arxiv: 2603.14869 · v2 · submitted 2026-03-16 · 💻 cs.AI

Recognition: no theorem link

A Self-Evolving Defect Detection Framework for Industrial Photovoltaic Systems

Haoyu He , Yu Duan , Wenzhen Liu , Hanyuan Hang , Boyu Qin , Qiantu Tuo , Xiaoke Yang , Rui Li

Authors on Pith no claims yet

Pith reviewed 2026-05-15 10:46 UTC · model grok-4.3

classification 💻 cs.AI

keywords self-evolving learningphotovoltaic defect detectionelectroluminescence imagingcontinual learningindustrial inspectionobject detectiondomain shiftsolar module defects

0 comments

The pith

A self-evolving framework lets defect detectors for solar panels adapt continuously to new conditions and defect types without manual retraining.

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

The paper introduces SEPDD to handle the practical difficulties of inspecting photovoltaic modules with electroluminescence images, where module shapes vary, images are low-resolution, defects are faint, classes are imbalanced, and data keeps shifting as inspection processes evolve. It combines automated model optimization with a continual self-evolving learning mechanism so the system can update itself over time while keeping earlier knowledge intact. Experiments on a public benchmark and a private industrial dataset show the approach reaching leading accuracy scores that exceed both standard automated systems and human experts. A reader would care because timely defect detection directly affects energy output, equipment lifespan, and maintenance expenses in operating solar installations.

Core claim

SEPDD integrates automated model optimization with a continual self-evolving learning mechanism, enabling the inspection system to progressively adapt to distribution shifts and newly emerging defect patterns during long-term deployment in industrial PV settings while achieving mAP50 of 91.4 percent on the public dataset and 49.5 percent on the private dataset.

What carries the argument

The continual self-evolving learning mechanism, which automatically updates the detection model to handle new data distributions and defect morphologies while retaining prior performance.

If this is right

Inspection systems become maintainable over years without repeated expert labeling or full retraining.
Performance holds up on severely imbalanced and domain-shifted industrial data.
The same pipeline produces higher detection rates than fixed autonomous models or human review on both benchmark and real-world PV images.
Long-term field operation gains robustness against evolving inspection conditions and labeling practices.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same adaptation loop could be tested on defect detection tasks in other continuous-production industries such as semiconductor or battery manufacturing.
Integration with additional sensor streams like thermal imaging might further reduce false negatives in outdoor PV arrays.
If adaptation stability holds across several annual cycles, total lifecycle costs for large solar farms could decrease through earlier intervention.

Load-bearing premise

The self-evolving process can keep incorporating new defect patterns and data shifts without causing instability or loss of accuracy on earlier patterns.

What would settle it

A multi-cycle deployment test on the private dataset where overall mAP50 drops below the initial baseline after several rounds of adaptation on newly labeled defects would show the mechanism fails to deliver stable improvement.

Figures

Figures reproduced from arXiv: 2603.14869 by Boyu Qin, Hanyuan Hang, Haoyu He, Qiantu Tuo, Rui Li, Wenzhen Liu, Xiaoke Yang, Yu Duan.

**Figure 1.** Figure 1: Overall architecture of the SEPDD framework. The evolution cycle is triggered by pre-defined indicators. The selfevolving search is tree-based with merge action, where a node represents an exploration. Each node orchestrates a self-contained evolution pipeline that produces high-quality and stable code with a deployment-ready model. datasets, deploying them in real industrial environments introduces seve… view at source ↗

**Figure 2.** Figure 2: Prompt template for Idea Generator. Algorithm 1: Workflow of one node (single expansion step). input : parent node, SEPDD input, journal. output: new node // Code generation 1 suggestions ← IdeaGenerator(parent_node) 2 code ← CodeCreator(parent_node, suggestions) // Code refinement 3 repeat 4 syntax_info, exec_output ← Validator(code) 5 buggy, analysis ← Analyzer(code, syntax_info, exec_output) 6 if not bu… view at source ↗

**Figure 3.** Figure 3: Defect distribution. TABLE I: PVEL-AD dataset. Bold: best per column; underline: second. Method Precision Recall mAP50 mAP50-95 AIDE 79.9 69.4 76.6 49.4 Expert-YOLO 79.9 80.4 86.7 55.7 Expert-FRCN 75.0 69.7 88.7 63.5 SEPDD (Qwen) 90.0 85.0 91.5 62.1 SEPDD (GPT-5.1) 88.2 84.8 90.3 62.9 This baseline is included to highlight the difficulty of crossdomain adaptation. Both expert baselines require substantial… view at source ↗

