A welding penetration prediction model for laser welding process based on self-supervised learning using physics-informed neural networks
SimPhysNet combines physics-informed neural networks with self-supervised contrastive learning to achieve 96.06% accuracy in laser welding penetration prediction using only 200 labeled images.
Key Findings
Methodology
This paper introduces SimPhysNet, a novel framework that integrates physics-informed neural networks (PINNs) with contrastive self-supervised learning to predict laser welding penetration states with minimal labeled data. The approach begins with a pretraining phase where the encoder, based on ResNet-18, is trained on unlabeled molten pool images using a contrastive loss that maximizes similarity between augmented views. To embed physical consistency, the PINN module incorporates PDE constraints related to heat conduction and energy conservation, regularizing the feature space. Additionally, three image augmentation tasks—rotation prediction, Gaussian blur estimation, and random cropping—are designed to enforce the learning of robust, hierarchical features relevant to weld pool morphology. In the fine-tuning phase, a prototypical network constructs class prototypes from a few labeled samples, enabling accurate classification of penetration status. Experimental results demonstrate that with only 200 labeled images, the model achieves 96.06% accuracy, comparable to models trained on full datasets, showcasing its effectiveness in data-scarce industrial scenarios.
Key Results
- The proposed SimPhysNet achieves a classification accuracy of 96.06% using only 200 labeled images, which is nearly equivalent to the performance of models trained on the entire labeled dataset (accuracy ~97%). This demonstrates the method’s high data efficiency.
- In diverse welding conditions, including variations in power, speed, and material thickness, the model maintains stable performance, indicating strong generalization and robustness.
- Incorporating PINN regularization improves the physical plausibility of learned features, leading to better transferability to unseen conditions and complex physical environments, outperforming purely data-driven models in few-shot scenarios.
Significance
This work addresses a critical bottleneck in industrial laser welding—dependence on large labeled datasets—by providing a highly accurate, low-cost, and physically consistent predictive model. The integration of physics-based constraints into self-supervised learning not only enhances accuracy but also improves interpretability, making it suitable for real-time quality control and process optimization. The approach paves the way for scalable, intelligent automation in manufacturing, reducing reliance on manual inspection and extensive data annotation. Its potential extends beyond welding, offering a blueprint for applying physics-guided deep learning to other complex physical systems with limited labeled data, thus significantly advancing the field of physics-informed AI in industry.
Technical Contribution
The key technical innovation of this paper lies in embedding PDE-based physical priors directly into the contrastive learning framework via PINN, serving as an implicit regularizer that enforces physical consistency of high-dimensional features. This contrasts with traditional supervised PINN applications that approximate specific solutions; here, PINN guides feature embedding to reflect underlying physics, improving robustness. The architecture combines a pretraining phase with contrastive loss, a PINN regularization module, and a few-shot classification stage based on prototypical networks. This multi-stage design effectively leverages unlabeled data, physical laws, and minimal labeled samples, achieving high accuracy with low annotation cost. The methodology also innovates in designing augmentation tasks tailored to weld pool morphology, ensuring hierarchical feature learning. Extensive ablation studies confirm the contributions of each component, demonstrating the framework’s superiority over baseline models.
Novelty
This study is the first to integrate PDE-based physics constraints directly into a self-supervised contrastive learning framework for industrial welding applications. Unlike prior works that focus solely on data-driven methods or supervised PINNs, this approach uses physics-informed regularization to guide feature learning from unlabeled data, ensuring physical plausibility. The combination of contrastive learning, PINN regularization, and few-shot classification in a unified architecture represents a significant advancement, enabling high-accuracy predictions with minimal labeled data. This paradigm shift opens new avenues for physics-guided AI in complex, data-scarce physical systems, setting a new benchmark for industrial predictive modeling.
Limitations
- The model’s physics constraints rely on known PDEs; for processes with poorly understood or highly nonlinear physics, the approach may not perform optimally. Its generalization to such scenarios requires further validation.
- Training complexity and computational cost are high, especially due to the PINN component, which may hinder real-time deployment in industrial settings without hardware acceleration.
- The current framework is validated primarily on controlled experimental data; real-world industrial environments with noise, vibrations, and optical disturbances may pose additional challenges that need addressing through further robustness studies.
