2 of 14

Introduction

Welding defect detection is essential for ensuring the safety and reliability of offshore structures, particularly in oil and gas piping systems.
Traditional Non-Destructive Testing (NDT) methods, such as ultrasonic and radiographic testing, often fail to detect internal or subtle defects, leading to structural failures and high maintenance costs.
Advanced AI-driven techniques offer enhanced accuracy and automation, but existing neural network-based approaches often lack domain-specific optimization and interpretability, limiting their real-world adoption.

INTERNATIONAL CONFERENCE ON AI FOR THE OCEANS 2025

ICAIO 2025

Images of defect types in the RIAWELC dataset

Crack Lack of Penetration Porosity No Defect

3 of 14

Problem Statement

Existing welding defect detection methods face the following challenges:

Accuracy Limitations: Traditional NDT techniques may not detect subtle defects in offshore environments.
Suboptimal Model Performance: Pre-trained models often underperform due to a lack of domain-specific optimization.
Lack of Interpretability: Black-box models hinder trust and safe deployment in high-stakes applications.

Objectives

To develop an adaptive AI framework Adapt-WeldNet for welding defect detection that:

Enhances model selection and parameter tuning through adaptive transfer learning.
Improves interpretability using Explainable AI (XAI) techniques, including Grad-CAM and LIME.
Incorporates domain expertise with the Defect Detection Interpretability Analysis (DDIA) framework, ensuring transparency, reliability, and human-in-the-loop validation.
Achieves trustworthy AI by combining adaptive modeling with explainability for safe and efficient welding defect detection.

INTERNATIONAL CONFERENCE ON AI FOR THE OCEANS 2025

ICAIO 2025

4 of 14

Literature Review

INTERNATIONAL CONFERENCE ON AI FOR THE OCEANS 2025

ICAIO 2025

Author(s)	Method/Model	Technique/Approach	Key Insights
Sun et al. (2019) [1]	Machine Vision Algorithm	Image-based defect classification	High accuracy but lacks interpretability and robustness in diverse conditions.
Vasan et al. (2024) [2]	Ensemble-Based Deep Learning	Combining multiple models for defect classification	Effective in NDT of submerged arc welds, but lacks domain-specific tuning.
Yang et al. (2018) [3]	Transfer Learning with GoogLeNet + MLP	Fine-tuning on spot-welding evaluations	High accuracy with transfer learning, effective with limited labeled data.
Kumar et al. (2023) [4]	VGG16 + Data Augmentation	Fine-tuned for X-ray weld defects	Effective despite data imbalance, improved through augmentation.
Zhang et al. (2024) [5]	Vision Transformer (ViT)	Deep Feature Extraction for X-ray weld defect detection	Enhanced feature extraction but computationally intensive.
Geng et al. (2024) [6]	VAE-DCGAN	Generative approach for weld defect feature representation	Improves defect detection performance by enhancing feature diversity.
Ajmi et al. (2024) [7]	Faster RCNN	Real-time weld defect detection	Suitable for practical, time-sensitive applications.
Perri et al. (2023) [8]	WelDeNet (CNN-based)	Deep CNN with 14 convolutional layers	Remarkable accuracy in real-world defect classification.

5 of 14

INTERNATIONAL CONFERENCE ON AI FOR THE OCEANS 2025

ICAIO 2025

Proposed Adapt-WeldNet framework

Adaptive AI Framework: Dynamically optimizes model selection, transfer learning strategies, and hyperparameters to achieve the best performance.

Multi-Model Evaluation: Integrates eight pre-trained architectures for robust defect detection.

Adaptive Optimizer Selection: Leverages seven optimizers to enhance convergence and model stability. Including various learning rate and batch size.

Flexible Transfer Learning Modes: Supports three training modes to adapt to domain-specific challenges.

Explainable AI Integration: Combines Grad-CAM and LIME to ensure transparency and interpretability.

Defect Detection Interpretability Analysis (DDIA) Framework

6 of 14

Adapt-WeldNet Framework:

The experiments utilized the RIAWELC dataset [9], X-ray weld images categorized into four defect types: Porosity (P), Lack of Penetration (LP), Crack (C), and No Defect.

Adaptive Model Selection:

8 Pre-trained Architectures:
ResNet18, DenseNet121, EfficientNet-B0, EfficientNet-V2-S

, MobileNet-V2, Wide ResNet50-2, ShuffleNet V2 X0.5,

SqueezeNet1.0.

Three learning strategies: Freeze Early Layers, Freeze All Layers, Fine-Tune All.

