Adapt-WeldNet: Adaptive Weld Defect Detection

• The paper introduces Adapt-WeldNet, an adaptive framework that systematically selects optimal neural architectures and hyperparameters for multi-class weld defect classification.
• It employs Optuna-guided search over pretrained backbones and tailored transfer-learning modes to enhance radiographic weld image analysis in harsh offshore conditions.
• The framework integrates expert validation via Grad-CAM and LIME under its DDIA layer, ensuring trustworthy AI interpretability and safety in deployment.

Adapt-WeldNet is an adaptive framework for weld defect detection in maritime and offshore operations that combines systematic model selection for radiographic X-ray weld classification with a companion interpretability and expert-audit procedure, Defect Detection Interpretability Analysis (DDIA). In its formal usage, the name refers to the framework introduced for multi-class classification of weld images into crack, lack of penetration, porosity, and no defect, using Optuna-guided selection over pretrained backbones, transfer-learning modes, optimizers, and hyperparameters, followed by Human-in-the-Loop validation of Grad-CAM and LIME explanations by certified ASNT NDE Level II professionals (S et al., 1 Aug 2025).

1. Definition and scope

Adapt-WeldNet was proposed to address two deficiencies identified in weld inspection for offshore and marine environments. First, offshore and maritime welding must tolerate harsh environments, dynamic loads, corrosion, and strict safety standards, while defects such as cracks, porosity, and lack of penetration can compromise the structural integrity of pipelines, rigs, and vessels. Second, the paper argues that traditional non-destructive testing methods, including ultrasonic, radiographic, and magnetic-particle methods, can struggle with subtle or internal defects under adverse imaging conditions such as overexposure, underexposure, and noise in X-ray films, and that prior neural approaches often rely on arbitrarily selected pretrained architectures and fixed training recipes without systematic optimization or interpretability support (S et al., 1 Aug 2025).

Within that setting, Adapt-WeldNet is defined as a principled adaptive search framework for weld-defect classification, while DDIA is defined as a Human-in-the-Loop, Trustworthy AI interpretability evaluation layer. The task is multi-class classification of radiographic weld images into four classes: crack (C), lack of penetration (LP), porosity (P), and no defect (ND). The framework therefore combines two distinct but coupled objectives: maximizing validation accuracy by systematic search over design choices, and validating the transparency and operational acceptability of the resulting model by expert review of explanation maps.

A central feature of the framework is that deployment is not treated as a purely statistical endpoint. The model is advanced to deployment only after expert validation of explainability outputs. This places Adapt-WeldNet in a safety-governed inspection pipeline rather than a classifier-only benchmark.

2. Data basis and workflow

The empirical basis of Adapt-WeldNet is the RIAWELC dataset of 24,407 radiographic X-ray weld images labeled into four classes: porosity, lack of penetration, crack, and no defects (S et al., 1 Aug 2025). The reported workflow begins with data ingestion, image resizing, and normalization, followed by augmentation used to correct class imbalance. The original training distribution was uneven, with $C=7008$, $LP=3712$, $ND=5352$, and $P=5768$. After augmentation, the training distribution was balanced to 7008 samples per class.

The workflow reported in the paper is explicitly end-to-end. After preprocessing and augmentation, candidate models are drawn from eight pretrained ImageNet backbones. Transfer learning mode, optimizer family, learning rate, and batch size are then selected adaptively through Optuna-guided search. The resulting classifier is evaluated using validation accuracy and a confusion matrix, and then passed to an interpretability stage in which Grad-CAM and LIME are applied across varied imaging conditions, including clear, noisy, overexposed, and underexposed films. DDIA then adds expert audits that record detection judgments, image quality, defect visibility, defect type, and confidence scores. Only after this stage is the model considered suitable for deployment.

The paper states that training and validation sets were formed with balanced representation per class, but a separate test set is not explicitly described. For the interpretability localization metric, 1,031 images were annotated by certified domain experts. Inter-rater reliability is not reported. No focal loss, class weighting, or cost-sensitive training is reported; class imbalance was addressed by augmentation alone.

