Predictive Failure Detection in Network Hardware Using Thermal Imaging and Deep Learning with Sensor Fusion

Ashly Joseph

REVIEW 4 major objections 3 minor 21 references

Predictive Failure Detection in Network Hardware Using Thermal Imaging and Deep Learning with Sensor Fusion

T0 review · 4 major / 3 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The paper reports that a CNN-LSTM fusing thermal images with power readings classifies simulated network-hardware failure states at 94% accuracy, and that ROI cropping plus normalization lifts ImageNet-pretrained CNNs from near 50% to…

desk verdict Honest writeup of a synthetic-data experiment, but the 94% number is just the simulator's threshold rule rediscovered by a neural net; needs a rule-based baseline and real data before the claims are credible. read the letter →

arxiv 2608.07582 v1 pith:EQCY54FG submitted 2026-08-05 cs.CV

Ashly Joseph This is my paper · ORCID

classification cs.CV

keywords predictivemaintenancethermalimagingCNN-LSTMfusionsensorregion-of-interestextractiontransferlearningnetworkhardwarefailuremulti-modalclassification

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

This paper argues that predictive maintenance for network hardware can be built from non-intrusive thermal images and power telemetry, and that how the inputs are prepared matters as much as which deep network is chosen. On a simulated dataset of routers, switches, and servers labeled Normal, Warning, and Critical, ImageNet-pretrained CNNs score only 47–52% accuracy on raw images, but the same architectures jump to 85–91% after cropping to device heat-generating regions and normalizing inputs. Adding a power-series branch to the image CNN in a CNN-LSTM fusion reaches 94% accuracy with precision and recall near 0.95. The practical claim is therefore that domain-specific preprocessing and multi-modal fusion are the main levers for early failure detection, at least on this simulation.

What carries the argument

The central mechanism is the multi-modal CNN-LSTM fusion pipeline with domain-specific preprocessing. ROI extraction crops each thermal frame to fixed bounding boxes around CPU heatsinks and PSU modules, removing background; normalization linearly scales thermal intensities to [0,1] and z-scores power readings to a common baseline. The image path is a four-layer 2D CNN with 32, 64, 128, and 256 filters, each followed by ReLU and max-pooling, then a 128-unit dense layer with dropout; the sensor path is a 50-unit LSTM over the normalized power series; the two branches are concatenated and classified into three states. The same preprocessing is applied to ImageNet-pretrained ResNet-50, InceptionV3, and VGG16 baselines, and it is the comparison across preprocessed and raw inputs that carries the paper's argument.

What would settle it

Run the same four models on empirically collected thermal image and power pairs from operating network equipment under induced fan, power-supply, and CPU failures. If accuracy falls well below the simulated 94%, or if a hand-coded threshold rule on temperature and power deviations already achieves near the model accuracy on the simulated data, the central claim that the deep-learning pipeline provides substantial predictive value is not supported.

Watch

Extended reading notes

Core claim

The central discovery is a measured comparison: on the simulated multi-modal dataset, input preparation dominates architecture choice. ResNet-50 goes from 52% to 91% accuracy when thermal images are ROI-cropped and normalized, while the best raw-image model, the CNN-LSTM fusion, reaches only 60%. Once preprocessed, the fusion model that concatenates a four-layer CNN image branch with a 50-unit LSTM over power readings achieves 94% accuracy and macro precision and recall near 0.95, outperforming the best image-only CNN (ResNet-50 at 91%) and every model on raw inputs. The paper states this demonstrates that ROI extraction and normalization, and the combination of thermal and power modalities, substantially improve classification of Normal, Warning, and Critical states on this dataset.

Load-bearing premise

The load-bearing assumption is that the simulated thermal images and power readings faithfully represent how real routers, switches, and servers look and degrade, so that the 94% accuracy would carry over to physical hardware; the paper itself says this still needs empirical testing.

Editorial extensions

If this is right

Pre-processing is the dominant lever in this pipeline: on this dataset it moved ResNet-50 from 52% to 91% accuracy, a gain of 39 points, compared with the 3-point gain from adding the power branch to the best CNN.
A practical monitoring pipeline should pair a thermal camera with existing PDU power telemetry, since the fusion of the two modalities outperformed the best image-only model (94% vs 91%).
The same preprocessing and fusion recipe can be transferred to other equipment types whose failure modes have thermal or power signatures, provided ROI boxes are redefined per device.
The three-state output (Normal, Warning, Critical) is a workable early-warning granularity; extending to named fault types would require a different labeling scheme and is flagged as future work.