Adaptive Hyperparameter Optimization:

Optimization, varying learning rates, batch sizes, and seven optimizers Adam, AdamW, SGD, RMSprop, Adagrad, Adadelta, and Adamax.

Explainability Layer:

Grad-CAM [10] and LIME [11] for defect visualization.

INTERNATIONAL CONFERENCE ON AI FOR THE OCEANS 2025

ICAIO 2025

Architecture

Proposed Adapt-WeldNet framework

Final Output: Optimal model selection with enhanced explainability and trustworthy AI through DDIA.

7 of 14

Proposed Methodology - Adapt-WeldNet Framework and Explainable AI

Data Preprocessing:

Image resizing, normalization, and data augmentation to balance defect classes.

Adaptive AI Framework for Welding Defect Detection
Integrates adaptive model selection, transfer learning, and optimizer tuning.
Dynamically optimizes model configurations for offshore environments.
Explainable AI (XAI) Techniques:
Utilizes Grad-CAM and LIME for visualizing decision-making processes.
Enhances transparency, interpretability, and reliability.
Benefits:
Identifies optimal model architectures and hyperparameters.

Improves model accuracy and explainability for critical maritime operations.

INTERNATIONAL CONFERENCE ON AI FOR THE OCEANS 2025

ICAIO 2025

Sample Images of Data Preprocessing

8 of 14

Proposed Methodology - Defect Detection Interpretability Analysis (DDIA) Framework

Defect Detection Interpretability Analysis (DDIA):

Goes beyond traditional XAI by incorporating a human-in-the-loop approach.
Human-in-the-Loop (HITL) Approach:

Involves domain experts (e.g., ASNT NDE Level II auditors) to assess model

outputs.

Ensures safe and transparent deployment in critical environments.

Expert-Guided Evaluation:

Combines Grad-CAM and LIME outputs with expert feedback.
Assesses detection accuracy, image quality, defect visibility, and confidence.

Impact:

Fosters trustworthy AI in maritime defect detection.
Balances model performance with practical, real-world validation.

Quantitative Metrics for Interpretability in Defect Localization :

To assess the interpretability and localization performance of Grad-CAM in

identifying welding defects, the Average Recall is calculated as:

INTERNATIONAL CONFERENCE ON AI FOR THE OCEANS 2025

ICAIO 2025

Defect Detection Interpretability Analysis (DDIA) Framework

9 of 14

Experimental Results - Adapt-WeldNet Framework

INTERNATIONAL CONFERENCE ON AI FOR THE OCEANS 2025

ICAIO 2025

Parallel coordinate plot showing the relationship between hyperparameters and the objective value in the adaptive AI model optimization process. Darker colors indicate higher objective values (Best viewed in color). The optimal configuration includes the DenseNet121 architecture fine-tuned with all layers updated, using the AdamW optimizer with a learning rate of 6.41 × 10⁻⁵ and a batch size of 16.

10 of 14

Experimental Results - Explainable AI

INTERNATIONAL CONFERENCE ON AI FOR THE OCEANS 2025

ICAIO 2025

The explainability analysis of welding defect detection such as Crack, Lack of Penetration (LP), and Porosity (P) using Grad-CAM and LIME demonstrates their effectiveness under different imaging conditions such as Noise, Underexposure, Overexposure, Clear Image.

Recall-based evaluation showcasing the Grad-CAM mask, ground truth mask, and recall output for a sample image. This analysis provides insights into the interpretability of Grad-CAM’s predictions and its ability to focus on relevant defect regions. The average recall, calculated across the entire dataset. Average Recall of 0.7722 indicates reliable defect localization across diverse conditions, fostering model transparency and confidence in defect detection outcomes.

11 of 14

Experimental Results - Domain expertise through a human-in-the-loop (HITL)

INTERNATIONAL CONFERENCE ON AI FOR THE OCEANS 2025

ICAIO 2025

Domain experts assess (a) X-ray films of varying quality (noisy, overexposed, underexposed, good) to ensure robustness. (b) Grad-CAM provides clearer defect localization with fewer errors compared to LIME, making it more reliable for defect detection. (c) Confidence distribution ranges from 1 (poor) to 5 (excellent).

Impact of Image Quality on XAI:

Good quality images (37.3%) yield reliable results, while noisy (14.7%), overexposed (21.7%), and underexposed (26.3%) images cause inconsistencies.

Defect Detection by Grad-CAM and LIME:

Grad-CAM: Clearer defect localization, reliable for prominent defects.
LIME: Fine-grained insights but sensitive to noise and underexposure.

