-RS performance and outperforms federated deep learning and other reservoir baselines on the SMD, SMAP, and PSM benchmarks, while using less communication than existing federated ESN methods.

What carries the argument

The load-bearing object is the precision matrix of reservoir states, obtained by inverting the regularized sum of covariance matrices. The reservoir is an echo state network with random input and recurrent weights shared by all clients, so its state update $x(t) = (1-\alpha)x(t-1) + \alpha\tanh(W_{\text{in}}u(t) + Wx(t-1))$ maps each input into a high-dimensional nonlinear feature space. MD-RS fits a zero-mean multivariate Gaussian to these states, and the Mahalanobis distance $x(t)^\top P_g x(t)$ is the anomaly score. The incremental update of the precision matrix via the Woodbury identity, $P(t+1) = P(t) - P(t)x(t)x(t)^\top P(t)/(1+x(t)^\top P(t)x(t))$, lets clients maintain local precision matrices online, while the server-side sum of covariance matrices gives the centralized-equivalent global model. Subsampling a random subset of reservoir nodes reduces both computation and the communication payload from $N_x^2$ to $\tilde{N}_x^2$ values per client.

What would settle it

Take one long normal time series, split it into $C$ contiguous blocks, have each client compute $\Phi_c$ with reservoir initial state $x(0)=0$, and compute $P_g = (\sum_c \Phi_c + \delta I)^{-1}$. Compare $P_g$ to the precision matrix obtained by running the same reservoir on the full series; a nonzero Frobenius-norm difference, or different Mahalanobis scores at test points near block boundaries, would show the claimed mathematical equivalence is exact only after controlling for reservoir initialization.

Watch

Extended reading notes

Core claim

The central claim is that a federated anomaly detector can be assembled losslessly from additive second-order statistics. Each client computes $\Phi_c = \sum_t x_c(t)x_c(t)^\top$ over its reservoir states for normal inputs; the server forms $\Phi_g = \sum_c \Phi_c$ and $P_g = (\Phi_g + \delta I)^{-1}$; and the squared anomaly score for a test state is $x(t)^\top P_g x(t)$. Because $\Phi_g$ is the sum of per-client covariance matrices, the paper asserts that $P_g$ is mathematically identical to the precision matrix of the centralized MD-RS method that would be computed if all data were pooled. In experiments on SMD, SMAP, and PSM with 24 clients, IncFed MD-RS produces the same evaluation scores as centralized MD-RS and surpasses FedAvg versions of TranAD, LSTM-AE, and ESN-based reconstruction, with the advantage largest when client data are short and heterogeneous.

Load-bearing premise

The equivalence to centralized MD-RS assumes each client's reservoir states are exactly the states a single model would compute on the pooled data; when one long time series is split across clients and each starts from zero, the state trajectories differ at segment boundaries, so the aggregated covariance is only approximately the centralized one.

Editorial extensions

If this is right

  • In the batch setting, the global anomaly detection model is assembled in one aggregation step after clients send their covariance matrices; no iterative gradient exchange is required.
  • The method's performance is independent of the number of clients, because aggregation is an exact sum rather than a model average, matching the paper's result that IncFed MD-RS and IncFed ESN-SRE do not degrade as the client count grows.
  • Communication cost scales as $O(N_x^2)$ per client, or $O(\tilde{N}_x^2)$ with subsampling, which is far smaller than transmitting full deep-network parameters.
  • Clients do not need deep learning hardware or backpropagation; training is a matrix accumulation over reservoir states, so the approach suits devices with limited computation.
  • Because the federated model is mathematically equivalent to the centralized model, its anomaly detection accuracy does not suffer from client-side approximation, unlike FedAvg-based federated deep models.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The exact-equivalence claim implicitly assumes each client's reservoir trajectory is the one a centralized model would produce. If a single long series is split contiguously and every client starts from $x(0)=0$, the first several states carry zero history from preceding segments, so $P_g$ is only approximately the centralized precision matrix; discarding initial transients or passing final states
  • The same sufficient-statistics aggregation could extend to other density-based anomaly detectors: any client-computable statistic that is additive over data, such as sufficient statistics for a Gaussian mixture, would yield a centralized-equivalent federated model, so the principle is not specific to Mahalanobis distance.
  • Sharing covariance matrices still reveals distributional information about each client's data, so formal privacy guarantees such as differential privacy are not automatic; adding calibrated noise to $\Phi_c$ would break exact equivalence, and the trade-off is not addressed in the paper.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes Incremental Federated Learning with Mahalanobis Distance of Reservoir States (IncFed MD-RS), a federated anomaly-detection method in which each client computes the covariance matrix of reservoir states for its local normal training data and sends only that matrix to a server. The server sums the local covariance matrices and computes a global precision matrix, which the authors claim is mathematically equivalent to the precision matrix of a centralized MD-RS model. The paper evaluates the method on SMD, SMAP, and PSM using AUC-ROC, AUC-PR, VUS-PR, and PATE, and reports that IncFed MD-RS outperforms FedAvg variants of TranAD, LSTM-AE, ESN-SRE, and MD-RS, while also showing robustness to short training data and to the number of clients. Source code is publicly available.

