VMamba-BiFPN Model for Laser Welding Penetration Estimation from Multi-Camera Molten Pool Images

Source abstract

Accurate estimation of laser welding penetration from molten-pool images is still a problem because of keyhole oscillation, metal-vapor plume and specular reflection obscuring view-dependent cues. This paper introduces a VMamba-BiFPN model to estimate the penetration depth from synchronous coaxial, lateral and rear-oblique molten-pool images. First, align the multi-camera image streams by exposure time and process distance; then, extract view-specific features using visual state-space blocks that selectively scan to preserve elongated thermal texture and pool-boundary information. A bidirectional feature pyramid fuses shallow boundary responses with deeper process-state representations further, and then a calibrated regression head predicts penetration depth. Experiments were conducted on 4,860 bead-on-plate laser welding samples obtained under five power-speed conditions, and metallographic cross-sections were used as reference labels. The proposed model achieved an MAE of 0.041 mm, an RMSE of 0.058 mm, an R² of 0.973, and 94.2% of predictions within +/-0.10 mm, and outperformed the ResNet-FPN and Swin-BiFPN baselines by 34.9% and 18.0% in MAE, respectively. Ablation results show that the multi-camera input, VMamba encoder and BiFPN fusion each reduce the error measurably, and the final configuration has an inference latency of 21.6 ms. According to the above results, state-space visual fusion can provide accurate and deployable penetration estimation for closed-loop laser welding monitoring.