PaperlayerView exact revision on arXiv

Source-linked AI summary

Robust Online Aero-Engine Blade Defect Detection via Dual-Alignment Test-Time Adaptation

Zhaoyang Wang, Haiyong Chen, Dongying Li, Yining Wang, Huapeng Wu, Xinwei Lv, Atik Shahariar

arXiv:2610.00067v1cs.CV

TL;DR

Domain shifts and sparse defects challenge reliable online aero-engine blade inspection, while existing adaptation approaches are not tailored to this setting. ABDD combines dual feature and pseudo-box alignment with uncertainty filtering and lightweight tuning. It improves detection across the evaluated datasets and is also integrated into a robotic-arm inspection system.

  • Problem

    Aero-engine blade detectors face changing target domains, and sparse defects make pseudo-box adaptation vulnerable to noisy or missing predictions.

    Source: §1 · p. 2

  • Method

    ABDD performs online teacher–student adaptation by combining pseudo-box and feature-statistics alignment, uncertainty-aware box filtering, and lightweight parameter tuning.

    Source: §1 · p. 3

  • Results

    ABDD reduces domain-shift degradation, with average mAP@50 improvements of 5.6%, 11.5%, and 2.8% on CD-AeBD Subsets I and II and HD-AeBD, respectively.

    Source: §5 · p. 12

  • Takeaways & Limitations

    ABDD was integrated into a robotic-arm inspection system deployed with a major aero-engine manufacturer.

    Source: §1 · p. 3

    ABDD is slower than feed-forward detectors, though its online throughput is acceptable for the inspection cycle specified by the industrial partner.

    Source: Quantitative Results · p. 9

Abstract

from arXiv · show

Reliable visual inspection is essential for quality assurance in aero-engine blade manufacturing, where defect appearance may vary across production lines, imaging conditions, blade poses, and surface backgrounds. Such domain shifts cause a mismatch between training and deployment data and degrade the reliability of deep defect detectors in online inspection. This problem is particularly challenging because aero-engine blade images usually contain sparse defects, making pseudolabel-based adaptation vulnerable to noisy or missing predictions. To address this issue, we propose Aero-engine Blade Defect Detector (ABDD), an online adaptive detection framework based on test-time adaptation. ABDD introduces a Dual-Alignment Strategy to jointly adapt global visual style and local defect morphology by combining feature-statistics alignment with pseudo-box alignment. To reduce error accumulation from unreliable pseudo labels, an Uncertainty-aware Box Filtering mechanism evaluates pseudo boxes using classification confidence, classification entropy, and localization entropy. In addition, a lightweight Sparse Dilated Mona module enables parameter-efficient delta tuning while limiting source-domain forgetting. ABDD is evaluated on CD-AeBD and HD-AeBD under multiple domain-shift scenarios, with TTA strategies compared under a unified RT-DETR + Swin-T architecture. Experiments show that ABDD consistently improves detection robustness under domain shifts, and its practicality is further validated on an industrial inspection platform.

1 Introduction

Domain shifts and sparse defects make reliable online aero-engine blade inspection difficult for existing detectors and test-time adaptation methods. ABDD combines pseudo-box and feature-statistics alignment with uncertainty filtering and lightweight tuning, and improves detection across the evaluated shift scenarios.

1 Introduction

  • Domain shifts degrade detector performance, while sparse defects leave fewer than one defect per image and make pseudo-box adaptation vulnerable to unreliable labels.Conventional manual and rule-based inspection is labor-intensive, subjective, and sensitive to illumination, surface textures, and complex defect morphologies.Source: §1 · p. 2
  • ABDD uses a teacher–student Dual-Alignment Strategy to adapt local defect morphology and global feature statistics, filters uncertain pseudo boxes, and tunes lightweight modules while freezing most parameters.The teacher is updated by exponential moving average; Sparse Dilated Mona is intended to reduce adaptation cost and avoid catastrophic forgetting.Source: §1 · p. 3
  • ABDD improves mAP@50 by 13.7%, 5.9%, and 2.6% under illumination, viewpoint, and background shifts on CD-AeBD Subset I; average gains are 11.5% on Subset II and 2.8% on HD-AeBD.On HD-AeBD, it also improves mAP@50 : 95 by 6.4%.Source: §1 · p. 3
  • The authors curate CD-AeBD and HD-AeBD and integrate ABDD into a robotic-arm inspection system deployed with a major aero-engine manufacturer.The introduction frames these datasets and experiments as evaluation of the proposed approach.Source: §1 · p. 3

2 Related work

Prior aero-engine blade detection work has explored traditional inspection and deep detectors, while test-time adaptation has mainly developed around feature alignment or pseudo-box learning. The review positions ABDD against gaps in changing-domain, sparse-defect inspection.