Significance. If the central equivalence claim holds, the method is a clean and communication-efficient way to federate a reservoir-computing anomaly detector: the aggregation step is a parameter-free linear sum of local covariance matrices, it supports incremental updates, and subsampling of reservoir nodes further reduces communication cost. The paper is also the first to apply federated reservoir computing to anomaly detection and evaluates on recent, temporally aware metrics (VUS-PR and PATE), which is a strength. The main weakness is that the exact-equivalence claim is not conditioned on how a single time series is split across clients, and the experimental section does not report variability over reservoir initializations; both issues need to be addressed before the central claims can be accepted.

major comments (3)
  1. [III-B, Eqs. (16)-(18)] The claim that the global precision matrix in (18) is 'mathematically equivalent' to the centralized precision matrix in (12) requires that each client's reservoir state trajectory x_c(t) be identical to the trajectory that a centralized model would compute. Under the stated recurrence (1) with x(0)=0, this holds only for the first chronological segment when a single long time series is split across clients. For PSM, one 132,481-point training series is divided among 24 clients (Section IV-A), so clients c>1 start from zero and their state sequences omit the history of earlier segments. Consequently, Phi_g in (17) is generally not equal to Phi_0 in (11). The phrase 'given that all data are available to the server' does not supply the missing condition; the authors must either add an explicit warm-up or state-transfer protocol that makes the states agree and prove the identity under that protocol, or reformulate the method as approximate and provide a bound or an empirical measurement of the discrepancy. Because Tables II-IV report IncFed MD-RS scores identical to centralized MD-RS, this issue is load-bearing for the paper's main claim.
  2. [IV-C, IV-D] All reported results are based on a single random reservoir initialization and a single data-partition realization, with no error bars, confidence intervals, or multiple seeds. Since MD-RS depends on the randomly generated matrices W_in and W (Section II-A), the conclusion that IncFed MD-RS 'outperforms other federated learning methods' is not robustly supported. The authors should report means and standard deviations over at least several reservoir realizations, or otherwise justify why a single realization is sufficient for the claimed comparison.
  3. [IV-A, IV-C] The description 'we divide the training time series data by the number of clients' is ambiguous for the multi-series datasets SMD and SMAP, which contain 28 and 55 time series respectively. It is not clear whether each complete time series is assigned wholly to a client or whether each time series is split chronologically across clients. The correctness of Eq. (17) depends on this distinction, because whole-series assignment and chronological splitting lead to different state trajectories and different equivalence conditions. The paper should specify the partitioning protocol for each dataset.
minor comments (6)
  1. [II-A, Eq. (1)] The text says 'with random initial condition x(0)=0', which is contradictory; it should read 'with the initial condition x(0)=0'.
  2. [III-A] There are spelling errors: 'Malahanobis' should be 'Mahalanobis' in the section title, and 'Guassian' should be 'Gaussian' in the first paragraph.
  3. [III-B, Eq. (16)] The symbol T0 is used for both the global training length in Eq. (11) and 'the length of client c's training data' in Eq. (16); if client lengths differ, a client-specific notation such as T_c should be used.
  4. [IV-C] The hyperparameter 'input scaling 0.001' is not defined in the model equations; the paper should state explicitly how input scaling enters Eq. (1), e.g., as a multiplicative factor on W_in.
  5. [III-B] The equivalence statement should explicitly mention that all clients share the same reservoir hyperparameters and the same realization of W_in and W; this is implied by the text but should be stated as a formal condition.
  6. [IV-D] The sentence 'It is noteworthy that the performance of IncFed MD-RS is the same as that of centralized MD-RS in tables for every dataset, validating the equivalence of them' is circular: if the implementation is based on the same covariance aggregation, identical scores are expected and do not independently validate the equivalence. A comparison on data with known different partitionings would be more informative.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: IncFed MD-RS's equivalence to centralized MD-RS is a linear sufficient-statistics identity; the unstated reservoir-initialization condition is a correctness caveat, not circular reasoning.

