One of the crucial challenges for Data Center (DC) operation is inefficient thermal management which leads to excessive energy waste. The information technology (IT) equipment and cooling systems of a DC are major contributors to power consumption. Additionally, failure of a DC cooling system leads to higher operating temperatures, causing critical electronic devices, such as servers, to fail which leads to significant economic loss. Improvements can be made in two ways, through (1) better design of a DC architecture and (2) optimization of the system for better heat transfer from hot servers. Row-based cooling is a suitable DC configuration that reduces energy costs by improving airflow distribution. Here, the IT equipment is contained within an enclosure that includes a cooling unit which separates cold and back chambers to eliminate hot air recirculation and cold air bypass, both of which produce undesirable airflow distributions. Besides, due to scalability, ease of implementation, and operational cost, row-based systems have gained in popularity for DC computing applications. However, a general thermal model is required to predict spatiotemporal temperature changes inside the DC and properly apply appropriate strategies. As yet, only primitive tools have been developed that are time-consuming and provide unacceptable errors during extrapolative predictions. We address these deficiencies by developing a rapid, adaptive, and accurate hybrid model by combining a DDM and the thermofluid transport relations to predict temperatures in a DC. Our hybrid model has low interpolative prediction errors below 0.7 oC and extrapolative errors less than one half of black-box models. Additionally, by changing the studied DC configuration such as cooling unit fans and severs locations, there are a few zones with prediction error more than 2 oC. Existing methods for cooling unit fault detection and diagnosis (FDD) are designed to successfully overcome individually occurring faults but have difficulty handling simultaneous faults. We apply a gray-box model involves a case study to detect and diagnose cooling unit fan and pump failure in a row-based DC cooling system. Fast detection of anomalous behavior saves energy and reduces operational costs by initiating remedial actions. Cooling unit fans and pumps are relatively low-reliability components, where the failure of one or more components can cause the entire system to overheat. Therefore, appropriate energy-saving strategies depend largely on the accuracy and timeliness of temperature prediction models. We used our gray-box model to produce thermal maps of the DC airspace for single as well as simultaneous failure conditions, which are fed as inputs for two different data-driven classifiers, CNN and RNN, to rapidly predict multiple simultaneous failures. Our FDD strategy can detect and diagnose multiple faults with accuracy as high as 100% while requiring relatively few simultaneous fault training data samples. / Thesis / Candidate in Philosophy

This thesis is a comparative study where the question is whether a neural network approach can outperform the principal component analysis (PCA) approach for predicting changes of interest rate curves. Today PCA is the industry standard model for predicting interest rate curves. Specifically the goal is to better understand the correlation structure between Swedish and European swap rates. The disadvantage with the PCA approach is that only the information contained in the covariance matrix can be used and not for example whether or not the curve might behave different depending on the current state. In other words, some information that might be quite important to the curve dynamic is lost in the PCA approach. This raises the question whether the lost information is important for prediction accuracy or not. As previously been shown by Alexei Kondratyev in the paper "Learning Curve Dynamics with Artificial Neural Networks", the neural network approach is able to use more information in the data and therefore has potential to outperform the PCA approach. Our thesis shows that the neural network approach is able to achieve the same or higher accuracy than PCA when performing long term predictions. The results show that the neural network model has potential to replace the PCA model, however, it is a more time consuming model. Higher accuracy can probably be achieved if the network is more optimized. / Det här är en jämförande studie där syftet är att undersöka hurvida noggrannare prediktioner kan uppnås genom att använda sig av artificiella neurala nätverk (ANN) istället för principalkomponentanalys (PCA) för att förutspå swapräntekurvor. PCA är idag industristandard för att förutspå räntekurvor. Specifikt är målet att bättre kunna förstå korrelationsstrukturen mellan de Svenska swapräntorna och de Europiska swapräntorna. En nackdel med PCA är att den enda tillgängliga informationen sparas i kovariansmatrisen. Det kan till exempel vara fallet att kurvan beter sig väldigt annorlunda beroende på om de nuvarande räntenivåerna är höga eller låga. Eftersom att sådan information går förlorad i PCA-modellen ligger intresset i att undersöka hur mycket noggrannare prediktionerna kan bli om man tar tillvara på ännu mer av informationen i datan. Som Alexei Kondratyev visar i rapporten "Learning Curve Dynamics with Artificial Neural Networks", så har ANN-modellen potential att ersätta PCA-modellen för att förutspå räntekurvor. I denna studie framgår det att ANN-modellen uppnår samma eller bättre resultat jämfört med PCA-modellen vid längre prediktioner.

