Abstract:

As critical arteries for energy transmission, the structural integrity of natural gas pipeline welds is paramount to national energy security and public safety. Conventional non-destructive testing (NDT) methods, such as manual film interpretation or single-modality instrument inspection, often exhibit inherent limitations in feature extraction completeness and cross-modal information fusion reliability when dealing with the complex geometry and harsh service conditions of pipeline girth welds. These limitations hinder their effectiveness in addressing feature mismatches and false alignment-a phenomenon in multi-modal data fusion where non-correlated features are incorrectly matched in space due to registration errors-primarily caused by data source heterogeneity, high-intensity environmental noise, and complex microstructural backgrounds. To overcome these challenges, this research developed and validated an intelligent diagnostic system for natural gas pipeline weld quality based on deep learning. This paper proposes an innovative closed-loop processing architecture: “Perception-Alignment-Detection-Reconstruction-Enhancement-Fusion.” The architecture commences with an image quality perception module, which generates pixel-level noise interference distribution maps via local signal-to-noise ratio assessment to quantify input data reliability. Subsequently, a robust feature alignment module utilizes these maps to guide a confidence factor-modulated deformable convolution mechanism-an operation that adaptively learns sampling locations to enhance feature extraction capability-achieving precise spatial registration of multi-modal data within credible regions. The system further integrates mapping deviation detection and alignment path reconstruction modules, forming an internal error correction loop to dynamically rectify residual misalignments. Furthermore, a prior-driven enhancement module is introduced, embedding physical priors-such as heat-affected zone (HAZ) boundaries, characteristic defect morphologies, and material microstructures-into the deep learning framework to bolster feature representation in critical regions. Finally, a defect identification and fusion module consolidates all optimized multi-modal features through an adaptive weighting strategy, outputting definitive defect classification and localization results. Experimental results on an ASME standard-compliant test set demonstrate that our system achieved a defect identification accuracy of 97.6%. Compared to the industry-leading OmniScan X3 system, our system significantly reduced the incidence of misleading false alignment errors from 15.8% to 4.3%, representing a substantial reduction of 72.8%. Under stringent low signal-to-noise ratio conditions (SNR < 5 dB), the system maintained a high recall rate of 92.4%, attesting to its exceptional noise robustness. In field validation within actual pipeline construction projects, the system performed consistently, sustaining an identification accuracy above 92% while drastically reducing the time required per inspection compared to traditional manual methods, underscoring its significant engineering application value. The principal conclusion of this study is that the proposed intelligent diagnostic system, through its closed-loop, perception-driven, and prior-infused architecture, effectively surmounts the core bottlenecks of traditional methods in multi-modal weld quality assessment. It thereby provides a transferable theoretical model and engineering paradigm for the reliable deployment of industrial artificial intelligence in complex industrial settings.

Key words: oil and gas pipeline, weld quality diagnosis, multimodal fusion, deformable convolution, false alignment suppression

摘要：

天然气管道作为能源输送的关键动脉，其焊接接头的质量直接关系到国家能源安全与公共安全。传统的无损检测(Non-Destructive Testing，NDT)方法，如人工评片或单模态仪器检测，在应对管道焊缝的复杂结构与恶劣工况时，往往存在特征提取不全面与跨模态信息融合不可靠的固有局限，难以有效解决因数据源异构性、高强度环境噪声及复杂微观组织背景所导致的特征误匹配与伪对齐(False Alignment)问题。为此，本研究开发并验证了一种基于深度学习的天然气管道焊缝质量智能诊断系统。本文提出了一种创新的”感知—对齐—检测—重构—增强—融合”闭环处理架构。该架构始于一个图像质量感知模块，通过局部信噪比评估生成像素级噪声干扰分布图以量化输入数据的可靠性。随后，一个鲁棒特征对齐模块利用该分布图导引一种基于置信因子调节的可变形卷积(Deformable Convolution)机制，实现多模态数据在可信区域内的精准空间配准(Registration)。系统进一步集成了映射偏差检测与对齐路径重构模块，形成内部纠偏循环以动态纠正残余误差。此外，引入了一个先验驱动增强模块，将热影响区(Heat-Affected Zone, HAZ)边界、典型缺陷形态及材料组织等物理先验知识嵌入深度学习框架，以增强对关键区域的特征表达能力。最终，一个缺陷识别融合模块通过自适应加权策略，整合所有优化后的多模态特征，输出缺陷分类与定位结果。实验结果表明，在基于ASME标准的测试集上，本系统实现了97.6%的缺陷识别准确率。与业界先进的OmniScan X3系统相比，本系统将具有误导性的伪对齐误差发生率从15.8%显著降低至4.3%，降幅达72.8%。在极低信噪比(SNR<5 dB)的严苛条件下，系统依然保持了92.4%的高召回率(Recall)，证明了其卓越的噪声鲁棒性(Robustness)。在真实的管道工程项目现场验证中，系统表现稳定，识别准确率持续保持在92%以上，同时单次检测耗时相较于传统人工方法大幅缩短，展现出极高的工程应用价值。本研究的主要结论是，所提出的智能诊断系统通过其闭环、感知驱动且融入先验知识的架构，有效克服了传统方法在多模态焊缝质量诊断中的核心瓶颈，为工业人工智能在复杂工业场景中的可靠落地提供了可借鉴的理论模型与工程范式。

关键词: 油气管道, 焊缝质量诊断, 多模态融合, 可变形卷积, 伪对齐抑制