The increasing occurrence of structural deformations in the electrodes of the jelly roll has raised quality concerns during battery manufacturing, emphasizing the need to detect them automatically with the advanced techniques. This thesis aims to explore and provide two models based on traditional computer vision (CV) and deep neural network (DNN) techniques using computed tomography (CT) scan images of jelly rolls to ensure that the product is of high quality. For both approaches, electrode peaks as keypoints of anodes and cathodes in prismatic lithium battery jelly rolls are detected to extract the geometric features to identify if a particular jelly roll has some structural deformations. For traditional CV methods, the images undergo some pre-processing steps, extraction of foreground through adaptive thresholding, and morphological operations to extract contour edges, followed by applying Harris corner detector to detect electrode peaks. However, this approach shows limitations in detecting small or negative distance differences in deformed images. Furthermore, this study proposes another approach based on supervised transfer learning using pre-trained deep learning models on annotated data. After exploring different architectures, the VGG19 model pre-trained on ImageNet dataset outperformed as compared to other architectures, even with insufficient training data, achieving a maximum accuracy of 93.13 % for 1-pixel distance, 98.87 % for 5-pixel distance and 99.29 % for 10-pixel distance on test data, where the performance metrics, such as Percentage of Correct Keypoint (PCK), Mean-Square Error and Huber loss are utilized. As a result, this baseline proves to be a valuable tool for detecting structural deformations in jelly rolls. Moreover, a GUI-based executable application is developed using both approaches for raising the OK or NG flags for detecting structural deformations in each jelly roll. / Den ökande förekomsten av strukturella deformationer av elektroderna i så kallade jelly rolls har väckt kvalitetsproblem under batteritillverkning, och betonat behovet av att upptäcka dem automatiskt med avancerade tekniker. Denna avhandling syftar till att utforska och tillhandahålla två modeller baserade på traditionell datorseende (CV) och djupa neurala nätverk (DNN) tekniker med hjälp av bilder från datortomografisk skanning (CT) av jelly rolls för att säkerställa att produkten är av hög kvalitet. För båda metoderna detekteras elektrodtoppar som nyckelpunkter på anoder och katoder i prismatiska litiumbatteriers jelly rolls för att extrahera de geometriska egenskaperna för att identifiera om en viss jelly roll har några strukturella deformationer. För traditionella CV-metoder genomgår bilderna några förbehandlingssteg, extraktion av förgrund genom adaptiv tröskling och morfologiska operationer för att extrahera konturkanter, följt av användning av Harris hörndetektor för att upptäcka elektrodtoppar. Denna metod visar dock begränsningar i att detektera små eller negativa avståndsskillnader i deformerade bilder. Vidare föreslår denna studie en annan metod baserad på övervakad överföringsinlärning med förtränade djupinlärningsmodeller på annoterade data. Efter att ha utforskat olika arkitekturer presterade VGG19-modellen förtränad på ImageNet-datasetet bättre jämfört med andra arkitekturer, även med otillräcklig träningsdata, och uppnådde en maximal noggrannhet på 91,56% för 1-pixels avstånd, 97,49% för 5-pixels avstånd och 98,91% för 10-pixels avstånd på testdata, där prestationsmått som procentandel av korrekta nyckelpunkter (PCK), medelkvadratfel och Huber-förlust används. Som ett resultat visar sig denna grundlinje vara ett värdefullt verktyg för att upptäcka strukturella deformationer i jelly rolls. Dessutom har exekverbar applikation med grafiskt gränssnitt utvecklats med båda metoderna för att höja OK/NG-flaggorna för att upptäcka strukturella deformationer i varje jelly roll.

