Deep Learning Classification and Optimization of Manufacturing Process Parameters

At Mercedes-Benz AG, “Die Casting” is an integral manufacturing process in which car parts are manufactured by forcing molten metal under high pressure and speed into a die cavity. The casting machines produce with a high number of process parameters, which are currently being tweaked by the engineers solely based on their experience to ensure that the process is as robust as possible. Hence, few defective parts are produced initially until the number of bad quality parts increases again and new adjustment is needed. Improvement of this procedure will lead to higher efficiency in this casting process. Owing to this, the thesis aims to build up a relationship between the process parameters and the quality measurements through data-driven modeling, and then later on use the learned mapping for finding an optimal set of parameters which maximizes the probability for producing a correct part.
From the last few decades, Artificial Neural Networks have been quite successful in capturing the complex, often non-linear, relationship which exists between the process parameters and the process output [1]. Therefore, the first goal in this thesis is fitting the given data with different neural network architectures such as Multi-Layer Perceptron [2], Residual Network [3], fine-tuning them and then selecting the best performing architecture out of them in terms of correctly mapping the process parameters to the ground-truth class labels.
Two major challenges in the first objective of the thesis are:
1. Pre-Processing of raw data which consists of a combination of categorical and numerical variables
2. Class Imbalance which poses a big challenge for unbiased training of neural networks [4]
Furthermore, the final step in the thesis would be to use our trained neural network for optimizing the input parameters in such a way that the probability of producing a good part is maximized by those parameters. One possible proposal to achieve this is by generating adversarial attacks on the neural network [5] where the input parameters are iteratively modified in the negative direction of the gradient of the loss function between the predicted class and the desired output class while the weights of the network are kept constant. Hence, instead of doing backpropagation onto the weights, backpropagation is performed onto the input parameters. Another popular approach in the literature that has shown satisfactory results in the manufacturing industry is to use the learned neural network mapping as the fitness function for evaluating candidate solutions of the optimal parameter set in a genetic algorithm [6].
In summary, the thesis will include the following points:
1. Training and Evaluation of different neural network architectures over the given data
2. Optimization of Process Parameters through the best-performing trained model
Supervisors: Prof. Dr. Andreas Maier, Olga Moreva, Lyubka Rund, Stephan Schwarz
Student: Saad Munir
Date: June 3, 2021 -November 30, 2021
References [1] Sukthomya, W., Tannock, J. The training of neural networks to model manufacturing processes. J Intell Manuf 16, 39–51 (2005). [2] Marius, Popescu & Balas, Valentina & Perescu-Popescu, Liliana & Mastorakis, Nikos. (2009). Multilayer perceptron and neural networks. WSEAS Transactions on Circuits and Systems. 8. [3] K. He, X. Zhang, S. Ren and J. Sun, “Deep Residual Learning for Image Recognition,” 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 2016, pp. 770-778. [4] Ling C.X., Sheng V.S. (2011) Class Imbalance Problem. In: Sammut C., Webb G.I. (eds) Encyclopedia of Machine Learning. Springer, Boston, MA. [5] Qiu, Shilin; Liu, Qihe; Zhou, Shijie; Wu, Chunjiang. 2019. “Review of Artificial Intelligence Adversarial Attack and Defense Technologies” Appl. Sci. 9, no. 5: 909.
[6] Julius Pfrommer, Clemens Zimmerling, Jinzhao Liu, Luise Kärger, Frank Henning, Jürgen Beyerer, Optimisation of manufacturing process parameters using deep neural networks as surrogate models, Procedia CIRP, Volume 72, 2018, Pages 426-431, ISSN 2212-8271.