Pith. sign in

REVIEW 3 major objections 4 minor 61 references

Seeing the Unseen: Towards Training-Free Inspection for Wind Turbine Blades Using Knowledge-Augmented Vision Language Models

T0 review · 3 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Wind turbine blade damage can be classified without task-specific training, the paper argues, by retrieving curated domain knowledge at inference time and feeding it to a vision-language model.

desk verdict Plausible engineering demonstration, but the 100% accuracy claim is unvalidated; the near-duplicate leakage between the knowledge base and the test set is a real, load-bearing weakness. read the letter →

arxiv 2510.22868 v2 pith:ZSX2725Y submitted 2025-10-26 cs.CV

Yang Zhang , Qianyu Zhou , Farhad Imani , Jiong Tang This is my paper
classification cs.CV
keywords windturbinebladeinspectionvision-languagemodelretrieval-augmentedgenerationzero-shotclassificationdamagedetectionmultimodalknowledgebaseexplainabilitysmall-sampleevaluation
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 a general vision-language model can inspect wind turbine blades for damage without any task-specific training, provided it can retrieve curated domain knowledge at inference time. On 30 labeled blade images spanning healthy, surface, environmental, and structural damage, the retrieval-grounded model classified all 30 correctly; the same model without retrieval reached 93.33% accuracy and 90.91% precision, recall, and F1. The authors' point is that knowledge injection—damage descriptions, blade material facts, maintenance-log precedents, and reference photos—can substitute for labeled training data, making inspection adaptable to rare or new defect types. Each decision is also traceable to specific retrieved evidence, which the authors treat as the framework's main advantage over supervised detectors and open-vocabulary baselines.

What carries the argument

The load-bearing component is the RAG-grounded VLM pipeline: a hybrid knowledge base holding damage descriptions, turbine background facts, simulated maintenance logs, and image–text reference pairs; dual encoders (a sentence-transformer for text and a contrastive vision-language model for images) that embed entries into a vector store; a two-stage retrieval process combining nearest-neighbor search with keyword-count and document-length reranking to retain the top three pieces of evidence per modality; and a dynamic prompt that injects that evidence into a vision-language model at inference time. Retrieval is the component that converts a generic VLM into a domain-grounded inspector without

What would settle it

Verify disjointness by computing image hashes or embedding distances between every query image and every knowledge-base entry and removing near-duplicates from the test set, then re-run the 30-image evaluation; if perfect accuracy persists on the provably unseen subset, the retrieval is augmenting, and if it drops, the reported gain is partly lookup.

Watch

Extended reading notes

Core claim

The central claim is that a multimodal retrieval-augmented vision-language model achieves perfect classification (100% accuracy; Clopper-Pearson 95% CI [88.4%, 100%]) on 30 wind turbine blade images, while the same VLM without retrieval achieves 93.33% accuracy and 90.91% precision/recall/F1. The mechanism is inference-time conditioning: a hybrid text-image retriever with keyword-aware reranking selects the three most relevant text chunks and three reference images, and a dynamic prompt asks the VLM to report blade count, damage presence, damage type from a fixed taxonomy, severity, and explanatory rationale. The ablation attributes the improvement to domain knowledge compensating for low-li

Load-bearing premise

The load-bearing premise is that the test images are genuinely unseen by the knowledge base, yet all knowledge-base reference images come from the same drone-footage dataset that supplies many test images, so some 'retrieval' could be recognizing near-duplicates rather than reasoning about unfamiliar input.

Editorial extensions

If this is right

  • If the central claim holds, new or rare damage types can be added to the system by writing descriptions and adding reference images to the knowledge base, with no retraining cycle.
  • Inspection decisions become auditable: every diagnosis can be traced to specific retrieved references, which the authors argue supports maintenance planning and trust.
  • The same architecture can be pointed at other visual inspection domains where labeled data are scarce—bridge inspection, corrosion monitoring, manufacturing defect detection—because no task-specific training is needed.
  • Without retrieval, the VLM failed on one low-light structural crack (false negative) and on surface contamination mistaken for cracks (false positive); retrieval corrected both, suggesting knowledge injection can compensate for visual ambiguity.
  • On the reported 30-image test set, the framework achieved perfect scores with a Clopper-Pearson lower bound of 88.4%, meaning even under small-sample uncertainty the lower confidence bound remains high.

