In the vast field of signal processing, machine learning is rapidly expanding its domain into all realms. As a constituent of this expansion, this thesis presents contributive work on advancements in machine learning algorithms by building on the shoulder of giants. The first chapter of this thesis contains enhancements to a CNN (convolutional neural network) for better classification of heartbeat arrhythmia. The network goes through a two stage development, the first being augmentations to the network and the second being the implementation of dropout. Chapter 2 involves the combination of CNN and LSTM (long short term memory) networks for the task of short-term energy use data regression. Exploiting the benefits of two of the most powerful neural networks, a unique, novel neural network is created to effectually predict future energy use. The final section concludes this work with directions for future works.

This research explores the temporal patterns of psychophysiological responses to musical excerpts by analyzing the expansive Emotion in Motion dataset, the most comprehensive of its kind. Utilizing the Dynamic Time Warping and T-test analysis techniques, we examined data from participants across seven countries who listened to three distinct musical pieces. During these listening sessions, Electrodermal Activity (EDA) and Pulse Oximetry (POX) readings were collected, complemented by qualitative feedback from the participants. Our analysis focused on detecting recurring patterns and extracting meaningful insights from the data. In addition to this, we compare several Deep Neural Networks to find the one that is best suited for prediction of emotional attributes with EDA and POX signals as input. To further facilitate a comprehensive visualization and analysis of the EDA, POX, and audio signals, we developed a dedicated platform, which features a coordinated multiple view interface, as an integral part of this work. / Master of Science / We explored how people's bodies react over time when they listen to music. We used a large collection of data called the ``Emotion in Motion'' dataset, which has information from people in seven countries who listened to three different music pieces. To understand this data, we used special tools that help detect patterns and changes in how the body responds. During the music sessions, the participant's skin's electrical activity and the amount of oxygen in the blood were recorded, which can give clues about emotional reactions. People also shared their feelings about the music. To make it easier to see and understand all this information together, we created a new web platform that simulates the experiment in real-time. This work aims to help us better understand the deep connection between music and human emotions.

Temperature regulation in buildings can be tricky and expensive. A common problem when heating buildings is that an unnecessary amount of energy is supplied. This waste of energy is often caused by a faulty regulation system. This thesis presents a machine learning ap- proach, using time series data, to predict the energy supply needed to keep the inside tem- perature at around 21 degrees Celsius. The machine learning models LSTM, Ensemble LSTM, AT-LSTM, ARIMA, and XGBoost were used for this project. The validation showed that the ensemble LSTM model gave the most accurate predictions with the Mean Absolute Error of 22486.79 (Wh) and Symmetric Mean Absolute Percentage Error of 5.41 % and was the model used for comparison with the current system. From the performance of the different models, the conclusion is that machine learning can be a useful tool to pre- dict the energy supply. But on the other hand, there exist other complex factors that need to be given more attention to, to evaluate the model in a better way. / Temperaturreglering i byggnader kan vara knepigt och dyrt. Ett vanligt problem vid upp- värmning av byggnader är att det tillförs onödigt mycket energi. Detta energispill orsakas oftast av ett felaktigt regleringssystem. Denna rapport studerar möjligheten att, med hjälp av tidsseriedata, kunna träna olika maskininlärningmodeller för att förutsäga den energitill- försel som behövs för att hålla inomhustemperaturen runt 21 grader Celsius. Maskininlär- ningsmodellerna LSTM, Ensemble LSTM, AT-LSTM, ARIMA och XGBoost användes för detta projekt. Valideringen visade att ensemble LSTM-modellen gav den mest exakta förut- sägelserna med Mean Absolute Error på 22486.79 (Wh) och Symmetric Mean Absolute Percentage Error på 5.41% och var modellen som användes för att jämföra med det befint- liga systemet. Från modellernas prestation är slutsatsen att maskininlärning kan vara ett an- vändbart verktyg för att förutsäga energitillförseln. Men å andra sidan finns det andra kom- plexa faktorer som bör tas hänsyn till så att modellen kan evalueras på ett bättre sätt.

No description available.

