Pith. sign in

REVIEW 4 major objections 5 minor 68 references

Multimodal Learning for Arcing Detection in Pantograph-Catenary Systems

T0 review · 4 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read The paper claims that combining synchronized pantograph-camera images with contact-force measurements detects electrical arcing in railway pantograph-catenary systems more accurately and robustly than either sensor alone, and it introduces

desk verdict Useful multimodal arc-detection paper, but the public benchmark's force channel is unvalidated synthetic data and the real-data results lack error bars; I'd send it to peer review. read the letter →

arxiv 2602.08792 v2 pith:IGILEHSK submitted 2026-02-09 cs.CV cs.AIcs.LG

Hao Dong , Eleni Chatzi , Olga Fink This is my paper
classification cs.CVcs.AIcs.LG
keywords pantograph-catenaryarcingdetectionmultimodalanomalysemi-superviseddeeplearningpseudo-anomalygenerationcontactforcesensingrailwayconditionmonitoringimage-forcefusionAUROCevaluation
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 tries to establish that pantograph-catenary arcing — the sparking from intermittent contact that erodes railway overhead equipment — is best detected by combining what a camera sees with what a contact-force sensor feels, rather than trusting either signal by itself. To make that case, the authors built two synchronized image-plus-force datasets: one from real diagnostic-train measurements and one from public videos paired with simulated force traces, the latter released as a public benchmark. They then introduce MultiDeepSAD, a multimodal semi-supervised anomaly detector that learns a compact normal region in a fused feature space, repels known anomalies with a smooth bounded penalty, and manufactures additional pseudo-anomalies from a handful of real arc examples. The model is reported to reach 93.49% and 95.03% AUROC on the two datasets, outperforming unimodal and classical baselines. If the results stand, railways could deploy a practical early-warning system for arcing using sensors that many trains already carry.

What carries the argument

The load-bearing mechanism is MultiDeepSAD, an extension of the DeepSAD objective to multiple modalities. DeepSAD is a semi-supervised anomaly detector that clusters normal samples around a center and repels known anomalies; MultiDeepSAD adds modality-specific encoders (a CNN for images, an MLP over the FFT magnitude spectrum for force) whose embeddings are concatenated and passed through an MLP. Training pushes normal fused representations toward a center while the anomaly penalty exp(-||phi(x)-c||^2) pushes labeled anomalies away with a smooth, bounded gradient. The second pillar is modality-specific pseudo-anomaly generation: arc regions cropped from real anomalies are pasted onto normal

What would settle it

Feed real force traces recorded during confirmed pantograph arcing events into a discriminator trained to distinguish real abnormal force from the paper's synthetic abnormal force; if the discriminator easily separates them, the generator is not representative, and the open-benchmark multimodal gain would need to be re-measured with real force data.

Watch

Extended reading notes

Core claim

The central discovery is that image and force modalities are complementary for pantograph-catenary arcing detection and that a semi-supervised multimodal anomaly detector can exploit that complementarity even when real arcing examples are extremely scarce. Concretely, the paper proposes MultiDeepSAD, which encodes each modality separately, fuses the embeddings, and optimizes a new loss that keeps normal samples compact around a hypersphere center while applying a bounded exponential penalty to labeled anomalous samples. Because labeled arcs are rare, the method fabricates training anomalies: arc regions cropped from real images are pasted onto normal pantograph images, and abnormal force tra

Load-bearing premise

The paper's own Section 3.2.2 generates the open dataset's force signals from a hand-crafted formula with chosen frequency ranges, amplitudes, and fault severity bands; the central claim's open-benchmark leg collapses if those synthetic signals do not capture how real arcing alters contact force, leaving the private real-measurement dataset as the only unexamined evidence.

Editorial extensions

If this is right

  • Multimodal fusion raises detection accuracy over either sensor alone: on the private real-data benchmark, image-only reaches 89.43% AUROC and force-only 70.66%, while the fused model reaches 93.49%.
  • Pseudo-anomaly generation makes the method data-efficient: with only two real arcing examples used for synthesis, AUROC jumps to 93.34% and stays above 92% up to five examples.
  • Frequency-domain FFT features are the most informative force representation, and late concatenation fusion outperforms gated and attention-based fusion in this setting.
  • The model retains high AUROC under image corruptions such as noise, blur, brightness change, and pixelation, with fog at strong levels as the main weakness.
  • A public dataset pairing internet-sourced pantograph videos with simulated force signals is released, providing a benchmark for future multimodal arcing research.