Reading between the lines

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

  • The headline 100% has an open caveat: knowledge-base reference images come from only one of the two source datasets, and some test images are drawn from the same video footage, so part of the apparent success could be retrieval of near-duplicate frames rather than genuine zero-shot generalization. The paper itself notes this provenance in §3.2.
  • The evaluation counts multi-label predictions as correct at the level of four mechanism-based groups (healthy, surface, environmental, structural) rather than the underlying 14-type taxonomy, and the ground-truth labels were assigned by the authors; a stricter per-type matching rule would be a harder test of whether the VLM truly distinguishes crack from corrosion.
  • A direct test of the mechanism would be to hold out entire damage types or source datasets from the knowledge base and re-measure accuracy; if performance survives, retrieval is generalizing; if it collapses, the method is in part an answer-key lookup.
  • The 88.4% lower confidence bound already signals that a 30-image perfect score is consistent with several real errors; reporting per-type counts rather than group-level accuracy would let readers judge practical value.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes a training-free ("towards zero-shot") wind turbine blade inspection framework that combines a vision-language model (Qwen-VL-Max) with retrieval-augmented generation. A multimodal knowledge base is built from three textual sources and an image-text metadata set, using dual text/image embeddings and keyword-aware reranking. At inference, the top retrieved text chunks and reference images are inserted into a dynamic prompt, and the VLM outputs a structured damage assessment (blade count, damage presence, type, severity, rationale, and retrieved context). The evaluation uses 30 labeled blade images from Chen and Foster et al. datasets, plus three supplemental images, organized into four mechanism-based groups. The paper reports 100% accuracy for the RAG-grounded VLM with Clopper-Pearson 95% CI [88.4%, 100%], versus 93.33% accuracy and 90.91% precision/recall/F1 for the same VLM without RAG, and states that RAG yields significant improvements, especially on a low-light crack case and a dirt-as-crack false-positive case. Additional comparisons against YOLO-World and a retrained YOLOv8n are used to argue for the zero-shot generalizability and explainability of the framework.

Significance. If the central empirical claim were established, the framework would be a useful contribution: it provides a data-efficient, interpretable inspection pipeline with provenance to retrieved knowledge items, and it ships open-sourced code, raw model responses, and uncertainty intervals. The no-RAG ablation is a genuine control, and the inclusion of full responses for several test images is a transparency strength. However, the current evidence is not yet sufficient to support the strong claim of a decisive RAG advantage. The test set is small, the 30/30 versus 28/30 difference is not statistically significant by standard tests, and the evaluation protocol has a possible data-leakage pathway because the knowledge-base reference images are drawn from the same Chen video dataset from which test frames were randomly sampled. These issues are load-bearing for the paper's central claim and need to be addressed before the results can be accepted at face value.

major comments (3)
  1. [§3.1 and §3.2] The evaluation does not rule out near-duplicate leakage between the knowledge base and the test set. §3.1 says test images were "randomly selected from both the training and testing datasets" of the Chen video dataset, while §3.2 says "the reference images included in the knowledge base are drawn only from the Chen datasets." Since the Chen data are videos, frames from the same clip can be near-duplicates of one another. If a test image's near-duplicate appears in the KB, the retrieved image-text metadata (which includes explicit damage descriptions, e.g., Table 6) effectively supplies the ground truth. The example in §4.1/Image 17 is concrete: the context used includes img_Crack_crk2-5, a Chen image. The paper must quantify frame overlap, exclude same-video or near-duplicate frames from the KB, or otherwise demonstrate that the two corrected cases are not lookup successes. Without this,
  2. [§4.1 and Abstract] The claim of "significant performance improvements" for RAG is not supported by the reported numbers. The comparison is 30/30 correct (Clopper-Pearson 95% CI [88.4%, 100%]) versus 28/30 correct (95% CI [77.9%, 99.2%]). These intervals overlap substantially, and a Fisher exact test on the 2x2 table gives p≈0.49. The entire difference comes from one missed structural crack and one healthy image misclassified as structural. The paper should report the exact statistical test and p-value, temper the language to remove "significant," and frame the RAG benefit as an illustrative case study rather than a demonstrated statistically significant gain. A larger evaluation or a preference-based analysis of the two failure cases would be needed to justify the current conclusion.
  3. [§3.3, Table 7, and §4.1] The evaluation has a design-level circularity that is not fully addressed. The authors authored the KB damage descriptions and the 14-type/4-group taxonomy, assigned the ground-truth labels, and then graded the model against those labels. More importantly, the multi-label matching rule is not stated. For example, Table 7 lists Image 7's ground truth as "peeling/corrosion/rust/erosion/crack" but the extracted type is "Corrosion, Rust, Crack," and Image 12's ground truth is "Lightning Strike/Burning/fracture" while the extracted type is "Burn Mark, Delamination, Lightning Strike/Burning." The paper says the confusion matrix is based on four groups but does not define how partial, extra, or missing fine-grained labels are counted as correct. Without an explicit matching rule and per-image label agreement, the reported perfect accuracy is not reproducible. Please state the matching rule, pro
minor comments (4)
  1. [Title and Abstract] The title says "Training-Free" while the abstract and body use "towards zero-shot" and define zero-shot as absence of task-specific training. Please harmonize the terminology. Also, the phrase "significantly" in the abstract should be removed or supported by a statistical test.
  2. [Equation (7)] Equation (7) is typeset with garbled subscripts and the Beta quantile notation is incomplete. It should define Beta^{-1}_p(a,b) or use standard quantile notation. As written, the piecewise cases are hard to read.
  3. [§4.2] The YOLO-World results are reported as fully failing (e.g., labeling blades as airplanes and damage as birds), but no quantitative detection metrics or sample boxes are shown in the text. Please provide a small quantitative summary or at least state the number of images where any valid detection occurred.
  4. [Data availability] The code repository is referenced only as a URL without a version/commit hash. For reproducibility, please include the exact commit or version, and specify the Qwen-VL-Max API model version/date used.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial evaluation circularity: the 100% accuracy claim is not shown to be free of near-duplicate lookup, because test images and knowledge-base references come from the same Chen video source; the decisive corrected case retrieves an entry named img_Crack_crk2-5.

