1. Introduction
Titanium-aluminiumis (TiAl) alloy a key engineering material for the future development of the aerospace and automotive industries. Its properties include low density (3.9–4.2 g/cm
3), high elastic modulus (170 GPa at room temperature, 150 GPa at 750 °C, equivalent to GH4169), high specific strength, high specific stiffness, low coefficient of expansion, high thermal conductivity, and resistance to oxidation, creep, and fatigue [
1,
2]. TiAl alloy is frequently prepared using near-isothermal rolling. It is impossible to avoid difficulties with the wrinkling, cracking, and mechanical characteristics of the alloy if the rolling parameters are not well controlled and near-isothermal rolling conditions are not satisfied [
3]. If we only change the material composition of the TiAl alloy without intelligently upgrading the near-isothermal rolling equipment, we cannot fundamentally solve the above problems. The finite element method has good predictive ability, but each working condition requires a lot of time and is not suitable for predicting constantly changing working conditions in industrial sites. However, neural networks can make predictions in less than 1 s, so they can play an important role in optimizing rolling plans and online monitoring.
Intelligent factories have steadily gained popularity [
4,
5] in recent years to overcome this issue, and realizing digital twin through simulating virtual factories has aroused great interest from all sectors of societyhas drawn significant interest from all spheres of society [
6,
7,
8]. By using information gathered from field equipment, digital twinning can perceive the existing situation and forecast the direction of future growth throughout the rolling process. State perception is the foundation of information physics systems. The exit thickness prediction model developed in this research is a crucial component of the rolling process state perception.
Neural networks have been extensively employed in the past several decades to anticipate rolling parameters. Rolling force plays an important role in the rolling process, and the accurate prediction of rolling force is extremely important for the forming performance of sheet metal. In order to increase the accuracy of the rolling force forecast using a neural network model, Zhou et al. [
9] continued the coordination adjustment to the deformation resistance and the friction force. Guo et al.’s [
10] modification of the material’s deformation resistance and friction coefficient by the parameter self-adaptation approach increased the rolling force prediction accuracy. Neural networks were utilized by Li Junhong et al. [
11] to forecast the mechanical characteristics of cold-rolled ribbed bars. To forecast rolling force and rolling torque under various rolling circumstances, Bisadi et al. [
12] created a neural network model. Using a Bayesian neural network, Wu et al. [
13] developed a mechanical model of carbon–manganese (C-Mn) steel. Fuquan Zhang et al. successfully predicted urban network traffic problems using the short-term learning ability of neural networks, which has an essential impact on the construction of smart cities [
14]. Hongbo Lin et al. used a prediction model based on a convolutional neural network to successfully learn and predict the trend of stock prices [
15]. In addition, many scholars at home and abroad are also studying rolling process prediction. Joon Sik Son et al. developed an online learning neural network for long-term and short-term learning, and successfully predicted rolling force data [
16]. Wang, Z.H et al. established an ELM regression model for particle swarm optimization (PSO) using a dataset obtained from finite element analysis [
17]. Jingyi Liu et al. proposed a rolling force prediction method based on a genetic algorithm, particle swarm optimization algorithm, and multi-hidden-layer extreme learning machine, and the prediction accuracy has been significantly improved [
18]. Xie et al. improved the hybrid mathematical rolling force model and adaptive neural network to predict the rolling force of strip steel by adjusting the adaptive learning algorithm [
19]. Jia et al. established the rolling force model of a tandem cold rolling mill by using the measured data and Elman dynamic recursive network method [
20]. Hwang R et al. used a depth neural network (DNN) and decision tree model based on gradient lifting to accurately predict rolling force [
21]. Tsang ECC et al. obtained a more reliable FNN by adjusting the knowledge representation parameter (KRP) [
22]. Guo et al. modified the deformation resistance and friction coefficient of materials by using adaptive methods, improving the prediction accuracy of rolling force [
10].
A large amount of research has also been conducted on the optimization of BP neural networks. Chen GY et al. [
23] established a machine learning model for accurately predicting the rebound of bent pipes, which considers multiple factors such as material properties and geometric parameters. Han Lili et al. [
24] increased the prediction accuracy of the width spread of the plate mill and sped up convergence by modifying the number of hidden layer neurons in the modified BP neural network. Wang et al. [
25] developed a combined laminar flow control system based on a feed-forward neural network and mathematical model feedback and accurately forecasted the crimp temperature using the added momentum BP neural network model. In order to increase the model’s capacity for generalization, Li et al. [
26] used BP neural networks with integrated learning to forecast the distribution of mechanical performance variations between samples.