full rationale

IncFed MD-RS's central equivalence claim is not circular. In Sec. III-B, Phi_g is defined as the sum of client covariance matrices (Eq. 17) and P_g as (Phi_g + delta I)^-1 (Eq. 18). If each client's reservoir-state sequence x_c(t) is the same as the sequence a centralized run would produce for the same data, then Phi_g = Phi_0 by the linearity of the sum-of-outer-products, and hence P_g = P_0 exactly. This is a sufficient-statistics aggregation, not a fit: no parameter is tuned to the centralized precision matrix or to the anomaly-detection scores, and the equivalence is not used to define the local covariances. The numerical agreement with Central.MD-RS in Tables II-IV is an independent empirical check, not a fitted prediction. The main caveat is a missing assumption rather than circularity: the paper's Eq. (1) fixes x(0)=0 per client, and Sec. IV splits a single long PSM sequence among 24 clients; without a state-transfer or warm-up protocol, later segments' states are not identical to the centralized trajectory, so the exact equality of Phi_g and Phi_0 is not guaranteed unless an unstated condition holds. The paper does not state this initialization condition or bound the discrepancy. Separately, Ref. [6] is a self-citation (Tamura and Tanaka overlap with the current authors) and MD-RS is the base method, but the federated aggregation proof in Eqs. (16)-(18) is derived in the present paper and does not depend on the correctness or acceptance of [6], so the self-citation is not load-bearing. No fitted input is renamed as a prediction, and no uniqueness theorem is imported from the authors' prior work.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central method adds no new physical or conceptual entities. It depends on several hand-chosen hyperparameters, on the zero-mean Gaussian assumption inherited from MD-RS, and on the unstated assumption that local reservoir states match centralized states for arbitrary data partitions. The last assumption is the most fragile element.

free parameters (6)
  • number of reservoir nodes N_x = 500
    Set in Sec IV-C; affects dimensionality of covariance and communication cost.
  • subsampling size N~_x = 200
    Set in Sec IV-C; trades communication and computation against accuracy.
  • leaking rate alpha = 1.0
    Set in Sec IV-C; controls reservoir memory and recurrence.
  • spectral radius rho = 0.95
    Set in Sec IV-C; controls echo state property and memory decay.
  • input scaling = 0.001
    Set in Sec IV-C; scales the nonlinearity of the reservoir input.
  • regularization delta = 0.0001
    Set in Sec IV-C; stabilizes the covariance inversion in Eqs (12) and (18).
assumptions (5)
  • domain assumption Echo state property: reservoir with spectral radius below 1 forgets its initial state over time.
    Standard reservoir computing assumption invoked for the state update in Eq (1).
  • domain assumption All clients share the same reservoir weights W_in, W and the same hyperparameters.
    Required for local covariance matrices to be summable into a coherent global model, as described in Sec II-C and III-B.
  • domain assumption Reservoir states for normal data follow a zero-mean multivariate Gaussian distribution.
    Assumed in the MD-RS formulation in Sec III-A, notably in the anomaly score of Eq (13).
  • domain assumption Training data contain only normal points.
    Semi-supervised anomaly detection setting stated in Sec III-A.
  • ad hoc to paper Local reservoir states for arbitrarily divided time series equal the states a centralized model would compute.
    Needed for the claimed exact equivalence in Sec III-B when a single time series is split among clients; not generally true because each segment starts from a zero initial state.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Federated Learning with Reservoir State Analysis for Time Series Anomaly Detection." pith.science (2026). https://pith.science/paper/CKNLS4RM

@misc{pith2026250205679,
  author       = {Pith},
  title        = {Pith review of: Federated Learning with Reservoir State Analysis for Time Series Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CKNLS4RM}},
  note         = {Machine review of arXiv:2502.05679}
}
read the original abstract

With a growing data privacy concern, federated learning has emerged as a promising framework to train machine learning models without sharing locally distributed data. In federated learning, local model training by multiple clients and model integration by a server are repeated only through model parameter sharing. Most existing federated learning methods assume training deep learning models, which are often computationally demanding. To deal with this issue, we propose federated learning methods with reservoir state analysis to seek computational efficiency and data privacy protection simultaneously. Specifically, our method relies on Mahalanobis Distance of Reservoir States (MD-RS) method targeting time series anomaly detection, which learns a distribution of reservoir states for normal inputs and detects anomalies based on a deviation from the learned distribution. Iterative updating of statistical parameters in the MD-RS enables incremental federated learning (IncFed MD-RS). We evaluate the performance of IncFed MD-RS using benchmark datasets for time series anomaly detection. The results show that IncFed MD-RS outperforms other federated learning methods with deep learning and reservoir computing models particularly when clients' data are relatively short and heterogeneous. We demonstrate that IncFed MD-RS is robust against reduced sample data compared to other methods. We also show that the computational cost of IncFed MD-RS can be reduced by subsampling from the reservoir states without performance degradation. The proposed method is beneficial especially in anomaly detection applications where computational efficiency, algorithm simplicity, and low communication cost are required.