  1. other [§3.1, §3.2 and §4.1 / Table 7 (Image 17)]
    "Since our goal is to achieve zero-shot inspection capability, we randomly selected images from both the training and testing datasets, rather than adhering to their original division. ... the reference images included in the knowledge base are drawn only from the Chen datasets [56-57] ... Context Used: ... Image Docs: img_Crack_crk2-5, img_Unspecified_hole4, img_Unspecified_hole3"

    The 100% claim presupposes that the test images are unseen by the knowledge base. The paper's setup fails to establish this: §3.1 samples test frames from both Chen train and test videos, while §3.2 restricts all KB reference images to those same Chen datasets; because Chen is video, near-duplicate frames from one clip can appear on both sides. For the decisive correction (Image 17), the model's predicted label 'crack' is explicitly grounded in the retrieved KB entry img_Crack_crk2-5, whose filename and metadata already contain that label. If the query is a near-duplicate of that frame, the output is answer lookup rather than zero-shot generalization. The no-RAG ablation is a real control and the framework has independent content, so the circularity is partial.

full rationale

No fitted parameter or closed-form derivation reduces to its input by construction: the similarity-search and reranking equations are standard RAG machinery, and there is no load-bearing self-citation or imported uniqueness theorem. The no-RAG ablation (93.33% vs 100%) and the three externally sourced test images give the framework independent content. The circularity is confined to the evaluation design: the paper never establishes that Chen-derived test frames are absent from the Chen-derived knowledge base, and it concedes only that 'part of the test images originates from different sources.' Together with the explicit use of label-bearing KB entries (img_Crack_crk2-5) to correct the two hard failures, this makes the headline 100% partly dependent on near-duplicate retrieval. This is design-level partial circularity rather than a fully forced result, hence score 4.

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

The central evaluation rests on two author-controlled assets: the knowledge-base texts (written by the authors to describe exactly the 14 damage types and 4 groups used in grading) and the ground-truth labels (assigned by the authors). No independent benchmark, inter-annotator check, or sensitivity analysis is provided; the proprietary VLM and encoder choices are assumed reliable. The most fragile premise is that the test images are genuinely unseen relative to the KB, since both draw on the same Chen drone-video dataset. These are the premises a reader should audit before accepting the 100% figure.

