Objective As a new advanced manufacturing technology, laser cladding rapid prototyping has been widely used to laser forming without any need for a mold or die. However, the traditional laser cladding usually adopts the low-power forming below 2000 W, which is inherent with problems such as low deposition efficiency, long-forming time, and insufficient material use. In contrast, the high-power hollow ring laser cladding can effectively improve the forming efficiency by optimizing the experimental process parameters using model analysis. In recent years, neural networks have been gradually applied to optimize multi-parameter objective in laser cladding, laser welding, and laser communication. The neural network prediction model is capable of fitting and modeling nonlinear data through iterative learning without the need to specify the function form in advance, demonstrating excellent ability to deal with multivariate nonlinear problems. However, the single neural network model may suffer from problems like slow and easy to fall into local extremum training speed. In this study, not only a neural network model optimized by a genetic algorithm is proposed to predict and optimize the laser cladding deposition efficiency, but also the parallel random search is employed to effectively solve the aforementioned two problems of the training process in the model. We expect that our research can contribute to improving deposition efficiency and shortening forming time in high-power laser cladding forming.

Methods The experimental equipment adopts the hollow ring internal light powder feeding cladding system, which mainly consists of a 6 kW Raycus laser, a 6-axis KUKA robot, and a hollow ring internal light powder feeding nozzle. The cladding material is Fe314 powder and the experimental substrate is 304 stainless steel. The effects of the laser power, scanning velocity, powder feeding velocity, and defocusing on the deposition rate of the cladding layer were studied systematically by an orthogonal experiment. The tissue differences of the samples under high and low deposition rates were consistently observed by scanning electron microscope. The Box-Behnken(BBD) experiment was then designed by the response surface method to study the influence of several interaction factors on the deposition rate. Meanwhile, the multiple regression model was also established to predict and optimize the deposition rate. Additionally, a series of randomized trials were conducted as a supplement. Both the results of the BBD experiment and the samples of the randomized trial were taken as the training data of the genetic neural network, and eventually, the genetic neural network model was trained to predict and optimize the deposition rate. By comparing the modeling ability, generalization ability, and optimization ability of the two models, the most suitable model was selected as the estimator for the following experiments to accomplish closed-loop control of the deposition rate.

Results and Discussions The range analysis method was adopted to analyze the orthogonal experimental results. It is indicated that powder feeding rate has the greatest influence on the cladding deposition rate followed by laser power and scanning speed, and the defocusing amount on the cladding deposition rate was of the least influence (Table 3). When it comes to the reciprocal influence of various factors on the deposition rate, the response surface methodology (RSM) model shows that the interaction effect between laser power and powder feeding rate is the most significant. This is probably because the laser energy density improves as the laser power enhances, resulting in the enlargement of the high-temperature range in the molten pool. Within a certain power range, the deposition rate increases significantly by increasing the molten powder in the molten pool [Fig. 5(a)]. The second interaction effect is between scanning speed and powder feeding rate [Fig. 5(b)], and the interaction effect between defocusing amount and laser power is the least significant [Fig. 5(c)]. Simultaneously, the comparison of the predicted experimental values of genetic algorithms-back propagation (GA-BP) and RSM models after training shows that both models have good fitting accuracy, but the value of R² in GA-BP is closer to 1 (Fig. 8). By comparing the generalization ability of GA-BP and RSM models, the absolute average deviation(AAD) values of RSM and GA-BP models were 8.762% and 4.938%, respectively (Fig. 9). The maximum deposition rate obtained by the optimized GA-BP model was 61.74 g/min, which was higher than the value of 53.55 g/min obtained by the response surface method (Table 7). The above studies show that the prediction, generalization, and optimization capabilities of the genetic neural network model are superior to those of the response surface model, and the neural network model optimized by the genetic algorithm can provide a more effective prediction method for the achievement of laser cladding forming with high deposition rate.

Conclusions In this study, considering the orthogonal experiment in the hollow ring internal light powder feeding cladding system, the BBD experiment designed by response surface method was adopted and an RSM model using deposition rate as the response target was established. Subsequently, the influence of laser power, powder feeding speed, scanning rate, and defocusing amount on the deposition rate of high-power laser cladding was systematically investigated, individually or interactively. Based on the results of the BBD experiment and the samples of the randomized trial, a GA-BP model was set up. Comparing the performance of the two models by analyzing the modeling and generalization abilities, the technological parameters of laser cladding with high deposition rate were optimized. In conclusion, the performance of the GA-BP model is slightly better than that of the RSM model as the root mean square error and the average absolute deviation of the GA-BP model are smaller than those of the RSM model, and the decision coefficient values obtained in the RSM model and GA-BP model are 0.9479 and 0.9726, respectively. The prediction abilities of GA-BP and RSM models are fairly close, but under the condition of high deposition rate, the optimal deposition rate in the GA-BP model is 61.74 g/min, which is higher than the value of 53.55 g/min optimized by the RSM model. Briefly, the GA-BP model can be considered an effective method to optimize the high deposition rate of cladding. Because of the high reliability, the model can find the optimal deposition rate within a specific process range, which can be used as an estimator for operating closed-loop control over the following forming system to obtain high-efficiency and high deposition rate in the later cladding.