Machine Learning-Based Anomaly Detection of Correlated Sensor Data: An Integrated Principal Component Analysis-Autoencoder Approach

Arnab Das; Mohammad R. Haider; Santanu Das; Srihari Varada; Tanish Baranwal

REVIEW 4 major objections 5 minor 25 references

Machine Learning-Based Anomaly Detection of Correlated Sensor Data: An Integrated Principal Component Analysis-Autoencoder Approach

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read For correlated sensor data, a PCA-triggered autoencoder matches the autoencoder's F1 score at 35% lower runtime.

desk verdict A plausible PCA-gated autoencoder with a real speedup number, but the headline result rests on one unstated operating point and one anomaly model. read the letter →

arxiv 2505.24044 v1 pith:ALSXOZBD submitted 2025-05-29 eess.SP

Tanish Baranwal , Arnab Das , Srihari Varada , Santanu Das , Mohammad R. Haider This is my paper

classification eess.SP

keywords IoTanomalydetectioncorrelatedsensordataprincipalcomponentanalysisautoencoderLSTMhybridmodelruntimeoptimizationfalsepositives

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

The paper tries to establish that anomaly detection on correlated sensor streams can be made cheaper without losing accuracy by chaining a fast linear detector (PCA) in front of a deep nonlinear one (an autoencoder). PCA watches every time step at low computational cost and wakes the autoencoder only when its pairwise sensor-distance signal exceeds a threshold. On real and simulated sensor data, the paper reports that this hybrid reaches an F1 score close to the standalone autoencoder (0.849 vs 0.856) while cutting per-step runtime by about 35% (1.84 ms vs 2.84 ms). If correct, this means resource-constrained IoT devices can afford continuous deep-learning-grade monitoring by running deep inference only on suspicion.

What carries the argument

The central object is the gating cascade: a cheap linear monitor (PCA) whose pairwise Euclidean distance matrix between sensors forms an anomaly score, triggering a deeper verifier (an LSTM autoencoder) only on flagged time steps. The autoencoder's reconstruction loss, compared against a training-distribution threshold, provides the confirmatory signal. The architecture's efficiency comes from making the expensive model a conditional consumer of the cheap model's output rather than running it every time step.

What would settle it

On the same 6000-step sensor dataset, sweep the PCA trigger threshold across a range and record hybrid F1 and per-step runtime at each setting; if no threshold yields both F1 within 1% of the autoencoder and a runtime saving near 35%, the paper's headline comparison collapses.

Watch

Extended reading notes

Core claim

The central discovery claim is that a cascade of PCA and an autoencoder yields most of the accuracy of the autoencoder while largely inheriting the speed of PCA on correlated multivariate sensor data. PCA continuously monitors the pairwise Euclidean distances between encoded sensor readings; when these distances deviate beyond a set threshold, the autoencoder, an LSTM-based encoder-decoder, is invoked on the flagged time step to confirm or reject the anomaly. In the paper's evaluation on 6000 time steps of real indoor temperature-sensor data, the hybrid achieves F1 = 0.84892 compared with the autoencoder's 0.85566, while per-time-step runtime drops from 2.84 ms to 1.84 ms on an AMD Ryzen 9 CPU. The paper also reports that the hybrid's response time scales with anomaly rarity, with a strong negative correlation (R = -0.9971) between runtime and anomaly percentage.

Load-bearing premise

The anomaly trigger rests on a threshold for pairwise sensor distances that the paper sets without reporting values or a selection procedure, so the speed and accuracy numbers describe one operating point rather than the method itself.

Editorial extensions

If this is right

A device running the hybrid can keep a deep anomaly detector effectively always-on at the energy and compute cost of PCA, because the autoencoder is invoked only when the linear monitor suspects a problem.
The hybrid inherits the autoencoder's robustness to small benign fluctuations, giving fewer false positives than PCA alone, while still detecting the anomalies the autoencoder would catch.
Because runtime depends on how often the autoencoder is triggered, the hybrid is fastest precisely when anomalies are rare, which is the common case in healthy sensor networks.
The F1 gap between hybrid and autoencoder (about 0.007) can be seen as the price of the speedup, a trade-off that could be tuned via the trigger threshold.