free parameters (6)
  • top_k (retrieval candidates per modality) = 5
    §2.2.3: initial retrieval count, hand-set 'to strike an optimal balance'; no sweep reported.
  • top_n (post-rerank documents) = 3
    §2.2.3: final context size per modality; no sensitivity analysis.
  • chunk_size / chunk_overlap = 1000 / 200 chars
    §2.2.1: document chunking parameters chosen without ablation.
  • Damage taxonomy in prompt = 14 named types
    Table 5: the closed list of 14 damage types given to the VLM was authored to cover the test classes.
  • Four-group mechanism-based collapse = healthy/surface/environmental/structural
    §3.1: coarse 4-group scoring drives the 100% accuracy; set by the authors, not by an external standard.
  • Severity extraction patterns and length-normalizer divisor = 1000
    §2.2.3 Eq. (4): heuristic constants not optimized or validated against a labeled severity set.
assumptions (5)
  • domain assumption Damage-description texts in the knowledge base (colors, shapes, locations of each defect) accurately describe the damage in the test images
    §3.2/Table 6: KB entries are authored by the paper's team; no independent source or inter-annotator check is given.
  • domain assumption Ground-truth labels for the 30 images (multi-label, grouped into 4 classes) are correct
    §3.3: 'ground-truth labels established by domain experts' — the authors are the experts; the full label table is not published.
  • standard math CLIP and Sentence-BERT embedding spaces are semantically meaningful for cross-modal retrieval of blade damage
    §2.2.2: off-the-shelf pretrained encoders; standard background ML assumption.
  • domain assumption Qwen-VL-Max produces reliable answers that follow the retrieved context, and rule-based post-processing faithfully extracts them
    §2.3/Table 7: extraction errors (severity 4→1; 'Wear' inserted where the model said 'closest to Ice and Snow') show this assumption is violated in at least one shown case.
  • ad hoc to paper Test images are not near-duplicates of knowledge-base reference images
    §3.2: KB references come from the Chen dataset while part of the test set also comes from Chen; overlap is acknowledged only qualitatively and never quantified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Seeing the Unseen: Towards Training-Free Inspection for Wind Turbine Blades Using Knowledge-Augmented Vision Language Models." pith.science (2026). https://pith.science/paper/ZSX2725Y

@misc{pith2026251022868,
  author       = {Pith},
  title        = {Pith review of: Seeing the Unseen: Towards Training-Free Inspection for Wind Turbine Blades Using Knowledge-Augmented Vision Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZSX2725Y}},
  note         = {Machine review of arXiv:2510.22868}
}
read the original abstract

Wind turbine blades operate in harsh environments, making timely damage detection essential for preventing failures and optimizing maintenance. Drone-based inspection and deep learning are promising, but typically depend on large, labeled datasets, which limit their ability to detect rare or evolving damage types. To address this, we propose a zero-shot-oriented inspection framework that integrates Retrieval-Augmented Generation (RAG) with Vision-Language Models (VLM). A multimodal knowledge base is constructed, comprising technical documentation, representative reference images, and domain-specific guidelines. A hybrid text-image retriever with keyword-aware reranking assembles the most relevant context to condition the VLM at inference, injecting domain knowledge without task-specific training. We evaluate the framework on 30 labeled blade images covering diverse damage categories. Although the dataset is small due to the difficulty of acquiring verified blade imagery, it covers multiple representative defect types. On this test set, the RAG-grounded VLM correctly classified all samples, whereas the same VLM without retrieval performed worse in both accuracy and precision. We further compare against open-vocabulary baselines and incorporate uncertainty Clopper-Pearson confidence intervals to account for the small-sample setting. Ablation studies indicate that the key advantage of the framework lies in explainability and generalizability: retrieved references ground the reasoning process and enable the detection of previously unseen defects by leveraging domain knowledge rather than relying solely on visual cues. This research contributes a data-efficient solution for industrial inspection that reduces dependence on extensive labeled datasets.

Figures

Figures reproduced from arXiv: 2510.22868 by the authors.

Figure 1
Figure 1. General flowchart for wind turbine blade inspection using RAG and VLM. [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Schematic diagram of retrieval augmented generation. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Structured output extraction process [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Schematic diagram of photography of drone for wind turbine [PITH_FULL_IMAGE:figures/full_fig_p018_4.png]
Figure 5
Figure 5. Figure 5: Testing wind turbine blade images from datasets [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: Confusion matrix (a) and evaluation metrics (b) for testing datasets [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]
Figure 7
Figure 7. Figure 7: Severity distributions (a) and response time distributions (b) among testing datasets. Another noteworthy aspect is the distribution of response times, as shown in [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]
Figure 8
Figure 8. Figure 8: Knowledge usages of testing image 1. This hybrid approach to knowledge retrieval offers several advantages. The textual knowledge provides valuable context about damage types, severity levels, and assessment challenges. The image references enable direct visual compari…
Figure 9
Figure 9. Figure 9: (a) Confusion matrix and (b) evaluation metrics for testing datasets. We further analyze the two incorrectly classified images as shown in [PITH_FULL_IMAGE:figures/full_fig_p027_9.png]
Figure 10
Figure 10. Figure 10: Images misidentified with the system without RAG. [PITH_FULL_IMAGE:figures/full_fig_p027_10.png]
Figure 11
Figure 11. Figure 11: Knowledge usages of testing image 17. The RAG approach enhances performance by enriching feature representation with contextualized domain knowledge. During inference, the model retrieves relevant documents and reference images, gaining access to specialized expertise…
Figure 12
Figure 12. Figure 12: Detection results on testing images from YOLO world model. [PITH_FULL_IMAGE:figures/full_fig_p030_12.png]
Figure 13
Figure 13. Figure 13: Detection results on testing images from retrained YOLO model. [PITH_FULL_IMAGE:figures/full_fig_p031_13.png]
Figure 14
Figure 14. Figure 14: Detection results on testing images from proposed approach. [PITH_FULL_IMAGE:figures/full_fig_p032_14.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 2 canonical work pages