Reading between the lines

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

A simple threshold classifier may already solve much of the simulated task, because Warning and Critical labels are defined by the same thermal-deviation and power-deviation thresholds that are used as model inputs; the paper does not include such a baseline.
The reported 94% is unlikely to transfer directly to real hardware, since the simulation assumes fixed ROI boxes and clean label thresholds; an object detector and calibration step would be needed in the field.
An ablation that feeds only the power series to the LSTM branch would isolate the image contribution; the paper itself notes this ablation is missing.
Per-class recall, especially on the Critical class, would matter more than macro-averaged numbers in a real maintenance setting; the paper defines Critical as the costly miss but does not report per-class matrices.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Quick read: this paper applies ImageNet CNNs and a CNN-LSTM fusion to simulated thermal images plus power telemetry for classifying Normal/Warning/Critical states in network hardware. The writing is clear, the method is standard, and the authors are unusually candid: they state in Section III that the fusion gain is 'an inference from the aggregate numbers rather than a measured attribution' and in the conclusion that empirical data is required. That honesty is real and should be credited.

What's actually new is small: a private simulated dataset and a comparison table. The individual ingredients—ROI cropping, normalization, transfer learning, LSTM fusion—are all borrowed from the cited literature. The accuracy gain from ROI preprocessing (roughly 39 points) is at least internally consistent across three architectures, which suggests the effect is robust within the simulation.

The load-bearing problem is the one the stress-test note identifies. The labels in Section II-A are deterministic threshold functions of the same thermal and power features the models receive. Normal is deviation ≤5°C, Warning is 6–12°C rise or ±1–2σ power, Critical is >12°C or >2σ. A rule-based classifier using max ROI temperature and max normalized power over a sequence would almost certainly recover these labels with high accuracy. The paper never evaluates a rule-based baseline, so the 94% CNN-LSTM accuracy is not evidence of learned features; it may just be fitting the simulator's own labeling rule. Without that baseline, the central claim that domain-specific pre-processing and sensor fusion substantially improve early failure prediction is unsupported.

Other soft spots are secondary but real: no error bars (single run per model?), no per-class confusion matrix (they emphasize Critical recall but don't report it), and the dataset is not released, so the result is not independently checkable. The authors flag the need for empirical validation, but that is exactly the missing piece: synthetic labels generated by thresholds on the inputs cannot demonstrate a capability on physical hardware.

Who is this for? Readers interested in a cautionary example of simulated-data circularity, or someone wanting a template for the experimental setup. As a research contribution it's a workshop-level proposal, not a demonstrated result. I would not cite it.

Recommendation: desk reject for a serious venue unless the authors add a rule-based baseline, report per-class metrics, release the synthetic data, and ideally evaluate on real equipment. The paper's own caveats confirm it is not ready.

Referee Report

4 major / 3 minor

Summary. The paper proposes a deep-learning pipeline for early failure detection in network hardware using thermal images and power sensor data. A synthetic dataset of 900 annotated sequences (400 Normal, 300 Warning, 200 Critical) is generated for a simulated rack of switches, routers, and servers. Three ImageNet-pretrained CNNs (ResNet-50, InceptionV3, VGG16) and a CNN-LSTM fusion model are evaluated with and without ROI extraction and normalization. The reported test-set results are 47–60% accuracy on raw inputs and 85–94% after preprocessing, with the CNN-LSTM reaching 94%. The paper concludes that domain-specific preprocessing and sensor fusion substantially improve early failure prediction.