Reading between the lines

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

The reported 35% speedup is tied to a single operating point; sweeping the PCA trigger threshold would trace out a runtime–F1 curve, and applications should choose a point on that curve rather than taking the headline numbers as intrinsic to the method.
The same gating idea could be applied with other cheap detectors in place of PCA, such as a simple statistical control chart, and with other expensive verification models in place of the autoencoder, making this a general 'cheap trigger, expensive verifier' pattern.
Because the paper does not report threshold values or a selection procedure, a direct replication would need to grid-search the threshold to verify whether the accuracy and runtime numbers hold outside the chosen setting.
A natural stress test is to measure the hybrid on anomaly types that are invisible to a linear monitor, such as subtle nonlinear correlation changes, to see whether the trigger still fires.

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.

Desk Editor's Note

read the letter

Worth a look if you work on lightweight IoT anomaly detection, but treat the headline numbers as a single data point, not a verified capability. The hybrid design is straightforward: PCA runs continuously, and when its pairwise-distance check fires, the LSTM autoencoder confirms the anomaly. That is a natural engineering move, and the paper is honest about what it is. What is genuinely new is the evaluation of that gate on real correlated sensor data: 6000 Intel Berkeley timesteps, four temperature sensors, and a runtime comparison across anomaly rates. The runtime plot showing a strong negative correlation between anomaly percentage and hybrid speedup is the most informative result; it makes the design logic concrete.

The paper also does some things well. The sensitivity comparison of PCA versus autoencoder on mean-shift and erasure anomalies is clear, and the finding that PCA is highly sensitive to mean shifts but insensitive to low-rate erasure is a real empirical observation that supports the gating idea. The F1 numbers are reported in a table, and the paper distinguishes between sensitivity and false positives in the ROPE analysis. That is more than many engineering papers do.

Where it gets soft: the headline claim of 'similar F1, 35% faster' is a single point. The PCA trigger threshold and the persistence counter in Figure 2 are never reported, so nobody can tell what fraction of steps wake the autoencoder. The runtime and F1 parity are properties of that operating point, not of the method. Worse, the hybrid is only evaluated on the mean-shift anomaly model. The paper itself shows PCA barely responds to 5% erasure, so a trigger tuned for mean shifts could miss exactly the erasure anomalies that the autoencoder would catch. The 'fewer false positives' claim in the abstract is never directly measured in the hybrid evaluation. There are no error bars or repeated runs, and no comparison with prior PCA-autoencoder hybrids, so the novelty claim is unsubstantiated even if the specific evaluation is new.

All of this is fixable with a threshold sweep, a second anomaly scenario, and a more careful comparison. The core idea is plausible and the paper deserves a serious referee, but it should not be accepted as is.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes a hybrid anomaly-detection pipeline for correlated IoT sensor streams in which PCA runs continuously and an LSTM autoencoder is invoked only when PCA flags a potential anomaly. The approach is evaluated on Intel Berkeley Research Labs temperature data and on synthetic Gaussian data, with F1 score and per-timestep runtime compared against standalone PCA, autoencoder, UMAP, and the proposed hybrid. The central claim, stated in Section IV.C.1 and Table II, is that the hybrid model achieves an F1 score comparable to the autoencoder (0.84892 versus 0.85566) while running 35% faster (1.84 ms versus 2.84 ms per timestep).

Significance. If the headline claim survives proper experimental characterization, the gating idea is practically valuable for resource-constrained IoT systems: a cheap linear trigger followed by selective deep-model verification could reduce average inference cost while retaining most of the autoencoder's accuracy. The paper uses a public dataset, provides a clear system flowchart, and reports numerical metrics, which are strengths. However, the current evidence is insufficient because the reported F1 and runtime values are properties of an unstated operating point, with no threshold sweep, no repeated runs, no confidence intervals, and no evaluation on anomaly types to which the PCA trigger is known to be insensitive.

major comments (4)