**Figure 4.** Figure 4: One-label-added regime: full label vs. one label re [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 5.** Figure 5: Detection comparison across public defect categories. Each row: defect type; each column: Original, Ground truth, AIDE, [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

**Figure 6.** Figure 6: Detection comparison on EF industrial defect types (microcracks, light and shade, black spots, broken gridlines, black [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗

**Figure 7.** Figure 7: Evolution tree from one EF industrial run (code-like [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗

read the original abstract

Reliable photovoltaic (PV) power generation requires timely detection of module defects that may reduce energy yield, accelerate degradation, and increase lifecycle operation and maintenance costs during field operation. Electroluminescence (EL) imaging has therefore been widely adopted for PV module inspection. However, automated defect detection in real operational environments remains challenging due to heterogeneous module geometries, low-resolution imaging conditions, subtle defect morphology, long-tailed defect distributions, and continual data shifts introduced by evolving inspection and labeling processes. These factors significantly limit the robustness and long-term maintainability of conventional deep-learning inspection pipelines. To address these challenges, this paper proposes SEPDD, a Self-Evolving Photovoltaic Defect Detection framework designed for evolving industrial PV inspection scenarios. SEPDD integrates automated model optimization with a continual self-evolving learning mechanism, enabling the inspection system to progressively adapt to distribution shifts and newly emerging defect patterns during long-term deployment. Experiments conducted on both a public PV defect benchmark and a private industrial EL dataset demonstrate the effectiveness of the proposed framework. Both datasets exhibit severe class imbalance and significant domain shift. SEPDD achieves a leading mAP50 of 91.4% on the public dataset and 49.5% on the private dataset. It surpasses the autonomous baseline by 14.8% and human experts by 4.7% on the public dataset, and by 4.9% and 2.5%, respectively, on the private dataset.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

SEPDD adds a self-evolving loop to PV defect detection and posts solid numbers on imbalanced data, but the abstract leaves the adaptation mechanism and its long-term stability too thin to judge the central claim.

read the letter

The paper's main move is to wrap standard detection with an automated optimization step plus a continual self-evolving loop meant to handle ongoing shifts in EL images of PV modules. That combination targets a real pain point: inspection pipelines that degrade when module types, lighting, or defect distributions change over months or years in the field. They evaluate on one public benchmark and one private industrial set, both with heavy imbalance and domain shift, and report mAP50 scores of 91.4% and 49.5% that beat the listed baselines and human experts by a few points. Those gains are the concrete result they put forward. The framing around long-term maintainability is also useful; most defect-detection papers stop at static test sets, so the emphasis on progressive adaptation is a step in the right direction for applied work. The soft spot is exactly where the stress-test note flags it. The abstract gives almost no description of the self-evolving algorithm itself, no count of adaptation cycles, no retention numbers on old defect classes, and no ablation that isolates the loop from ordinary fine-tuning. Without those, the reported margins could come from better one-shot training rather than stable continual learning. The private-set numbers are also low enough that small changes in protocol could move them a lot. This is the kind of paper that belongs in a reading group focused on applied computer vision for energy or manufacturing systems. Readers who care about deployment longevity will find the motivation and the two-dataset setup worth seeing, even if they end up wanting more experimental detail. It should go to peer review; the problem is practical, the empirical setup uses both public and private data, and the gaps are fixable with clearer methods and longer-horizon tests rather than fatal.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes SEPDD, a Self-Evolving Photovoltaic Defect Detection framework that integrates automated model optimization with a continual self-evolving learning mechanism to adapt to distribution shifts and emerging defect patterns in electroluminescence imaging of PV modules. Experiments on a public benchmark and a private industrial dataset (both with severe class imbalance and domain shift) report leading mAP50 scores of 91.4% and 49.5%, respectively, with gains of 14.8% and 4.9% over an autonomous baseline and smaller margins over human experts.