Significance. If the results transferred to physical hardware, the approach would be practically valuable for data-center maintenance. The paper is transparent about key limitations: it states in §III that the fusion contribution is 'an inference from the aggregate numbers rather than a measured attribution,' and in §IV that empirical validation is required. However, the central quantitative claim is undermined by the design of the simulation: the labels are threshold functions of the same thermal and power features used as model inputs, and no rule-based baseline is reported. Consequently, the high accuracies may reflect reproduction of the simulator's labeling rule rather than a general capability. The paper provides a detailed description of the simulated setup and a clean comparison framework, but it does not establish the claimed predictive-maintenance capability; at best it demonstrates high accuracy on a self-defined synthetic benchmark.

major comments (4)

[II-A, Tables I–II] The labels Normal, Warning, and Critical are deterministic functions of the very features the models receive as input. In §II-A, Normal is thermal deviation ≤5°C, Warning is a 6–12°C hotspot rise or a ±1–2σ power anomaly, and Critical is a >12°C thermal rise or >2σ power deviation. Because the same temperature and power signals are used as inputs, a threshold-based classifier on maximum ROI temperature rise and maximum normalized power deviation can in principle reproduce the labels. The paper never reports such a rule-based baseline. Without it, the reported 35–39 percentage point gains from ROI preprocessing and the 94% CNN-LSTM accuracy cannot be attributed to learned feature extraction; they may simply reflect the simulator's labeling rule.
[III] The fusion contribution is asserted but not measured. The text states, 'This is an inference from the aggregate numbers rather than a measured attribution; an ablation that feeds the power series to a classifier on its own would separate the two contributions.' No such ablation is reported. The CNN-LSTM (Table II, 0.94 accuracy) and ResNet-50 (0.91 accuracy) differ in architecture, input modality, and preprocessing, so the comparison cannot isolate the effect of sensor fusion. The authors should report a power-only classifier, an image-only version of the same CNN-LSTM architecture, and a fusion model with a matched CNN branch and training schedule.
[IV (also Abstract)] The conclusion concedes that 'the method should be evaluated on empirical data from operating network equipment to test whether the simulated results transfer.' Since the title and abstract claim predictive failure detection in network hardware and state that preprocessing and fusion 'substantially improve early failure prediction,' the current evidence from a synthetic dataset does not support the central claim as stated. The claims need to be explicitly restricted to the simulator, or empirical validation must be supplied.
[II-B, Tables I–II] ROI extraction and normalization are applied jointly, so the individual contribution of each preprocessing step is unidentified. A proper ablation would compare raw inputs, normalization only, ROI only, and ROI with normalization under the same models. As presented, the 35–39 percentage point improvements are joint effects and cannot be attributed to 'domain-specific preprocessing' as a single mechanism.

minor comments (3)

[II-D] Per-class metrics are missing. Since §II-D states that recall on the Critical class is the most important maintenance metric, the paper should report at least the Critical-class recall and a confusion matrix, not only macro-averaged values.
[II-A] No confidence intervals, repeated-seed experiments, or statistical tests are reported; with 900 sequences and a 15% test split (135 samples), the differences between ResNet-50 (0.91) and CNN-LSTM (0.94) may not be significant.
[II-A] The simulation description does not specify the temperature-to-image rendering model, noise model, or baseline temperature distributions, and no dataset or code release is mentioned. This limits reproducibility and assessment of the synthetic data's realism.

Circularity Check

1 steps flagged · score 6.0 of 10

The simulated failure labels are defined by thresholding the same thermal and power signals the models receive as input, so the 94% accuracy mainly reflects learning the simulator's labeling rule.

self definitional [Section II-A (Dataset Simulation and Setup) and Section III (Results and Discussion)]
""Normal (N): steady-state operation, thermal deviation≤5◦C from baseline." "Warning (W): gradual hotspot rise of 6–12 ◦C, or power anomalies of±1–2σdeviation." "Critical (C): rapid thermal rise>12 ◦C, or shutdown-like power behavior (>2σdeviation or dips)." "Each simulated observation consists of a thermal image, which is a two-dimensional temperature map of a device, paired with a power reading sampled at the same instant.""

The target labels are deterministic threshold functions of the thermal deviation and power z-score, which are exactly the signals encoded in the model inputs: the thermal image is a temperature map, and the power series is z-score normalized in Section II-B using the same baseline statistics as the label thresholds. Classifying these states from these inputs is therefore equivalent to recovering the simulator's labeling rule. A simple rule-based classifier on max ROI temperature rise and normalized power deviation could likely reproduce the labels, yet no such baseline is reported. The high CNN-LSTM accuracy and the 35–39 percentage point preprocessing gain thus measure how well the models learn the simulation's own thresholds, not how well they detect real hardware failures.