[IV.C.1, Table II, Fig. 2] The headline comparison is a single operating point whose defining parameters are never reported. The paper does not state the PCA distance threshold, the persistence-counter threshold in Fig. 2, the autoencoder reconstruction-loss threshold, the sliding-window length, or the autoencoder architecture and training hyperparameters. Since the hybrid runtime is approximately p_trigger * t_autoencoder + (1 - p_trigger) * t_PCA, the claimed 35% speedup and the F1 parity both depend directly on p_trigger, the fraction of timesteps at which PCA wakes the autoencoder. Without reporting the threshold values and the actual trigger rate, and without showing how F1 and runtime vary as the PCA threshold is swept, the central quantitative claim cannot be verified or generalized.
[IV.C.1 versus IV.A.2] The hybrid is evaluated only under a single anomaly model: a mean shift of 10 applied to the fourth sensor in alternating 1000-step bins, making 50% of the 6000 timesteps anomalous. Earlier, Section IV.A.2 demonstrates that PCA responds strongly to 20% erasure errors but only weakly to 5% erasure errors, whereas the autoencoder responds to both. Because the hybrid trigger is PCA-based, a threshold tuned for mean-shift anomalies may simply miss the erasure anomalies that the autoencoder would have caught. The paper should evaluate the hybrid on erasure errors and on other anomaly types, and should report the trigger rate and F1 under each scenario; otherwise the claim of F1 parity with the autoencoder is established only in the easiest regime.
[III.B, IV.B] The anomaly thresholds appear to be fitted to the same data used for evaluation, and the ground-truth labels are defined by the same injected transformations used to compute the metrics. Section III.B states that readings are flagged when pairwise Euclidean distances 'exceeded a set threshold' and that autoencoder anomalies are detected using 'a loss threshold approach,' but no threshold-selection procedure is described. The reported F1 values therefore partly reflect the chosen operating points rather than an independent property of the methods. The authors should specify how thresholds are selected (for example, on a separate validation set), and should report results on held-out time segments or held-out anomaly scenarios.
[IV.C.1, Table II] All quantitative results are based on a single run with no repeated trials, seeds, or error bars. The F1 gap between the hybrid (0.84892) and the autoencoder (0.85566) is only 0.00674, which is likely within run-to-run variability for a deep model with stochastic training. Similarly, the runtime comparison gives no information about variance, implementation details, or software/hardware stack beyond 'AMD Ryzen 9 CPU.' The authors should provide multiple runs with confidence intervals, report the autoencoder training procedure and number of seeds, and make the evaluation script or at least the threshold and trigger-rate measurements available.

minor comments (5)

[IV.C.1] The text says 'As can be seen in Figure II' when the results are presented in Table II; this cross-reference should be corrected.
[III.A and III.C] Section III.A describes an experiment with 2000 timesteps and an anomaly introduced at timestep 1000, while Section III.C describes the hybrid evaluation on 6000 timesteps with alternating 1000-step bins. The relationship between these two experimental setups should be clarified so the reader can map the methods to the results.
[Table II and Fig. 11] The runtime values in Table II have no units in the header row, and Fig. 11 shows percent reduction of response time without error bars or a description of how the percentage is computed for each anomaly rate; adding units and error bars would improve interpretability.
[References] The reference list contains inconsistencies: reference [18] duplicates reference [16] (both are van der Maaten and Hinton, t-SNE), and the numbering used in the text (e.g., the LSTM citation) does not align cleanly with the reference list. These should be corrected before publication.
[II, Table I] Table I lists Autoencoder complexity as 'Constant' and PCA as 'O(input × output),' but the autoencoder cost is not constant in practice and the table's notation is not fully defined; a brief explanation of the complexity measure would avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

Measured empirical comparison with no derivation chain; missing threshold reporting is a reproducibility issue, not circularity.

full rationale

The paper makes no analytical derivation whose output is equivalent to its input. The central claims—that the hybrid model has an F1 score comparable to the Autoencoder (0.84892 vs. 0.85566) and runs 35% faster (1.84 ms vs. 2.84 ms)—are presented as measured results on Intel Berkeley sensor data (Table II), not as predictions derived from fitted parameters. The anomaly-detection procedure uses a Euclidean-distance threshold, but the paper does not report threshold values or characterize how the hybrid runtime and F1 vary with the threshold; this is an evaluation-completeness limitation, not a circular reduction. No equation in the paper sets a claimed result equal to an input by construction, no fitted parameter is renamed as a prediction, and none of the cited references are self-citations carrying argumentative weight. The reader's concern about operating-point dependence is a legitimate reproducibility criticism, but under the required standard—quoting a specific reduction—no circular step can be identified.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper introduces no invented physical or mathematical entities. Its central empirical claim depends on several unreported or loosely specified choices: the PCA gate threshold, the autoencoder loss threshold, the window size, and the autoencoder architecture. The structural assumptions are that pairwise Euclidean distances in a reduced space are adequate anomaly scores and that the chosen Intel Berkeley sensors plus injected faults represent correlated IoT sensing.