Reading between the lines

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

  • The open benchmark's force signals are generated by the paper's hand-coded simulator; if that simulator is not faithful to real arc-induced force dynamics, the multimodal gain measured on that benchmark is an artifact of the generator, and the private real-force dataset carries the evidential weight.
  • A deployment-ready system would likely need to add spatiotemporal tracking of repeated events into a cumulative damage index; the paper itself notes that single detections do not directly inform maintenance decisions.
  • The bounded exponential anomaly penalty is a generic device; it may improve other multimodal anomaly-detection tasks with scarce labeled anomalies, a hypothesis worth testing independently of arcing.
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

4 major / 5 minor

Summary. The paper addresses arcing detection in pantograph-catenary systems using a multimodal framework that fuses images and contact-force measurements. The authors construct two datasets: SBB-AD, a private dataset with real synchronized image/force data from Swiss Federal Railways, and Open-AD, a public dataset combining internet-sourced videos with synthetic force signals generated from hand-crafted rules. Methodologically, they propose MultiDeepSAD, an extension of DeepSAD to multimodal inputs with a new exponential anomaly penalty (Eq. 6) and modality-specific pseudo-anomaly generation: image paste for visuals and Mixup-style interpolation for force. On both datasets, the method achieves the highest AUROC (93.49% on SBB-AD, 95.03% on Open-AD), outperforming unimodal variants and several baselines. Extensive ablations cover corruption robustness, low-shot anomaly counts, force representations, fusion methods, backbones, and multimodal LLMs.

Significance. If the claims hold, the paper would make a useful contribution: a publicly released multimodal benchmark for arcing detection and a semi-supervised method that appears to exploit complementarity between visual and mechanical signals. The paper is well-structured and the ablation coverage is broad, including force representation choices, fusion mechanisms, and comparisons with multimodal LLMs. The pseudo-anomaly generation strategy is pragmatic and clearly described. However, the central evidence is weakened by two issues: the Open-AD force modality is entirely synthetic and generated from the same hand-crafted simulator that supplies the training pseudo-anomalies, and all performance numbers are reported as single AUROC values without error bars or significance tests. The real-data evidence from SBB-AD is limited to one private dataset with 164 positive test samples, and the paper does not provide details of the train/test split protocol. These gaps must be addressed before the main claims can be accepted.

major comments (4)
  1. [Sections 3.2.2, 5 and Tables 1–2] The Open-AD force modality is synthesized from Eq. (1) with hand-coded normal parameters and abnormal categories (30–90% force drops, 40–150% increases, burst noise). The same generator supplies abnormal force test labels and, through Force Mixup (Eq. 5), the pseudo-anomalies used for training. Consequently, the Open-AD force-only (83.30%) and multimodal (95.03%) results may largely reflect the ability of the network to memorize the simulator's class-conditional statistics rather than to detect real arc-induced force dynamics. The paper provides no validation that the statistical properties (PSD, transient duration, SNR) of these synthetic signals resemble the real SBB-AD force data. This is load-bearing for the abstract's claim of the "first publicly available dataset that integrates image and force data" and for the general claim of multimodal benefit. The authors should either validat
  2. [Section 5, Implementation Details and Tables 1–4, 6–8] Every AUROC in the paper is reported as a single number, with no standard deviations, number of random seeds, or statistical significance tests. The core claim that multimodal fusion outperforms image-only by ~4 AUROC points (SBB-AD: 89.43→93.49; Open-AD: 91.64→95.03) is evaluated on a test set of only 328 samples (SBB-AD). Without repeated runs, these differences may fall within run-to-run variability. The same issue affects Table 1 (MultiDeepSAD* vs MultiDeepSAD, supporting the new loss in Eq. 6) and Table 4 (sensitivity to anomaly count). At minimum, the authors should report mean ± std over multiple seeds and, for key comparisons, a paired significance test.
  3. [Section 3.1 and Table 2] The SBB-AD test set contains 164 normal and 164 abnormal one-second windows, but the paper does not describe how the train/test split was performed (e.g., by time, by run, or random), how arcing labels were obtained, or how the image and force streams were synchronized. If windows from the same continuous recordings are split randomly, temporal autocorrelation can inflate AUROC; if the arcing events come from only a few runs, the effective sample size is much smaller than 328. The real-data evidence is a central part of the validation of the multimodal approach, so the paper must provide a clear data-splitting and labeling protocol, including the number of distinct events and runs.
  4. [Section 4.3, Figure 5, and Table 6] Image pseudo-anomalies are generated by pasting cropped real arcing regions onto normal images. The paper does not analyze whether this introduces detectable visual shortcuts (e.g., sharp boundaries, brightness discontinuities, or mismatched context) that the model may exploit instead of learning the intrinsic appearance of an arc. Since the performance improvement from pseudo-anomaly generation is a key contribution (Table 6), the authors should include a control experiment—for example, pasting non-arc patches or evaluating on a held-out set of natural arcs—to demonstrate that the model is not overfitting to artifacts of the pasting operation.
minor comments (5)
  1. [Eq. (5) and Section 5] The Mixup hyperparameter α is defined in the method but its value is never specified in the experiments. Please state the chosen α and, ideally, its sensitivity.
  2. [Table 3] Corruption levels 1 and 3 are mentioned in the table and Figure 7, but no definition of the levels is given in the text. Please clarify how severity is quantified.
  3. [Table 5] The abbreviations GAF, MTF, and RP are used in the table without being defined in the caption. They are introduced in the text, but a caption-level definition would improve readability.
  4. [Section 1 and throughout] There are minor typographical issues, e.g., "witha new loss formulation" in Section 1; a copyedit pass is recommended.
  5. [Section 4.3, Eq. (5)] In Eq. (5), the notation d=λa1+(1−λ)a2 uses lowercase d for the mixed signal and a1/a2 for arrays; please use consistent vector notation and clarify indexing.