full rationale

The paper contains no problematic self-citation chain and does not import a uniqueness theorem from prior work; the references are external. The core circularity is in the evaluation design: Section II-A defines the ground-truth classes Normal, Warning, and Critical as threshold rules on thermal deviation and power z-score, and the models are fed the same thermal images (temperature maps) and power readings. Consequently, the classification task is self-definitional: the label is a function of the input features, so test accuracy only demonstrates that the learned models can approximate the simulator's labeling thresholds. The paper is admirably transparent about the simulation and explicitly calls for empirical validation, which prevents a higher severity score, but the central claims that ROI preprocessing and sensor fusion "substantially improve early failure prediction" are not supported against an external ground truth or even against a simple threshold baseline. The 94% accuracy is a measurement on a dataset whose labels are constructed from the predictors, so the result reduces by construction to the simulation's labeling rule.

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

The central result rests on the synthetic data generation: the thresholds defining the labels are also the features the models see, and the simulation is assumed to mimic real devices. No external data, benchmark, or rule-based baseline is provided, so the paper's evidence is self-contained but not independently grounded.

free parameters (4)

Label thresholds for Normal/Warning/Critical = 5°C, 6-12°C, >12°C; 1-2σ, >2σ
Chosen by hand to define the simulated states. The accuracy claim depends on these thresholds, which also appear as the features used by the model.
Class distribution = 400 Normal, 300 Warning, 200 Critical
Ad hoc choice of sequence counts. Affects macro-averaged metrics and the balance of the problem.
ROI bounding boxes per device type = Not specified numerically
Fixed boxes around CPU heatsinks and PSU modules. The 85 to 91 percent accuracy after cropping depends on these boxes being informative; the actual coordinates are not given.
Simulation camera parameters = FLIR A65 class: 640x512, 50 mK, 30 Hz, 1.2 m standoff, 45 deg FOV
Chosen to match a commercial camera. These parameters define the synthetic image distribution and are not derived from data.

assumptions (4)

domain assumption Synthetic thermal images and power readings faithfully represent real device failure signatures.
The entire experiment uses simulated data. The paper's conclusion notes that empirical transfer is untested (Conclusion, first extension).
domain assumption Labels derived from thermal and power thresholds correspond to meaningful failure states.
Warning and Critical are defined by thresholds in Section II-A; there is no independent evidence that these thresholds match actual precursor conditions in operating hardware.
domain assumption ImageNet pretrained features transfer to thermal images of network hardware.
The models are initialized from ImageNet weights (Section II-C). This is a standard assumption in transfer learning but is not validated for thermal imagery of this domain.
standard math Single random split (70/15/15) gives a reliable test estimate.
No repeated splits or confidence intervals are reported (Section II-D).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Predictive Failure Detection in Network Hardware Using Thermal Imaging and Deep Learning with Sensor Fusion." pith.science (2026). https://pith.science/paper/EQCY54FG

@misc{pith2026260807582,
  author       = {Pith},
  title        = {Pith review of: Predictive Failure Detection in Network Hardware Using Thermal Imaging and Deep Learning with Sensor Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EQCY54FG}},
  note         = {Machine review of arXiv:2608.07582}
}

read the original abstract

Unplanned network hardware malfunctions can interrupt services and result in expensive downtime in data centers. A deep learning-based predictive maintenance strategy is presented that utilizes thermal imaging and power sensor data to detect early indicators of equipment breakdown in routers, switches, and servers. A simulated dataset was generated comprising annotated thermal pictures and power readings indicative of three operating states: Normal, Warning, and Critical. Three ImageNet-pretrained convolutional neural network (CNN) models ResNet-50, InceptionV3, and VGG16 were assessed together with a multi-modal CNN-LSTM fusion model that integrates visual and sensor time-series information. Experiments were performed with and without pre-processing procedures, including region-of-interest (ROI) extraction and normalization. In the absence of pre-processing, CNNs attained moderate accuracy (e.g., ResNet-50 at 52%), but ROI-based pre-processing significantly enhanced performance (ResNet-50 accuracy reaching 91%). The CNN-LSTM model attained the greatest accuracy of 94%, with precision and recall approaching 95%, illustrating the effectiveness of multi-modal fusion. The results validate that domain-specific pre-processing and sensor fusion substantially improve early failure prediction, providing a potential foundation for proactive maintenance of network hardware through non-intrusive monitoring.

