No description available.

No description available.

No description available.

No description available.

[pt] Os polissacarídeos constituintes da biomassa lignocelulósica podem ser beneficiados através de processos industriais. Entretanto, para manipulá-los é necessário que a biomassa seja submetida ao processo de pré-tratamento. Esta é uma das etapas mais caras e relevantes para a disposição e aplicação das frações lignocelulósicas. O presente estudo consiste em uma investigação detalhada do processo de pré-tratamento da biomassa lignocelulósica com H2O2, a qual foi realizada através de tecnologias inteligentes que viabilizaram a otimização deste processo. Ferramentas de inteligência artificial revelam-se vantajosas na solução dos gargalos associados aos avanços tecnológicos. Possibilitam a modelagem matemática de um processo com máxima eficiência, otimizando sua produtividade, transformando dados experimentais em informações úteis e demonstrando as infinitas possibilidades das relações das variáveis envolvidas. As variáveis independentes estudadas foram a temperatura (25 – 45 graus Celsius) e a concentração de peróxido de hidrogênio (1.5 – 7.5 porcento m/v). Técnicas analíticas qualitativas (Raman e FTIR) e quantitativa (Método de Klason) foram aplicadas para produzir um banco de dados referente a extração da lignina com H2O2, o qual foi utilizado no desenvolvimento de modelos neurais aplicando Redes Neurais Artificiais (ANN, do inglês Artificial Neural Networks) e Sistema de Inferência Adaptativa Neuro-Difusa (ANFIS, do inglês Adaptive neuro fuzzy inference system). E modelos polinomiais, os quais tiveram seus parâmetros estimados por Algoritmos Genéticos (GA, do inglês Genetic Algorithms). Os modelos desenvolvidos conseguiram predizer: o Teor de Lignina Extraída (porcento) por Espectroscopia Raman, o Teor de Lignina Oxidada (porcento) por FTIR, o Teor de Lignina Residual (porcento) pelo Método de Klason, e por último, dois modelos para a comparação da resposta analítica qualitativa com a resposta analítica quantitativa. Os modelos polinomiais, que tiveram seus parâmetros estimados por GA foram avaliados estatisticamente através da ANOVA e pelo coeficiente de correlação (R2). E os modelos neurais desenvolvidos foram avaliados pelo coeficiente de correlação (R2), número de parâmetros e índices de erro (SSE, MSE e RMSE). Para cada modelo polinomial e neural proposto, quando coerente, superfícies de resposta e curvas de contorno foram plotadas permitindo a identificação da região operacional mais indicada para a realização do pré-tratamento com H2O2. Dentre as estratégias inteligentes propostas, os modelos desenvolvidos com ANN mostraram-se mais eficientes para as predições relacionadas à extração da lignina. / [en] Industrial processes benefit the polysaccharides constituting the lignocellulosic biomass. However to manipulate them it is necessary that the biomass is submitted to the pre-treatment process. This is one of the most expensive and relevant steps for the arrangement and application of lignocellulosic fractions. The present study consists of a detailed investigation of the pretreatment process of lignocellulosic biomass with H2O2, applying intelligent technologies that enabled the optimization of this process. Artificial intelligence tools prove to be advantageous in solving the bottlenecks associated with technological advances. They enable the mathematical modeling of a process with maximum efficiency, optimizing its productivity, transforming experimental data into useful information and demonstrating the infinite possibilities of the relationships of the variables involved. The independent variables studied were the temperature (25-45 Celsius degrees) and the concentration of hydrogen peroxide (1.5 - 7.5 percent m / v). Qualitative analytical techniques (Raman and FTIR) and quantitative (Klason method) were applied to produce a database for the extraction of lignin with H2O2, which was used in the development of neural models applying Artificial Neural Networks (ANN) and Adaptive Neuro-Fuzzy Inference System (ANFIS). And polynomial models, which had their parameters estimated by Genetic Algorithms (GA). The models developed were able to predict: the Extracted Lignin Content (percent) by Raman Spectroscopy, the Oxidized Lignin Content (percent) by FTIR, the Residual Lignin Content (percent) by the Klason Method, and lastly, two models for the comparison of the qualitative analytical response with the quantitative analytical response. The polynomial models, which had their parameters estimated by GA, were statistically evaluated using ANOVA and correlation coefficient (R2) evaluated the polynomial models developed by GA statistically. And the neural models developed were evaluated by the coefficient of correlation (R2), number of parameters and error indexes (SSE, MSE and RMSE). For each proposed polynomial and neural model, when coherent, response surfaces and contour curves were plotted allowing the identification of the most suitable operational region for the pretreatment with H2O2. Among the proposed intelligent strategies, the models developed with ANN proved to be more efficient for the predictions related to lignin extraction.