Figures

Figures reproduced from arXiv: 2502.05679 by the authors.

Figure 1
Figure 1. Schematic illustrations of federated learning schemes. (a) Federated Averaging. Each client [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overview of Mahalanobis Distance of Reservoir States. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Distributions of the evaluation metric scores over time series in the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 6
Figure 6. Figure 6: (a) Performance of IncFed MD-RS with different subsampling sizes ˜ [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 5
Figure 5. Figure 5: Various performance comparisons on Pooled Server Metrics (PSM). [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 16 canonical work pages

  1. [1]

    Communication-Efficient Learning of Deep Networks from Decentralized Data,

    B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y. Arcas, “Communication-Efficient Learning of Deep Networks from Decentralized Data,” in Proceedings of the 20th International Conference on Artificial Intelligence and Statistics , ser. Proceedings of Machine Learning Research, A. Singh and J. Zhu, Eds., vol. 54. PMLR, 20–22 Apr 2017, pp. 1273–1282....

    2017

  2. [2]

    The “echo state

    H. Jaeger, “The “echo state” approach to analysing and training recurrent neural networks,” GMD-Report 148, German National Research Institute for Computer Science , 01 2001

    work page 2001

  3. [3]

    Federated reservoir computing neural networks,

    D. Bacciu, D. Di Sarli, P. Faraji, C. Gallicchio, and A. Micheli, “Federated reservoir computing neural networks,” in 2021 International Joint Conference on Neural Networks (IJCNN) , 2021, pp. 1–7

    work page 2021

  4. [4]

    Decentralized incremental federated learning with echo state networks,

    G. Pompei, P. Dazzi, V . De Caro, and C. Gallicchio, “Decentralized incremental federated learning with echo state networks,” in 2024 International Joint Conference on Neural Networks (IJCNN) , 2024, pp. 1–8

    work page 2024

  5. [5]

    Decentralized federated learning for industrial iot with deep echo state networks,

    W. Qiu, W. Ai, H. Chen, Q. Feng, and G. Tang, “Decentralized federated learning for industrial iot with deep echo state networks,”IEEE Transactions on Industrial Informatics , vol. 19, no. 4, pp. 5849–5857, 2023

    work page 2023

  6. [6]

    Mahalanobis distance of reservoir states for online time-series anomaly detection,

    H. Tamura, K. Fujiwara, K. Aihara, and G. Tanaka, “Mahalanobis distance of reservoir states for online time-series anomaly detection,” Apr. 2023. [Online]. Available: http://dx.doi.org/10.36227/techrxiv. 22678774.v1

    doi:10.36227/techrxiv 2023

  7. [7]

    Anomaly detection in time series: a comprehensive evaluation,

    S. Schmidl, P. Wenig, and T. Papenbrock, “Anomaly detection in time series: a comprehensive evaluation,” Proceedings of the VLDB Endowment, vol. 15, no. 9, pp. 1779–1797, 2022

    2022

  8. [8]

    Deep learning for anomaly detection in time-series data: Review, analysis, and guidelines,

    K. Choi, J. Yi, C. Park, and S. Yoon, “Deep learning for anomaly detection in time-series data: Review, analysis, and guidelines,” IEEE access, vol. 9, pp. 120 043–120 065, 2021

    work page 2021

Show all 24 references
  1. [9]

    Reservoir computing approaches to recurrent neural network training,

    M. Luko ˇseviˇcius and H. Jaeger, “Reservoir computing approaches to recurrent neural network training,” Computer science review , vol. 3, no. 3, pp. 127–149, 2009

    2009

  2. [10]

    Optimization and applications of echo state networks with leaky-integrator neurons,

    H. Jaeger, M. Luko ˇseviˇcius, D. Popovici, and U. Siewert, “Optimization and applications of echo state networks with leaky-integrator neurons,” Neural networks, vol. 20, no. 3, pp. 335–352, 2007

    2007

  3. [11]

    A survey on federated learning: The journey from centralized to distributed on-site learning and beyond,

    S. Abdulrahman, H. Tout, H. Ould-Slimane, A. Mourad, C. Talhi, and M. Guizani, “A survey on federated learning: The journey from centralized to distributed on-site learning and beyond,” IEEE Internet of Things Journal, vol. PP, 10 2020

    2020

  4. [12]

    Federated learning for beginners: Types, simulation environments, and open challenges,

