Heuristic techniques to optimize neural network architecture in manufacturing applications

TLDR: Four different approaches have been used to increase the generalization abilities of a neural network, based on the use of genetic algorithms, Taguchi, tabu search and decision trees, and the results show that a suitable technique for determining the architecture of a Neural network can generate a significant performance improvement compared to a trial-and-error approach.

Abstract: Nowadays application of neural networks in the manufacturing field is widely assessed even if this type of problem is typically characterized by an insufficient availability of data for a robust network training. Satisfactory results can be found in the literature, in both forming and machining operations, regarding the use of a neural network as a predictive tool. Nevertheless, the research of the optimal network configuration is still based on trial-and-error approaches, rather than on the application of specific techniques . As a consequence, the best method to determine the optimal neural network configuration is still a lack of knowledge in the literature overview. According to that, a comparative analysis is proposed in this work. More in detail four different approaches have been used to increase the generalization abilities of a neural network. These methods are based, respectively, on the use of genetic algorithms, Taguchi, tabu search and decision trees. The parameters taken into account in this work are the training algorithm, the number of hidden layers, the number of neurons and the activation function of each hidden layer. These techniques have been firstly tested on three different datasets, generated through numerical simulations in the Deform2D environment, in an attempt to map the input---output relationship for an extrusion, a rolling and a shearing process. Subsequently, the same approach has been validated on a fourth dataset derived from the literature review for a complex industrial process to widely generalize and asses the proposed methodology in the whole manufacturing field. Four tests were carried out for each dataset modifying the original data with a random noise with zero mean and standard deviation of one, two and five per cent. The results show that the use of a suitable technique for determining the architecture of a neural network can generate a significant performance improvement compared to a trial-and-error approach.