  1. [1]

    and Fæster, S., 2021

    Mishnaevsky Jr, L., Hasager, C.B., Bak, C., Tilg, A.M., Bech, J.I., Rad, S.D. and Fæster, S., 2021. Leading edge erosion of wind turbine blades: Understanding, prevention and protection. Renewable Energy, 169, pp.953-969

    2021

  2. [2]

    and Weaver, P.M., 2023

    Kong, K., Dyer, K., Payne, C., Hamerton, I. and Weaver, P.M., 2023. Progress and trends in damage detection methods, maintenance, and data -driven monitoring of wind turbine blades –A review. Renewable Energy Focus, 44, pp.390-412

    2023

  3. [3]

    and See, J., 2025

    Gohar, I., Yew, W.K., Halimi, A. and See, J., 2025. Review of state-of-the-art surface defect detection on wind turbine blades through aerial imagery: Challenges and recommendations. Engineering Applications of Artificial Intelligence, 144, p.109970

    2025

  4. [4]

    and Zhou, L., 2025

    Jiang, L., Zhang, S.P., Shen, G.Q. and Zhou, L., 2025. Acoustic Emission- based wind turbine blade icing monitoring using deep learning technology. Renewable Energy, p.122980

    2025

  5. [5]

    and Ong, C.P .K., 2017

    Meng, M., Chua, Y .J., Wouterson, E. and Ong, C.P .K., 2017. Ultrasonic signal classification and imaging system for composite materials via deep convolutional neural networks. Neurocomputing, 257, pp.128-135

    2017

  6. [6]

    and Lanzagorta, J.L., 2025

    Mendikute, J., Carmona, I., Aizpurua, I., Bediaga, I., Castro, I., Galdos, L. and Lanzagorta, J.L., 2025. Defect detection in wind turbine blades applying convolutional neural networks to ultrasonic testing. NDT & E International, 154, p.103359

    2025

  7. [7]

    and Gu, Y ., 2022

    Wang, C. and Gu, Y ., 2022. Research on infrared nondestructive detection of small wind turbine blades. Results in Engineering, 15, p.100570

    2022

  8. [8]

    and Masoum, M.A.S., 2024

    Memari, M., Shakya, P., Shekaramiz, M., Seibi, A.C. and Masoum, M.A.S., 2024. Review on the advancements in wind turbine blade inspection: Integrating drone and deep learning technologies for enhanced defect detection. IEEE Access, 12, pp.33236-33282

    2024

Show all 61 references
  1. [9]

    and Zhou, B., 2024

    Zhang, S., He, Y ., Gu, Y ., He, Y ., Wang, H., Wang, H., Yang, R., Chady, T. and Zhou, B., 2024. UA V based defect detection and fault diagnosis for static and rotating wind turbine blade: A review. Nondestructive Testing and Evaluation, pp.1-39

    2024

  2. [10]

    and Chauhan, A.P.S., 2025

    Aggarwal, A.K. and Chauhan, A.P.S., 2025. Robust feature extraction from omnidirectional outdoor images for computer vision applications . International Journal of Instrumentation and Measurement, 10. 34

    2025

  3. [11]

    and Zhu, S., 2024

    Liu, L., Li, P., Wang, D. and Zhu, S., 2024. A wind turbine damage detection algorithm designed based on YOLOv8. Applied Soft Computing, 154, p.111364

    2024

  4. [12]

    and Wang, G., 2024

    Wu, Z., Zhang, Y ., Wang, X., Li, H., Sun, Y . and Wang, G., 2024. Algorithm for detecting surface defects in wind turbines based on a lightweight YOLO model. Scientific Reports, 14(1), p.24558

    2024

  5. [13]

    and Liu, L., 2024

    Hang, X., Zhu, X., Gao, X., Wang, Y . and Liu, L., 2024. Study on crack monitoring method of wind turbine blade based on AI model: Integration of classification, detection, segmentation and fault level evaluation. Renewable Energy, 224, p.120152

    2024

  6. [14]

    and Chen, X., 2024

    Jia, X. and Chen, X., 2024. AI -based optical -thermal video data fusion for near real -time blade segmentation in normal wind turbine operation. Engineering Applications of Artificial Intelligence, 127, p.107325

