Industrial-Grade 3D Detection for Laser Cleaning of Metallic Components Based on Neural Radiance Fields and Temporal-Spatial Learning

With the in-depth development of the Industry 4.0 technology toward the intelligent manufacturing, laser cleaning technology has gradually become an important method in the field of metal surface treatment due to its advantages such as high efficiency, green and non-contact, and the related technological innovation has also been a research focus in recent years. However, the laser cleaning process is highly dependent on the precise control of the process parameters, and the real-time detection of dynamic changes on the metal surface is still challenging. This paper proposes an industrial-grade 3D detection algorithm based on Neural Radiance Field (NeRF) and spatiotemporal consistency learning. The algorithm is used for accurate monitoring of metal laser cleaning process. By designing a multi-modal sensing system and combining the NeRF optimization strategy of Octree scene partitioning, high-fidelity reconstruction of the rust edge and micro-cracks on the metal surface is achieved. At the same time, the spherical harmonic coding and cross-scale attention fusion mechanism are further introduced to improve the spatio-temporal consistency modeling capability between multi-view images. The main innovations of this study include: 1. Build a data acquisition system that integrates event cameras and RGB cameras to give full play to the advantages of event cameras in capturing rapid changes in light intensity; 2. Propose a NeRF optimization strategy based on Octree scene partitioning to achieve high-fidelity reconstruction of micro-structures such as rust layer edges and micro-cracks; 3. Improve the collaborative modeling ability between multi-temporal images through cross-scale attention fusion; 4. Introduce spherical harmonic functions to replace traditional position encoding functions, enhance the anisotropic reflectance modeling ability and significantly reduce the network parameter scale. Experimental results show that the proposed method achieves a peak signal-to-noise ratio of 33.56 dB and an edge accuracy of 93.2% in metal surface 3D reconstruction, and the training efficiency is 35% higher than that of traditional methods. In addition, the cleaning cleanliness assessment accuracy based on HSV and 3D fusion images is as high as 99.0%. This breakthrough innovation provides a new paradigm of non-contact, high-precision visual inspection for industrial scenarios such as aircraft skin paint stripping and rail transit rust removal.