L'analyse de la démarche clinique est généralement subjective, étant effectuée par des cliniciens observant la démarche des patients. Des alternatives à une telle analyse sont les systèmes basés sur les marqueurs et les systèmes basés sur les plates-formes au sol. Cependant, cette analyse standard de la marche nécessite des laboratoires spécialisés, des équipements coûteux et de longs délais d'installation et de post-traitement. Les chercheurs ont fait de nombreuses tentatives pour proposer une alternative basée sur la vision par ordinateur pour l'analyse de la demarche. Avec l'apparition de caméras 3D bon marche, le problème de l'évaluation qualitative de la démarche a été re-examiné. Les chercheurs ont réalisé le potentiel des dispositifs de cameras 3D pour les applications d'analyse de mouvement. Cependant, malgré des progrès très encourageants dans les technologies de détection 3D, leur utilisation réelle dans l'application clinique reste rare.Cette thèse propose des modèles et des techniques pour l'évaluation du mouvement à l'aide d'un capteur Microsoft Kinect. En particulier, nous étudions la possibilité d'utiliser différentes données fournies par une caméra RGBD pour l'analyse du mouvement et de la posture. Les principales contributions sont les suivantes. Nous avons réalisé une étude de l'etait de l'art pour estimer les paramètres importants de la démarche, la faisabilité de différentes solutions techniques et les méthodes d'évaluation de la démarche existantes. Ensuite, nous proposons un descripteur de posture basé sur un nuage de points 3D. Le descripteur conçu peut classer les postures humaines statiques a partir des données 3D. Nous construisons un système d'acquisition à utiliser pour l'analyse de la marche basée sur les donnees acquises par un capteur Kinect v2. Enfin, nous proposons une approche de détection de démarche anormale basée sur les données du squelette. Nous démontrons que notre outil d'analyse de la marche fonctionne bien sur une collection de données personnalisées et de repères existants. Notre méthode d'évaluation de la démarche affirme des avances significatives dans le domain, nécessite un équipement limité et est prêt à être utilisé pour l'évaluation de la démarche. / Clinical Gait analysis is traditionally subjective, being performed by clinicians observing patients gait. A common alternative to such analysis is markers-based systems and ground-force platforms based systems. However, this standard gait analysis requires specialized locomotion laboratories, expensive equipment, and lengthy setup and post-processing times. Researchers made numerous attempts to propose a computer vision based alternative for clinical gait analysis. With the appearance of commercial 3D cameras, the problem of qualitative gait assessment was reviewed. Researchers realized the potential of depth-sensing devices for motion analysis applications. However, despite much encouraging progress in 3D sensing technologies, their real use in clinical application remains scarce.In this dissertation, we develop models and techniques for movement assessment using a Microsoft Kinect sensor. In particular, we study the possibility to use different data provided by an RGBD camera for motion and posture analysis. The main contributions of this dissertation are the following. First, we executed a literature study to estimate the important gait parameters, the feasibility of different possible technical solutions and existing gait assessment methods. Second, we propose a 3D point cloud based posture descriptor. The designed descriptor can classify static human postures based on 3D data without the use of skeletonization algorithms. Third, we build an acquisition system to be used for gait analysis based on the Kinect v2 sensor. Fourth, we propose an abnormal gait detection approach based on the skeleton data. We demonstrate that our gait analysis tool works well on a collection of custom data and existing benchmarks. Weshow that our gait assessment approach advances the progress in the field, is ready to be used for gait assessment scenario and requires a minimum of the equipment.

Le domaine du traitement automatique de la parole regroupe un très grand nombre de tâches parmi lesquelles on trouve la reconnaissance de la parole, l'identification de la langue ou l'identification du locuteur. Ce domaine de recherche fait l'objet d'études depuis le milieu du vingtième siècle mais la dernière rupture technologique marquante est relativement récente et date du début des années 2010. C'est en effet à ce moment qu'apparaissent des systèmes hybrides utilisant des réseaux de neurones profonds (DNN) qui améliorent très notablement l'état de l'art. Inspirés par le gain de performance apporté par les DNN et par les travaux d'Alex Graves sur les réseaux de neurones récurrents (RNN), nous souhaitions explorer les capacités de ces derniers. En effet, les RNN nous semblaient plus adaptés que les DNN pour traiter au mieux les séquences temporelles du signal de parole. Dans cette thèse, nous nous intéressons tout particulièrement aux RNN à mémoire court-terme persistante (Long Short Term Memory (LSTM) qui permettent de s'affranchir d'un certain nombre de difficultés rencontrées avec des RNN standards. Nous augmentons ce modèle et nous proposons des processus d'optimisation permettant d'améliorer les performances obtenues en segmentation parole/non-parole et en identification de la langue. En particulier, nous introduisons des fonctions de coût dédiées à chacune des deux tâches: un simili-WER pour la segmentation parole/non-parole dans le but de diminuer le taux d'erreur d'un système de reconnaissance de la parole et une fonction de coût dite de proximité angulaire pour les problèmes de classification multi-classes tels que l'identification de la langue parlée. / Automatic speech processing is an active field of research since the 1950s. Within this field the main area of research is automatic speech recognition but simpler tasks such as speech activity detection, language identification or speaker identification are also of great interest to the community. The most recent breakthrough in speech processing appeared around 2010 when speech recognition systems using deep neural networks drastically improved the state-of-the-art. Inspired by this gains and the work of Alex Graves on recurrent neural networks (RNN), we decided to explore the possibilities brought by these models on realistic data for two different tasks: speech activity detection and spoken language identification. In this work, we closely look at a specific model for the RNNs: the Long Short Term Memory (LSTM) which mitigates a lot of the difficulties that can arise when training an RNN. We augment this model and introduce optimization methods that lead to significant performance gains for speech activity detection and language identification. More specifically, we introduce a WER-like loss function to train a speech activity detection system so as to minimize the word error rate of a downstream speech recognition system. We also introduce two different methods to successfully train a multiclass classifier based on neural networks for tasks such as LID. The first one is based on a divide-and-conquer approach and the second one is based on an angular proximity loss function. Both yield performance gains but also speed up the training process.

