Search

The use of Automated Fiber Placement is being widespread in the aerospace industry. The need of manufacturing large and complex structural composite components, it makes the use of this technology much more efficient than the conventional hand lay-up manufacturing. However, these components are still being manually inspected and the effect of the defects found is calculated with a simulation software. The scope of this thesis is to create a Machine Learning model that is able to calculate the effect on the effective stiffness for different defect configuration. This Machine Learning model should be provided with the geometrical defect characteristics in the laminate and it has to be able to predict, with a high level of accuracy, the effective stiffness of the laminate. Training this model with a big amount of different configuration defects generates the need to create a parametrized FE model of a composite laminate on the coupon level. The results show that a Multi Layer Perceptron architecture with two hidden layers. The first one with 281 nodes and the second one with 76 nodes which is able to predict the effective stiffness of a defective laminate coupon with an accuracy of 0,1 GPa, L'ús del Automated Fiber Placement està estenent-se en la indústria aeroespacial. La necessitat de fabricar components estructurals compostos grans i complexes, fa que l'ús d'aquesta tecnologia sigui molt més eficient que la fabricació convencional amb col·locació manual. No obstant això, aquests components encara s'estan inspeccionant manualment i es calcula l'efecte dels defectes trobats amb software de simulació. L'abast d'aquesta tesi és crear un model de Machine Lerning que sigui capaç de calcular l'efecte en la rigidesa efectiva per diferents configuracions de defectes. Aquest model d'aprenentatge automàtic hauria de rebre les característiques geomètriques dels defectes en el laminat i de ser capaç de predir, amb un alt nivell de precisió, la rigidesa efectiva del laminat. Entrenar aquest model amb una gran quantitat de configuracions de defectes diferents genera la necessitat de crear un model FE parametritzat d'una laminació composta en el nivell de cupó. Els resultats mostren que una arquitectura de Multilayer Perceptron amb dues hidden layers. La primera amb 281 nodes i la segona amb 76 nodes, és capaç de predir la rigidesa efectiva d'un laminat defectuós amb una precisió de 0,1 GPa, El uso del Automated Fiber Placement se está expandiendo en la industria aeroespacial. La necesidad de fabricar grandes y complejos componentes estructurales de materiales compuestos, hace que el uso de esta tecnología sea mucho más eficiente que la fabricación manual convencional. Sin embargo, estos componentes siguen siendo inspeccionados manualmente y se calcula el efecto de los defectos encontrados con un software de simulación. El objetivo de esta tesis es crear un modelo de Machine Learning que sea capaz de calcular el efecto sobre la rigidez efectiva para diferentes configuraciones de defectos. A este modelo de aprendizaje automático se le deben proporcionar las características geométricas del defecto en el laminado y tiene que ser capaz de predecir, con un alto nivel de precisión, la rigidez efectiva del laminado. El entrenamiento de este modelo se debe de realizar con una gran cantidad de configuraciones de defectos diferentes. Este hecho genera la necesidad de crear un modelo de elementos finitos parametrizado de un laminado a nivel de cupón. Los resultados muestran que una arquitectura Multilayer Perceptron con dos hidden layers. La primera con 281 nodos y la segunda con 76 nodos que es capaz de predecir la rigidez efectiva de un coupon laminado defectuoso con una precisión de 0,1 GPa

This thesis addresses the challenge of sub synchronous oscillations (SSO) in the electrical grid. Although there are clear protocols in place for SSO screening, the increasing complexity and wider range of operating conditions in the energy system is increasing the risk of new modes of SSO occurring. For this reason, sophisticated monitoring solutions are needed to enhance system security. In this thesis, a real-time detection tool for early SSO risk identification is developed. In addition, the thesis explores how artificial intelligence can be used to support the development of the proposed detection tool. The applied methodology to build the detection tool consists of three steps. The first step is to develop a synthetic data set of SSO oscillations in in Matlab Simulink ® and then extract the current waveform which is the power system measurement where SSO seems to manifest itself the most prominently. Then, in the second step, two signal processing algorithms, Fast Fourier Transform (FFT) and Prony’s method, are applied to the synthetically generated SSO waveforms to extract the signals’ features. The methods are implemented in Matlab ®, tested, validated and compared for SSO detection performance. Due to their different merits and draw- backs, both methods are implemented in parallel to provide complementary analysis. In the third and final step of the tool development in this thesis, the data set containing the system operating conditions and the signal features extracted with FFT and Prony’s method is analysed making use of machine learning. Specifically, K-means clustering is applied on the data set in Python. Although the results of this exploratory approach are not conclusive, the data set seems to contain around four distinguishable clusters which can be linked to SSO risk levels. It is, however, recommended to focus on other predictive machine learning applications to further improve the tool’s predictive SSO detection capabilities. The validated perfor, Objectius de Desenvolupament Sostenible::7 - Energia Assequible i No Contaminant::7.b - Per a 2030, ampliar la infraestructura i millorar la tecnologia per tal d’oferir serveis d’energia moderns i sos­tenibles per a tots els països en desenvolupament, en particular els països menys avançats, els petits estats insulars en desenvolupament i els països en desenvolupament sense litoral, d’acord amb els programes de suport respectius

The Cloud as computing paradigm has become nowadays crucial for most Internet business models. Managing and optimizing its performance on a moment-by-moment basis is not easy given as the amount and diversity of elements involved (hardware, applications, workloads, customer needs...). Here we show how a combination of scheduling algorithms and data mining techniques helps improving the performance and profitability of a data-center running virtualized web-services. We model the data-center's main resources (CPU, memory, IO), quality of service (viewed as response time), and workloads (incoming streams of requests) from past executions. We show how these models to help scheduling algorithms make better decisions about job and resource allocation, aiming for a balance between throughput, quality of service, and power consumption., Peer Reviewed, Postprint (published version)