| dc.description.abstract | Additive Manufacturing (AM) technologies allow researchers and companies to develop and build components in which the complexity is much higher compared to the components that traditional manufacturing processes, like milling and turning, can obtain. Nowadays AM technologies are used to deal with components which include a high rate of complexity when it comes to their analysis in the engineering field. One of these are complex notch components, for which a brief literature review is summarized in the second chapter of this study. In the mentioned chapter, several types of notches are considered, but all of them are common and constant notches. One of the most important parts in the analysis of those components is in the stress field analysis, which is often time-consuming to perform, even though finite element (FE) analyses are used. As a consequence, a comparison between many different notched geometries could not be feasible for researchers and companies. However, and as this paper proves, this problem can be solved with data mining and in particular with a specific data-driven approach, which works thanks to a machine learning algorithm.
Machine learning, in conjunction with large datasets, has seen many success stories in recent years, leading to increased interest in the field. One popular class of machine learning models is deep neural networks, where stacked layers of neurons are used to learn approximate representations of data. One particular model, the Convolutional Neural Network (CNN), is useful when input geometries have to be studied.
In this study, an overall dataset equal to 13890 different geometries are generated with a “step building blocks” technique (see chapter three). In order to compare all these possible geometries, an energy-based benchmark parameter is used: the averaged strain energy density (ASED). Then, a data-driven approach is proposed: a finite element method (FEM) and an ML algorithm are embedded together in order to prove that this approach is able to evaluate the fatigue behaviour of nonconstant notched geometries. After an introduction to the neural networks and CNNs, the optimized CNN model is investigated through a parametric study. The result is a six-layer deep convolutional neural network model with five convolution layers and one fully connected layer. In the fifth chapter, the results of the proposed data-driven approach are shown: a prediction accuracy higher than 99.0% is achieved when the CNN model is trained with a number of training data equal to 80% of the overall dataset. In order to understand whether the prediction of the CNN is correct, a fatigue simulation is then performed for the best, the worst and a common V-notch geometry. The result obtained for the best geometry proves that the CNN model can predict which is the best geometry. This best notch configuration is therefore insensitive or, at least, optimized against the fatigue failure, proving the concept of fatigueless structures.
As a result, the proposed data-driven approach can be successfully applied when thousands of different notches have to be compared, highlighting the notches that can provide a bad, good or an optimized (i.e. fatigueless) behaviour in terms of fatigue life.
In conclusion, and to deepen the research, the CNN model is trained with only 20% of the data randomly selected from the dataset. In this way, a prediction accuracy higher than 96.0% is reached. The accuracy is lower compared to the CNN that is trained over the entire dataset but it is a matter of cost-efficiency: the choice of testing only on a small percentage of data using the proposed data-driven approach can save a huge amount of computational time. | |