Real-time semantic scene understanding is a challenging computer vision task for autonomous vehicles. A limited amount of research has been done regarding forestry and off-road scene understanding, as the industry focuses on urban and on-road applications. Studies have shown that Deep Convolutional Neural Network architectures, using parameters trained on large datasets, can be re-trained and customized with smaller off-road datasets, using a method called transfer learning and yield state-of-the-art classification performance. This master’s thesis served as an extension of such existing off-road semantic segmentation studies. The thesis focused on detecting and visualizing the general trade-offs between classification performance, classification time, and the network’s number of available classes. The results showed that the classification performance declined for every class that got added to the network. Misclassification mainly occurred in the class boundary areas, which increased when more classes got added to the network. However, the number of classes did not affect the network’s classification time. Further, there was a nonlinear trade-off between classification time and classification performance. The classification performance improved with an increased number of network layers and a larger data type resolution. However, the layer depth increased the number of calculations and the larger data type resolution required a longer calculation time. The network’s classification performance increased by 0.5% when using a 16-bit data type resolution instead of an 8-bit resolution. But, its classification time considerably worsened as it segmented about 20 camera frames less per second with the larger data type. Also, tests showed that a 101-layered network slightly degraded in classification performance compared to a 50-layered network, which indicated the nonlinearity to the trade-off regarding classification time and classification performance. Moreover, the class constellations considerably impacted the network’s classification performance and continuity. It was essential that the class’s content and objects were visually similar and shared the same features. Mixing visually ambiguous objects into the same class could drop the inference performance by almost 30%. There are several directions for future work, including writing a new and customized source code for the ResNet50 network. A customized and pruned network could enhance both the application’s classification performance and classification speed. Further, procuring a task-specific forestry dataset and transferring weights pre-trained for autonomous navigation instead of generic object segmentation could lead to even better classification performance. / Se filen

This master thesis focuses on applying deep neural networks (DNNs) in image-based condition monitoring of air-jet spinning machines, specifically focusing on the spinning pressure parameter. The study aims to develop a sensor system to detect structural defects in yarns and assign them to specific machine conditions. The research explores using DNNs to analyze images of yarns generated at different spinning pressures within the spinning box to create a rich dataset for training deep learning models. The study also evaluates the effectiveness of the DNN-based approach in detecting and classifying structural defects in yarns and determining the corresponding machine conditions. The outcomes of this research could potentially help textile enterprises improve the quality and efficiency of their yarn manufacturing processes.

Les différences entre conditions d'apprentissage et conditions de test peuvent considérablement dégrader la qualité des transcriptions produites par un système de reconnaissance automatique de la parole (RAP). L'adaptation est un moyen efficace pour réduire l'inadéquation entre les modèles du système et les données liées à un locuteur ou un canal acoustique particulier. Il existe deux types dominants de modèles acoustiques utilisés en RAP : les modèles de mélanges gaussiens (GMM) et les réseaux de neurones profonds (DNN). L'approche par modèles de Markov cachés (HMM) combinés à des GMM (GMM-HMM) a été l'une des techniques les plus utilisées dans les systèmes de RAP pendant de nombreuses décennies. Plusieurs techniques d'adaptation ont été développées pour ce type de modèles. Les modèles acoustiques combinant HMM et DNN (DNN-HMM) ont récemment permis de grandes avancées et surpassé les modèles GMM-HMM pour diverses tâches de RAP, mais l'adaptation au locuteur reste très difficile pour les modèles DNN-HMM. L'objectif principal de cette thèse est de développer une méthode de transfert efficace des algorithmes d'adaptation des modèles GMM aux modèles DNN. Une nouvelle approche pour l'adaptation au locuteur des modèles acoustiques de type DNN est proposée et étudiée : elle s'appuie sur l'utilisation de fonctions dérivées de GMM comme entrée d'un DNN. La technique proposée fournit un cadre général pour le transfert des algorithmes d'adaptation développés pour les GMM à l'adaptation des DNN. Elle est étudiée pour différents systèmes de RAP à l'état de l'art et s'avère efficace par rapport à d'autres techniques d'adaptation au locuteur, ainsi que complémentaire. / Differences between training and testing conditions may significantly degrade recognition accuracy in automatic speech recognition (ASR) systems. Adaptation is an efficient way to reduce the mismatch between models and data from a particular speaker or channel. There are two dominant types of acoustic models (AMs) used in ASR: Gaussian mixture models (GMMs) and deep neural networks (DNNs). The GMM hidden Markov model (GMM-HMM) approach has been one of the most common technique in ASR systems for many decades. Speaker adaptation is very effective for these AMs and various adaptation techniques have been developed for them. On the other hand, DNN-HMM AMs have recently achieved big advances and outperformed GMM-HMM models for various ASR tasks. However, speaker adaptation is still very challenging for these AMs. Many adaptation algorithms that work well for GMMs systems cannot be easily applied to DNNs because of the different nature of these models. The main purpose of this thesis is to develop a method for efficient transfer of adaptation algorithms from the GMM framework to DNN models. A novel approach for speaker adaptation of DNN AMs is proposed and investigated. The idea of this approach is based on using so-called GMM-derived features as input to a DNN. The proposed technique provides a general framework for transferring adaptation algorithms, developed for GMMs, to DNN adaptation. It is explored for various state-of-the-art ASR systems and is shown to be effective in comparison with other speaker adaptation techniques and complementary to them.