Around the world bridges are ageing. In Europe approximately two thirds of all railway bridges are over 50 years old. As these structures age, it becomes increasingly important that they are properly maintained. If damage remains undetected this can lead to premature replacement which can have major financial and environmental costs. It is also imperative that bridges are kept safe for the people using them. Thus, it is necessary for damage to be detected as early as possible. This research investigates an unsupervised, model-free damage detection method which could be implemented for continuous structural health monitoring. The method was based on past research by Gonzalez and Karoumi (2015), Neves et al. (2017) and Chalouhi et al. (2017). An artificial neural network (ANN) was trained on accelerations from the healthy structural state. Damage sensitive features were defined as the root mean squared errors between the measured data and the ANN predictions. A baseline healthy state could then be established by presenting the trained ANN with more healthy data. Thereafter, new data could be compared with this reference state. Outliers from the reference data were taken as an indication of damage. Two outlier detection methods were used: Mahalanobis distance and the Kolmogorov-Smirnov test. A model steel bridge with a span of 5 m, width of 1 m and height of approximately 1.7 m was used to study the damage detection method. The use of an experimental model allowed damaged to be freely introduced to the structure. The structure was excited with a 12.7 kg rolling mass at a speed of approximately 2.1 m/s (corresponding to a 20.4 ton axle load moving at 47.8 km/h in full scale). Seven accelerometers were placed on the structure and their locations were determined using an optimal sensor placement algorithm. The objectives of the research were to: identify a number of single damage cases, distinguish between gradual damage cases and identify the location of damage. The proposed method showed promising results and most damage cases were detected by the algorithm. Sensor density and the method of excitation were found to impact the detection of damage. By training the ANN to predict correlations between accelerometers the sensor closest to the damage could be detected, thus successfully localising the damage. Finally, a gradual damage case was investigated. There was a general increase in the damage index for greater damage however, this did not progress smoothly and one case of ‘greater’ damage showed a decrease in the damage index.

Global sea surface temperature (SST) anomalies have a demonstrable effect on terrestrial climate dynamics throughout the continental U.S. SST variations have been correlated with greenness (vegetation densities) and precipitation via ocean-atmospheric interactions known as climate teleconnections. Prior research has demonstrated that teleconnections can be used for climate prediction across a wide region at sub-continental scales. Yet these studies tend to have large uncertainties in estimates by utilizing simple linear analyses to examine chaotic teleconnection relationships. Still, non-stationary signals exist, making teleconnection identification difficult at the local scale. Part 1 of this research establishes short-term (10-year), linear and non-stationary teleconnection signals between SST at the North Atlantic and North Pacific oceans and terrestrial responses of greenness and precipitation along multiple pristine sites in the northeastern U.S., including (1) White Mountain National Forest - Pemigewasset Wilderness, (2) Green Mountain National Forest - Lye Brook Wilderness and (3) Adirondack State Park - Siamese Ponds Wilderness. Each site was selected to avoid anthropogenic influences that may otherwise mask climate teleconnection signals. Lagged pixel-wise linear teleconnection patterns across anomalous datasets found significant correlation regions between SST and the terrestrial sites. Non-stationary signals also exhibit salient co-variations at biennial and triennial frequencies between terrestrial responses and SST anomalies across oceanic regions in agreement with the El Nino Southern Oscillation (ENSO) and North Atlantic Oscillation (NAO) signals. Multiple regression analysis of the combined ocean indices explained up to 50% of the greenness and 42% of the precipitation in the study sites. The identified short-term teleconnection signals improve the understanding and projection of climate change impacts at local scales, as well as harness the interannual periodicity information for future climate projections. Part 2 of this research paper builds upon the earlier short-term study by exploring a long-term (30-year) teleconnection signal investigation between SST at the North Atlantic and Pacific oceans and the precipitation within Adirondack State Park in upstate New York. Non-traditional teleconnection signals are identified using wavelet decomposition and teleconnection mapping specific to the Adirondack region. Unique SST indices are extracted and used as input variables in an artificial neural network (ANN) prediction model. The results show the importance of considering non-leading teleconnection patterns as well as the known teleconnection patterns. Additionally, the effects of the Pacific Ocean SST or the Atlantic Ocean SST on terrestrial precipitation in the study region were compared with each other to deepen the insight of sea-land interactions. Results demonstrate reasonable prediction skill at forecasting precipitation trends with a lead time of one month, with r values of 0.6. The results are compared against a statistical downscaling approach using the HadCM3 global circulation model output data and the SDSM statistical downscaling software, which demonstrate less predictive skill at forecasting precipitation within the Adirondacks.