3. Adaptive model selection and optimization protocol

The adaptive core of Adapt-WeldNet is a structured search over backbone architecture, transfer-learning regime, optimizer, and training hyperparameters (S et al., 1 Aug 2025). The candidate ImageNet-initialized backbones are ResNet18, DenseNet121, EfficientNet-B0, EfficientNet-V2-S, MobileNet-V2, WideResNet50-2, ShuffleNet V2 X0.5, and SqueezeNet1.0. The transfer-learning modes are Freeze Early Layers, Freeze All Layers except classifier, and Fine-tune All. The optimizer candidates are Adam, AdamW, SGD, RMSprop, Adagrad, Adadelta, and Adamax. The learning rate is sampled log-uniformly in $[10^{-5}, 10^{-2}]$, and the batch size is selected from $\{16,32,64\}$.

Training uses Cross-Entropy Loss,

$L = -\sum_i y_i \log p_i,$

for up to 100 epochs with early stopping of patience 5. At the outer optimization level, Optuna maximizes validation accuracy over the search space. Conceptually, the selection problem is

$$\max \mathrm{Acc}_{val}(M,\mathrm{mode},\mathrm{opt},lr,bs),$$

subject to the reported search space over model $M$, training mode, optimizer, learning rate, and batch size.

The reported ablations indicate that Fine-tune All generally outperformed Freeze Early Layers and Freeze All Layers, with higher objective values and lower variance. EfficientNet-B0, EfficientNet-V2-S, and DenseNet121 performed best at low learning rates of approximately $10^{-5}$ with AdamW or Adam. Batch sizes 16 and 64 tended to yield higher objective values, while 32 showed more variability. DenseNet121 and WideResNet50-2 often achieved superior objective values relative to MobileNet-V2 and ShuffleNet V2 X0.5.

The final selected configuration was DenseNet121 with Fine-tune All, AdamW, learning rate $LP=3712$0, and batch size 16. The paper presents a confusion matrix for this classifier, but numeric confusion-matrix entries are not provided.

4. Defect Detection Interpretability Analysis

DDIA is the interpretability and governance component associated with Adapt-WeldNet. It combines post hoc XAI techniques with domain-specific evaluation by certified ASNT NDE Level II professionals and is explicitly framed as Human-in-the-Loop and aligned with Trustworthy AI principles of transparency, accountability, and reliability (S et al., 1 Aug 2025).

The two XAI methods used are Grad-CAM and LIME. For Grad-CAM, the paper gives the class-specific weights

$LP=3712$1

and the heatmap

$LP=3712$2

stated in the paper as

$LP=3712$3

For LIME, the locality-aware surrogate objective is

$LP=3712$4

with

$LP=3712$5

A novel DDIA contribution is a recall-based localization metric for Grad-CAM on expert-annotated weld defects. The per-image recall is

$LP=3712$6

where $LP=3712$7 is the Grad-CAM heatmap and $LP=3712$8 is the binary ground-truth defect region. The average recall is

$LP=3712$9

The reported result is $ND=5352$0 on 1,031 expert-annotated images.

The DDIA audit schema records, for both Grad-CAM and LIME, Detection (Yes/No), Image quality (clear, underexposed, overexposed, noisy), Defect visibility (clearly visible, partially visible, not visible), Defect type (C, LP, P), and Confidence scores from 1 to 5. The image-quality distribution in the XAI test set is reported as noisy 14.7%, overexposed 21.7%, underexposed 26.3%, and good/clear 37.3%.

Qualitatively, Grad-CAM was reported to robustly localize defect regions across clear, overexposed, underexposed, and noisy conditions, providing broader coverage. LIME provided fine-grained, boundary-focused explanations, but showed less consistent coverage under noise and underexposure. Experts awarded higher confidence to Grad-CAM on average and reported fewer misdetections relative to LIME.

5. Reported findings, deployment logic, and limitations

The principal quantitative outputs reported for Adapt-WeldNet are validation accuracy as the Optuna objective, the non-numeric confusion matrix for the best classifier, and the DDIA localization recall of 0.7722 (S et al., 1 Aug 2025). The paper also reproduces general metric definitions for accuracy, precision, recall, and F1:

$ND=5352$1

$ND=5352$2

However, it does not report numeric values for accuracy, precision, recall, F1, ROC-AUC, or per-class scores for the final classifier. No statistical significance testing is reported. No external baselines beyond the eight-candidate architecture set are included.

The deployment logic is governed by expert acceptance rather than by accuracy alone. DDIA requires per-image expert review of XAI outputs, and expert feedback is used to identify gaps such as ambiguous regions in noisy films, recommend data refinements, and motivate retraining strategies. This feedback loop makes expert validation a prerequisite for deployment in safety-critical offshore inspection systems.