La mémoire de traduction (MT) joue un rôle décisif lors de la traduction et constitue une base de données idéale pour la plupart des professionnels de la langue. Cependant, une MT est très sujète au bruit et, en outre, il n’y a pas de source spécifique. Des efforts importants ont été déployés pour nettoyer des MT, en particulier pour former un meilleur système de traduction automatique. Dans cette thèse, nous essayons également de nettoyer la MT mais avec un objectif plus large : maintenir sa qualité globale et la rendre suffisament robuste pour un usage interne dans les institutions. Nous proposons un processus en deux étapes : d’abord nettoyer une MT institutionnelle (presque propre), c’est-à-dire éliminer le bruit, puis détecter les textes traduits à partir de systèmes neuronaux de traduction. Pour la tâche d’élimination du bruit, nous proposons une architecture impliquant cinq approches basées sur l’heuristique, l’ingénierie fonctionnelle et l’apprentissage profond. Nous évaluons cette tâche à la fois par annotation manuelle et traduction automatique (TA). Nous signalons un gain notable de +1,08 score BLEU par rapport à un système de nettoyage état de l’art. Nous proposons également un outil Web qui annote automatiquement les traductions incorrectes, y compris mal alignées, pour les institutions afin de maintenir une MT sans erreur. Les modèles neuronaux profonds ont considérablement amélioré les systèmes MT, et ces systèmes traduisent une immense quantité de texte chaque jour. Le matériel traduit par de tels systèmes finissent par peuplet les MT, et le stockage de ces unités de traduction dans TM n’est pas idéal. Nous proposons un module de détection sous deux conditions: une tâche bilingue et une monolingue (pour ce dernier cas, le classificateur ne regarde que la traduction, pas la phrase originale). Nous rapportons une précision moyenne d’environ 85 % en domaine et 75 % hors domaine dans le cas bilingue et 81 % en domaine et 63 % hors domaine pour le cas monolingue en utilisant des classificateurs d’apprentissage profond. / Translation Memory (TM) plays a decisive role during translation and is the go-to database for most language professionals. However, they are highly prone to noise, and additionally, there is no one specific source. There have been many significant efforts in cleaning the TM, especially for training a better Machine Translation system. In this thesis, we also try to clean the TM but with a broader goal of maintaining its overall quality and making it robust for internal use in institutions. We propose a two-step process, first clean an almost clean TM, i.e. noise removal and then detect texts translated from neural machine translation systems. For the noise removal task, we propose an architecture involving five approaches based on heuristics, feature engineering, and deep-learning and evaluate this task by both manual annotation and Machine Translation (MT). We report a notable gain of +1.08 BLEU score over a state-of-the-art, off-the-shelf TM cleaning system. We also propose a web-based tool “OSTI: An Open-Source Translation-memory Instrument” that automatically annotates the incorrect translations (including misaligned) for the institutions to maintain an error-free TM. Deep neural models tremendously improved MT systems, and these systems are translating an immense amount of text every day. The automatically translated text finds a way to TM, and storing these translation units in TM is not ideal. We propose a detection module under two settings: a monolingual task, in which the classifier only looks at the translation; and a bilingual task, in which the source text is also taken into consideration. We report a mean accuracy of around 85% in-domain and 75% out-of-domain for bilingual and 81% in-domain and 63% out-of-domain from monolingual tasks using deep-learning classifiers.