Circularity Check

1 steps flagged · score 6.0 of 10

Pseudo-anomaly training appears to draw labeled arcing samples from the test set, making the reported test AUROC partly in-sample; the Open-AD force channel is also an unvalidated synthetic benchmark.

  1. fitted input called prediction [Section 3.1 (SBB-AD split) and Section 4.3 (Pseudo-anomaly Image/Force Sample Generation); Table 4]
    "The dataset consists of 3,107 normal samples for training and a balanced test set of 328 samples, comprising 164 normal, 164 arcing anomaly instances. ... We first manually crop the arcing regions in the labeled anomaly samples ... For the force modality, we assume access to a small set of real anomalous force signals along with a substantial amount of normal samples."

    The dataset description places all 164 arcing instances in the test set and gives no labeled arcing samples for training. The pseudo-anomaly construction nevertheless requires 'labeled anomaly samples' (cropped arcing regions) and 'real anomalous force signals' to build the training set Da. On the stated split, these can only be taken from the test set. The model is therefore trained on test-derived anomaly content and then evaluated on the same test set, so the reported AUROC is partly an in-sample fit: for samples whose cropped/mixed versions entered training, the 'prediction' is forced by construction.

full rationale

The central empirical claim (MultiDeepSAD achieves 93.49% AUROC on SBB-AD and 95.03% on Open-AD) is not supported by a clean held-out evaluation as described. Section 3.1 defines the training set as 3,107 normal samples and a test set containing all 164 arcing positives; Section 4.3 then uses 'labeled anomaly samples' to generate pseudo-anomalies added to Da. If those labeled anomalies are the test arcs, the test AUROC is contaminated. The SBB-AD real-force results are the only independent force evidence, but they inherit the same pseudo-anomaly protocol and are reported without confidence intervals. Separately, the Open-AD force channel is entirely synthetic (Section 3.2.2) and never validated against real SBB force data; this limits external validity and makes the Open-AD force results a test of the authors' simulator rather than real force-based arcing, although I do not count that as a separate formal circularity because training and test are distinct draws from the same generator. No load-bearing self-citation or uniqueness-import argument appears. Score 6 reflects the central test-set contamination; it is not 8-10 because the model is not uniquely forced by a self-citation chain, and most test anomalies are unseen even under the leakage scenario.

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

The Open-AD benchmark's force channel is generated by a hand-crafted simulator (Eq. 1 and Section 3.2.2), so the force modality's test distribution is defined by the paper's own parameters. This makes the public benchmark only partially external; SBB-AD provides real force data but is private. The learning method itself rests on standard DeepSAD assumptions plus ad hoc pseudo-anomaly constructions.

free parameters (4)
  • Open-AD synthetic force generator parameters (normal and abnormal) = Fmean~U(20,30)N; f1,f2,f3 ranges; amplitudes; noise sigma=1.0/0.6N; drift d; transient probabilities; abnormal severity
    Hand-crafted in Eq. (1) and Section 3.2.2 to mimic real force; not fitted to or validated against real SBB force data, so Open-AD force labels are defined by these choices.
  • Anomaly penalty weight eta = 1
    Set in Section 5 Implementation Details with no sensitivity analysis; it controls the balance of the pseudo-anomaly term in Eq. (6).
  • Mixup Beta parameter alpha = not reported
    Pseudo-anomaly force generation in Eq. (5) uses lambda~Beta(alpha,alpha); alpha is not specified, affecting pseudo-anomaly hardness.
  • Number of pseudo-anomalies generated = not reported
    Image paste augmentations and force mixup counts are not given; training data size affects results and is unreported.
assumptions (4)
  • domain assumption A small set of labeled real anomalies D_a is representative of the true arcing distribution and sufficient to guide pseudo-anomaly generation.
    Used throughout Section 4.3 to justify pasting cropped arcs and mixup; if not representative, pseudo-anomalies mislead training.
  • ad hoc to paper Synthetic force signals in Eq. (1) and the abnormal generators capture the distinguishing physical signatures of real pantograph-catenary arcing.
    Section 3.2.2 states it models the physical reality but no validation against real force measurements; Open-AD force test set is generated from these rules.
  • domain assumption FFT magnitude spectrum is a sufficient force representation for arcing detection.
    Section 5 Implementation Details; chosen after ablation on SBB-AD, with no theoretical justification.
  • ad hoc to paper Pasting a cropped arcing region onto a normal image produces a valid pseudo-anomaly without introducing shortcuts.
    Section 4.3 pseudo-anomaly image generation; no check for paste artifacts leaking into test-time performance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal Learning for Arcing Detection in Pantograph-Catenary Systems." pith.science (2026). https://pith.science/paper/IGILEHSK