Significance. If the self-evolving component can be shown to deliver stable long-term adaptation, the framework would address a genuine industrial need for maintainable inspection systems under evolving data conditions, potentially lowering O&M costs in PV plants. The empirical results on imbalanced, shifted datasets are promising and the absence of circularity in the metrics is a strength, but the central claim of progressive, stable evolution currently rests on insufficient longitudinal evidence.

major comments (3)

[§3] §3 (Method): The continual self-evolving learning mechanism is described only at a high level; no concrete description is given of the adaptation algorithm, the mechanism for avoiding catastrophic forgetting, the criteria for initiating evolution cycles, or any regularization terms used. This detail is load-bearing for the paper's core claim of progressive adaptation during long-term deployment.
[§4] §4 (Experiments): No ablation studies isolate the contribution of the self-evolving loop from standard fine-tuning or one-shot transfer learning. Consequently the reported 14.8% and 4.9% gains over the autonomous baseline cannot be confidently attributed to the proposed continual mechanism rather than to additional training data or hyper-parameter tuning.
[§4.3] §4.3 (Results): The evaluation contains no longitudinal metrics—such as retention accuracy on prior defect classes after multiple adaptation cycles, performance drift curves, or simulated multi-year deployment length—directly leaving the weakest assumption (stable adaptation without instability or forgetting) untested.

minor comments (2)

[Abstract] Abstract: A single sentence quantifying the number of evolution cycles or simulated deployment duration tested would strengthen the self-evolving claim without lengthening the abstract.
[Table 2] Table 2 (or equivalent results table): Provide full implementation details (epochs, learning-rate schedules, data-augmentation policies) for every baseline, including the autonomous baseline, to support reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback. The comments highlight important areas where the manuscript can be strengthened, particularly regarding technical details of the self-evolving mechanism and supporting empirical evidence. We address each major comment point by point below, with plans to revise the manuscript accordingly.

read point-by-point responses

Referee: [§3] §3 (Method): The continual self-evolving learning mechanism is described only at a high level; no concrete description is given of the adaptation algorithm, the mechanism for avoiding catastrophic forgetting, the criteria for initiating evolution cycles, or any regularization terms used. This detail is load-bearing for the paper's core claim of progressive adaptation during long-term deployment.

Authors: We agree that §3 currently presents the self-evolving mechanism at a high level. In the revised manuscript, we will expand this section with a concrete description of the adaptation algorithm (including pseudocode for the update procedure), the specific approach to mitigating catastrophic forgetting (e.g., selective replay of prior samples combined with elastic weight consolidation), the criteria for initiating evolution cycles (performance drop thresholds on a validation buffer and distribution shift detection via KL divergence), and the regularization terms incorporated in the loss function. These additions will directly support the claims of stable progressive adaptation. revision: yes
Referee: [§4] §4 (Experiments): No ablation studies isolate the contribution of the self-evolving loop from standard fine-tuning or one-shot transfer learning. Consequently the reported 14.8% and 4.9% gains over the autonomous baseline cannot be confidently attributed to the proposed continual mechanism rather than to additional training data or hyper-parameter tuning.

Authors: We acknowledge the absence of targeted ablations. In the revision, we will add ablation experiments in §4 that compare the full SEPDD framework against (i) the autonomous baseline with standard fine-tuning on the same incremental data and (ii) one-shot transfer learning variants. These controlled comparisons will isolate the contribution of the continual self-evolving loop and clarify that the reported gains stem from the proposed mechanism rather than data volume or tuning alone. revision: yes
Referee: [§4.3] §4.3 (Results): The evaluation contains no longitudinal metrics—such as retention accuracy on prior defect classes after multiple adaptation cycles, performance drift curves, or simulated multi-year deployment length—directly leaving the weakest assumption (stable adaptation without instability or forgetting) untested.

Authors: We recognize that the current results lack explicit longitudinal evaluation of stability. While the reported experiments demonstrate adaptation under distribution shift, they do not track retention across cycles. In the revised manuscript, we will add simulated longitudinal experiments in §4.3, including retention accuracy on previously seen defect classes after successive adaptation cycles, performance drift curves over multiple iterations, and simulated multi-year deployment scenarios using sequential data streams. These additions will directly test and support the assumption of stable adaptation without significant forgetting. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical performance claims rest on held-out test measurements, not self-referential definitions or fitted inputs.

full rationale

The paper describes SEPDD as integrating automated optimization with a continual self-evolving mechanism and reports mAP50 scores (91.4% public, 49.5% private) on public and private datasets. These are presented as direct experimental outcomes on held-out data exhibiting class imbalance and domain shift. No equations, parameters, or self-citations are shown that would make the reported metrics equivalent to inputs by construction. The self-evolving component is asserted to handle distribution shifts, but its effectiveness is evaluated empirically rather than derived tautologically. This is a standard empirical ML paper with no load-bearing self-definitional or fitted-prediction steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The framework rests on standard assumptions from deep learning and continual learning literature; no explicit free parameters, axioms, or invented entities are detailed in the abstract.