    M. Panigrahi, S. Bharti, and A. Sharma, “Federated learning for beginners: Types, simulation environments, and open challenges,” in 2023 International Conference on Computer, Electronics & Electrical Engineering & their Applications (IC2E3) , 2023, pp. 1–6

    2023

  5. [13]

    On the convergence of federated optimization in heterogeneous networks,

    A. K. Sahu, T. Li, M. Sanjabi, M. Zaheer, A. Talwalkar, and V . Smith, “On the convergence of federated optimization in heterogeneous networks,” CoRR, vol. abs/1812.06127, 2018. [Online]. Available: http://arxiv.org/abs/1812.06127

    2018 arXiv

  6. [14]

    SCAFFOLD: stochastic controlled averaging for on-device federated learning,

    S. P. Karimireddy, S. Kale, M. Mohri, S. J. Reddi, S. U. Stich, and A. T. Suresh, “SCAFFOLD: stochastic controlled averaging for on-device federated learning,” CoRR, vol. abs/1910.06378, 2019. [Online]. Available: http://arxiv.org/abs/1910.06378

    1910 arXiv

  7. [15]

    Model-contrastive federated learning,

    Q. Li, B. He, and D. Song, “Model-contrastive federated learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021

    2021

  8. [16]

    Fedtadbench: Federated time-series anomaly detection bench- mark,

    F. Liu, C. Zeng, L. Zhang, Y . Zhou, Q. Mu, Y . Zhang, L. Zhang, and C. Zhu, “Fedtadbench: Federated time-series anomaly detection bench- mark,” in 2022 IEEE 24th Int Conf on High Performance Computing & Communications; 8th Int Conf on Data Science & Systems; 20th Int Conf on ...

    2022

  9. [17]

    Robust anomaly detection for multivariate time series through stochastic recurrent neural network,

    Y . Su, Y . Zhao, C. Niu, R. Liu, W. Sun, and D. Pei, “Robust anomaly detection for multivariate time series through stochastic recurrent neural network,” in Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , ser. KDD ’19. New Yo...

    2019 doi

  10. [18]

    Detecting spacecraft anomalies using lstms and nonparametric dynamic thresholding,

    K. Hundman, V . Constantinou, C. Laporte, I. Colwell, and T. Soder- strom, “Detecting spacecraft anomalies using lstms and nonparametric dynamic thresholding,” arXiv preprint arXiv:1802.04431 , 2018

    2018 arXiv

  11. [19]

    Practical approach to asynchronous multivariate time series anomaly detection and localization,

    A. Abdulaal, Z. Liu, and T. Lancewicki, “Practical approach to asynchronous multivariate time series anomaly detection and localization,” in Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining , ser. KDD ’21. New York, NY , USA: Association for C...

    2021

  12. [20]

    V olume Under the Surface: A New Accuracy Evaluation Measure for Time-Series Anomaly Detection,

    J. Paparrizos, P. Boniol, T. Palpanas, R. S. Tsay, A. Elmore, and M. J. Franklin, “V olume Under the Surface: A New Accuracy Evaluation Measure for Time-Series Anomaly Detection,” Proceedings of the VLDB Endowment, vol. 15, no. 11, pp. 2774–2787, 2022

    2022

  13. [21]

    Pate: Proximity- aware time series anomaly evaluation,

    R. Ghorbani, M. J. Reinders, and D. M. Tax, “Pate: Proximity- aware time series anomaly evaluation,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , ser. KDD ’24. New York, NY , USA: Association for Computing Machinery, 2024, p. 872–88...

    2024

  14. [22]

    TranAD: Deep Transformer Networks for Anomaly Detection in Multivariate Time Series Data,

    S. Tuli, G. Casale, and N. R. Jennings, “TranAD: Deep Transformer Networks for Anomaly Detection in Multivariate Time Series Data,” Proceedings of VLDB , vol. 15, no. 6, pp. 1201–1214, 2022

    2022

  15. [23]

    Lstm-based encoder-decoder for multi-sensor anomaly de- tection,

    P. Malhotra, A. Ramakrishnan, G. Anand, L. Vig, P. Agarwal, and G. Shroff, “Lstm-based encoder-decoder for multi-sensor anomaly de- tection,” 07 2016

    2016

  16. [24]

    Reconstructive reservoir computing for anomaly detection in time-series signals,

    J. Kato, G. Tanaka, R. Nakane, and A. Hirose, “Reconstructive reservoir computing for anomaly detection in time-series signals,” Nonlinear Theory and Its Applications, IEICE , vol. 15, no. 1, pp. 183–204, 2024

    2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.