@misc{pith2026260208792,
  author       = {Pith},
  title        = {Pith review of: Multimodal Learning for Arcing Detection in Pantograph-Catenary Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IGILEHSK}},
  note         = {Machine review of arXiv:2602.08792}
}
read the original abstract

The pantograph-catenary interface is essential for ensuring uninterrupted and reliable power delivery in electrified rail systems. However, electrical arcing at this interface poses serious risks, including accelerated wear of contact components, degraded system performance, and potential service disruptions. Detecting arcing events at the pantograph-catenary interface is challenging due to their transient nature, noisy operating environment, data scarcity, and the difficulty of distinguishing arcs from other similar transient phenomena. To address these challenges, we propose a novel multimodal framework that combines high-resolution image data with force measurements to more accurately and robustly detect arcing events. First, we construct two arcing detection datasets comprising synchronized visual and force measurements. One dataset is built from data provided by the Swiss Federal Railways (SBB), and the other is derived from publicly available videos of arcing events in different railway systems and synthetic force data that mimic the characteristics observed in the real dataset. Leveraging these datasets, we propose MultiDeepSAD, an extension of the DeepSAD algorithm for multiple modalities with a new loss formulation. Additionally, we introduce tailored pseudo-anomaly generation techniques specific to each data type, such as synthetic arc-like artifacts in images and simulated force irregularities, to augment training data and improve the discriminative ability of the model. Through extensive experiments and ablation studies, we demonstrate that our framework significantly outperforms baseline approaches, exhibiting enhanced sensitivity to real arcing events even under domain shifts and limited availability of real arcing observations.

Figures

Figures reproduced from arXiv: 2602.08792 by the authors.

Figure 1
Figure 1. Proposed solution for arcing detection in pantograph-catenary systems with image and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Representative samples from the SBB-AD dataset, consisting of time-synchronized panto [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Examples from the Open-AD dataset, constructed from internet and simulation data. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Illustration of the DeepSAD objective: normal samples are embedded close to a hypersphere [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Illustration of pseudo-anomaly image samples generation, where we randomly paste arcing [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Illustration on pseudo-anomaly force sample generation, where we randomly combine [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Illustration of different corruptions on the original image (level 3). [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 6 linked inside Pith

  1. [1]

    Gated multimodal units for information fusion.arXiv preprint arXiv:1702.01992, 2017

    John Arevalo, Thamar Solorio, Manuel Montes-y Gómez, and Fabio A González. Gated multimodal units for information fusion.arXiv preprint arXiv:1702.01992, 2017

    arXiv 2017

  2. [2]

    Anomaly detection using a modified kernel-based tracking in the pantograph–catenary system.Expert Systems with Applications, 42(2):938–948, 2015

    Ilhan Aydin, Mehmet Karakose, and Erhan Akin. Anomaly detection using a modified kernel-based tracking in the pantograph–catenary system.Expert Systems with Applications, 42(2):938–948, 2015

    2015

  3. [3]

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan ...

    arXiv 2025

  4. [4]

    Arc detection in pantograph- catenary systems by the use of support vector machines-based classification.IET Electrical Systems in Transportation, 4(2):45–52, 2014

    Sami Barmada, Marco Raugi, Mauro Tucci, and Francesco Romano. Arc detection in pantograph- catenary systems by the use of support vector machines-based classification.IET Electrical Systems in Transportation, 4(2):45–52, 2014

    2014

  5. [5]

    Use of advanced signal processing techniques for arcing detection on ac pantograph catenary systems

    Sami Barmada, Mauro Tucci, et al. Use of advanced signal processing techniques for arcing detection on ac pantograph catenary systems. InIEEE International Conference on International Conference on Pantograph-Catenary Interaction Framework for Intelligent Control, 2011

    2011

  6. [6]

    Lof: identifying density-based local outliers

    Markus M Breunig, Hans-Peter Kriegel, Raymond T Ng, and Jörg Sander. Lof: identifying density-based local outliers. InProceedings of the 2000 ACM SIGMOD international conference on Management of data, pages 93–104, 2000

    2000

  7. [7]

    Pantograph– catenary interaction: recent achievements and future research challenges.International Journal of Rail Transportation, 6(2):57–82, 2018

    Stefano Bruni, Giuseppe Bucca, Marco Carnevale, Andrea Collina, and Alan Facchinetti. Pantograph– catenary interaction: recent achievements and future research challenges.International Journal of Rail Transportation, 6(2):57–82, 2018

    2018

  8. [8]

    High-speed railway pantograph-catenary anomaly detection method based on depth vision neural network.IEEE Transactions on Instrumentation and Measurement, 71:1–10, 2022

    Richeng Chen, Yunzhi Lin, and Tao Jin. High-speed railway pantograph-catenary anomaly detection method based on depth vision neural network.IEEE Transactions on Instrumentation and Measurement, 71:1–10, 2022