2.1 Defect Detection of Aero-engine Blades

  • Aero-engine blade detectors address minute or irregular defects through varied architectures, but most assume fixed data distributions; Zhang et al. do not address online adaptation for defect detection under changing target domains.Earlier approaches also include manual inspection, traditional machine learning, edge or contour analysis, and ultrasonic imaging.Source: §2.1 · pp. 3–4

2.2 Test-Time Adaptation for Object Detectors

  • Detection TTA uses feature alignment or teacher–student pseudo-box learning, but prior methods largely target dense-label driving or common-object settings rather than sparse AeB defects and coupled shifts.Prior strategies include statistic matching, lightweight adapters, memory prototypes, overlap-based filtering, and human-in-the-loop weak supervision.Source: §2.2 · p. 4

3 Methodology

ABDD uses an RT-DETR-style detector with a Swin/SSDM backbone and adapts it online through dual pseudo-box and feature-statistics alignment. Uncertainty-aware filtering selects pseudo boxes for supervision, while Sparse Dilated Mona modules support parameter-efficient adaptation.

3.1 Aero-engine Blade Defect Detector

  • The detector combines a Swin backbone with SSDM modules, an RT-DETR encoder and decoder, and separate classification and FCOS-style localization branches.The localization branch predicts distances from sampling points to box edges, supporting uncertainty-aware box filtering.Source: §3.1 · p. 4

3.2 Dual-Alignment Strategy

3.4 Sparse Dilated Mona

  • Three parallel SDM convolution branches provide 3×3, 5×5, and 7×7 receptive fields with 54n rather than 166n branch parameters, lowering adaptation cost and mitigating forgetting.Only SDM modules are updated while the backbone is frozen.Source: §3.4 · p. 7

4 Experimental evaluation

Across CD-AeBD I, CD-AeBD II, and HD-AeBD, ABDD shows stronger domain-shift robustness than the compared detectors and TTA methods. Ablations report the most stable performance from combining alignment strategies, while SDM offers a parameter-efficient tuning trade-off.

4.1 Datasets and Experimental Settings

  • Evaluation uses CD-AeBD I, CD-AeBD II, and HD-AeBD, with HD-AeBD collected from real industrial inspection scenes across production batches.All TTA methods use RT-DETR with a Swin-T backbone, and target images are processed sequentially in a streaming inference-and-adaptation setup.Source: §4.1 · pp. 7–8

4.2 Main Results

  • On CD-AeBD I, ABDD exceeds CMT by 6.7% in mean mAP@50 and improves over the baseline by 2.6%, 5.9%, and 13.7% on Back., View., and Illum., respectively.It achieves at least 84.9% AP on each of the four defect categories.Source: Table 2 · p. 9, Quantitative Results · p. 9
  • Although ABDD is slower than feed-forward detectors, its online throughput is acceptable for the inspection cycle specified by the industrial partner.Validation-platform and timing details are reported in Supplementary Material II.Source: Quantitative Results · p. 9
  • On CD-AeBD II, ABDD reaches 79.9% mean mAP@50, with 78.3% on D2 and 75.8% on D3; on HD-AeBD, it leads at 44.4% mean mAP@50 : 95.The reported drops on D2 and D3 are below 10%, and ABDD has the smallest drops on both HD-AeBD target batches.Source: Table 3 · p. 10, Quantitative Results · p. 10
  • In qualitative examples from CD-AeBD-I/II and HD-AeBD, ABDD produces more accurate detections than the non-adaptive baseline and suppresses defect-irrelevant shallow features.The examples associate baseline misses with illumination and viewpoint changes, and misclassifications with defect-morphology variation; shallow-feature suppression is linked to focus on tiny defects and defects in complex backgrounds.Source: Qualitative Results · p. 10, Figure 3 · p. 11

4.3 Ablation Study

  • Across CD-AeBD I, CD-AeBD II, and HD-AeBD, combining pseudo-box and feature-distribution alignment gives the most stable ablation performance; SDM has the lowest degradation on CD-AeBD I and HD-AeBD with 0.64M trainable parameters.Adding positional and classification entropy to confidence filtering improves pseudo-box quality, while SDM uses 46.4% of Mona's trainable parameters.Source: Table 4 · p. 11, §4.3 · p. 12

5 Discussion and Conclusion

ABDD combines dual alignment, uncertainty-aware pseudo-box filtering, and sparse dilated tuning to address domain-shift degradation in aero-engine blade inspection. It improves reported mAP across CD-AeBD and HD-AeBD, while its adaptation cost and false-negative handling constrain its scope.

5 Discussion and Conclusion

  • Average mAP@50 improves by 5.6% on CD-AeBD Subset I, 11.5% on Subset II, and 2.8% on HD-AeBD; HD-AeBD mAP@50 : 95 gains 6.4%.Source: §5 · p. 12
  • ABDD combines DAS feature-statistic and pseudo-box alignment, UBF reliable pseudo-label selection, and SDM parameter-efficient transformer adaptation.Source: §5 · p. 12
  • The deployed inspection system further validates ABDD’s practical value for rigorous blade screening and flight-safety-related manufacturing inspection.Source: §5 · p. 12
  • Adaptation requires two forward and one backward propagation, potentially limiting highly real-time use; UBF cannot explicitly recover false-negative defects under extremely large domain gaps.Source: §5 · p. 12

Related papers

Loading 2610.00067v1…