Future Work
Future research will focus on extending the physics model to incorporate multiple physical phenomena, such as plasma dynamics and vaporization effects, to improve accuracy in complex welding environments. Additionally, integrating online learning and transfer learning techniques could enhance adaptability across different equipment and materials. Efforts to optimize PINN training algorithms will aim to reduce computational costs, enabling real-time deployment. Exploring explainability methods will also be crucial to improve trust and interpretability for industrial operators. Ultimately, the goal is to develop a comprehensive, physics-guided, real-time welding quality monitoring system that can be seamlessly integrated into automated manufacturing lines.
AI Executive Summary
Laser welding has revolutionized modern manufacturing with its ability to produce high-quality, precise joints rapidly. However, ensuring the weld penetrates fully into the material remains a persistent challenge. Traditional methods for predicting weld penetration rely heavily on extensive labeled datasets, which involve laborious post-process inspections—an expensive and time-consuming process. This bottleneck limits the scalability and real-time applicability of predictive models in industrial settings.
Recent advances in deep learning have shown promise in weld quality assessment, but their dependence on large labeled datasets hampers widespread adoption. To address this, the authors propose SimPhysNet, a novel framework that combines physics-informed neural networks (PINNs) with self-supervised contrastive learning. This hybrid approach leverages the abundant unlabeled data generated during welding processes and embeds physical laws directly into the learning process, ensuring the extracted features are physically meaningful.
The core idea involves pretraining the encoder network on unlabeled molten pool images using contrastive loss, which encourages the model to learn invariant features. Simultaneously, PINNs incorporate PDE constraints related to heat transfer and energy conservation, regularizing the feature space to adhere to physical principles. To further enhance robustness, three image augmentation tasks—rotation prediction, Gaussian blur estimation, and random cropping—are employed, forcing the model to learn hierarchical and invariant features relevant to weld morphology.
In the fine-tuning phase, a few labeled samples are used to construct class prototypes via a prototypical network, enabling accurate classification of penetration states with minimal annotation. Experimental results demonstrate that with only 200 labeled images, the model achieves 96.06% accuracy, comparable to models trained on full datasets, highlighting its efficiency and robustness.
This approach significantly reduces the need for costly data annotation, accelerates the deployment of predictive models in industrial environments, and enhances the interpretability of the predictions through physical constraints. The integration of physics and machine learning in SimPhysNet marks a substantial step toward intelligent, autonomous welding systems. Future work will focus on extending the physical models, improving computational efficiency, and validating the framework in real-world industrial settings, ultimately paving the way for fully automated, high-quality laser welding processes.