The use of genetic algorithm (GA) for optimizing neural networks has also made good progress. A Neural Network Model Optimized by GA is a new prediction model for predicting the rolling parameters. Domestic and foreign scholars have carried out many studies to evaluate its performance in the prediction of different parameters, such as determining the metallurgical transformation of steel [
27,
28], predicting the rolling force and rolling moment of cold rolling [
29,
30], flatness prediction for hot strip mills [
31], and flow stress prediction under hot deformation conditions [
32,
33,
34]. Wouter M. Geerdes et al. combined the knowledge embedded in the heat transfer model and the temperature prediction ability of the artificial neural network: three different temperature prediction capabilities were realized, but they could only be used in an offline way and could not be monitored online [
35]. Hwang, R. used machine learning methods to predict rolling force and temperature in hot rolling [
21]. Azadi et al. used the finite element method to solve the governing equations of heat conduction and plastic deformation, and used a neural network model to predict the flow stress of the rolling stock [
36]. Hosein Alaei et al. developed a neural network to predict the thermal expansion of work rolls during rolling, and implemented online monitoring using a validated 3D analytical model to guide and supervise the learning process [
37]. Mian Jiang et al. proposed an online model for accurately predicting the thermal crown of the roll during hot rolling, and described the heat conduction process of the work roll temperature with a nonlinear partial differential equation (PDE) on cylindrical coordinates [
38]. Li Cuiling et al. established a deep belief network, conducted unsupervised training on the restricted Boltzmann machine, and fine-tuned the entire network, which made the prediction accuracy better than the traditional temperature calculation formula, and the error fluctuation range was less than 8 °C [
39]. YZ Zheng et al. [
40] optimized the BP neural network through GA and ant colony algorithm, and deeply preprocessed the data, thereby successfully predicting the ship’s operating trajectory, RB Li [
41] accurately predicted surface roughness during the turning process of nickel-based high-temperature alloys using an adaptive GA. CM Zhu [
42] successfully predicted the traffic flow on the highway using a neural network model optimized by a GA, solving the problem of traffic congestion. MJ Cao [
43] used a GA to predict the fatigue life of 304 stainless steel, and the predicted correlation reached over 98%. K. Tajziehchi et al. [
44] used a genetic algorithm to optimize earthquake control through research on the displacement heat transfer of nanofluids with baffles.
There is no research on the prediction of exit thickness of near-isothermal rolled TiAl alloy, despite the fact that an artificial neural network (ANN) is frequently employed in the prediction of the rolling process. Compared with finite element methods, neural networks require a large amount of data for learning. Therefore, in order to train high-quality neural networks, finite element models are used to expand the training dataset, ensuring the neural network has good generalization ability [
12]. The availability and accuracy of the neural networks are confirmed by contrasting the field test results with the trained neural network prediction data. After comparing the prediction outcomes of a single-layer neural network (SNN), deep neural network (DNN), deep belief network (DBN), genetic algorithm neural network (GANN), and T-S fuzzy neural network (FNN), it is confirmed that the GANN has the advantages of low calculation cost, good prediction accuracy, and robustness, and successfully solves the control precision problem of the exit thickness of near-isothermal rolled TiAl alloy. It offers a potent tool for further enhancing the mechanical qualities of TiAl alloys [
3]. It also provides strong support for solving the bottleneck problem in the manufacturing of large-sized TiAl alloys.
5. Analysis and Discussion of Prediction Model
The root mean square error (RMSE) results predicted by the five algorithms are shown in
Table 4. The execution time is shown in
Table 5. The prediction process of outlet thickness is shown in
Figure 9. If the training error reaches the set value or the training frequency reaches the upper limit, the training will be stopped, as shown in
Figure 10.
Figure 11 displays a comparison of the five models’ errors. From
Figure 10 and
Figure 11, it can be seen that, although the training performance of the DBN is slightly better than that of the GANN, the GANN performs better in terms of prediction accuracy.
Table 6 provides an overview of the greatest forecast error and overall forecast error. It is clear that the FNN model has the biggest prediction error. Although the errors of the DBN and GANN models are comparable, the GANN model is better able to avoid the local minimum issue with the DBN model. The GANN model’s maximum forecast error is merely 0.18 mm, and its average prediction error is 0.05 mm.
Figure 12 depicts the fitness curve for the GANN model, which has a fitness curve termination algebra of 10 generations.
Figure 13 depicts the mean square error diagram for the GANN model during training, with the ideal performance value being 0.0026781 and approaching it quickly.
Figure 14 displays the distribution of the predicted values along the target line, with a correlation coefficient of 0.99658. In
Figure 15, the prediction curve is displayed. The GANN model uses a 4-6-1 network configuration. The transfer functions of the hidden layer and the output layer are tansig and purelin, respectively. The population size is 10 generations, the cross probability is 0.2, and the mutation probability is 0.1 in the evolutionary algebra.
The near-isothermal rolling exit thickness of the TiAl alloy is accurately predicted using the GANN model, and the rolling parameters are tuned to prevent cracking. The rolling effect is depicted in
Figure 16. In conclusion, the TiAl alloy exit thickness during rolling may be predicted using the GANN model.