This paper mainly focuses on the laser cladding repair process of the shield machine main bearing. Aiming at the problem that the process parameters are difficult to characterize the cladding area mapping, a main bearing laser cladding process morphology prediction model based on BP neural network is proposed to predict the morphology size of the single-pass cladding area through process parameters to further improve the overall cladding area performance. First, a 3-factor 4-level orthogonal test was carried out with 42CrMo bearing steel as the matrix and steelite6 as the cladding powder material, and the melt height, melt width and melt depth were used as indicators; then, the BP neural network model was designed and trained based on the indicators obtained from the orthogonal test and the process parameters; finally, the predicted morphology size was compared with the actual size and the error between the two was calculated. By analyzing the test results, it was found that the prediction error of the model for melt height and melt width was within 2%, while the prediction effect for melt depth was poor. The analysis results show that the weight of the network model is trapped in a local optimal solution due to data limitations.

In recent years, the domestic rail transit mileage has increased rapidly, and the shield machine has become the main tool for tunnel construction with its advantages of high automation, high efficiency, compact structure, and high construction safety. Its domestic demand has also increased with the growth of rail transit mileage [1]. The main bearing is the main slewing bearing in the tunneling process of the shield machine. The working environment is complex and it bears most of the load in the tunneling process [2-3]. Therefore, the failure rate is relatively high. At present, most of the main bearings installed in the shield machines in service in China are imported from abroad. They are expensive. Once a failure occurs, the replacement cost is extremely high [4]. Therefore, the remanufacturing and repair process for the main bearing has gradually become the research focus of domestic scholars.

As a surface repair process, laser cladding technology has the advantages of concentrated energy, small heat-affected area, and high material utilization rate [5]. Therefore, it has become the best process for repairing main bearing raceway cracks, fatigue spalling and other damages. The laser cladding process has extremely strict requirements on process parameters. Unreasonable parameters will cause cracking, pores, low bonding strength and other problems in the cladding area. In the bearing raceway repair process, multi-layer and multi-pass cladding is mainly used. In multi-layer and multi-pass cladding, the melting height, melting depth and melting width of the single-pass cladding area will directly affect the overall strength of the final repair area and the surface quality of the cladding layer [6]. By exploring the relationship between process parameters and the morphology and size of the final cladding area, it can effectively guide the selection of the optimal process parameters.

At present, domestic and foreign scholars have a lot of research foundations on the relationship between process parameters and cladding area morphology and size. The existing research methods mainly include multivariate regression analysis, variance analysis, BP neural network, etc. [7-9]. Among them, BP neural network has a strong ability to deal with complex nonlinear relationships. The processing ability of the laser cladding machine is good, and it has good stability, and it can achieve a high degree of mapping between the morphology size and the process parameters [10]. This paper will use the BP neural network algorithm and combine the orthogonal test to explore the influence of laser power, powder feeding speed and scanning speed on the cladding morphology. The network model is trained based on the experimental data to achieve accurate prediction of the morphology size of the cladding area.

1 Experimental equipment and methods

The laser cladding equipment used in the study mainly includes 6000YLS fiber laser, KUKA robot, HUIRUI laser powder feeder, HUIRUI laser cladding head, water cooler, workbench, etc. During the experiment, argon was used as the protective gas and powder delivery carrier.

The substrate is 42CrMo bearing steel with a size of 100 mm × 80 mm × 16 mm. Before the test, the substrate surface was polished with sandpaper to remove surface oxides and impurities that may affect the cladding. Then, the surface was cleaned with anhydrous ethanol and dried with a hair dryer. The cladding material powder was stellite6 powder with a particle size distribution of 150 ~ 300 mesh. Before the test, the powder was placed in a drying oven for 2 hours to remove the moisture in the powder. The element distribution of stellite6 powder is shown in Table 1.

There are many process parameters that affect the morphology of the cladding layer, including but not limited to laser power, powder feeding speed, scanning speed, defocus, spot diameter, etc. Among these parameters, the laser power, scanning speed and powder feeding speed have a more significant impact on the morphology. Therefore, this study will focus on the three key parameters of laser power, scanning speed and powder feeding speed. As a research variable, in order to deeply explore the influence of morphology in the cladding process, the prediction of morphology size is realized through BP neural network training data.

