TIG Weld Penetration Prediction Using Capsule Network Architecture

Literature Overview

The paper by Wang Ying, Gao Sheng, and Wu Liming, published in "Welding" (2023, No. 4, pp. 15-20), presents a capsule network-based model for predicting weld penetration states in TIG welding operations. Conducted jointly by Northeast Petroleum University and Daqing Oilfield Co., Ltd., this research introduces a squeeze-excitation (SE) module into the capsule network architecture to enhance feature extraction and classification accuracy. The model takes real-time molten pool frontal images as input and classifies the weld penetration state into three categories: under-penetration, adequate penetration, and over-penetration. This work represents a significant advancement in real-time weld quality monitoring for pipeline construction and related industries.

Technical Background and Problem Statement

Weld penetration is one of the most critical quality parameters in TIG welding, particularly for pipeline construction where incomplete penetration can lead to catastrophic failures under pressure. Traditional methods of verifying penetration include destructive testing (cross-sectional examination), radiographic testing (RT), or ultrasonic testing (UT), all of which are performed after welding is complete and therefore cannot prevent defective welds from being produced. Real-time monitoring of penetration during welding would enable immediate corrective action, reducing rework costs and improving production efficiency.

The challenge of real-time penetration prediction lies in the indirect nature of the measurement. The weld penetration depth is not directly observable during welding; instead, it must be inferred from observable indicators such as the appearance of the molten pool, arc characteristics, and welding parameters. Visual observation of the molten pool frontal surface provides valuable information about the weld geometry, but the interpretation of these visual cues requires expertise that is difficult to automate.

Capsule Network Architecture and SE Module Integration

The capsule network (CapsNet) is a data analysis architecture that processes visual information through hierarchical capsules, where each capsule represents the presence of a specific visual entity with its associated pose parameters. Unlike traditional convolutional neural networks that use scalar activations, capsules output vectors that encode both the presence and the properties of detected features. The dynamic routing algorithm iteratively updates the coupling coefficients between capsules, allowing the network to learn which lower-level capsules contribute to the formation of higher-level capsules.

The squeeze-excitation (SE) module, originally proposed for convolutional neural networks, introduces channel-wise attention by learning to recalibrate channel features. The SE module operates in two stages: a squeeze stage that produces a channel descriptor through global average pooling, and an excitation stage that generates channel-wise weights through a fully connected network with ReLU and sigmoid activations. These weights are then applied to the original feature maps, emphasizing informative channels and suppressing less relevant ones.

The integration of the SE module into the capsule network architecture creates a hybrid model that combines the spatial feature extraction capability of convolutional layers, the attention mechanism of the SE module, and the hierarchical representation learning of the capsule layers. This architecture is particularly well-suited to the weld penetration prediction task, where different regions of the molten pool image may contain complementary information about the penetration state.