A Hybrid Neural Network for Prediction of Surface Residual Stress in MQL Face Turning

Xia Ji; Alexander H. Shih; Manik Rajora; Ya Min Shao; Steven Y. Liang

doi:10.4028/www.scientific.net/AMM.633-634.574

Paper Titles

Impact of Fillers on Mechanical Properties of Biomass Composites
p.554

Effects of Polysaccharides from Radix astragali on Oxidative Stress Induced by Exhaustive Swimming Exercise in Liver and Muscle of Mice
p.558

Electrosprayed PVP/Shellac Composite Medicated Microparticles for Providing Biphasic Drug Release Profile
p.562

3D Printing Technology and the Adaptability of Printing Material
p.569

A Hybrid Neural Network for Prediction of Surface Residual Stress in MQL Face Turning
p.574

A Hybrid Neural Network for Prediction of Surface Roughness in Machining
p.579

Bonding Performance of Wood Treatment by Oxygen and Nitrogen Cold Plasma
p.583

Calculation of Axial Rolling Torque in Radial-Axial Ring Rolling Process
p.589

Comparison for the Test Methods of Fatigue Life of Welded Structures
p.593

HomeApplied Mechanics and MaterialsApplied Mechanics and Materials Vols. 633-634A Hybrid Neural Network for Prediction of Surface...

A Hybrid Neural Network for Prediction of Surface Residual Stress in MQL Face Turning

Abstract:

Surface integrity, such as surface roughness and residual stress, is an aspect of surface quality on machined parts. Residual stress in the machined surface and subsurface is affected by materials, machining conditions, and tool geometry. These residual stresses could affect the service qualify and component life significantly. Residual stress can be determined by empirical or numerical experiments for selected configurations, even if both are expensive procedures. This paper presents a hybrid neural network that is trained using Simulated Annealing (SA) and Levenberg-Marquardt Algorithm (LM) in order to predict the values of residual stresses in cutting and radial direction after the MQL face turning process accurately. To verify the performance of the proposed approach, the predicted results are compared with the results obtained by training an ANN using SA and LM separately. The results have shown that the hybrid neural network outperforms SA and LM in predicting machining induced surface integrity that is critical to determine the fatigue life of the components.

You might also be interested in these eBooks

View Preview

Info:

Periodical:

Applied Mechanics and Materials (Volumes 633-634)

Pages:

574-578

DOI:

https://doi.org/10.4028/www.scientific.net/AMM.633-634.574

Citation:

Cite this paper

Online since:

September 2014

Authors:

Xia Ji*, Alexander H. Shih, Manik Rajora, Ya Min Shao, Steven Y. Liang

Keywords:

Artificial Neural Network (ANN), Face Turning, Hybrid Neural Networks, Levenberg-Marquardt (LM), Minimum Quantity Lubrication (MQL), Residual Stress, Simulated Annealing (SA)

Export:

RIS, BibTeX

Price:

Permissions CCC:

Request Permissions

Permissions PLS:

Request Permissions

Сopyright:

Citation:

* - Corresponding Author

References

[14] Meanwhile, gradient based algorithms have drawbacks such as they are too slow for practical problems and the faster algorithms lack generalization for noisy and small data [10]. SA is a random displacement based algorithm that can essentially be used to optimize functions of any kind with any number of variables. It resembles the annealing process in metallurgical practice where molten metal tries to attain the crystalline structure with the minimum possible free energy through cooling [10]. It has the ability to avoid getting trapped in local minimum as it can accept worse function values depending on a random number generated during its iterations. In order to take advantage of both SA and LM a hybrid SA-LM algorithm was created. First, SA was used to train the ANN weight and bias values for 1000 iterations. The results from SA were given to LM, which had an epoch limit of 1000, as a starting point in order to further improve upon the results. LM was also run for 1000 iterations and the MSE was used as a stopping criteria. The ANN had a single hidden layer with a 3-5-2 structure i. e. 3 input neurons, 5 hidden neurons and 2 output neurons. The activation function of the hidden layer was the tan-sigmoid function. The results obtained by using the hybrid SA-LM algorithm were also compared to the results obtained by using just SA and LM that had a 2000 iteration stopping criteria. Results and Discussion Once an ANN was trained, it was then presented with the testing set to validate the training. The MSE was as a stopping criterion for the three algorithms and also as the objective function for SA and LM. Out of the 51 experimental data available, 44 were used to train the ANN and 7 were used to test it. Five different scenarios were created in which the data in the training and testing sets were altered. For each of the five scenarios, the three algorithms were used five times to predict the residual stress in radial and cutting directions and their average was used as the final output for that particular scenario. The prediction capability of the trained ANN was assessed by calculating the relative error between the average of the predicted and the actual values of the residual stress and taking the average of those errors. Table 1 shows the input and output parameters of the 5 different simulations and also the values predicted by the different algorithms. Fig. 4 and Fig. 5 show the average relative % error for each scenario in cutting and radial directions. Table 1: Training data and prediction results Scenario Cutting Speed.