Orthogonal test method is a powerful experimental design method. Orthogonal test ensures that each level of each factor is evenly combined with each level of other factors by using orthogonal arrays. This means that in the experiment, various levels of various factors are fully covered, which is helpful to obtain comprehensive information, with high efficiency and saving experimental resources. After a series of preliminary tests, the relatively reasonable process parameter range of the experiment is determined. The orthogonal test designed in this study has a total of 3 factors, 4 levels, and 16 groups of experiments. The specific parameter settings are shown in Table 2.

The spot emitted by the laser used in this experiment is a circular spot with an adjustable spot diameter, which is 4 mm by default. According to the experimental parameters, 16 groups of experiments are carried out. The macroscopic morphology of the cladding layer obtained in the single-layer single-pass cladding test is shown in Figure 1. It can be seen that the surface of the cladding layer obtained from the 16 groups of tests is relatively smooth, without obvious defects such as cracks and pores. After the cladding is completed, the cladding layer is cut along the vertical direction of the laser scanning using a wire cutting machine to obtain the cross-section of the cladding layer. The specimen is made into a sample of 8 mm × 8 mm × 10 mm, and then it is mounted for subsequent grinding. This study uses a fully automatic metallographic sample mounting machine, sets the heating temperature to 135 ℃, the heating time to 960 s, and the holding time to 1 280 s. The specimen and the mounting material are placed, and the pressure cover is covered to complete the mounting. Subsequently, the specimen is mounted using 240, 400, 600, 800, 1 000, 1 500, and 1 The sample was polished with 800 and 2000 mesh sandpaper, and then placed on an automatic polishing machine for polishing. Before etching, it was cleaned with anhydrous ethanol to remove the residual polishing liquid. Next, 10% volume fraction nitric acid alcohol was used for etching until the cross section of the cladding layer showed light gray. After etching, it was ultrasonically cleaned with anhydrous ethanol and dried to ensure that the sample surface was completely clean. Finally, the cross-sectional morphology of the cladding layer after grinding and polishing was observed using an ultra-depth of field three-dimensional microscope. The cross-sectional morphology of the cladding layer is shown in Figure 2.

The width, height and depth of the cladding layer were measured using image measurement software. The results of the measurements were statistically analyzed and the results are shown in Table 3.

2 BP neural network prediction model and prediction results

2.1 Establishment of BP neural network prediction model

There is a nonlinear relationship between the laser cladding process parameters and the morphology and size of the cladding layer. In order to effectively simulate and analyze this nonlinear relationship, BP neural network was selected as a research tool. BP neural network is widely regarded as a universal function approximator. Its unique ability to adapt to various nonlinear relationships makes it an ideal tool to reveal the complex and changeable relationship between laser cladding process parameters and the morphology and size of the cladding layer. BP neural network is trained through supervised learning and can learn from a large number of labeled samples and adjust weights to improve the accuracy of the model.

The morphology and size of the laser cladding layer are affected by multiple process parameters. Laser power (P), scanning speed (V) and powder feeding speed (N) are the process parameters that have a significant impact on the morphology and size of the cladding layer. In order to more accurately model and predict this complex relationship, this paper chooses the above three variables as the input variables of the BP neural network model, and the output variables are the corresponding melt width (W), melt height (H), and melt depth (D) under each process parameter combination. The internal structure of the BP neural network model is shown in Figure 3. The number of neurons in the input layer and the output layer jointly affects the number of neurons in the hidden layer. The formula for selecting the number of neurons in the hidden layer [11] is: See formula (1) in the figure
Where: K, n, m are the number of neurons in the input layer, hidden layer, and output layer respectively; the value of a is often found by debugging the neural network to find a suitable value.

2.2 Training and testing of BP neural network prediction model

In order to train and compare parameters of different dimensions together, it is necessary to normalize the data. Convert all data between [0, 1] [12], which helps to eliminate the problem of uneven weight distribution caused by different numerical ranges and improve the training effect and performance of the neural network. The processing formula is: See formula (2) in the figure

Where: X is the original data; Xmax and Xmin are the maximum and minimum values ​​in the original data respectively; X′ is the normalized data.