Deep Analysis
Background
Laser welding作为一种高效、精确的金属连接技术，已在航空航天、核能、船舶、汽车等行业得到广泛应用。早期研究主要集中在工艺参数优化和焊接质量检测，如Kang等利用光学发射光谱（OES）实现实时焊缝几何预测，Li等基于多任务卷积神经网络实现焊缝深宽预测。然而，深度学习模型的成功依赖于大量高质量标注数据，尤其是在焊接穿透深度的判定上，标注工作繁琐且成本高昂。近年来，少样本学习和自监督学习逐渐成为研究热点，旨在降低对标注数据的依赖。物理信息神经网络（PINN）作为结合物理规律的深度学习工具，也在焊接质量预测中展现出潜力，但多用于单一任务的正向或逆向问题，缺乏在复杂多物理场景中的系统应用。本文基于此背景，提出融合PINN和自监督对比学习的SimPhysNet，旨在突破工业焊接中的数据瓶颈，提升穿透预测的准确性和鲁棒性。
Core Problem
激光焊接中的穿透深度预测是确保焊接质量的关键环节。传统方法多依赖于大量标注数据，通过后续检测确认焊缝是否达到全穿透状态，过程繁琐且成本高。工业环境中，焊接参数变化复杂，材料多样，导致模型在实际应用中难以普适。现有深度学习模型虽然性能优异，但对标注数据的依赖限制了其推广。如何在少量标注数据条件下实现高精度、鲁棒的穿透预测，成为亟待解决的核心问题。这不仅关系到生产效率，也影响到焊接缺陷的早期识别和工艺优化。本文试图通过引入物理约束和自监督机制，解决数据不足带来的模型泛化能力不足的问题，为工业焊接自动化提供技术支撑。
Innovation
本研究的创新点主要体现在以下几个方面：
- �� 物理信息正则化：引入偏微分方程（PDE）约束到对比学习中，确保特征的物理合理性，提升模型的泛化能力。这是将PINN作为特征正则器的首次尝试，突破了传统PINN仅用于单一任务的局限。
- �� 自监督对比学习：采用Simsiam算法，通过未标注数据的增强视图学习稳定的特征表示，极大减少对标注数据的依赖。
- �� 图像增强任务设计：结合旋转预测、高斯模糊和随机裁剪三项预任务，强化模型对焊池形态和细节特征的理解，提升鲁棒性。
- �� 少样本分类策略：结合原型网络（Prototypical Network），在少量标注样本基础上快速构建类别原型，实现高精度分类。
- �� 多阶段训练架构：预训练阶段利用未标注数据进行特征学习，微调阶段通过少样本学习实现分类，整体架构高效且具有良好的迁移能力。
Methodology
- �� 数据采集：利用激光焊接设备和高速相机实时采集焊池图像，构建标注与未标注数据集。
- �� 预处理：对图像进行归一化、增强（旋转、高斯模糊、裁剪）以丰富训练样本。
- �� 自监督训练：采用Simsiam框架，输入增强视图，训练编码器（ResNet-18）和投影头，最大化视图间的相似性，学习全局和局部特征。
- �� 物理正则化：在对比损失中引入PINN，利用热传导和能量守恒的偏微分方程，确保特征的物理一致性。
- �� 图像任务：设计旋转角度预测、模糊程度预测和裁剪位置预测任务，强化模型对焊池形态和细节的理解。
- �� 微调阶段：利用少量标注样本，通过原型网络构建类别原型，进行分类训练。
- �� 评估：在测试集上计算准确率、召回率等指标，验证模型性能和鲁棒性。
Experiments
实验采用的主要数据集包括：由激光焊接实验采集的焊池图像，标注为穿透（类别0）和非穿透（类别1），共计1400张训练样本和600张验证样本。未标注数据集包含8000张未标注焊池图像。模型训练分两个阶段：预训练阶段使用未标注数据，结合对比学习和PINN正则化，训练时间约为48小时；微调阶段用少量标注样本（200张）进行少样本分类，训练时间约为12小时。模型性能通过准确率（96.06%）、F1分数、ROC-AUC等指标评估。还进行了消融实验，验证物理正则化和图像增强任务对性能的提升作用。对比了纯数据驱动模型和引入PINN的模型，结果显示后者在少样本条件下表现更优，鲁棒性更强。
Results
SimPhysNet在仅用200个标注样本时，达到了96.06%的分类准确率，几乎等同于利用全部标注数据训练的模型（准确率97%）。引入PINN正则化后，模型的泛化能力显著增强，尤其在复杂焊接参数变化时表现稳定。消融实验显示，图像增强任务提升了模型对焊池微小特征的敏感性，减少了过拟合风险。模型在不同材料厚度和焊接参数下均保持高性能，验证了其广泛适用性。这些结果充分证明了SimPhysNet在工业少样本环境中的优越性，为焊接质量自动检测提供了可靠工具。
Applications
该模型可直接应用于焊接过程的实时监控与质量控制，帮助操作员快速判断焊缝是否达到全穿透状态，减少返工和废品率。其低标注需求使得在不同工厂和设备上快速部署成为可能，适应多变的生产环境。未来还可结合工业机器人实现全自动化焊接质量检测，提升生产效率。长远来看，模型的物理基础和少样本能力为其他制造环节的缺陷检测、工艺优化提供了借鉴，有望推动智能制造的全面升级。
Limitations & Outlook
尽管SimPhysNet在少样本条件下表现优异，但在极端工况或未覆盖的材料类型下的泛化能力仍需验证。PINN引入的物理模型依赖于已知的物理规律，对于未知或非线性复杂物理过程的适应性有限。此外，模型训练对计算资源的需求较高，尤其是在大规模未标注数据和复杂PINN正则化的情况下，训练时间较长，影响工业部署的效率。未来需要优化模型结构和训练策略，以降低成本和提升效率。
Abstract
The laser welding full-penetration is of critical importance, as it constitutes one of the fundamental factors in achieving defect-free welded joints. Accurate prediction of the penetration state is therefore essential for ensuring weld quality. To this end, this paper introduces SimPhysNet, a novel algorithm that achieves high classification accuracy in laser welding penetration prediction using only a limited number of labelled images. This approach effectively overcomes the limitations of supervised learning classification algorithms, which are hindered in industrial applications by their dependence on extensive, high-quality labelled data. The core of SimPhysNet is a unique self-supervised learning paradigm that embeds physical priors into a contrastive learning framework. By incorporating a physics-informed neural network (PINN), the model is guided to extract physically meaningful features of the