Expert Confidence:

Grad-CAM receives higher confidence scores (4-5), while LIME scores slightly lower, indicating less trust in critical applications.

12 of 14

Research Findings

Enhanced Accuracy and Reliability:

Integrates adaptive learning and hyperparameter tuning to optimize performance.

Improved Interpretability and Trust:

Incorporates XAI techniques (Grad-CAM and LIME) to ensure decision transparency.
The DDIA framework bridges the gap between AI predictions and human validation, fostering Trustworthy AI.

High Practical Impact:

Adapt-WeldNet supports critical applications in maritime and offshore environments, ensuring safe and reliable operations.

First-of-its-Kind Approach:

Combines adaptive model optimization with domain-specific interpretability, setting a new standard in welding defect detection.

INTERNATIONAL CONFERENCE ON AI FOR THE OCEANS 2025

ICAIO 2025

13 of 14

Conclusion and Future Scope

Conclusion:

Developed Adapt-WeldNet: An innovative adaptive AI framework designed to optimize model selection and hyperparameter tuning, ensuring robust and accurate defect detection in challenging offshore environments.

Introduced the Defect Detection Interpretability Analysis (DDIA) Framework: A pioneering approach that integrates domain expertise through a human-in-the-loop (HITL) mechanism, enhancing the validation of explainability and fostering trustworthy AI deployment.

Proposed a Novel Recall-Based Evaluation Metric: Specifically designed to quantitatively assess the interpretability of XAI methods in defect localization, setting a new standard for reliable evaluation in safety-critical applications.

Future Work:

Exploring Vision Transformers for improved feature extraction and classification.
Enhancing real-time defect detection by optimizing computational efficiency.
Developing automated expert feedback loops for continuous model improvement.

INTERNATIONAL CONFERENCE ON AI FOR THE OCEANS 2025

ICAIO 2025

14 of 14

References

INTERNATIONAL CONFERENCE ON AI FOR THE OCEANS 2025

ICAIO 2025

[1] Jun Sun, Chao Li, Xiao-Jun Wu, Vasile Palade, and Wei Fang. An effective method of weld defect detection and classification based on machine vision. IEEE Transactions on Industrial Informatics, 15(12):6322–6333, 2019.

[2] Vinod Vasan, Naveen Venkatesh Sridharan, Rebecca Jeyavadhanam Bala sundaram, and Sugumaran Vaithiyanathan. Ensemble-based deep learning model for welding defect detection and classification. Engineering Applications of Artificial Intelligence, 136:108961, 2024.

[3] Ye Yang, Peng Zheng, Hao He, Tianyu Zheng, Lei Wang, and Shan He. An evaluation method of acceptable and failed spot welding products based on image classification with transfer learning technique. In Proceedings of the 2nd International Conference on Computer Science and Application Engineering, pages 1–6, 2018

[4] Dheeraj Dhruva Kumar, Cheng Fang, Yue Zheng, and Yuqing Gao. Semi supervised transfer learning-based automatic weld defect detection and visual inspection. Engineering Structures, 292:116580, 2023.

[5] Rui Zhang, Donghao Liu, Qiaofeng Bai, Liuhu Fu, Jing Hu, and Jinlong Song. Research on x-ray weld seam defect detection and size measurement method based on neural network self-optimization. Engineering Applications of Artificial Intelligence, 133:108045, 2024.

[6] Chen Geng, Sheng Buyun, Fu Gaocai, Chen Xiangxiang, and Zhao Guangde. A gan-based method for diagnosing bodywork spot welding defects in response to small sample condition. Applied Soft Computing, 157:111544, 2024.

[7] Chiraz Ajmi, Juan Zapata, Sabra Elferchichi, and Kaouther Laabidi. Advanced faster-rcnn model for automated recognition and detection of weld defects on limited x-ray image dataset. Journal of Nondestructive Evaluation, 43(1):14, 2024.

[8] Stefania Perri, Fanny Spagnolo, Fabio Frustaci, and Pasquale Corsonello. Welding defects classification through a convolutional neural network. Manufacturing Letters, 35:29–32, 2023.

[9] Benito Totino, Fanny Spagnolo, and Stefania Perri. Riawelc: a novel dataset of radiographic images for automatic weld defects classification. International Journal of Electrical and Computer Engineering Research, 3(1):13–17, 2023.

[10] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017.

[11] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. ”why should i trust you?”: Explaining the predictions of any classifier, 2016.