Bipolar mood disorder is a severe mental condition that has multiple episodesof either of two types: manic or depressive. These phases can lead patients tobecome hyperactive, hyper-sexual, lethargic, or even commit suicide — all ofwhich seriously impair the quality of life for patients. Predicting these phaseswould help patients manage their lives better and improve our ability to applymedical interventions. Traditionally, interviews are conducted in the evening topredict potential episodes in the following days. While machine learningapproaches have been used successfully before, the data was limited tomeasuring a few self-reported parameters each day. Using biometrics recordedat short intervals over many months presents a new opportunity for machinelearning approaches. However, phases of unrest and hyperactivity, which mightbe predictive signals, are not only often experienced long before the onset ofmanic or depressive phases but are also separated by several uneventful days.This delay and its aperiodic occurrence are a challenge for deep learning. In thisthesis, a fictional dataset that mimics long and irregular delays is created andused to test the effects of such long delays and rare events. LSTMs, RNNs, andGRUs are the go-to models for deep learning in this situation. However, theydiffer in their ability to be trained over a long time. As their acronym suggests,LSTMS are believed to be easier to train and to have a better ability to remember(as their name suggests) than their simpler RNN counterparts. GRUs representa compromise in complexity between RNNs and LSTMs. Here, I will show that,contrary to the common assumption, LSTMs are surprisingly forgetful and thatRNNs have a much better ability to generalize over longer delays with shortersequences. At the same time, I could confirm that LSTMs are easily trained ontasks that have more prolonged delays.

The demand for high data rates communication and scarcity of spectrum in existing microwave bands has been the key aspect in 5G. To fulfill these demands, the millimeter wave (mmWave) with large bandwidths has been proposed to enhance the efficiency and the stability of the 5G network. In mmWave communication, the concentration of the transmission signal from the antenna is conducted by beamforming and beam tracking. However, state-of-art methods in beam tracking lead to high resource consumption. To address this problem, we develop 2 machine-learning-based solutions for overhead reduction. In this paper, a scenario configuration simulator is proposed as the data collection approach. Several LSTM based time series prediction models are trained for experiments. Since the overhead is reduced by decreasing the number of sweeping beams in solutions, multiple data imputation methods are proposed to improve the performance of the solution. These methods are based on Multiple Imputation by Chained Equations (MICE) and generative adversarial networks. Both qualitative and quantitative experimental results on several types of datasets demonstrate the efficacy of our solution. / Efterfrågan på hög datahastighetskommunikation och brist på spektrum i befintliga mikrovågsband har varit nyckelaspekten i 5G. För att uppfylla dessa krav har millimetervåg (mmWave) med stora bandbredder föreslagits för att förbättra effektiviteten och stabiliteten i 5G-nätverket. I mmWavekommunikation utförs koncentrationen av överföringssignalen från antennen genom strålformning och strålspårning. Toppmoderna metoder inom strålspårning leder dock till hög resursförbrukning. För att lösa detta problem utvecklar vi två maskininlärningsbaserade lösningar för reduktion av omkostnader. I det här dokumentet föreslås en scenariokonfigurationssimulator som datainsamlingsmetod. Flera LSTM-baserade modeller för förutsägelse av tidsserier tränas för experiment. Eftersom omkostnaderna reduceras genom att minska svepstrålarna i lösningar föreslås flera datainputeringsmetoder för att förbättra lösningens prestanda. Dessa metoder är baserade på Multipel Imputation by Chained Equations (MICE) och generativa kontroversiella nätverk. Både kvalitativa och kvantitativa experimentella resultat på flera typer av datamängder visar effektiviteten i vår lösning.

