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

arXiv:2610.00067 · cs.CV · Submitted 2026-09-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "Robust Online Aero-Engine Blade Defect Detection via Dual-Alignment Test-Time Adaptation".

Tom: 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.

Jane: First, who's behind it and why it matters.

Title and authors: Tom: So, we're looking at the paper titled "Robust Online Aero-Engine Blade Defect Detection via Dual-Alignment Test-Time Adaptation," and the authors are Zhaoyang Wang, Haiyong Chen, Dongying Li, Yining Wang, Huapeng Wu, Xinwei Lv, and ATIK SHAHARIAR. This research focuses on making deep defect detectors work reliably in real-time when inspecting aero-engine blades.

Jane: Exactly; the title tells us the main goal is robustness during online inspection through test-time adaptation, which is a key concept here. It's about keeping the detection accurate even when things like the background or lighting shift between different testing conditions.

Lu: The authors are clearly tackling a very practical problem in aerospace engineering where failure can be catastrophic, so their focus on reliable visual inspection is extremely important for safety protocols.

Meng: I see why they focused on online adaptation; if we have to stop the line to retrain the model every time the lighting changes, that defeats the purpose of real-time monitoring. It’s about keeping things flowing while maintaining quality checks.

Lalam: It’s impressive how they've structured this work, combining local and global adaptation strategies into one cohesive framework for this specific application, which shows a deep understanding of the domain challenges.

The paper's summary: Tom: The paper explains that conventional deep defect detectors suffer from domain shifts at inference time because the target data differs from the training data, violating the standard assumption of being identically distributed. ABDD proposes an online adaptive detection framework using a teacher-student model to fix this by adapting both local and global domain shifts simultaneously.

Jane: To break it down simply, they suggest that local variations in defect shape can be handled by pseudo-box supervision, while global variations like background changes are addressed through feature statistics alignment using both normal and defective target images.

Lu: That dual hypothesis is the core strength; they hypothesize that these two adaptation methods work together to compensate for the different types of shifts we see in production data.

Meng: So, they aren't just trying one thing; they have a mechanism for local morphological variations and another mechanism for large-scale visual style differences, which makes it more comprehensive than single-strategy TTA methods.

Lalam: It’s fascinating how they specifically address the issue where sparse defects make pseudo labels unreliable by incorporating an Uncertainty-aware Box Filtering mechanism to filter out those noisy predictions before they influence the student model.

The paper's improvements: Tom: The methodology introduces several specific techniques to achieve this, including Pseudo-Box Alignment which uses classification and regression losses with GIoU and L1 losses, coupled with that UBF mechanism we just touched on. Plus, they have Feature Distribution Alignment which builds a source-domain memory bank to retrieve statistics from the nearest source images during adaptation via KL divergence loss.

Jane: And then they make sure the model itself is efficient by using Sparse Dilated Mona, which allows them to perform parameter-efficient delta tuning by freezing most parameters and only updating lightweight modules, saving quite a few parameters.

Lu: The parameter efficiency aspect with Sparse Dilated Mona is interesting because it directly addresses the cost of adaptation; reducing trainable parameters from 166n down to 54n shows a real focus on practical implementation constraints in deployment.

Meng: That reduction in trainable parameters is significant for deployment, and the teacher update mechanism using an exponential moving average helps control the forgetting problem, which is crucial when you’re constantly adapting online.

Lalam: The combination of these parts—the local pseudo-label supervision, the global feature statistics alignment, and the parameter efficiency via SDM—is what makes their proposed Dual-Alignment Strategy so robust across different domain shift scenarios like illumination or viewpoint changes.

Conclusion: Tom: So, to wrap up on this paper about Robust Online Aero-Engine Blade Defect Detection via Dual-Alignment Test-Time Adaptation, the main point is that combining pseudo-box supervision for local morphology and feature statistics alignment for global style creates a very stable online detection framework.

Jane: They successfully show that this dual strategy improves detection robustness across different shifts, achieving gains like thirteen point seven percent improvement under illumination variation on one dataset, which really validates the approach in practice.

Lu: The authors confirm that the combination of DAS—DAS Pseudo-Box and Feature Distribution Alignment—achieves the most stable performance when tested across CD-AeBD I, CD-AeBD II, and HD-AeBD datasets.

Meng: From a practical standpoint, even though it requires two forward propagations during test time, their research validates the value of online adaptive detection for industrial inspection settings where things are constantly changing.

Lalam: This paper really pushes us toward building more reliable AI systems that can handle real-world noise and variation in high-stakes environments like aerospace manufacturing, which is a huge step forward for dependable quality assurance.

Zhaoyang Wang, Haiyong Chen, Dongying Li, Yining Wang, Huapeng Wu, Xinwei Lv, ATIK SHAHARIAR

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

cs.CV

Submitted: 2026-09-04

Updated: 2026-09-04

Importance score: 81/100

The gist: 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

Key concepts

Dual-Alignment Strategy (DAS)
This strategy tackles two types of domain shifts simultaneously. Pseudo-Box Alignment handles local variations in defect shape using uncertainty filtering, while Feature Distribution Alignment addresses global shifts like changes in lighting or background by aligning feature statistics with stored source domain statistics.
Pseudo-Box Alignment
This component uses pseudo-labels generated during test time to supervise the student model's predictions. A crucial part is the Uncertainty-aware Box Filtering (UBF) mechanism, which filters out unreliable pseudo boxes based on classification and localization entropy before they influence training.
Feature Distribution Alignment
To combat global shifts in appearance, this technique aligns intermediate feature statistics of the target image with a pre-stored memory bank of source domain statistics. This alignment is performed using KL divergence loss, ensuring the model learns features that are invariant to these global visual changes.

