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

cs.CVSep 4, 2026

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

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

Organizations: School of Artificial Intelligence, Hebei University of Technology, Tianjin, China · Lappeenranta-Lahti University of Technology, Lappeenranta, Finland

Abstract

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.

Figures & tables

Figure 1: Aero-engine blade defect detection under cross-domain shifts. Source and target images may differ in background, viewpoint, and illumination, resulting in feature-distribution discrepancies and detection performance degradation.

Figure 2: Overview of the proposed ABDD for aero-engine blade defect detection.

Setting	Images	Boxes	Def.	Domains	Source $\rightarrow$ Target
CD-AeBD I	1,639	2,371	4	Same, Back., View., Illum.	Same $\rightarrow$ Back./View./Illum.
CD-AeBD II	2,703	4,050	4	D1, D2, D3	D1 $\rightarrow$ D2/D3
HD-AeBD	1,857	2398	3	B1, B2, B3	B1 $\rightarrow$ B2/3

Table 1: Summary of the datasets and domain-shift settings. Def. denotes defect categories.

Method	Same	Back.	View.	Illum.	Ab	Br	Fr	Gr	Mean	F1	FPS
Yolov10 [ 25 ]	87.9	82.0 ↓5.9	77.4 ↓10.5	69.7 ↓18.2	67.7	86.5	83.3	80.8	79.3	70.7	42.5
Yolov13 [ 9 ]	86.8	80.8 ↓6.0	77.9 ↓8.9	67.4 ↓19.4	78.2	85.6	72.0	78.9	78.2	69.9	43.8
RT-DETR [ 38 ]	89.8	84.2 ↓5.6	79.7 ↓10.1	68.3 ↓21.5	68.8	87.2	83.1	82.0	80.5	71.4	22.9
MI-DETR [ 14 ]	89.8	85.5 ↓4.3	79.0 ↓10.8	75.5 ↓14.3	80.7	83.2	86.8	77.7	82.5	73.6	20.2
Dfine [ 16 ]	90.4	86.5 ↓3.9	80.9 ↓9.5	71.1 ↓19.3	73.6	87.8	84.0	83.2	82.2	72.3	36.0
MR-DETR [ 35 ]	92.1	87.0 ↓5.1	83.2 ↓8.9	76.6 ↓15.5	77.1	87.6	85.7	86.9	84.7	76.7	10.8

Table 2: Results (

mAP@50

) of detectors and TTA methods across different domains in CD-AeBD I. Same , Back. , View. , and Illum. denote source, background-shift, viewpoint-shift, and illumination-shift domains, while Ab , Br , Fr , and Gr denote ablation, breakdown, fracture, and groove. The mean denotes the average score across four domains, representing overall performance. Best scores are highlighted in bold .

Method

CD-AeBD II (

mAP@50

)

HD-AeBD (

mAP@50

)

HD-AeBD (

mAP@50:95

)

Mean

Yolov10 [ 25 ]

85.2

63.1 ↓22.1

65.4 ↓19.8

71.2

88.4

75.0 ↓13.4

78.8 ↓9.6

80.7

50.8

36.1 ↓14.7

35.6 ↓15.2

40.8

Yolov13 [ 9 ]

84.9

66.0 ↓18.9

68.7 ↓16.2

73.2

85.7

68.9 ↓16.8

71.2 ↓14.5

75.3

45.8

29.7 ↓16.1

26.8 ↓19.0

34.1

RT-DETR [ 38 ]

85.7

60.3 ↓25.4

59.3 ↓26.4

68.4

86.3

75.2 ↓11.1

77.3 ↓9.0

79.6

47.8

32.3 ↓15.5

33.9 ↓13.9

38.0

Dfine [ 16 ]

84.8

64.9 ↓19.9

66.0 ↓18.8

71.9

83.7

74.4 ↓9.3

69.9 ↓13.8

76.0

47.4

34.0 ↓13.4

31.4 ↓16.0

37.6

MR-DETR [ 35 ]

87.9

68.8 ↓19.1

70.8 ↓17.1

75.8

89.0

80.8 ↓8.2

79.2 ↓9.8

83.0

56.1

29.6 ↓26.5

30.8 ↓25.3

38.8

Table 3: Results of state-of-the-art detectors and TTA methods on CD-AeBD II and HD-AeBD. CD-AeBD II reports

mAP@50

across D1–D3. HD-AeBD reports

mAP@50

and

mAP@50:95

across B1–B3. Subscripted downward arrows indicate performance drops from the source domain. The best scores in each column are highlighted in bold .