    2024

  7. [15]

    and Wang, L., 2023

    Zhou, W., Wang, Z., Zhang, M. and Wang, L., 2023. Wind turbine actual defects detection based on visible and infrared image fusion. IEEE Transactions on Instrumentation and Measurement, 72, pp.1 - 8

    2023

  8. [16]

    and Khoder, J., 2024

    Rizk, P., Rizk, F., Karganroudi, S.S., Ilinca, A., Younes, R. and Khoder, J., 2024. Advanced wind turbine blade inspection with hyperspectral imaging and 3D convolutional neural networks for damage detection. Energy and AI, 16, p.100366

    2024

  9. [17]

    and Pal, M., 2024

    Dwivedi, D., Babu, K.V .S.M., Yemula, P.K., Chakraborty, P. and Pal, M., 2024. Identification of surface defects on solar PV panels and wind turbine blades using attention based deep learning model. Engineering Applications of Artificial Intelligence, 131, p.107836

    2024

  10. [18]

    and Du, Y ., 2025

    Li, W., Zhao, W. and Du, Y ., 2025. Large-scale wind turbine blade operational condition monitoring based on UA V and improved YOLOv5 deep learning model. Mechanical Systems and Signal Processing, 226, p.112386

    2025

  11. [19]

    and Chen, X., 2025

    Jia, X. and Chen, X., 2025. Unsupervised wind turbine blade damage detection with memory- aided denoising reconstruction. IEEE Transactions on Industrial Informatics, 21(1), pp.762-770

    2025

  12. [20]

    and Sun, L., 2024

    Lei, X., Sun, M., Zhao, R., Wu, H., Zhou, Z., Dong, Y . and Sun, L., 2024. Unsupervised vision‐based structural anomaly detection and localization with reverse knowledge distillation. Structural Control and Health Monitoring, 2024(1), p.8933148

    2024

  13. [21]

    and Sharma, S., 2022

    Foster, A., Best, O., Gianni, M., Khan, A., Collins, K. and Sharma, S., 2022. Drone footage wind turbine surface damage detection. In 2022 IEEE 14th Image, Video, and Multidimensional Signal Processing Workshop (IVMSP), pp.1-5. IEEE

    2022

  14. [22]

    and Ataei, S., 2025

    Ataei, S.T., Zadeh, P.M. and Ataei, S., 2025. Vision -based autonomous structural damage detection using data-driven methods. arXiv preprint, arXiv:2501.16662. 35

    2025 arXiv

  15. [23]

    and Luo, X., 2025

    Zhang, Y ., Wang, L., Huang, C. and Luo, X., 2025. Wind turbine blade defect detection based on the genetic algorithm-enhanced YOLOv5 algorithm using synthetic data. IEEE Transactions on Industry Applications, 61(1), pp.653-665. doi: 10.1109/TIA.2024.3481190

    2025

  16. [24]

    and Luo, X., 2024

    Ye, X., Wang, L., Huang, C. and Luo, X., 2024. Wind turbine blade defect detection with a semi - supervised deep learning framework. Engineering Applications of Artificial Intelligence, 136, p.108908

    2024

  17. [25]

    and Chen, X., 2024

    Sheiati, S., Jia, X., McGugan, M., Branner, K. and Chen, X., 2024. Artificial intelligence-based blade identification in operational wind turbines through similarity analysis aided drone inspection. Engineering Applications of Artificial Intelligence, 137, p.109234

    2024

  18. [26]

    and Khushu, S., 2022

    Iyer, A., Nguyen, L. and Khushu, S., 2022. Learning to identify cracks on wind turbine blade surfaces using drone-based inspection images. arXiv preprint, arXiv:2207.11186

    2022 arXiv

  19. [27]

    and Khushu, S., 2022

    Nguyen, L., Iyer, A. and Khushu, S., 2022. An automated system for detecting visual damages of wind turbine blades. arXiv preprint, arXiv:2205.10954

    2022 arXiv

  20. [28]

    and See, J., 2023

    Yang, C., Liu, X., Zhou, H., Ke, Y . and See, J., 2023. Towards accurate image stitching for drone-based wind turbine blade inspection. Renewable Energy, 203, pp.267-279

    2023

  21. [29]

    and Zhang, W., 2024

    Li, T., Luan, Y ., Pang, Z. and Zhang, W., 2024. Structural digital twin modeling and adaptive pretrain- finetune learning for dynamic impact identification on wind turbine blades. IEEE Transactions on Industrial Informatics, 20(8), pp.10292-10303