Terminology

Summary

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. The gist: ABDD introduces an online adaptive detection framework based on test-time adaptation that utilizes a Dual-Alignment Strategy to jointly adapt global visual style and local defect morphology.

Problem Statement

Aero-engine blades (AeBs) are critical components whose subtle surface defects can lead to structural degradation or catastrophic failures, yet conventional deep defect detectors suffer from domain shifts at inference time. These shifts arise because target-domain data differs from the source-domain training distribution, violating the independent and identically distributed assumption. Existing detection-oriented Test-Time Adaptation (TTA) methods are insufficient for fine-grained AeB defect detection because they often rely on abundant and reliable pseudo labels, which are unreliable when AeB images contain sparse defects, making normal images uninformative or erroneous pseudo boxes harmful. Furthermore, many feature-alignment methods are designed for architectures like Faster R-CNN with ROI pooling, limiting their compatibility with modern YOLO and DETR-based detectors.

Proposed Framework: Dual-Alignment Strategy (DAS)

ABDD proposes a framework based on a teacher–student model employing a Dual-Alignment Strategy (DAS) to complementarily adapt local and global domain shifts. This strategy involves two main components:

  1. Pseudo-Box Alignment: This addresses local shifts caused by defect morphology variations, where pseudo-label supervision is useful. The loss function, defined as Equation (2), combines classification loss with regression losses using GIoU and L1 losses for the student predictions. Crucially, an Uncertainty-aware Box Filtering (UBF) mechanism evaluates pseudo boxes using classification confidence, classification entropy, and localization entropy to remove unreliable predictions before they supervise the student.

  2. Feature Distribution Alignment: This component mitigates global shifts caused by background or illumination variations by aligning intermediate feature statistics with source-domain statistics. A source-domain memory bank is constructed offline to store multilevel channel-wise feature statistics, and during test-time adaptation, the teacher extracts the target image's statistic vector as a query to retrieve its k nearest source statistics from the memory bank for alignment via KL divergence loss (Equation 6).

Parameter Efficiency and Stability

To improve efficiency and reduce catastrophic forgetting, ABDD employs two key mechanisms:

  1. Sparse Dilated Mona (SDM): This module enables parameter-efficient delta tuning by freezing most parameters and updating only lightweight modules. SDM is inspired by Mona, inserting two modules into each Swin block to form the SSDM backbone. Compared to explicit convolutions, SDM reduces the trainable parameters of the parallel convolution branches from 2(32 + 52 + 72)n = 166n to 2 × 3 × 32n = 54n, saving 112n parameters.

  2. Teacher Update: The teacher model is updated using an exponential moving average of the student parameters (Equation 9).

Experimental Validation and Results

ABDD was evaluated on two self-collected datasets, CD-AeBD and HD-AeBD, under multiple domain-shift scenarios. Quantitative results show that ABDD consistently improves detection robustness under domain shifts. For example, on CD-AeBD Subset I, ABDD improved mAP@50 by 13.7%, 5.9%, and 2.6% under illumination variation, viewpoint change, and background shift, respectively. Furthermore, in CD-AeBD II (where viewpoint and illumination change simultaneously), ABDD achieved an average 11.5% mAP@50 gain. On HD-AeBD, ABDD improved mAP@50 by 2.8% and mAP@50:95 by 6.4%. Qualitative results confirm that ABDD produces more accurate detections than the baseline detector, effectively adapting to the target domain at test time and suppressing defect-irrelevant features in shallow feature maps. Ablation studies show that the combination of DAS (DAS PB + FD) achieves the most stable performance across CD-AeBD I, CD-AeBD II, and HD-AeBD.

Conclusion

ABDD successfully combines pseudo-box supervision with feature-statistics alignment and efficient delta tuning to jointly exploit local and global domain shifts during test-time adaptation. This approach demonstrates that combining these strategies improves both domain robustness and defect-level reliability, achieving significant performance gains across various domain shift scenarios for aero-engine blade defect detection. Despite its computational cost requiring two forward propagations, ABDD validates the practical value of online adaptive detection in industrial inspection settings. Future work will focus on faster adaptation strategies and vision-language guidance to improve false-negative discovery under extremely large domain gaps.

How it works

Improvements for AI systems

Based on the provided paper, here are the specific improvements to existing AI systems that can be achieved by implementing or adapting their proposed framework (ABDD):

  1. The improved system can perform real-time, online defect detection on aero-engine blades directly from a production line camera feed without requiring periodic offline retraining.

  2. It achieves significant robustness against domain shifts (changes in background, viewpoint, and illumination) that plague current deep learning detectors during deployment in industrial settings.

  3. The system can reliably detect sparse defects (like scratches or dents) even when the defect appearance varies significantly across different production batches or inspection conditions.

  4. It provides high-precision localization for defects, specifically showing superior performance on stringent metrics like mAP@50:95, which is critical for safety-related manufacturing quality assurance.

  5. The system offers a parameter-efficient adaptation mechanism (Sparse Dilated Mona) that minimizes computational overhead during test-time updates while effectively mitigating catastrophic forgetting from source domain knowledge.

  6. It leverages a dual approach—combining pseudo-box supervision (for local morphology) and feature statistics alignment (for global style)—to ensure both instance-level and image-level adaptation are effective, leading to more stable performance than methods relying on only one strategy.

Sources

Related papers