Of the 16 groups of data, 14 groups are set as training sets and trained using the BP neural network, and 2 groups of data are set as test sets to detect the accuracy of the fit. In order to reduce randomness, the randperm function is used to disrupt the order of the data set and randomly select the training set. The training parameters of the neural network have a great influence on the prediction effect. After debugging, the maximum number of iterations selected is 1,000, the minimum error is 0.000 001, and the learning rate is 0.01. Table 4 shows the data when the prediction results are good. It can be found that the relative error of the cladding layer width is within 1.023%, the relative error of the cladding layer height is within 1.818%, and the relative error of the cladding layer depth is within 19.17%. Table 4 is a comparison between the test values ​​of the width and height of the test set and the BP neural network prediction values. The comparison between the test values ​​of the cladding layer width, height, and depth of the test set and the BP prediction values ​​are shown in Figures 4 to 6 respectively.

It can be seen from Figure 4 that among the 14 groups of training data, except for the fourth group of samples, the test values ​​of the cladding layer width, height, and depth are all within 1.023%, 1.818%, and 1.818% respectively. The actual values ​​of the two test sets are very close to the predicted values, and the relative error is only 1.818%, indicating that the BP neural network training effect is good and the prediction accuracy is high.

As shown in Figure 5, among the 14 training data, except for the fourth group of sample data, which may have deviations due to the accidental nature of the experiment, the deviation of the melt height is smaller than the deviation of the melt width, and the fitting effect of the other groups of samples is good. The actual values ​​of the two test sets are very close to the predicted values, and the relative error is only 1.023%, indicating that the BP neural network training effect is good and the prediction accuracy is high.

As shown in Figure 6, among the 14 training data, except for the fifth and ninth groups of sample data, which may have deviations due to the accidental nature of the experiment, the fitting effect of the other groups of samples is good. The relative error between the real value and the predicted value of the two test set data is large, which is 19.17%, and the relative error of the second group is small, which is 10.52%, indicating that the BP neural network training effect is good and the prediction accuracy is high. In summary, the established BP neural network model has certain errors in the prediction of the morphology and size of the cladding layer. The relative errors of the prediction of the width and height of the cladding layer are small, both within 2%, but the relative error of the prediction of the depth of the cladding layer exceeds 10%. One of the reasons is that the value of the depth of the cladding layer is too small, which is easier to expand the relative error.

2.3 Analysis of prediction results

After analyzing the errors, it is believed that the weights and thresholds of the BP neural network are relatively random, and it is easy to converge locally and end the training prematurely. The predicted results obtained are greatly deviated from the actual results; the amount of sample data is relatively small, and only 14 sets of data are used for training, and there may be contingencies that make the prediction effect poor; the structure of the network, including the number of layers and the number of neurons in each layer, will also affect the performance of the model. If the network structure is not appropriate, it may not be able to capture the complex relationship in the data.

In order to improve the prediction effect of the neural network, genetic algorithm can be considered to optimize it in the future. Genetic algorithm is an optimization algorithm based on natural selection and genetic mechanism. It gradually optimizes the solution by simulating the evolutionary process. The basic idea of ​​combining genetic algorithm with BP neural network is to search for the optimal solution of neural network weight and structure through the evolutionary process to improve the prediction performance. Genetic algorithm can search the solution space globally, which helps to jump out of the local optimal solution and improve the generalization performance of the model. In addition, genetic algorithm can also be used to search for the structural hyperparameters of neural network, such as the number of hidden layer nodes and the number of layers, so that the neural network is more suitable for specific problems.

3 Conclusions

(1) This paper constructs a model of the morphology size of the laser cladding layer. The BP neural network prediction model predicts the width, height and depth of the cladding layer by laser power, scanning speed and powder feeding rate. The relative errors of the prediction results of the cladding layer height and width are within 2%, and the relative error of the prediction results of the cladding layer depth is within 20%.

(2) Subsequently, other optimization algorithms can be combined to optimize the initial weights and thresholds of the neural network to avoid falling into the local optimal solution and improve the prediction accuracy.