The paper emphasizes qualitative robustness under varied X-ray quality conditions, but several standard generalization and reliability analyses are not explicitly reported. Cross-site generalization, out-of-distribution detection, calibration, threshold selection, Platt scaling, and ECE are absent. Compute resources, deep-learning framework versions, model footprint, runtime throughput, quantization, pruning, and code availability are also not provided in the pre-publication version. The selected model is DenseNet121, but no model-size or latency claims are reported.

The limitations discussed in the paper include possible failure under extreme exposure issues or unseen defect patterns, sensitivity to preprocessing and augmentation, the coarse nature of Grad-CAM, instability of LIME under heavy noise, and the fact that explanations are not causal. No inter-rater reliability is reported for expert annotations; the paper notes that future work could incorporate measures such as Cohen’s kappa. Proposed extensions include multi-modal sensor fusion such as X-ray plus ultrasonic or PAUT, active learning, semi-supervised learning, calibration, out-of-distribution detection, and advanced XAI methods tailored to industrial radiography.

6. Broader uses and interpretive extensions of the name

Only one paper formally introduces a method named Adapt-WeldNet. Several subsequent papers use the term only interpretively, as a shorthand for adaptation-oriented welding AI rather than as an official method name. This distinction is central to the current literature.

Paper	Status of the name	Core adaptive mechanism
"Advancing Welding Defect Detection in Maritime Operations via Adapt-WeldNet and Defect Detection Interpretability Analysis" (S et al., 1 Aug 2025)	Official named framework	Optuna-based adaptive classifier selection plus DDIA expert-validated XAI
"Learning and Adaptation in Wire Arc Additive Manufacturing Bead Geometry Control" (Lu et al., 27 May 2026)	Not an official name; mapped interpretively to Adapt-WeldNet	Layer-wise adaptive fine-tuning of a simple RNN and one-step-ahead predictive control
"A cross-process welding penetration status prediction algorithm based on unsupervised domain adaptation in laser and TIG welding" (Li et al., 24 Jun 2026)	Not an official name; described as UDA + GSDE	ResNet-50 feature extraction, multiple bottlenecks, CDAN-style adversarial alignment, semantic centroid alignment, MixMatch, and gradual source domain expansion
"Data-Driven Model Reduction using WeldNet: Windowed Encoders for Learning Dynamics" (Dahal et al., 11 Dec 2025)	Not an official name; discussed as “Toward Adapt-WeldNet”	Windowed encoder-decoder charts, latent propagators, and transcoders with adaptation strategies for regime changes
"Improved Training Strategies for Physics-Informed Neural Networks using Real Experimental Data in Aluminum Spot Welding" (Zak et al., 6 Aug 2025)	Not an official name; framed as a guide to designing an Adapt-WeldNet framework	Progressive inclusion of experimental losses, rolling-window scheduling, and conditional material-parameter updates in a PINN

These papers span substantially different problem classes. In WAAM bead geometry control, the adaptive RNN controller updates its model using previous-layer prediction error and yields reported height-SD reductions of approximately 33% with edges included and approximately 41% with edges excluded relative to the constant-input baseline. In cross-process TIG/laser penetration prediction, the UDA+GSDE framework reaches average accuracies of 90.65% on TIGFH and 90.72% on LSPS in same-process settings, and 80.48% for TIG to Laser and 81.13% for Laser to TIG in cross-process transfer. In WeldNet, the name refers originally to “Windowed Encoders for Learning Dynamics,” and the paper develops a windowed autoencoder–propagator–transcoder framework for nonlinear model reduction; “Adapt-WeldNet” appears there as an extension concept rather than a formal algorithm. In aluminum spot welding PINNs, Adapt-WeldNet is again a design interpretation rather than the authors’ official method name.

This suggests that “Adapt-WeldNet” now has two levels of meaning in the arXiv literature. At the narrow level, it denotes the 2025 maritime weld-defect framework with DDIA. At the broader, interpretive level, it functions as a label for welding-related neural systems whose defining property is adaptation: across architectures and hyperparameters, across thermal layers in WAAM, across domains in TIG/laser transfer, across temporal windows in latent surrogate modeling, or across physics and data regimes in PINN-based resistance spot welding.