Electrocardiogram (ECG) signal processing has been the subject of intense research in the past years, due to its strategic place in the detection of several cardiac pathologies. However, ECG signal is frequently corrupted with different types of noises such as 60Hz power line interference, baseline drift, electrode movement and motion artifact, etc. In this thesis, a hybrid two-stage model based on the combination of wavelet decomposition and artificial neural network is proposed for ECG noise reduction based on excellent localization features: wavelet transform and the adaptive learning ability of neural network. Results from the simulations validate the effectiveness of this proposed method. Simulation results on actual ECG signals from MIT-BIH arrhythmia database [30] show this approach yields improvement over the un-filtered signal in terms of signal-to-noise ratio (SNR).

Electrochemical Energy Storage Systems are a viable and popular solution to fulfill energy storage requirements for energy generated through sustainable energy resources. With the increasing demand for Electrical Vehicles (EVs), Lithium-ion batteries (LIB) are being widely and getting popular compared to other battery technologies due to their energy storage capacity. However, LIBs suffer from disadvantages such as battery life and the degradation of electrode material with time, that can be improved by understanding these mechanisms using experimental and computational techniques. Further, it has been experimentally observed and numerically determined that lithium-ion intercalation induced stress and thermal loading can cause capacity fading and local fractures in the electrode materials. These fractures are one of the major degradation mechanisms in Lithium-ion batteries. With LixMn2O4 as a cathode material, stress values differ widely especially for intermediate State Of Charge (SOC), and very few attempts have been made to understand the stress distribution as a function of SOC at molecular level. Therefore, the estimates of mechanical properties such as Young’s modulus, diffusion coefficient etc. differ, especially for partially charged states. Further, the effect of temperature, particularly elevated temperatures, have not been taken into the consideration. Studying these parameters at the atomic scale can provide insight information and help in improving these materials lifetime. Hence, molecular/atomic level mathematical modelling has been used to understand capacity fade due to Lithium-ion intercalation/de-intercalation induced stress. Large-scale Atomic/Molecular Massively Parallel Simulator (LAMMPS) [1], that is widely used for atomic simulations, has been used for the simulation studies of this dissertation. Thus, the objective of this study is to understand the fracture mechanisms in the Lithium Manganese Oxide (LiMn2O4) electrode at the molecular level by studying mechanical properties of the material at different SOC values using the principles of molecular dynamics (MD). As part of the model validation, the lattice parameter and volume changes of LixMn2O4 as a function of SOC (0 < x < 1) has been studied and validated with respect to the experimental data. This validated model has been used for a parametric study involving the SOC value, strain-rate (charge and discharge rate), and temperature. Based on the validated MD setup, doping and co-doping studies have been undertaken to design and develop new and novel cathode materials with enhanced properties. In the absence of experimental data for the new engineered structures, validation with Quantum Mechanics generated lattice structures has been done. The results suggest that lattice constant values obtained from both MD and QM simulations are in good agreement (∼ 99%) with experimental values. Further, Single Particle Model (SPM) based macro scale Computational Fluid Dynamics findings show that co-doping has improved the material properties especially for Yttrium and Sulfur doped structures which can improve the cycle life anywhere between 600-7000 cycles. Further in order to reduce the required computational time to obtain minimum potential energy ionic configuration out of millions of scenario, Artificial Neural Network (ANN) technique is being used. It improved the processing time by more than 88%. / Thesis / Doctor of Philosophy (PhD)

The purpose of this study was to investigate the effect of varying optimization parameters on the proposed optimum of a tablet coating formulation requiring minimization of crack velocity and maximization of film opacity. An artificial neural network (ANN) comprising six input and two output nodes separated by a single hidden layer of five nodes was trained using 100 pseudo-randomly distributed records and optimized by guided evolutionary simulated annealing (GESA). GESA was unable to identify a formulation that satisfied both a crack velocity of 0 m s¿1 and a film opacity of 100% due to conflict centred on the response of the properties to variation in pigment particle size. Constraining film thickness exacerbated the property conflict. By adjusting property weights (i.e. the relative importance of each property), GESA was able to propose formulations that were either crack resistant or that were fully opaque. Reducing the stringency of the performance criteria (crack velocity >0 m s¿1, film opacity <100%) enabled GESA to propose optima that met or exceeded the looser targets. Under these conditions, starting GESA from different locations within model space resulted in the proposal of different optima. Therefore, application of loose targets resulted in the identification of an optimal zone within which all formulations satisfied these less stringent performance criteria. It is concluded that application of the most stringent performance criteria and selection of appropriate property weights is necessary for unequivocal identification of the true optimum. A strategy for optimization experiments is proposed.