Embedded platforms are often used as a sensor data processing node to collect data and transmit the data to the remote server. Due to the poor performance and power limitation, data processing was often left to the remote server. With the improvement of the computation ability, it is becoming possible to do some partial data processing on the embedded platforms, which would reduce the power and time consumption on the data transmission. Moreover, processing the data locally on the embedded platforms could reduce the dependence on the network. The platform could even do some tasks offline. This project aims to explore effective data analysis methods, especially for anomaly detection, which could be implemented on the embedded platform to be analyzed and detected locally. In this project, we select four methods: Seasonal and Trend Decomposition Using Loess (STL), Autoregressive Integrated Moving Average Model (ARIMA), Vector Autoregression (VAR), Long ShortTerm Memory (LSTM), to implement on the embedded platform ESP32. To test which methods could better fit the platform, we evaluate and compare the result from two aspects: the time overhead and the accuracy. The results show that the STL has the highest detection accuracy, but its time overhead is significantly higher than all other methods. ARIMA has the smallest time overhead and higher accuracy than LSTM and VAR. For LSTM, the method performs better with univariable input than multivariable input. Finally, we discuss the factors that may influence the result and future works. / Inbäddade plattformar används ofta som en sensor databehandlingsnod för att samla in och sedan överföra data till fjärrservern. Databehandling lämnades ofta till fjärrservern på grund av den dåliga prestandan och effektbegränsningen. Med förbättrad beräkningsförmåga blir det framkomligt att göra en del databehandling på de inbäddade plattformarna, vilket skulle minska ström och tidsförbrukningen för dataöverföringen. För övrigt kan lokal behandling av data på de inbäddade plattformarna minska beroendet av nätverket. Plattformen kan till och med utföra vissa uppgifter I nedkopplat läge. Detta projekt avser att utforska effektiva dataanalysmetoder särskilt för avvikelsedetektering, som kan verkställas på den inbäddade plattformen för att analyseras och upptäckas lokalt. I det här projektet väljer vi fyra metoder för att införa på den inbäddade plattformen ESP32: Seasonal and Trend Decomposition Using Loess (STL), Autoregressive Integrated Moving Average Model (ARIMA), Vector Autoregression (VAR), Long Short-Term Memory (LSTM). För att testa vilka metoder som bättre passar plattformen utvärderar och jämför vi resultatet med hänsyn till två aspekter: tidsomkostnaderna och noggrannheten. Resultaten visar att STL har den högsta detektionsnoggrannheten, men dess tidsomkostning är betydligt högre än alla andra metoder. ARIMA har den minsta tidsomkostningen och högre noggrannhet än LSTM och VAR. För LSTM fungerar metoden bättre med univariable input än multivariable input. Slutligen diskuterar vi faktorerna som möjligtvis påverkar resultatet och framtida arbeten.

Thunderstorms are both dangerous as well as important rain-bearing structures for large parts of the world. The prediction of thunderstorm trajectories is however difficult, especially in tropical regions. This is largely due to their smaller size and shorter lifespan. To overcome this issue, this thesis investigates how well a neural network composed of long short-term memory (LSTM) units can predict the trajectories of thunderstorms, based on several years of lightning strike data. The data is first clustered, and important features are extracted from it. These are used to predict the mean position of the thunderstorms using an LSTM network. A random search is then carried out to identify optimal parameters for the LSTM model. It is shown that the trajectories predicted by the LSTM are much closer to the true trajectories than what a linear model predicts. This is especially true for predictions of more than 1 hour. Scores commonly used to measure forecast accuracy are applied to compare the LSTM and linear model. It is found that the LSTM significantly improves forecast accuracy compared to the linear model. / Åskväder är både farliga och livsviktiga bärare av vatten för stora delar av världen. Det är dock svårt att förutsäga åskcellernas banor, främst i tropiska områden. Detta beror till större delen på deras mindre storlek och kortare livslängd. Detta examensarbete undersöker hur väl ett neuralt nätverk, bestående av long short-term memory-lager (LSTM) kan förutsäga åskväders banor baserat på flera års blixtnedlslagsdata. Först klustras datan, och viktiga karaktärsdrag hämtas ut från den. Dessa används för att förutspå åskvädrens genomsnittliga position med hjälp av ett LSTMnätverk. En slumpmässig sökning genomförs sedan för att identifiera optimala parametrar för LSTM-modellen. Det fastslås att de banor som förutspås av LSTM-modellen är mycket närmare de sanna banorna, än de som förutspås av en linjär modell. Detta gäller i synnerhet för förutsägelser mer än 1 timme framåt. Värden som är vanliga för att bedöma prognosers träffsäkerhet beräknas för att jämföra LSTM-modellen och den linjära. Det visas att LSTM-modellen klart förbättrar förutsägelsernas träffsäkerhet jämfört med den linjära modellen.