Deep learning-based prediction of process parameters for laser cleaning of sluice gate coatings

To protect the sluice gate substrate during laser cleaning of aged coatings, and to evaluate the cleaning effectiveness and surface quality after cleaning, this paper proposes a deep learning-based method for predicting process parameters for laser cleaning of sluice gate coatings. The proposed method introduces a learnable weighted decision layer (LWDL) based on fuzzy theory, which converts classification results obtained from images into multidimensional laser-cleaning process parameters. First, a sample dataset consisting of 194 images with a resolution of 2992×2992 pixels was established for the composite coating surface of the sluice gate. The samples were divided into 10 categories according to the laser-cleaning effectiveness, and data augmentation techniques were applied to expand the dataset and improve the model's generalization capability. Second, based on several existing high-performance classification network models, four network architectures suitable for this task were selected to construct an ensemble network framework. Subsequently, LWDL was introduced after the original classification networks, and the network models were trained using laser-cleaning process parameters obtained through theoretical calculations. Finally, the proposed model was deployed on an edge computing platform, and the end-to-end response time was tested to verify the real-time feasibility of the proposed scheme. In addition, the proposed method was compared with laser-cleaning approaches reported by different researchers in terms of cleaning effectiveness. Confocal microscopy was employed to systematically characterize the microscopic surface morphology of the cleaned sluice gate coating, thereby evaluating both the cleaning effectiveness and the degree of damage to the substrate material. The experimental results show that only a small amount of coating remains on the substrate surface, and the proposed continuous laser parameter adjustment scheme achieves a relatively better cleaning effect.