Yes / The current UK rail network is managed by Network Rail, which requires an investment of £5.2bn per year to cover operational costs [1]. These expenses include the maintenance and repairs of the railway rails. This paper aims to create a proof of concept for an autonomous health monitoring system of the rails using an integrated finite element analysis (FEA) and artificial neural network (ANN) approach. The FEA is used to model worn profiles of a standard rail and predict the stress field considering the material of the rail and the loading condition representing a train travelling on a straight line. The generated FEA data is used to train an ANN model which is utilised to predict the stress field of a worn rail using optically scanned data. The results showed that the stress levels in a rail predicted with the ANN model are in an agreement with the FEA predictions for a worn rail profile. These initial results indicate that the ANN can be used for the rapid prediction of stresses in worn rails and the FEA-ANN based approach has the potential to be applied to autonomous health monitoring of rails using fast scanners and validated ANN models. However, further development of this technology would be required before it could be used in the railway industry, including: real time data processing of scanned rails; improved scanning rates to enhance the inspection efficiency; development of fast computational methods for the ANN model; and training the ANN model with a large set of representative data representing application specific scenarios.

The chargeback process is costly for the merchant. Not only does the merchant lose the revenue from the purchase, but it must also pay handling fees to the bank and risks never getting paid for provided service. The purpose of this study is to examine and investigate how to prognosticate future chargebacks by using machine learning in form of Artificial Neural Network on transaction data. Doing so can be used to minimize and decrease financial costs for the merchant. The study indicates that it’s complex to prognosticate chargebacks, but illuminates that it’s possible under certain circumstances. The created model has been concluded to be more suitable as a compliment, rather than a substitute for the current Rule-Based classification system. The model should be implemented based on economic analysis since it can be used to reduce costs and contribute to profitability over time. Furthermore, the study highlight lessons learned and complementary research areas for future studies. / Återbetalningar medför stora kostnader för handlare i form av förlorade intäkter vid återbetalning av transaktionssumman, samt tillkommande handläggningsavgifter i processen. Syftet med rapporten är att utvärdera och undersöka möjligheten att prognostisera framtida återbetalningar, genom att applicera maskininlärning i form av Artificiellt Neuralt Nätverk på transaktionsdata. På så sätt kan återbetalningar minimeras och reducera finansiella kostnader hos handlaren. Studien påvisar att det är komplext att predicera återbetalningar, men att det är möjligt under särskilda omständigheter. Modellen som skapats har konstaterats mer lämpad som ett komplement till det aktuella regelbaserade klassificeringssystemet än ett substitut. Utifrån en ekonomisk analys klargörs att algoritmen bör implementeras för att reducera kostnader och på sikt bidra till lönsamhet. Studien belyser även lärdomar, samt kompletterande forskningsområden för framtida studier.

Steel is an important engineering material used in a variety of applications due to its mechanical properties and durability. With increasing demand for higher performance, complex structures, and the need for cost reduction within manufacturing processes, there are numerous challenges with traditional steel design options and production methods with manufacturing cost being the most significant. In this research, this challenge is addressed by developing a micro-genetic algorithm to minimize the manufacturing cost while designing steel with the desired mechanical properties. The algorithm was integrated with machine learning models to predict the mechanical properties and microstructure for the generated alloys based on their chemical compositions and heat treatment conditions. Through this, it was demonstrated that new steel alloys with specific mechanical property targets could be generated at an optimal cost. The research’s contribution lies in the development of a different approach to optimize steel production that combines the advantages of machine learning and evolutionary algorithms while increasing the number of input parameters. Additionally, it uses a small dataset illustrating that it can be used in applications where data is lacking. This approach has significant implications for the steel industry as it provides a more efficient way to design and produce new steel alloys. It also contributes to the overall material science field by demonstrating its ability in a material’s design and optimization. Overall, the proposed framework highlights the potential of utilizing machine learning and evolutionary algorithms in material design and optimization. / Thesis / Master of Applied Science (MASc) / This research aims to develop an AI-based functional integrated with a heuristic algorithm that optimizes parameters to meet desired mechanical properties and cost for steels. The developed algorithm generates new alloys which meet desired mechanical property targets by considering alloy composition and heat treatment condition inputs. Used in combination with machine learning models for the mechanical property and microstructure prediction of new alloys, the algorithm successfully demonstrates its ability to meet specified targets while achieving cost savings. The approach presented has significant implications for the steel industry as it offers a quick method of optimizing steel production, which can reduce overall costs and improve efficiency. The integration of machine learning within the algorithm offers a different way of designing new steel alloys which has the potential to improve manufactured products by ultimately improving their performance and quality.