    2022

Show all 68 references
  1. [9]

    A reasoning approach-based pattern graph for analyzing the risk level of correlations among catenary components considering time distribution

    Liu Diyang, Gao Shibin, Luo Jiaming, Wei Xiaoguang, and Shi Jian. A reasoning approach-based pattern graph for analyzing the risk level of correlations among catenary components considering time distribution. Reliability Engineering & System Safety, 245:110035, 2024. 13

    2024

  2. [10]

    Liu Diyang, Gao Shibin, Wei Xiaoguang, Luo Jiaming, and Shi Jian. Impactability and susceptibility assessment based on ds evidence theory for analyzing the risk of fault propagation among catenary components.Reliability Engineering & System Safety, 251:110389, 2024

    2024

  3. [11]

    Nng-mix: Improving semi-supervised anomaly detection with pseudo-anomaly generation.IEEE Transactions on Neural Networks and Learning Systems, 2024

    Hao Dong, Gaëtan Frusque, Yue Zhao, Eleni Chatzi, and Olga Fink. Nng-mix: Improving semi-supervised anomaly detection with pseudo-anomaly generation.IEEE Transactions on Neural Networks and Learning Systems, 2024

    2024

  4. [12]

    Advances in multimodal adaptation and generalization: From traditional approaches to foundation models

    Hao Dong, Moru Liu, Kaiyang Zhou, Eleni Chatzi, Juho Kannala, Cyrill Stachniss, and Olga Fink. Advances in multimodal adaptation and generalization: From traditional approaches to foundation models. arXiv preprint arXiv:2501.18592, 2025

    2025

  5. [13]

    Multiood: Scaling out-of-distribution detection for multiple modalities.Advances in Neural Information Processing Systems, 37:129250–129278, 2024

    Hao Dong, Yue Zhao, Eleni Chatzi, and Olga Fink. Multiood: Scaling out-of-distribution detection for multiple modalities.Advances in Neural Information Processing Systems, 37:129250–129278, 2024

    2024

  6. [14]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    2010 arXiv

  7. [15]

    A novel algorithm for fast dc electric arc detection.Energies, 14(2):288, 2021

    Michał Doł˛ egowski and Mirosław Szmajda. A novel algorithm for fast dc electric arc detection.Energies, 14(2):288, 2021

    2021

  8. [16]

    Short-time fourier transform: two fundamental properties and an optimal implementation.IEEE Transactions on Signal Processing, 51(5):1231–1242, 2003

    Lutfiye Durak and Orhan Arikan. Short-time fourier transform: two fundamental properties and an optimal implementation.IEEE Transactions on Signal Processing, 51(5):1231–1242, 2003

    2003

  9. [17]

    Condition monitoring of the railway line and overhead equipment through onboard train measurement-an italian experience

    M Elia, Giorgio Diana, M Bocciolone, Stefano Bruni, Federico Cheli, Andrea Collina, and Ferruccio Resta. Condition monitoring of the railway line and overhead equipment through onboard train measurement-an italian experience. InIET International Conference on Railway Condition...

    2006

  10. [18]

    From physics to machine learning and back: Part ii-learning and observational bias in phm.arXiv preprint arXiv:2509.21207, 2025

    Olga Fink, Ismail Nejjar, Vinay Sharma, Keivan Faghih Niresi, Han Sun, Hao Dong, Chenghao Xu, Amaury Wei, Arthur Bizzi, Raffael Theiler, et al. From physics to machine learning and back: Part ii-learning and observational bias in phm.arXiv preprint arXiv:2509.21207, 2025

    2025

  11. [19]

    From physics to machine learning and back: Part i-learning with inductive biases in prognostics and health management.Reliability Engineering & System Safety, page 112213, 2026

    Olga Fink, Vinay Sharma, Ismail Nejjar, Leandro V on Krannichfeldt, Sergei Garmaev, Zepeng Zhang, Amaury Wei, Gaetan Frusque, Florent Forest, Mengjie Zhao, et al. From physics to machine learning and back: Part i-learning with inductive biases in prognostics and health managem...

    2026

  12. [20]

    Pantograph–catenary arcing detection based on electromagnetic radiation.IEEE Transactions on Electromagnetic Compatibility, 61(4):983–989, 2018

    Guoqiang Gao, Xin Yan, Zefeng Yang, Wenfu Wei, Yi Hu, and Guangning Wu. Pantograph–catenary arcing detection based on electromagnetic radiation.IEEE Transactions on Electromagnetic Compatibility, 61(4):983–989, 2018

    2018

  13. [21]

    Detection and monitoring system of the pantograph-catenary in high-speed railway (6c)

    SB Gao, ZG Liu, and Long Yu. Detection and monitoring system of the pantograph-catenary in high-speed railway (6c). In2017 7th International conference on power electronics systems and applications-smart mobility, power transfer & security (PESA), pages 1–7, 2017