Figures

Figures reproduced from arXiv: 2608.07582 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 20 canonical work pages

[1]

A cyber-physical systems architec- ture for industry 4.0-based manufacturing systems,

J. Lee, B. Bagheri, and H. A. Kao, “A cyber-physical systems architec- ture for industry 4.0-based manufacturing systems,”Manuf. Lett., vol. 3, pp. 18–23, 2015

work page 2015
[2]

Thermal image-based fault detection using segmentation and SVM classification,

J. Liu, H. Zhang, and Y . Wang, “Thermal image-based fault detection using segmentation and SVM classification,”J. Electron. Imaging, vol. 30, no. 2, pp. 150–165, 2023

work page 2023
[3]

Proactive IT management with a predictive maintenance system,

K. Patel, M. Deshmukh, and S. Arora, “Proactive IT management with a predictive maintenance system,”Preprints.org, Feb. 2025. [Online]. Available: https://www.preprints.org/manuscript/202502.2062

work page arXiv 2025
[4]

Predictive maintenance using deep learning: Enhancing reliability and reducing electrical system downtime,

J. Basit and A. Zeb, “Predictive maintenance using deep learning: Enhancing reliability and reducing electrical system downtime,”Int. J. Innov. Sci. Technol., vol. 6, no. 3, pp. 1120–1136, 2024

work page 2024
[5]

Fault detection in electrical systems using pre-trained CNNs and thermal imagery,

C. Ukiwe, A. Sharma, and Y . Liu, “Fault detection in electrical systems using pre-trained CNNs and thermal imagery,” inProc. IEEE Int. Conf. Ind. Cyber-Phys. Syst. (ICPS), 2024

work page 2024
[6]

Deep learning and its applications to machine health monitoring,

R. Zhao, R. Yan, Z. Chen, K. Mao, P. Wang, and R. X. Gao, “Deep learning and its applications to machine health monitoring,”Mech. Syst. Signal Process., vol. 115, pp. 213–237, 2019

work page 2019
[7]

How transferable are features in deep neural networks?

J. Yosinski, J. Clune, Y . Bengio, and H. Lipson, “How transferable are features in deep neural networks?” inAdv. Neural Inf. Process. Syst., vol. 27, 2014

work page 2014
[8]

ImageNet classification with deep convolutional neural networks,

A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep convolutional neural networks,” inAdv. Neural Inf. Process. Syst., vol. 25, 2012

work page 2012

Show all 21 references

[9]

Thermal profiling of high-density servers for proactive fault detection,

N. Ukidwe, S. Joshi, and R. Parekh, “Thermal profiling of high-density servers for proactive fault detection,”IEEE Trans. Ind. Informat., vol. 19, no. 7, pp. 9420–9431, Jul. 2023

2023
[10]

Multimodal fusion for multimedia analysis: A survey,

P. K. Atrey, M. A. Hossain, A. El Saddik, and M. S. Kankanhalli, “Multimodal fusion for multimedia analysis: A survey,”Multimedia Syst., vol. 16, no. 6, pp. 345–379, 2010

2010
[11]

Long short term memory networks for anomaly detection in time series,

P. Malhotra, L. Vig, G. Shroff, and P. Agarwal, “Long short term memory networks for anomaly detection in time series,” inProc. Eur . Symp. Artif. Neural Netw. (ESANN), 2015

2015
[12]

A systematic analysis of performance measures for classification tasks,

M. Sokolova and G. Lapalme, “A systematic analysis of performance measures for classification tasks,”Inf. Process. Manag., vol. 45, no. 4, pp. 427–437, 2009

2009
[13]

Transfer learning for electrical fault detection using thermal images,

J. Perez, R. Castillo, and T. Kim, “Transfer learning for electrical fault detection using thermal images,”IEEE Trans. Ind. Electron., vol. 68, no. 5, pp. 4293–4303, May 2021

2021
[14]

CNN-LSTM hybrid deep learning model for remaining useful life estimation,

G. Muthukumar and J. Philip, “CNN-LSTM hybrid deep learning model for remaining useful life estimation,” arXiv:2412.15998, 2024

2024 arXiv
[15]

Thermal image- based fault diagnosis in induction machines via self-organized opera- tional neural networks,