Figure 3: Visualizations of shallow/deep feature maps and detection results. Rows 1–3 are target domains in CD-AeBD-I, rows 4–5 are target domains in CD-AeBD-II, and rows 6–7 are target domains in HD-AeBD.

Comp.	Variant	CD-I	CD-II	HD	Aux.
DAS	w/o PB, FD	13.1	25.7	10.1	–
DAS	PB only	6.9	24.4	6.5	–
DAS	FD only	7.4	18.9	11.2	–
DAS	PB + FD	5.1	8.5	5.9	–
UBF	Conf.	10.9	38.3	12.5	35.4/18.8/33.9
UBF	PE + CE	6.2	17.6	8.8	60.2/39.7/41.2

Table 4: Ablation study of ABDD. CD-I, CD-II, and HD denote average performance drops, where lower is better. Aux. reports GT p for UBF and trainable parameters for tuning. Best results are in bold .

Fault Detection

Fault detection identifies abnormal behavior in machines, sensors, infrastructure, software, and autonomous systems before it causes harm or disruption.

See all 105 papers

BladeYOLO: Wind Turbine Blade Defect Detection with Limited Annotations and Weak-Saliency Awareness
Jul 30, 2026Yabin Xu, Fangtao Zhang, Fan Wang +5Object Detection Structural Health Monitoring
School of Mechanical Engineering, Zhejiang Sci-Tech University, Hangzhou 310018, China · Department of Artificial Intelligence and Robotics, Zhejiang Energy Digital Technology Co., Ltd, Hangzhou 311100, China · School of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing, China +1
NFAD: Nuisance-Filtered Anomaly Detection Under Distribution Shift
Aug 29, 2026Dat Cao, Son Nghiem, Phan Nguyen +2Anomaly Detection Industrial Anomaly Detection
Korea Advanced Institute of Science and Technology (KAIST), South Korea · Sony Computer Science Laboratories (Sony CSL), Japan · The University of Tokyo, Japan +1
FogTTA: Online Test-Time Adaptation for Robust Transformer-based Object Detection in Foggy Weather
Oct 21, 2026Ali Abedi, Q. M. Jonathan Wu, Ning Zhang +1Object Detection Continual Test-Time Adaptation
University of Windsor · University of Toronto Mississauga

Jul 30, 2026J/K move · Enter open · S save

BladeYOLO: Wind Turbine Blade Defect Detection with Limited Annotations and Weak-Saliency Awareness

Yabin Xu, Fangtao Zhang, Fan Wang, Zhan Wang+4

School of Mechanical Engineering, Zhejiang Sci-Tech University, Hangzhou 310018, China · Department of Artificial Intelligence and Robotics, Zhejiang Energy Digital Technology Co., Ltd, Hangzhou 311100, China · School of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing, China+1

Object Detection Structural Health Monitoring

Open paper AskPDF arXiv

Wind turbine blade defect detection remains highly challenging in real-world inspection scenarios due to limited on-site data and the subtle visual characteristics of defects. In practice, blade defects are often small-scale, low-contrast, and difficult to distinguish from complex backgrounds, which significantly limits the robustness of existing detectors. To address these challenges, we propose BladeYOLO, a defect detection framework for wind turbine blades. Specifically, we integrate a Vision Transformer (ViT) backbone initialized with DINOv3 self-supervised pre-trained weights into YOLOv12-L, enabling the transfer of large-scale generic visual priors to blade defect detection and improving feature representation under limited training annotations. To enhance the perception of subtle defects, we further develop a Mamba-guided Weak-Defect Enhancement module, which consists of a Detail-Enhanced Multi-scale Branch for preserving high-frequency structural cues and a Cross-Mamba module for progressively propagating high-level semantic guidance to shallow features. In addition, we introduce a lightweight Style-Injector module that captures environment-related style information via Fourier decomposition and injects it into selected ViT self-attention layers, thereby improving robustness against environment-induced appearance variations. Extensive experiments demonstrate that BladeYOLO achieves superior performance on the WTBlade-Defect dataset, with additional annotation-budget experiments showing its favorable performance under reduced training annotations. Evaluation on the public Wind Surface Defect dataset further provides supportive evidence for the cross-dataset robustness of BladeYOLO. In particular, on this public dataset, BladeYOLO outperforms the best competing method by 3.5% in mAP

_{50}

and 2.5% in mAP

_{50-95}

Abstract

Fault Detection

Explore similar work

BladeYOLO: Wind Turbine Blade Defect Detection with Limited Annotations and Weak-Saliency Awareness

NFAD: Nuisance-Filtered Anomaly Detection Under Distribution Shift

FogTTA: Online Test-Time Adaptation for Robust Transformer-based Object Detection in Foggy Weather