pith-pipeline@v0.9.0 · 5575 in / 1151 out tokens · 39116 ms · 2026-05-15T10:46:23.719363+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages · 1 internal anchor

[1]

Three decades, three climates: environ- mental and material impacts on the long-term reliability of photovoltaic modules,

E. Özkalay, H. Quest, A. Gassner, A. Virtuani, G. C. Eder, C. Buerhop- Lutz, G. Friesen, and C. Ballif, “Three decades, three climates: environ- mental and material impacts on the long-term reliability of photovoltaic modules,”EES Solar, 2025

work page 2025
[2]

Understanding and reducing the risk of extreme photovoltaic degradation,

Y . Tang, S. Poddar, M. Kay, and F. Rougieux, “Understanding and reducing the risk of extreme photovoltaic degradation,”IEEE Journal of Photovoltaics, 2025

work page 2025
[3]

An efficient cnn-based detec- tor for photovoltaic module cells defect detection in electroluminescence images,

Q. Liu, M. Liu, C. Wang, and Q. M. J. Wu, “An efficient cnn-based detec- tor for photovoltaic module cells defect detection in electroluminescence images,”Solar Energy, vol. 266, p. 112245, 2023

work page 2023
[4]

Advancing photovoltaic cells defect detection in electroluminescence images through exploring multiple object detectors,

M. W. Akram, J. Bai, C. Xuan, X. Xu, J. Hu, and S. Wu, “Advancing photovoltaic cells defect detection in electroluminescence images through exploring multiple object detectors,”Solar Energy Materials and Solar Cells, vol. 285, p. 113777, 2025

work page 2025
[5]

Quantification of damages and classification of flaws in mono-crystalline photovoltaic cells through the application of vision transformers,

J. P. C. Barnabé, L. P. Jiménez, G. Fraidenraich, E. R. D. De Lima, and H. F. Santos, “Quantification of damages and classification of flaws in mono-crystalline photovoltaic cells through the application of vision transformers,”IEEE Access, 2023

work page 2023
[6]

Polycrystalline silicon photovoltaic cell defects detection based on global context information and multi- scale feature fusion in electroluminescence images,

S. Chen, Y . Lu, G. Qin, and X. Hou, “Polycrystalline silicon photovoltaic cell defects detection based on global context information and multi- scale feature fusion in electroluminescence images,”Materials Today Communications, vol. 42, p. 110627, 2024

work page 2024
[7]

A pv cell defect detector combined with transformer and attention mechanism,

D. Lang and Z. Lv, “A pv cell defect detector combined with transformer and attention mechanism,”Scientific Reports, vol. 14, p. 72019, 2024

work page 2024
[8]

Learning from imbalanced data,

H. He and E. Garcia, “Learning from imbalanced data,”IEEE Transac- tions on Knowledge and Data Engineering, vol. 21, no. 9, pp. 1263–1284, 2009

work page 2009
[9]

Class-balanced loss based on effective number of samples,

Y . Cui, M. Jia, T. Y . Lin, Y . Song, and S. Belongie, “Class-balanced loss based on effective number of samples,” inCVPR, 2019, pp. 9268–9277

work page 2019
[10]

Equalization loss for long-tailed object recognition,

J. Tan, X. Lu, G. Zhang, and J. Yin, “Equalization loss for long-tailed object recognition,” inCVPR, 2020, pp. 11 662–11 671

work page 2020
[11]

Automatic classification of defective photovoltaic module cells in electroluminescence images,

J. Deitsch, V . Christlein, S. Bergeret al., “Automatic classification of defective photovoltaic module cells in electroluminescence images,” Solar Energy, vol. 185, pp. 455–468, 2019

work page 2019
[12]

Deep-learning-based automatic detection of photovoltaic cell defects in electroluminescence images,

J. Wang, L. Bi, P. Sunet al., “Deep-learning-based automatic detection of photovoltaic cell defects in electroluminescence images,”Sensors, vol. 23, no. 1, p. 297, 2022

work page 2022
[13]

Long-term reliability of photovoltaic c-si modules – a detailed assessment based on the first italian bipv project,