    2024

  22. [30]

    and Li, X., 2023

    Jiang, G., Yue, R., He, Q., Xie, P. and Li, X., 2023. Imbalanced learning for wind turbine blade icing detection via spatio-temporal attention model with a self-adaptive weight loss function. Expert Systems with Applications, 229, p.120428

    2023

  23. [31]

    and Chen, J., 2023

    Zhou, Q., Pang, G., Tian, Y ., He, S. and Chen, J., 2023. AnomalyClip: Object-agnostic prompt learning for zero-shot anomaly detection. arXiv preprint, arXiv:2310.18961

    2023

  24. [32]

    and Wang, J., 2024a

    Gu, Z., Zhu, B., Zhu, G., Chen, Y ., Li, H., Tang, M. and Wang, J., 2024a. FILO: Zero-shot anomaly detection by fine-grained description and high-quality localization. In Proceedings of the 32nd ACM International Conference on Multimedia, pp.2041-2049

  25. [33]

    and Taciroglu, E., 2023

    Soleimani-Babakamali, M.H., Soleimani-Babakamali, R., Nasrollahzadeh, K., Avci, O., Kiranyaz, S. and Taciroglu, E., 2023. Zero-shot transfer learning for structural health monitoring using generative adversarial networks and spectral mapping. Mechanical Systems and Signal Proc...

    2023

  26. [34]

    and Taciroglu, E., 2025

    Soleimani-Babakamali, M.H., Soleimani -Babakamali, R., Kashfi -Yeganeh, A., Nasrollahzadeh, K., Avci, O., Kiranyaz, S. and Taciroglu, E., 2025. Multi-source transfer learning for zero -shot structural damage detection. Applied Soft Computing, 169, p.112519

    2025

  27. [35]

    and Xia, Y ., 2024

    Xiong, Q., Kong, Q., Xiong, H., Chen, J., Yuan, C., Wang, X. and Xia, Y ., 2024. Zero-shot knowledge transfer for seismic damage diagnosis through multi -channel 1D CNN integrated with autoencoder - based domain adaptation. Mechanical Systems and Signal Processing, 217, p.111535. 36

    2024

  28. [36]

    and Jeon, J.S., 2024a

    Chen, M., Mangalathu, S. and Jeon, J.S., 2024a. Rapid damage state identification of structures using generalized zero‐shot learning method. Earthquake Engineering & Structural Dynamics, 53(14), pp.4269-4286

  29. [37]

    and Wang, J., 2024b

    Gu, Z., Zhu, B., Zhu, G., Chen, Y ., Tang, M. and Wang, J., 2024b. AnomalyGPT: Detecting industrial anomalies using large vision- language models. In Proceedings of the AAAI Conference on Artificial Intelligence, 38(3), pp.1932-1940

    1932

  30. [38]

    and Ying-you, W., 2024b

    Chen, W., Yan-yi, L., Tie -zheng, G., Da -peng, L., Tao, H., Zhi, L., Qing- wen, Y ., Hui-han, W. and Ying-you, W., 2024b. Systems engineering issues for industry applications of large language model. Applied Soft Computing, 151, p.111165

  31. [39]

    and Dethlefs, N., 2024a

    Walker, C., Rothon, C., Aslansefat, K., Papadopoulos, Y . and Dethlefs, N., 2024a. SafeLLM: Domain- specific safety monitoring for large language models: A case study of offshore wind maintenance. arXiv preprint, arXiv:2410.10852

    arXiv

  32. [40]

    and Nandakumar, D.,

    Pastoriza, S., Yousfi, I., Redino, C., Vucovich, M., Rahman, A., Aguinaga, S. and Nandakumar, D.,

  33. [41]

    and Doan, B.L., 2024

    Thimonier, H., Popineau, F., Rimmel, A. and Doan, B.L., 2024. Retrieval augmented deep anomaly detection for tabular data. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pp.2250-2259

    2024

  34. [42]

    and Li, Y .F., 2023

    Wang, H. and Li, Y .F., 2023. Large language model empowered by domain-specific knowledge base for industrial equipment operation and maintenance. In 2023 5th International Conference on System Reliability and Safety Engineering (SRSE), pp.474-479. IEEE

    2023

  35. [43]

    and Dethlefs, N., 2024b

    Walker, C., Rothon, C., Aslansefat, K., Papadopoulos, Y . and Dethlefs, N., 2024b. Using large language models to recommend repair actions for offshore wind maintenance. In Journal of Physics: Conference Series, 2875(1), p.012025. IOP Publishing

  36. [44]

    and Veeramachaneni, K., 2024

    Alnegheimish, S., Nguyen, L., Berti-Equille, L. and Veeramachaneni, K., 2024. Large language models can be zero-shot anomaly detectors for time series?. arXiv preprint, arXiv:2405.14755