Drinking water is one of our most important re- sources, so the ability to reliably monitor harmful contaminations in our water distribution network is vital. In order to minimize false alarms for water monitoring, while keeping a high sensitivity, a machine learning approach was evaluated in this project. Measurement data captured with a new kind of sensor, an electronic tongue, was provided by Linköping university. The solution was an artificial neural network, in the structure of an Autoencoder, which could learn the dynamic behaviour of natural deviations and with a false alarm rate of approximately one false alarm per week. This was done by evaluating the data and assembling an input structure to account for daily cyclic phenomena, which then was used to train the neural network. The solution could detect anomalies as small as 1.5% by comparing the input with the reconstructed vector, and raise an alarm. In conclusion, an Autoencoder is a viable method for detecting anomalies in water quality. / Drickvatten är en av våra mest värdefulla tillgångar, det är därför mycket viktigt att det finns sätt att pålitligt övervaka om dricksvattennätet blivit förorenat. För att kunna minimera antalet falsklarm och samtidigt ha hög känslighet mot dessa föroreningar undersöktes och implementerades en lösning med maskininlärningsalgoritmer. Mätdata tillhandahölls av Linköpings universitet och kom från en ny sensor kallad elektronisk tunga. Lösningen var ett artificiellt neuralt nätverk i form av en Autoencoder, som kunde lära sig det dynamiska beteende som ofarliga avvikelser utgjorde. Detta gav en lösning som i medel gav ett falsklarm per sju dagar. Detta gjordes genom att utvärdera rådata och konstruera en struktur på indata som tar hänsyn till dygnsbunda naturliga fenomen. Denna struktur användes sedan för att träna det neurala nätverket. Lösningen kunde upptäcka fel ner till 1.5% genom att jämföra indata med den rekonstruerade vektorn, och på så sätt ge ett alarm. / Kandidatexjobb i elektroteknik 2020, KTH, Stockholm

Reinforcement learning methods allows self-learningagents to play video- and board games autonomously. Thisproject aims to study the efficiency of the reinforcement learningalgorithms Q-learning and deep Q-learning for dynamical multi-agent problems. The goal is to train robots to optimally navigatethrough a warehouse without colliding.A virtual environment was created, in which the learning algo-rithms were tested by simulating moving agents. The algorithms’efficiency was evaluated by how fast the agents learned to performpredetermined tasks.The results show that Q-learning excels in simple problemswith few agents, quickly solving systems with two active agents.Deep Q-learning proved to be better suited for complex systemscontaining several agents, though cases of sub-optimal movementwere still possible. Both algorithms showed great potential fortheir respective areas however improvements still need to be madefor any real-world use. / Förstärkningsinlärningsmetoder tillåter självlärande enheter att spela video- och brädspel autonomt. Projektet siktar på att studera effektiviteten hos förstärkningsinlärningsmetoderna Q-learning och deep Q-learning i dynamiska problem. Målet är att träna upp robotar så att de kan röra sig genom ett varuhus på bästa sätt utan att kollidera. En virtuell miljö skapades, i vilken algoritmerna testades genom att simulera agenter som rörde sig. Algoritmernas effektivitet utvärderades av hur snabbt agenterna lärde sig att utföra förutbestämda uppgifter. Resultatet visar att Q-learning fungerar bra för enkla problem med få agenter, där system med två aktiva agenter löstes snabbt. Deep Q-learning fungerar bättre för mer komplexa system som innehåller fler agenter, men fall med suboptimala rörelser uppstod. Båda algoritmerna visade god potential inom deras respektive områden, däremot måste förbättringar göras innan de kan användas i verkligheten. / Kandidatexjobb i elektroteknik 2020, KTH, Stockholm