free parameters (4)

PCA distance threshold = not reported
Used in Section III.B to convert pairwise Euclidean distances into binary anomaly labels and in the hybrid to decide when to invoke the autoencoder. No selection procedure or tradeoff curve is given.
Autoencoder reconstruction loss threshold = not reported
Used to flag anomalies based on elevated reconstruction loss; the threshold approach is cited but its value, percentile, and validation split are not specified.
Sliding window length = 100 samples
Chosen in Section II for all experiments; affects the input dimension, the distance matrix, and the runtime comparison.
Autoencoder architecture and training hyperparameters = not reported
Layer sizes, latent dimension, learning rate, epochs, and optimizer are not specified; these determine the learned normal distribution and the reported F1.

assumptions (3)

domain assumption Euclidean distance between sensor encodings is a valid anomaly score, and a fixed threshold can separate anomalous from normal sensor pairs.
Section III.B defines anomaly detection as distances exceeding a set threshold; no distributional justification or calibration is given.
domain assumption The autoencoder is trained on data for which a reconstruction-loss threshold represents normal behavior, and anomalies are detected by elevated loss.
Section III.B relies on this standard assumption but does not state the training data regime or the loss threshold.
domain assumption The four selected Intel Berkeley temperature sensors are representative of correlated sensors in resource-constrained IoT, and injected mean shifts and erasures model real faults.
Sections III.A and IV generalize from this single dataset and synthetic anomalies to broad IoT deployments without evidence of transferability.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Machine Learning-Based Anomaly Detection of Correlated Sensor Data: An Integrated Principal Component Analysis-Autoencoder Approach." pith.science (2026). https://pith.science/paper/ALSXOZBD

@misc{pith2026250524044,
  author       = {Pith},
  title        = {Pith review of: Machine Learning-Based Anomaly Detection of Correlated Sensor Data: An Integrated Principal Component Analysis-Autoencoder Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ALSXOZBD}},
  note         = {Machine review of arXiv:2505.24044}
}

read the original abstract

The growing adoption of IoT systems in industries like transportation, banking, healthcare, and smart energy has increased reliance on sensor networks. However, anomalies in sensor readings can undermine system reliability, making real-time anomaly detection essential. While a large body of research addresses anomaly detection in IoT networks, few studies focus on correlated sensor data streams, such as temperature and pressure within a shared space, especially in resource-constrained environments. To address this, we propose a novel hybrid machine learning approach combining Principal Component Analysis (PCA) and Autoencoders. In this method, PCA continuously monitors sensor data and triggers the Autoencoder when significant variations are detected. This hybrid approach, validated with real-world and simulated data, shows faster response times and fewer false positives. The F1 score of the hybrid method is comparable to Autoencoder, with much faster response time which is driven by PCA.

Figures

Figures reproduced from arXiv: 2505.24044 by the authors.

**Figure 1.** Setup for simulation of normal and anomalous behavior [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 4.** PCA simulation results for anomaly model (a) [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Figures from the paper (5 more)

**Figure 5.** Figure 5: Autoencoder simulation results for anomaly model (b) [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

**Figure 10.** Figure 10: Sensitivity and accuracy of the PCA with erasure errors (for the [PITH_FULL_IMAGE:figures/full_fig_p005_10.png]

**Figure 8.** Figure 8: Sensitivity and accuracy of the PCA (for the fourth sensor) [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]

**Figure 9.** Figure 9: Sensitivity and accuracy of the Autoencoder with erasure errors (for [PITH_FULL_IMAGE:figures/full_fig_p005_9.png]

**Figure 11.** Figure 11: Percent reduction of the response time of the hybrid model versus [PITH_FULL_IMAGE:figures/full_fig_p006_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 21 canonical work pages

[1]