molten pool and keyhole from a large set of unlabelled data, while three image augmentation tasks further enhance its generalization capabilities. Subsequently, a few-shot learning strategy, based on prototypical networks, enables robust classification by constructing class representations from a minimal set of labelled images. Experimental results demonstrate that SimPhysNet achieves a classification accuracy of 96.06% using only 200 labelled images (approximately 5% of the total labelled dataset), which is comparable to the performance of conventional supervised learning algorithms that utilize the entire labelled dataset. This work presents a new, efficient, and highly accurate method, providing the way for the intelligent automation of laser welding.
References (20)
Intelligent welding system technologies: State-of-the-art review and perspectives
Baicun Wang, S. Hu, Lei Sun et al.
Nearest neighbor pattern classification
T. Cover, P. Hart
Random Forests
L. Breiman
Potentials of few-shot learning for quality monitoring in laser welding of hairpin windings
Tim Raffin, A. Mayr, Marcel Baader et al.
Support-Vector Networks
Corinna Cortes, V. Vapnik
Improvements to and limitations of Latin hypercube sampling
D. Huntington, C. Lyrintzis
Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization
Ramprasaath R. Selvaraju, Abhishek Das, Ramakrishna Vedantam et al.
Prototypical Networks for Few-shot Learning
Jake Snell, Kevin Swersky, R. Zemel
Learning to Compare: Relation Network for Few-Shot Learning
Flood Sung, Yongxin Yang, Li Zhang et al.
A hybrid back-propagation neural network and intelligent algorithm combined algorithm for optimizing microcellular foaming injection molding process parameters
W. Guo, Feng Deng, Z. Meng et al.
Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning
Jean-Bastien Grill, Florian Strub, Florent Altch'e et al.
Big Self-Supervised Models are Strong Semi-Supervised Learners
Ting Chen, Simon Kornblith, Kevin Swersky et al.
Real-time penetration depth prediction via physics-informed learning from molten pool surface morphology in laser filler wire welding
Rundong Lu, M. Lou, Yujun Xia et al.
Application of sensing techniques and artificial intelligence-based methods to laser welding real-time monitoring: A critical review of recent literature
Wang Cai, Jianzhuang Wang, P. Jiang et al.
Conditional Self-Supervised Learning for Few-Shot Classification
Yuexuan An, H. Xue, Xingyu Zhao et al.
In-situ monitoring system for weld geometry of laser welding based on multi-task convolutional neural network model
Huaping Li, Hang Ren, Zhenhui Liu et al.
A Deep Meta-Metric Learning Method for Few-Shot Weld Seam Visual Detection
Tianchen Zhu, Shiqiang Zhu, Jiakai Zhu et al.
On the numerical treatment of heat sources in laser beam welding processes
Philipp Hartwig, L. Scheunemann, J. Schröder
Evolutions of temperature field and stress field in narrow gap oscillating laser welding process based on equivalent heat source
Guodong Liang, Guoliang Qin, Peize Cao et al.
Heat source models for numerical simulation of laser welding processes – a short review
M. Behúlová, E. Babalová
Cited By (2)
FSTL-Net: A few-shot welding defect detection method based on temporal modeling and metric learning
A damage detection algorithm for aero-engine blade with texture enhancement