S. Kilickaya, C. Celebioglu, L. Eren, and M. Askar, “Thermal image- based fault diagnosis in induction machines via self-organized opera- tional neural networks,” arXiv:2412.05901, 2024

2024 arXiv
[16]

Wavelet CNN-LSTM model for elec- tricity generation prediction in biomass thermal systems,

H. Sun, Y . Xu, and B. Zhang, “Wavelet CNN-LSTM model for elec- tricity generation prediction in biomass thermal systems,”IEEE Access, vol. 12, pp. 38012–38024, 2024

2024
[17]

Advanced thermal image pre-processing for defect segmentation using pulsed thermography,

F. Li, C. Zhang, and T. Liu, “Advanced thermal image pre-processing for defect segmentation using pulsed thermography,”Materials, vol. 17, no. 7, 2024

2024
[18]

Multimodal CNN-LSTM model for real-time predictive maintenance in data centers,

A. Kumar, S. Rao, and D. George, “Multimodal CNN-LSTM model for real-time predictive maintenance in data centers,” inProc. Int. Conf. Comput. Ind. Eng. (CIE), 2024

2024
[19]

Multimodal transformer for early alarm prediction,

W. Strem, D. S. Dhami, B. Schmidt, and K. Kersting, “Multimodal transformer for early alarm prediction,”Eng. Appl. Artif. Intell., vol. 127, art. no. 107349, 2024

2024
[20]

Machine learning-aided thermography for autonomous heat loss detection in buildings,

A. Waqas and M. T. Araji, “Machine learning-aided thermography for autonomous heat loss detection in buildings,”Energy Convers. Manag., vol. 304, art. no. 118243, 2024

2024
[21]

A U- Net CNN model application for identification of energy loss in infrared thermographic images,

D. Gertsvolf, M. Horvat, D. Aslam, A. Khademi, and U. Berardi, “A U- Net CNN model application for identification of energy loss in infrared thermographic images,”Appl. Energy, vol. 360, art. no. 122696, 2024

2024

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

A cyber-physical systems architec- ture for industry 4.0-based manufacturing systems,

J. Lee, B. Bagheri, and H. A. Kao, “A cyber-physical systems architec- ture for industry 4.0-based manufacturing systems,”Manuf. Lett., vol. 3, pp. 18–23, 2015

work page 2015

[2] [2]

Thermal image-based fault detection using segmentation and SVM classification,

J. Liu, H. Zhang, and Y . Wang, “Thermal image-based fault detection using segmentation and SVM classification,”J. Electron. Imaging, vol. 30, no. 2, pp. 150–165, 2023

work page 2023

[3] [3]

Proactive IT management with a predictive maintenance system,

K. Patel, M. Deshmukh, and S. Arora, “Proactive IT management with a predictive maintenance system,”Preprints.org, Feb. 2025. [Online]. Available: https://www.preprints.org/manuscript/202502.2062

work page arXiv 2025

[4] [4]

Predictive maintenance using deep learning: Enhancing reliability and reducing electrical system downtime,

J. Basit and A. Zeb, “Predictive maintenance using deep learning: Enhancing reliability and reducing electrical system downtime,”Int. J. Innov. Sci. Technol., vol. 6, no. 3, pp. 1120–1136, 2024

work page 2024

[5] [5]

Fault detection in electrical systems using pre-trained CNNs and thermal imagery,

C. Ukiwe, A. Sharma, and Y . Liu, “Fault detection in electrical systems using pre-trained CNNs and thermal imagery,” inProc. IEEE Int. Conf. Ind. Cyber-Phys. Syst. (ICPS), 2024

work page 2024

[6] [6]

Deep learning and its applications to machine health monitoring,

R. Zhao, R. Yan, Z. Chen, K. Mao, P. Wang, and R. X. Gao, “Deep learning and its applications to machine health monitoring,”Mech. Syst. Signal Process., vol. 115, pp. 213–237, 2019

work page 2019

[7] [7]

How transferable are features in deep neural networks?

J. Yosinski, J. Clune, Y . Bengio, and H. Lipson, “How transferable are features in deep neural networks?” inAdv. Neural Inf. Process. Syst., vol. 27, 2014

work page 2014

[8] [8]

ImageNet classification with deep convolutional neural networks,

A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep convolutional neural networks,” inAdv. Neural Inf. Process. Syst., vol. 25, 2012

work page 2012