This report explores whether machine learning methods such as regression and classification can be used with the goal of estimating the resolution time of trouble tickets in a telecommunications network. Historical trouble ticket data from Telenor were used to train different machine learning models. Three different machine learning classifiers were built: a support vector classifier, a logistic regression classifier and a deep neural network classifier. Three different machine learning regressors were also built: a support vector regressor, a gradient boosted trees regressor and a deep neural network regressor. The results from the different models were compared to determine what machine learning models were suitable for the problem. The most important features for estimating the trouble ticket resolution time were also investigated. Two different prediction scenarios were investigated in this report. The first scenario uses the information available at the time of ticket creation to make a prediction. The second scenario uses the information available after it has been decided whether a technician will be sent to the affected site or not. The conclusion of the work is that it is easier to make a better resolution time estimation in the second scenario compared to the first scenario. The differences in results between the different machine learning models were small. Future work can include more information and data about the underlying root cause of the trouble tickets, more weather data and power outage information in order to make better predictions. A standardised way of recording and logging ticket data is proposed to make a future trouble ticket time estimation easier and reduce the problem of missing data. / Den här rapporten undersöker om maskininlärningsmetoder som regression och klassificering kan användas för att uppskatta hur lång tid det tar att lösa en felanmälan i ett telenät. Data från tidigare felanmälningar användes för att träna olika maskininlärningsmodeller. Tre olika klassificerare byggdes: en support vector-klassificerare, en logistic regression-klassificerare och ett neuralt nätverk-klassificerare. Tre olika regressionsmodeller byggdes också: en support vector-regressor, en gradient boosted trees-regressor och ett neuralt nätverk-regressor. Resultaten från de olika modellerna jämfördes för att se vilken modell som är lämpligast för problemet. En undersökning om vilken information och vilka datavariabler som är viktigast för att uppskatta tiden det tar att lösa felanmälan utfördes också. Två olika scenarion för att uppskatta tiden har undersökts i rapporten. Det första scenariot använder informationen som är tillgänglig när en felanmälan skapas. Det andra scenariot använder informationen som finns tillgänglig efter det har bestämts om en tekniker ska skickas till den påverkade platsen. Slutsatsen av arbetet är att det är lättare att göra en bra tidsuppskattning i det andra scenariot jämfört med det första scenariot. Skillnaden i resultat mellan de olika maskininlärningsmodellerna var små. Framtida arbete inom ämnet kan använda information och data om de bakomliggande orsakerna till felanmälningarna, mer väderdata och information om elavbrott. En standardiserad metod för att samla in och logga data för varje felanmälan föreslås också för att göra framtida tidsuppskattningar bättre och undvika problemet med datapunkter som saknas.

Purpose This project aims to explore the classification method of kinase inhibitors with five-channel cell painting image data based on the deep learning model. Methods A ResNet50 transfer learning model was used as the starting point to build the deep neural network (DNN) model, where different DNN parameters were selected to make the deep learning model more suitable for the cell painting data. Two different adaptive layers (adaptive average pooling 3D and convolution 2D) were added separately before the ResNet50 transfer learning model to adapt the five-layer cell painting image to the neural network. In addition, the skimage.transform.resize function was used to compress the five-layer cell painting image. Results The proposed deep learning model demonstrates the effectiveness in all three classification experiments. The proposed model performs particularly well in classifying among control, EGFR, PIKK and CDK kinase inhibitors families. It achieves an F1-score of 0.7764 on all four targets and has a 93\% accuracy rate in the PIKK kinase inhibitors family. The adaptive average pooling 3D layer successfully adapts the five-layer images to the model, resulting in an improved effect. The training time of the model is significantly reduced to one-fortieth by compressing the image size. Conclusion The proposed model achieved convincing effectiveness in classifying families, which showed progress in building the deep learning model to classify kinase inhibitors on five-channel cell painting data. This study also proved the feasibility of directly inputting five-channel cell painting images to DNN. In addition, the speed of the model increased sharply by compressing the image size without an obvious loss of data information.