A number of non-invasive fault detection and diagnosis (FDD) techniques have been researched and have proven to have worked well in classifying faults in internal combustion engines (ICE) and other mechanical and electrical systems. These techniques are an integral step to creating more robust and accurate methods of determining where or how a fault has or will occur in such systems. These FDD techniques have the potential to not only save time avoiding a tear-down of a costly machine, but could potentially add another layer of safety in detecting and diagnosing a fault much earlier than was possible before. Looking at the previous research methods and the systems they used to acquire this data, it is a natural progression to try and make a system which is able to encapsulate all of these ideologies into one inexpensive module capable of integrating itself into the advanced set of FDD. This thesis follows along with the development of a new wireless sensor that is developed specifically for the use in FDD for ICE and other mechanical systems. A new set of software and firmware is created for the system to be able to be incorporated into previously designed algorithms. After creating and manufacturing the sensor it is put to the test by incorporating it into several Artificial Neural Networks (ANN) and comparing the results to previous experiments done with previous research equipment. Using vibration data acquired from a running engine to train a neural network, the wireless sensor was able to perform equally as well as its expensive counter parts. It proved to have the ability to achieve 100% accuracy in classifying specific engine faults. The performance of three ANN training algorithms, Levenberg-Marquardt (LM), extended Kalman Filter (EKF), and Smooth Variable Structure filter (SVSF), were tested and compared. Adding to the feasibility of a standalone system the wireless sensor was tested in a live environment as a method of instant ICE fault detection. / Thesis / Master of Applied Science (MASc)

Counterparty risk calculations have gained importance after the latest financial crisis. The bankruptcy of Lehman Brothers showed that even large financial institutiones face a risk of default. Hence, it is important to measure the risk of default for all the contracts written between financial institutions. Credit Value Adjustment, CVA, is an industry standard method for such calculations. Nevertheless, the implementation of this method is contract dependent and the necessary computer simulations can be very intensive. Monte Carlo simulations have for a long time been known as a precise but slow technique to evaluate the cash flows for contracts of all kinds. Measuring the exposure of a contract written on structured products might require half a day of calculations if the implementation is written without significant optimization. Several ideas have been presented by researchers and applied in the industry, the idea explored and implemented in this thesis was based on using Artificial Neural Networks in Python. This procedure require a decomposition of the Expected Exposure calculation within the CVA and generating a large data set using a standard Monte Carlo simulation. Three network architectures have been tested and the final performance was compared with using standard techniques for the very same calculation. The performance gain was significant, a portfolio of 100 counterparties with 10 contracts each would take 20 minutes of calculations in total when using the best performing architecture whereas a parallel C++ implementation of the standard method would require 2.6 days.

Multidisciplinary design optimization (MDO) requires many designs to be evaluated while searching for an optimum. As a result, the calculations done to evaluate the designs must be quick and simple to have a reasonable turn-around time. This makes aerodynamic calculations in the transonic regime difficult. Running computational fluid dynamics (CFD) calculations within the MDO code would be too computationally expensive. Instead, CFD is used outside the MDO to find two-dimensional aerodynamic properties of a chosen airfoil shape, BACJ, at a number of points over a range of thickness-to-chord ratios, free-stream Mach numbers, and lift coefficients. These points are used to generate surrogate models which can be used for the two-dimensional aerodynamic calculations required by the MDO computational design environment. Strip theory is used to relate these two-dimensional results to the three-dimensional wing. Models are developed for the center of pressure location, the lift curve slope, the wave drag, and the maximum allowable lift coefficient before buffet. These models have good agreement with the original CFD results for the airfoil. The models are integrated into the aerodynamic and aeroelastic sections of the MDO code. / Master of Science