    2017

  14. [22]

    Automatic detection and monitoring system of pantograph–catenary in china’s high-speed railways.IEEE Transactions on Instrumentation and Measurement, 70:1–12, 2020

    Shibin Gao. Automatic detection and monitoring system of pantograph–catenary in china’s high-speed railways.IEEE Transactions on Instrumentation and Measurement, 70:1–12, 2020

    2020

  15. [23]

    Deep anomaly detection using geometric transformations.Advances in neural information processing systems, 31, 2018

    Izhak Golan and Ran El-Yaniv. Deep anomaly detection using geometric transformations.Advances in neural information processing systems, 31, 2018

    2018

  16. [24]

    Assessment of catenary condition monitoring by means of pantograph head acceleration and artificial neural networks.Mechanical Systems and Signal Processing, 202:110697, 2023

    S Gregori, M Tur, J Gil, and FJ Fuenmayor. Assessment of catenary condition monitoring by means of pantograph head acceleration and artificial neural networks.Mechanical Systems and Signal Processing, 202:110697, 2023

    2023

  17. [25]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  18. [26]

    Reducing the dimensionality of data with neural networks

    Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313(5786):504–507, 2006

    2006

  19. [27]

    Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  20. [28]

    Chen Hong-ming, Zhou Ning, Cheng Yao, and Zhang Wei-hua. Non-contact detection method for pantograph-catenary contact force based on image processing.Proceedings of the Institution of Mechanical Engineers, Part F: Journal of Rail and Rapid Transit, 238(7):745–756, 2024

    2024

  21. [29]

    Arc detection and recognition in the pantograph-catenary system based on multi-information fusion.Transportation Research Record, 2674(10):229–240, 2020

    Shize Huang, Wei Chen, Bo Sun, Ting Tao, and Lingyu Yang. Arc detection and recognition in the pantograph-catenary system based on multi-information fusion.Transportation Research Record, 2674(10):229–240, 2020

    2020

  22. [30]

    Cluster analysis based arc detection in pantograph-catenary system.Journal of Advanced Transportation, 2018(1):1329265, 2018

    Shize Huang, Liangliang Yu, Fan Zhang, Wei Zhu, and Qiyi Guo. Cluster analysis based arc detection in pantograph-catenary system.Journal of Advanced Transportation, 2018(1):1329265, 2018

    2018

  23. [31]

    Arc detection and recognition in pantograph–catenary system based on convolutional neural network.Information Sciences, 501:363–376, 2019

    Shize Huang, Yachan Zhai, Miaomiao Zhang, and Xiaoxue Hou. Arc detection and recognition in pantograph–catenary system based on convolutional neural network.Information Sciences, 501:363–376, 2019. 14

    2019

  24. [32]

    Deep learning based arc detection in pantograph- catenary systems

    Gulsah Karaduman, Mehmet Karakose, and Erhan Akin. Deep learning based arc detection in pantograph- catenary systems. In2017 10th International Conference on Electrical and Electronics Engineering (ELECO), pages 904–908, 2017

    2017

  25. [33]

    Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

    Diederik P Kingma. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  26. [34]

    Dynamic distinction learning: adaptive pseudo anomalies for video anomaly detection

    Demetris Lappas, Vasileios Argyriou, and Dimitrios Makris. Dynamic distinction learning: adaptive pseudo anomalies for video anomaly detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3961–3970, 2024

    2024

  27. [35]

    Theoretical modeling and fault diagnosis of pantograph–catenary arc under high-speed airflow.Scientific Reports, 15(1):9233, 2025

    Bin Li, Jiahui Shu, and Diansong Du. Theoretical modeling and fault diagnosis of pantograph–catenary arc under high-speed airflow.Scientific Reports, 15(1):9233, 2025

    2025

  28. [36]

    Cutpaste: Self-supervised learning for anomaly detection and localization

    Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9664–9674, 2021

    2021

  29. [37]

    Virtual point tracking method for online detection of relative wheel-rail displacement of railway vehicles

    Haoqian Li, Yong Wang, Jing Zeng, Fansong Li, Zhenhuan Yang, Guiming Mei, and Yunguang Ye. Virtual point tracking method for online detection of relative wheel-rail displacement of railway vehicles. Reliability Engineering & System Safety, 246:110087, 2024

    2024

  30. [38]

    Isolation forest

    Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation forest. In2008 eighth ieee international conference on data mining, pages 413–422, 2008

    2008

  31. [39]

    Extremely simple multimodal outlier synthesis for out-of-distribution detection and segmentation.arXiv preprint arXiv:2505.16985, 2025

    Moru Liu, Hao Dong, Jessica Kelly, Olga Fink, and Mario Trapp. Extremely simple multimodal outlier synthesis for out-of-distribution detection and segmentation.arXiv preprint arXiv:2505.16985, 2025

    2025

  32. [40]