    2024 arXiv

  37. [45]

    and Tahan, A., 2024

    Jose, S., Nguyen, K.T., Medjaher, K., Zemouri, R., Lévesque, M. and Tahan, A., 2024. Advancing multimodal diagnostics: Integrating industrial textual data and domain knowledge with large language models. Expert Systems with Applications, 255, p.124603

    2024

  38. [46]

    and Bianco, S., 2025

    Bonomo, M. and Bianco, S., 2025. Visual RAG: Expanding MLLM visual knowledge without fine - tuning. arXiv preprint arXiv:2501.10834

    2025 arXiv

  39. [47]

    and Sarkar, S., 2025, July

    Bhat, N.N., Mondal, J. and Sarkar, S., 2025, July. ExpertNeurons at SciVQA -2025: Retrieval Augmented VQA with Vision Language Model (RA VQA-VLM). In Proceedings of the Fifth Workshop on Scholarly Document Processing (SDP 2025) (pp. 221-229). 37

    2025

  40. [48]

    and Oyamada, M., 2025

    Dong, Y ., Ueda, N., Boros, K., Ito, D., Sera, T. and Oyamada, M., 2025. SCAN: Semantic Document Layout Analysis for Textual and Visual Retrieval -Augmented Generation. arXiv preprint arXiv:2505.14381

    2025

  41. [49]

    and Elhoseiny, M., 2025

    Khan, F.F., Chen, J., Mohamed, Y ., Feng, C.M. and Elhoseiny, M., 2025. VR-RAG: Open-vocabulary Species Recognition with RAG-Assisted Large Multi-Modal Models. arXiv preprint arXiv:2505.05635

    2025

  42. [50]

    and Hu, X., 2025

    Zheng, X., Weng, Z., Lyu, Y ., Jiang, L., Xue, H., Ren, B., Paudel, D., Sebe, N., Van Gool, L. and Hu, X., 2025. Retrieval augmented generation and understanding in vision: A survey and new outlook. arXiv preprint arXiv:2503.18016

    2025 arXiv

  43. [51]

    LangChain

    Chase, H., 2022. LangChain. Available at: https://github.com/langchain-ai/langchain

    2022

  44. [52]

    and Gurevych, I., 2019

    Reimers, N. and Gurevych, I., 2019. Sentence -BERT: Sentence embeddings using Siamese BERT - networks. arXiv preprint, arXiv:1908.10084

    2019 arXiv

  45. [53]

    and Krueger, G., 2021

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J. and Krueger, G., 2021. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pp.8748-8763. PMLR

    2021

  46. [54]

    and Jégou, H., 2024

    Douze, M., Guzhva, A., Deng, C., Johnson, J., Szilvasy, G., Mazaré, P.E., Lomeli, M., Hosseini, L. and Jégou, H., 2024. The FAISS library. arXiv preprint, arXiv:2401.08281

    2024 arXiv

  47. [55]

    and Zhou, J., 2023

    Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C. and Zhou, J., 2023. Qwen-VL: A versatile vision -language model for understanding, localization, text reading, and beyond. arXiv preprint, arXiv:2308.12966

    2023 arXiv

  48. [56]

    Drone-based optical and thermal videos of rotor blades taken in normal wind turbine operation

    Chen, X., 2023. Drone-based optical and thermal videos of rotor blades taken in normal wind turbine operation. IEEE Dataport. doi: https://dx.doi.org/10.21227/yzs5-1067

    2023 doi

  49. [57]

    Dataset for AI -based optical -thermal video data fusion for near real -time blade segmentation in normal wind turbine operation

    Chen, X., 2024. Dataset for AI -based optical -thermal video data fusion for near real -time blade segmentation in normal wind turbine operation. Mendeley Data, V1. doi: 10.17632/9rcf5p89zn.1

    2024 doi

  50. [58]

    and Chen, X., 2018

    Shihavuddin, A. and Chen, X., 2018. DTU - Drone inspection images of wind turbine

    2018

  51. [59]

    and Zhao, Y ., 2022

    Wang, W., Xue, Y ., He, C. and Zhao, Y ., 2022. Review of the typical damage and damage-detection methods of large wind turbine blades. Energies, 15(15), p.5672

    2022

  52. [60]

    Ultralytics. (2023). YOLOv8. GitHub repository. https://github.com/ultralytics/ultralytics

    2023

  53. [2025]

    arXiv preprint, arXiv:2502.19534

    Retrieval augmented anomaly detection (RAAD): Nimble model adjustment without retraining. arXiv preprint, arXiv:2502.19534

    arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.