Chamilov

A. Chamilov. 6 Cool Examples of Internet of Things Applica- tion, Eastern Peak (https://easternpeak.com/blog/6-cool-examples-of- internet-ofthings-applications-and-how-to-develop-one/), Oct 1, 2020

work page 2020
[2]

A. Ukil, S. Bandyoapdhyay, C. Puri and A. Pal. IoT Healthcare Analytics: The Importance of Anomaly Detection, In Proc. 2016 IEEE 30th International Conference on Advanced Information Networking and Applications (AINA), Crans-Montana, March 2016

work page 2016
[3]

Abebe Diro ,Naveen Chilamkurti, Van-Doan Nguyen, and Will Heyne: A Comprehensive Study of Anomaly Detection Schemes in IoT Networks Using Machine Learning Algorithms, 13 December 2021, https://www.mdpi.com/1424-8220/21/24/8320

work page 2021
[4]

Menachem Domb, Sujata Joshi and Arulmozhi Khn: Anomaly Detec- tion in IoT: Recent Advances, AI and ML Perspectives and Applica- tions, 29 July 2023, https://www.intechopen.com/chapters/87783

work page 2023
[5]

A. A. Cook, G. Mısırlı and Z. Fan. Anomaly Detection for IoT Time- Series Data: A Survey, in IEEE Internet of Things Journal, July 2020

work page 2020
[6]

”Pearson’s correlation coeffi- cient”

Stewart, Ken. ”Pearson’s correlation coeffi- cient”. Encyclopedia Britannica, 16 Feb. 2024, https://www.britannica.com/topic/Pearsonscorrelation-coefficient. Accessed 8 April 2024

work page 2024
[7]

Redhwan, R

A. Redhwan, R. Murugesan, M. Man, A. Abdulateef, M.A. Al-Sharaf, A. Alkahtani. A Review of Machine learning and Deep Learning Techniques for Anomaly Detection in IoT Data, Appl. Sci. June 2021

work page 2021
[8]

A. B. Nassif, M. A. Talib, Q. Nasir and F. M. Dakalbab. Machine Learning for Anomaly Detection: A Systematic Review, in IEEE Access, vol. 9, May 2021

work page 2021

Show all 25 references

[9]

Sharma, L

B. Sharma, L. Sharma and C. Lal. Anomaly Detection Techniques using Deep Learning in IoT: A Survey, 2019 International Conference on Computational Intelligence and Knowledge Economy (ICCIKE), Dubai, December 2019

2019
[10]

A. Diro, N. Chilamkurti, V .D. Nguyen and W. Heyne. A Com- prehensive Study of Anomaly Detection Schemes in IoT Networks Using Machine Learning Algorithms. Sensors (Basel). 2021 Dec 13;21(24):8320. doi:10.3390/s21248320. PMID: 34960414; PMCID: PMC8708212

2021 doi
[11]

Simulation and Modeling for Anomaly Detection in IoT Network Using Machine Learning

Indrajit Mukherjee, Nilesh Kumar Sahu and Sudip Kumar Sahana. Simulation and Modeling for Anomaly Detection in IoT Network Using Machine Learning. 05 January 2022

2022
[12]

A Survey of AI-Based Anomaly Detection in IoT and Sensor Networks

Kyle DeMedeiros, Abdeltawab Hendawi, and Marco Alvarez. A Survey of AI-Based Anomaly Detection in IoT and Sensor Networks. Sensors 2023, 23(3), 1352; https://doi.org/10.3390/s23031352

2023 doi
[13]

Jolliffe and J

I.T. Jolliffe and J. Cadima. Principal component analysis: a review and recent developments. Phil. Trans. R. Soc. A 374:, January 2016. http://dx.doi.org/10.1098/rsta.2015.0202

2016
[14]

P. Baldi. Autoencoders, Unsupervised Learning, and Deep Architec- tures In. Proc. of ICML Workshop on Unsupervised and Transfer Learning, in Proceedings of Machine Learning Research. 2012

2012
[15]

McInnes, J

L. McInnes, J. Healy, N. Saul, and L. Großberger. UMAP: Uniform Manifold Approximation and Projection, Journal of Open Source Software, 2019

2019
[17]