    Xiaohong Liu, Xiaoyu Wang, Wei Quan, Guoxin Gu, Xiaoqian Xu, and Shibin Gao. A pantograph-catenary arcing detection model for high-speed railway based on semantic segmentation and generative adversarial network.International Journal of Rail Transportation, pages 1–22, 2024

    2024

  33. [41]

    A novel arcing detection model of pantograph–catenary for high-speed train in complex scenes.IEEE Transactions on Instrumentation and Measurement, 72:1–13, 2023

    Yang Liu, Wei Quan, Xuemin Lu, Xiaohong Liu, Shibin Gao, Haiquan Zhao, Long Yu, and Jie Zheng. A novel arcing detection model of pantograph–catenary for high-speed train in complex scenes.IEEE Transactions on Instrumentation and Measurement, 72:1–13, 2023

    2023

  34. [42]

    The electrical behaviour of railway pantograph arcs.Energies, 16(3):1465, 2023

    Andrea Mariscotti. The electrical behaviour of railway pantograph arcs.Energies, 16(3):1465, 2023

    2023

  35. [43]

    Knowledge-embedding deep interpretable graph model for wear prediction: Application in pantograph-catenary systems.Reliability Engineering & System Safety, 261:111140, 2025

    Yutao Mo, Yizhen Peng, Jing Wu, and Kangbo Fan. Knowledge-embedding deep interpretable graph model for wear prediction: Application in pantograph-catenary systems.Reliability Engineering & System Safety, 261:111140, 2025

    2025

  36. [44]

    Recall and refine: A simple but effective source-free open-set domain adaptation framework.arXiv preprint arXiv:2411.12558, 2024

    Ismail Nejjar, Hao Dong, and Olga Fink. Recall and refine: A simple but effective source-free open-set domain adaptation framework.arXiv preprint arXiv:2411.12558, 2024

    2024 arXiv

  37. [45]

    Springer, 1982

    Henri J Nussbaumer and Henri J Nussbaumer.The fast Fourier transform. Springer, 1982

    1982

  38. [46]

    Contact force control in multibody pantograph/catenary systems.Proceedings of the Institution of Mechanical Engineers, Part K: Journal of Multi-body Dynamics, 230(4):307–328, 2016

    Carmine M Pappalardo, Mohil D Patel, Brian Tinsley, and Ahmed A Shabana. Contact force control in multibody pantograph/catenary systems.Proceedings of the Institution of Mechanical Engineers, Part K: Journal of Multi-body Dynamics, 230(4):307–328, 2016

    2016

  39. [47]

    Arc- mask: a robust and fast image-based method for high-speed railway pantograph-catenary arcing instance segmentation.Neural Computing and Applications, 35(9):6875–6890, 2023

    Wei Quan, Shaopeng Guo, Xuemin Lu, Guoxin Gu, Wenjing Cheng, Ning Zhou, and Long Yu. Arc- mask: a robust and fast image-based method for high-speed railway pantograph-catenary arcing instance segmentation.Neural Computing and Applications, 35(9):6875–6890, 2023

    2023

  40. [48]

    Arcse: A dual-branch semantic segmentation model for robust pantograph-catenary arcing detection in complex environments.IEEE Transactions on Instrumentation and Measurement, 2024

    Wei Quan, Xiaoqian Xu, Xiaolin Liu, Xiaoyu Wang, and Shibin Gao. Arcse: A dual-branch semantic segmentation model for robust pantograph-catenary arcing detection in complex environments.IEEE Transactions on Instrumentation and Measurement, 2024

    2024

  41. [49]

    Efficient algorithms for mining outliers from large data sets

    Sridhar Ramaswamy, Rajeev Rastogi, and Kyuseok Shim. Efficient algorithms for mining outliers from large data sets. InProceedings of the 2000 ACM SIGMOD international conference on Management of data, pages 427–438, 2000

    2000

  42. [50]

    Deep one-class classification

    Lukas Ruff, Robert Vandermeulen, Nico Goernitz, Lucas Deecke, Shoaib Ahmed Siddiqui, Alexander Binder, Emmanuel Müller, and Marius Kloft. Deep one-class classification. InInternational conference on machine learning, pages 4393–4402. PMLR, 2018

    2018

  43. [51]

    Deep semi-supervised anomaly detection.arXiv preprint arXiv:1906.02694, 2019

    Lukas Ruff, Robert A Vandermeulen, Nico Görnitz, Alexander Binder, Emmanuel Müller, Klaus-Robert Müller, and Marius Kloft. Deep semi-supervised anomaly detection.arXiv preprint arXiv:1906.02694, 2019

    1906 arXiv

  44. [52]

    f-anogan: Fast unsupervised anomaly detection with generative adversarial networks.Medical image analysis, 54:30–44, 2019

    Thomas Schlegl, Philipp Seeböck, Sebastian M Waldstein, Georg Langs, and Ursula Schmidt-Erfurth. f-anogan: Fast unsupervised anomaly detection with generative adversarial networks.Medical image analysis, 54:30–44, 2019

    2019

  45. [53]

