Process Parameter Optimization and Crack Density Prediction in Laser Cladding Using a BP Neural Network Model

Abstract

To address the crack defects arising from the complex nonlinear mapping relationship between key laser cladding process parameters and coating preparation, this study conducted a theoretical analysis and experimental verification by employing a BP neural network to predict crack density based on orthogonal experimental data. Initially, a three-factor four-level orthogonal experiment was performed to obtain fundamental sample data for the neural network, followed by dataset expansion using kernel density estimation (KDE). The dataset was then preprocessed through min-max normalization, ultimately establishing a three-layer predictive neural network model that correlates laser cladding layer process parameters (powder feeding rate, overlap rate, and scanning speed) with crack susceptibility. The results demonstrate that the BP neural network model achieves crack density predictions with relative errors fluctuating within ±5%, while maintaining an average error of 0.85% and a mean square error of 1.11%, indicating high prediction accuracy and stable performance. Furthermore, a comparative analysis of various regression methods, including KNN, Ridge, and Random Forest, was conducted in terms of R², RMSE, and MAE metrics, revealing that the BPNN exhibits superior comprehensive performance. These findings validate the feasibility of applying BP neural networks for crack density prediction through process parameters in laser cladding applications, which holds significant importance for fabricating crack-free nickel-based cladding layers.