Aggarwal, A

C. Aggarwal, A. Hinneburg, and D. Keim. On the Surprising Behavior of Distance Metric in High-Dimensional Space, Database theory, ICDT 200, 8th International Conference, London, UK, January 4 6, 2001

2001
[18]

van der Maaten and G

L. van der Maaten and G. Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research. 9. 2579-2605, 2008

2008
[19]

Sherstinsky

A. Sherstinsky. Fundamentals of Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) Network, CoRR, http://arxiv.org/abs/1808.03314, 2018

2018 arXiv
[20]

Bodik, W

P. Bodik, W. Hong, C. Guestrin, S. Madden, M. Paskin, and R. Thibaux. Intel Berkeley Research Labs Dataset, http://db.csail.mit.edu/labdata/labdata.html, 2004

2004
[21]

Mica2Dot Sensors, (http://www.cmtgmbh.de/Produkte/WirelessSensor Networks/Datenblaetter/MICA2DOT Datasheet.pdf)
[22]

TinyDB, https://tinydb.readthedocs.io/en/latest/index.html
[23]

Tiny OS, (https://en.wikipedia.org/wiki/TinyOS)
[24]

Sokolova, N

M. Sokolova, N. Japkowicz, S. Szpakowicz. (2006). Beyond Ac- curacy, F-Score and ROC: A Family of Discriminant Measures for Performance Evaluation. In: Sattar, A., Kang, Bh. (eds) AI 2006: Advances in Artificial Intelligence. AI 2006. Lecture Notes in Computer Science(), vol 4...

2006 doi
[25]

Huang, J

C. Huang, J. Cao, F. Ye, M. Li, Y . Zhang, C. Lu. (2019). InverseTrans- form AutoEncoder for Anomaly Detection. ArXiv, abs/1911.10676

2019 arXiv
[26]

Kruschke, J. K. (2018). Rejecting or Accepting Parameter Values in Bayesian Estimation. Advances in Methods and Practices in Psycho- logical Science. https://doi.org/10.1177/2515245918771304

2018 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Chamilov

A. Chamilov. 6 Cool Examples of Internet of Things Applica- tion, Eastern Peak (https://easternpeak.com/blog/6-cool-examples-of- internet-ofthings-applications-and-how-to-develop-one/), Oct 1, 2020

work page 2020

[2] [2]

A. Ukil, S. Bandyoapdhyay, C. Puri and A. Pal. IoT Healthcare Analytics: The Importance of Anomaly Detection, In Proc. 2016 IEEE 30th International Conference on Advanced Information Networking and Applications (AINA), Crans-Montana, March 2016

work page 2016

[3] [3]

Abebe Diro ,Naveen Chilamkurti, Van-Doan Nguyen, and Will Heyne: A Comprehensive Study of Anomaly Detection Schemes in IoT Networks Using Machine Learning Algorithms, 13 December 2021, https://www.mdpi.com/1424-8220/21/24/8320

work page 2021

[4] [4]

Menachem Domb, Sujata Joshi and Arulmozhi Khn: Anomaly Detec- tion in IoT: Recent Advances, AI and ML Perspectives and Applica- tions, 29 July 2023, https://www.intechopen.com/chapters/87783

work page 2023

[5] [5]

A. A. Cook, G. Mısırlı and Z. Fan. Anomaly Detection for IoT Time- Series Data: A Survey, in IEEE Internet of Things Journal, July 2020

work page 2020

[6] [6]

”Pearson’s correlation coeffi- cient”

Stewart, Ken. ”Pearson’s correlation coeffi- cient”. Encyclopedia Britannica, 16 Feb. 2024, https://www.britannica.com/topic/Pearsonscorrelation-coefficient. Accessed 8 April 2024

work page 2024

[7] [7]

Redhwan, R

A. Redhwan, R. Murugesan, M. Man, A. Abdulateef, M.A. Al-Sharaf, A. Alkahtani. A Review of Machine learning and Deep Learning Techniques for Anomaly Detection in IoT Data, Appl. Sci. June 2021

work page 2021

[8] [8]

A. B. Nassif, M. A. Talib, Q. Nasir and F. M. Dakalbab. Machine Learning for Anomaly Detection: A Systematic Review, in IEEE Access, vol. 9, May 2021

work page 2021