    Estimating the support of a high-dimensional distribution.Neural computation, 13(7):1443–1471, 2001

    Bernhard Schölkopf, John C Platt, John Shawe-Taylor, Alex J Smola, and Robert C Williamson. Estimating the support of a high-dimensional distribution.Neural computation, 13(7):1443–1471, 2001

    2001

  46. [54]

    Unseen visual anomaly generation

    Han Sun, Yunkang Cao, Hao Dong, and Olga Fink. Unseen visual anomaly generation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 25508–25517, 2025. 15

    2025

  47. [55]

    Defect diagnosis of rigid catenary system based on pantograph vibration performance

    Mengying Tan, Haishangyang Li, and Lei Nie. Defect diagnosis of rigid catenary system based on pantograph vibration performance. InActuators, volume 13, page 162. MDPI, 2024

    2024

  48. [56]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  49. [57]

    Uncertainty- aware trustworthy weather-driven failure risk predictor for overhead contact lines.Reliability Engineering & System Safety, 242:109734, 2024

    Jian Wang, Shibin Gao, Long Yu, Xingyang Liu, Ferrante Neri, Dongkai Zhang, and Lei Kou. Uncertainty- aware trustworthy weather-driven failure risk predictor for overhead contact lines.Reliability Engineering & System Safety, 242:109734, 2024

    2024

  50. [58]

    Jian Wang, Shibin Gao, Long Yu, Chaoqun Ma, Dongkai Zhang, and Lei Kou. A data-driven integrated framework for predictive probabilistic risk analytics of overhead contact lines based on dynamic bayesian network.Reliability Engineering & System Safety, 235:109266, 2023

    2023

  51. [59]

    Review of research development of pantograph- catenary arc model and electrical characteristics.Journal of the China Railway Society, 35(8):35–43, 2013

    Ying Wang, Zhigang Liu, Fuqiang Fan, and S Bin Gao. Review of research development of pantograph- catenary arc model and electrical characteristics.Journal of the China Railway Society, 35(8):35–43, 2013

    2013

  52. [60]

    Encoding time series as images for visual inspection and classification using tiled convolutional neural networks

    Zhiguang Wang, Tim Oates, et al. Encoding time series as images for visual inspection and classification using tiled convolutional neural networks. InAAAI Workshops, 2015

    2015

  53. [61]

    Diagnosis and detection of service performance of pantograph and catenary.The Electrical Contact of the Pantograph-Catenary System: Theory and Application, pages 221–278, 2019

    Guangning Wu, Guoqiang Gao, Wenfu Wei, Zefeng Yang, Guangning Wu, Guoqiang Gao, Wenfu Wei, and Zefeng Yang. Diagnosis and detection of service performance of pantograph and catenary.The Electrical Contact of the Pantograph-Catenary System: Theory and Application, pages 221–278, 2019

    2019

  54. [62]

    Causal correlation inference between fault catenary components based on kolmogorov complexity.Reliability Engineering & System Safety, page 111805, 2025

    Wu Xinyuan, Gao Shibin, Wei Xiaoguang, Liu Diyang, and Ling Weize. Causal correlation inference between fault catenary components based on kolmogorov complexity.Reliability Engineering & System Safety, page 111805, 2025

    2025

  55. [63]

    Research on multimodal techniques for arc detection in railway systems with limited data.Structural Health Monitoring, page 14759217251336797, 2025

    Jingke Yan, Yao Cheng, Fan Zhang, Mudi Li, Ning Zhou, Bo Jin, Hui Wang, Haonan Yang, and Weihua Zhang. Research on multimodal techniques for arc detection in railway systems with limited data.Structural Health Monitoring, page 14759217251336797, 2025

    2025

  56. [64]

    A novel arc detection and identification method in pantograph-catenary system based on deep learning.Scientific Reports, 15(1):3511, 2025

    Yue Yan, Hu Liu, Linfeng Gan, and Runtong Zhu. A novel arc detection and identification method in pantograph-catenary system based on deep learning.Scientific Reports, 15(1):3511, 2025

    2025

  57. [65]

    Arc recognition for pantograph-catenery system of railway using current.Measurement, 251:117304, 2025

    Gang Yang, Guowei Kong, and Xin Shen. Arc recognition for pantograph-catenery system of railway using current.Measurement, 251:117304, 2025

    2025

  58. [66]

    Identification method for series arc faults based on wavelet transform and deep neural network.Energies, 13(1):142, 2019

    Qiongfang Yu, Yaqian Hu, and Yi Yang. Identification method for series arc faults based on wavelet transform and deep neural network.Energies, 13(1):142, 2019

    2019

  59. [67]

    Wavelet transform

    Dengsheng Zhang. Wavelet transform. InFundamentals of image data mining: Analysis, Features, Classification and Retrieval, pages 35–44. Springer, 2019

    2019

  60. [68]

    mixup: Beyond empirical risk minimization

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. InInternational Conference on Learning Representations, 2018. 16

    2018

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.