C. Del Pero, N. Aste, F. Leonforte, and F. Sfolcini, “Long-term reliability of photovoltaic c-si modules – a detailed assessment based on the first italian bipv project,”Solar Energy, vol. 274, p. 112074, 2023

work page 2023
[14]

An assessment of photovoltaic module degradation for life expectancy: A comprehensive review,

A. Kumar, H. Ganesan, V . Saini, S. Sharma, and A. Agrawal, “An assessment of photovoltaic module degradation for life expectancy: A comprehensive review,”Engineering Failure Analysis, vol. 152, p. 107863, 2023

work page 2023
[15]

Review of degradation and failure phenomena in photovoltaic modules,

M. Aghaei, A. Fairbrother, A. Goket al., “Review of degradation and failure phenomena in photovoltaic modules,”Renewable and Sustainable Energy Reviews, vol. 149, p. 112160, 2022

work page 2022
[16]

A review of the degradation of photovoltaic modules for life expectancy,

J. Kim, M. Rabelo, S. Padiet al., “A review of the degradation of photovoltaic modules for life expectancy,”Energies, vol. 14, no. 14, p. 4278, 2021

work page 2021
[17]

Deep long-tailed learning: A survey,

Y . Zhang, B. Kang, B. Hooiet al., “Deep long-tailed learning: A survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021

work page 2021
[18]

Deep domain adaptation for machine learning: A survey,

M. Wang and W. Deng, “Deep domain adaptation for machine learning: A survey,”Neurocomputing, vol. 338, pp. 78–94, 2020

work page 2020
[19]

Continual lifelong learning with neural networks: A review,

G. Parisi, R. Kemker, J. Part, C. Kanan, and S. Wermter, “Continual lifelong learning with neural networks: A review,”Neural Networks, vol. 113, pp. 54–71, 2019

work page 2019
[20]

Neural architecture search: A survey,

T. Elsken, J. Metzen, and F. Hutter, “Neural architecture search: A survey,”Journal of Machine Learning Research, vol. 20, pp. 1–21, 2019

work page 2019
[21]

Aide: Ai-driven exploration in the space of code.arXiv preprint arXiv:2502.13138,

Z. Jiang, D. Schmidt, D. Srikanth, D. Xu, I. Kaplan, D. Jacenko, and Y . Wu, “Aide: Ai-driven exploration in the space of code,”arXiv preprint arXiv:2502.13138, 2025

work page arXiv 2025
[22]

M., et al

E. Toledo, K. Hambardzumyan, M. Josifoski, R. Hazra, N. Baldwin, A. Audran-Reiss, M. Kuchnik, D. Magka, M. Jiang, A. M. Lupidi et al., “Ai research agents for machine learning: Search, exploration, and generalization in mle-bench,”arXiv preprint arXiv:2507.02554, 2025

work page arXiv 2025
[23]

Automlgen: Navigating fine- grained optimization for coding agents.arXiv preprint arXiv:2510.08511, 2025a

S. Du, X. Yan, D. Jiang, J. Yuan, Y . Hu, X. Li, L. He, B. Zhang, and L. Bai, “Automlgen: Navigating fine-grained optimization for coding agents,”arXiv preprint arXiv:2510.08511, 2025

work page arXiv 2025
[24]

Pvel-ad: A large-scale open-world dataset for photovoltaic cell anomaly detection,

B. Su, Z. Zhou, and H. Chen, “Pvel-ad: A large-scale open-world dataset for photovoltaic cell anomaly detection,”IEEE Transactions on Industrial Informatics, vol. 19, no. 1, pp. 404–413, 2022

work page 2022
[25]

Ultralytics yolo11,

G. Jocher and J. Qiu, “Ultralytics yolo11,” 2024. [Online]. Available: https://github.com/ultralytics/ultralytics

work page 2024
[26]

YOLOv12: Attention-Centric Real-Time Object Detectors

Y . Tian, Q. Ye, and D. Doermann, “Yolo12: Attention-centric real-time object detectors,”arXiv preprint arXiv:2502.12524, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[27]

Frus- tratingly simple few-shot object detection,

X. Wang, T. E. Huang, T. Darrell, J. E. Gonzalez, and F. Yu, “Frus- tratingly simple few-shot object detection,” inProceedings of the 37th International Conference on Machine Learning, 2020, pp. 9919–9928

work page 2020