Sound source separation is a popular and active research area, especially with modern machine learning techniques. In this thesis, the focus is on single-channel separation of two speakers into individual streams, and specifically considering the case where two speakers are also accompanied by background noise. There are different methods to separate speakers and in this thesis three different methods are evaluated: the Conv-TasNet, the DPTNet, and the FaSNetTAC.  The methods were used to train models to perform the sound source separation. These models were evaluated and validated through three experiments. Firstly, previous results for the chosen separation methods were reproduced. Secondly, appropriate models applicable for NFC's datasets and applications were created, to fulfill the aim of this thesis. Lastly, all models were evaluated on an independent dataset, similar to datasets from NFC. The results were evaluated using the metrics SI-SNRi and SDRi. This thesis provides recommended models and methods suitable for NFC applications, especially concluding that the Conv-TasNet and the DPTNet are reasonable choices.

Recently, Deep Learning has revolutionized many fields, where one such area is Voice Activity Detection (VAD). This is of great interest to sectors of society concerned with detecting speech in sound signals. One such sector is the police, where criminal investigations regularly involve analysis of audio material. Convolutional Neural Networks (CNN) have recently become the state-of-the-art method of detecting speech in audio. But so far, understanding the impact of noise and sampling rates on such methods remains incomplete. Additionally, there are evaluation metrics from neighboring fields that remain unintegrated into VAD. We trained on four different sampling rates and found that changing the sampling rate could have dramatic effects on the results. As such, we recommend explicitly evaluating CNN-based VAD systems on pertinent sampling rates. Further, with increasing amounts of white Gaussian noise, we observed better performance by increasing the capacity of our Gated Recurrent Unit (GRU). Finally, we discuss how careful consideration is necessary when choosing a main evaluation metric, leading us to recommend Polyphonic Sound Detection Score (PSDS).

With the accessibility to generous spectrum and development of high gain antenna arrays, wireless communication in higher frequency bands providing multi-gigabit short range wireless access has become a reality. The directional antennas have proven to reduce losses due to interfering signals but are still exposed to blockage events. These events impede the overall user connectivity and throughput. A mobile blocker such as a moving vehicle amplifies the blockage effect. Modelling the blockage effects helps in understanding these events in depth and in maintaining the user connectivity. This thesis proposes the use of a four state channel model to describe blockage events in high-frequency communication. Two deep learning architectures are then designed and evaluated for two possible tasks, the prediction of the signal strength and the classification of the channel state. The evaluations based on simulated traces show high accuracy, and suggest that the proposed models have the potential to be extended for deployment in real systems. / Med tillgängligheten till generöst spektrum och utveckling av antennmatriser med hög förstärkning har trådlös kommunikation i högre frekvensband som ger multi-gigabit kortdistans trådlös åtkomst blivit verklighet. Riktningsantennerna har visat sig minska förluster på grund av störande signaler men är fortfarande utsatta för blockeringshändelser. Dessa händelser hindrar den övergripande användaranslutningen och genomströmningen. En mobil blockerare såsom ett fordon i rörelse förstärker blockeringseffekten. Modellering av blockeringseffekter hjälper till att förstå dessa händelser på djupet och bibehålla användaranslutningen. Denna avhandling föreslår användning av en fyrstatskanalmodell för att beskriva blockeringshändelser i högfrekvent kommunikation. Två djupinlärningsarkitekturer designas och utvärderas för två möjliga uppgifter, förutsägelsen av signalstyrkan och klassificeringen av kanalstatusen. Utvärderingarna baserade på simulerade spår visar hög noggrannhet och föreslår att de föreslagna modellerna har potential att utökas för distribution i verkliga system.