Fusiform Attention Network for Online Penetration Monitoring in Laser Welding

01 Introduction

As a core process in high-end manufacturing, the penetration state of laser welding directly determines weld quality and structural reliability. However, interference from strong plasma, spatter, and intense light at the welding site makes traditional visual monitoring methods difficult to implement accurately. Conventional CNN and ViT models use square receptive fields that do not match the geometric characteristics of the molten pool, leading to loss of detail and insufficient recognition accuracy. Furthermore, existing algorithms struggle to balance global morphology perception with local feature extraction, and their inference speeds often fail to meet industrial online demands. Therefore, constructing a dedicated deep learning model that fits the physical morphology of the molten pool for high-precision, low-latency online monitoring is of significant engineering necessity to ensure welding process stability and product qualification rates.

02 Overview

Addressing the need for online visual monitoring of laser welding penetration status, this paper proposes a novel Fusiform Attention Network (FANet). Using visual images of the molten pool as input, the network designs a Ternary Multi-head Linear Attention (TMLA) mechanism that fuses window attention with horizontal and vertical axial attention, combined with an overlapping patch strategy to match the elongated distribution of the molten pool. It also introduces Scalar Gated Attention (SGA) to achieve linear computational complexity. A laser welding visual acquisition platform was established, and a dataset containing four types of samples (full penetration, partial penetration, humping, and burn-through) was constructed. The accuracy, efficiency, and robustness of the algorithm were verified through comparative and ablation experiments, ultimately achieving real-time and reliable penetration status monitoring in industrial scenarios.

03 Figure Analysis

Figure 1 displays the overall composition of the laser welding visual monitoring platform, including a fiber laser, welding robot, welding camera, argon shielding, and calibration modules. The camera captures images of the molten pool area at a 30°–35° pitch angle, using a bandpass filter to reduce plasma interference. The images are input into FANet in real-time for penetration status classification. Research indicates that the system can stably capture visual features strongly correlated with penetration, such as the molten pool, keyhole, spatter, and weld morphology. Conclusion: This visual acquisition scheme effectively suppresses intense light noise and completely preserves the geometric and dynamic features of the molten pool, serving as the hardware foundation for high-precision penetration monitoring.

Figure 1. Schematic of the visual monitoring system for laser welding penetration status.

Figure 2 presents the four-stage pyramid structure of the Fusiform Attention Network. With the molten pool image as input, TMLA modules are continuously stacked after convolutional embedding. The core innovation is the TMLA block, which parallelly fuses window attention, horizontal axial attention, and vertical axial attention, balancing local details with long-range dependencies. Research shows that this architecture allows the attention field to precisely match the molten pool morphology. Through structural design driven by physical priors, FANet breaks through the geometric mismatch limitations of general models, achieving efficient extraction and expression of molten pool features.

Figure 2. Overall network architecture of FANet.

Figure 3 explains the TMLA attention allocation rules: 1/2 is used for window attention to capture local spatter and plasma details; 1/4 is used for horizontal axial attention to perceive the overall length and contour of the molten pool; and 1/4 is used for vertical axial attention to supplement information on the edges and weld tail. The results from the three paths are concatenated and output via linear projection. Experiments prove that horizontal axial attention is most sensitive to molten pool morphology, while the vertical branch improves directional robustness. TMLA achieves unified modeling of local details and global morphology, which is the core reason why FANet outperforms traditional attention architectures.

Figure 3. Diagram of the Ternary Multi-head Linear Attention (TMLA) mechanism.

Figure 4 visualizes the four types of samples: full penetration, partial penetration, humping, and burn-through. Window attention focuses on local features like spatter and plasma; horizontal strip attention accurately covers the main area of the molten pool, responding to morphological differences; and vertical strip attention assists in capturing the molten pool tail and the solidified weld. The final activation is concentrated on key areas of the molten pool without background noise interference. The heatmaps intuitively verify that FANet can precisely focus on effective information, with the TMLA branches having a clear division of labor and stable, interpretable perception capabilities for different penetration states.

Figure 4. Activation heatmaps for each attention branch.

04 Conclusion

The FANet proposed in this paper achieves precise alignment of the receptive field with the molten pool morphology through the TMLA mechanism. Coupled with overlapping patching and efficient SGA computation, it reaches a 93.24% accuracy and a 91.55% F1 score in penetration status recognition, significantly outperforming general models such as ResNet and Swin Transformer. With an end-to-end latency of only 5.41ms, it meets the requirements for 100FPS online monitoring. Ablation experiments confirm that ternary attention, overlapping patching, and SGA are key to the performance improvement. This research integrates physical priors into the network structure, breaking the geometric mismatch bottleneck of traditional vision algorithms and providing an efficient intelligent solution for closed-loop quality control in automotive, aerospace, and shipbuilding laser welding.

Original link: https://doi.org/10.1016/j.measurement.2026.121107