Narrow your search

Library

ULiège (3)


Resource type

dissertation (3)


Language

English (2)

French (1)


Year
From To Submit

2022 (2)

2020 (1)

Listing 1 - 3 of 3
Sort by

Dissertation
Long Short-Term Memory neural networks and Support Vector Data Description for anomaly detection
Authors: --- --- --- ---
Year: 2020 Publisher: Liège Université de Liège (ULiège)

Loading...
Export citation

Choose an application

Bookmark

Abstract

Anomaly detection refers to the problem of finding rare patterns in data which raise suspicions because they do not comply with an expected behavior. We can consider different kinds of applications like intrusion detection, image processing, system health monitoring and sensor networks. For example, an anomalous pattern coming from sensors on a machine could mean that the machine is ready to break. 
Most of the current studies on anomaly detection do not consider recent/past events to detect possible new incoming outliers. The use of Long Short-Term Memory (LSTM) networks is then proposed to deal with time dependent data related with anomaly detection problems.
The goal of Support Vector Data Description (SVDD) is to describe a realistic domain for the data, excluding superfluous space. The resulting boundary can then be used to detect outliers.

In this master thesis, we consider a LSTM-based prediction model for sensor readings coming from a pulp and paper manufacturing machine. Anomalies will then result from too large prediction errors. We compare the SVDD and a discrimination rule based on the assumption of normality for the errors. In the final chapter, we show that for a real world applications the Gaussian distribution for the errors cannot hold and that the need of a non-parametric data descriptions using kernels is real.


Dissertation
Master thesis : A climate forecasting model to assist Belgian wine-growers against bud freezing
Authors: --- --- --- ---
Year: 2022 Publisher: Liège Université de Liège (ULiège)

Loading...
Export citation

Choose an application

Bookmark

Abstract

In Belgium, the global warming offers new opportunities for the wine culture. In the past fifteen years, the area dedicated for vineyards became eight times bigger, rising from 72 hectares in 2006 to 587 hectares in 2020.
Nevertheless, the Belgian climate is not always indulgent with vineyards throughout the year. In the spring particularly, when the bud has already grew and became more sensitive to frost, a few cold nights may arise and lead to the destruction of the grapes.

Hopefully, few methods exist to fight bud freezing. From one vineyard to another, different techniques will be used depending on, amongst other, the size of the plantation, the intensity of the frost or the financial means. 
In 2017, half of the plantations have a size below one hectare. For this kind of small area, the construction of expensive infrastructures is not viable and more adapted methods must be chosen. The most suitable solution, and often the only one conceivable in small vineyards, consists in lightning large candles all over the place to increase the temperature around the grapes.
Although this solution has the advantages to be simple and relatively cheap, it also has two drawbacks. The first one, common to other more expensive methods, is the temperature monitoring, which consist in determining when to take action. The second one, inherent to this particular method, is the need of a workforce available in a hurry to light all the candles in the middle of the night.

In our work, we propose to assist Belgian wine makers by means of acrfull{ml} techniques in the two aspects aforementioned. We achieved to accurately forecast the temperatures around the grapes up to 36 hours ahead with an error below 1 degree Celsius. Additionally, we also provide a live supervising model capable of detecting negative temperatures with an error below 0.5 degree. In doing so, we avoid the wine growers to unnecessarily wake up during night to monitor the temperatures and we also allow them to organize a workforce sufficiently early enough.


Dissertation
Automatisation de la reconnaissance d'espèces animales dans des vidéos de pièges photographiques installés dans les forêts tropicales en Afrique centrale, grâce à l'apprentissage profond
Authors: --- --- --- --- --- et al.
Year: 2022 Publisher: Liège Université de Liège (ULiège)

Loading...
Export citation

Choose an application

Bookmark

Abstract

The world today is threatened by a dramatic biodiversity crisis. It is therefore becoming essential to monitor the animal and plant populations that inhabit the earth's ecosystems. In this sense, camera traps are cameras that capture images or videos when they detect movement. These cameras are increasingly used in the scientific world and could become an essential tool in wildlife monitoring systems. They have the advantage of being very low-intrusive and of being able to be installed in remote and difficult-to-access places. The main weakness of this technology is that it generates a huge amount of data. The analysis of this data by humans is therefore very time-consuming and tedious. A solution to this problem could be found in the use of deep learning. This allows deep neural networks to be trained to automate a task usually performed by humans. Some deep learning approaches have achieved better results in solving complex problems. The main objective of this work is therefore to use deep learning to automate the recognition of animal species in videos of camera traps installed in the tropical rainforests of Central Africa. To this end, three datasets were created and 22 classes were defined. Different model architectures were then tested. These are composed of convolutional neural networks (two-dimensional and three-dimensional ResNet) and recurrent neural networks (convolutional long short-term memory (ConvLSTM) and long short-term memory (LSTM)). This work also discusses the comparison of different tools that have been developed to automatically classify camera traps data. The best trained models achieved, on a test dataset, an overall accuracy of 67,93 % for multispecies classification and 84,89 % for binary classification (animal/background). These models performed better than the other tested tools for the multispecies classification but not for the binary classification. Finally, the models developed could be used under certain conditions to assist in the analysis of camera traps data. The results obtained are promising. Le monde actuel est menacé par une crise de la biodiversité dramatique. Il devient donc primordial de surveiller les populations animales et végétales qui habitent les écosystèmes de la Terre. Dans ce sens, les pièges photographiques sont des caméras qui capturent des images ou des vidéos lorsqu'elles détectent un mouvement. Ces caméras sont de plus en plus utilisées dans le monde scientifique et pourraient devenir un outil essentiel dans des systèmes de surveillance de la faune et de la flore. Elles possèdent notamment l'avantage d'être très peu intrusives et de pouvoir être installées dans des endroits reculés et difficilement accessibles. Le point faible de cette technologie est qu'elle génère une quantité très importante de données. L'analyse de ces dernières par l'homme est donc très chronophage et fastidieuse. Une solution pourrait être apportée à ce problème grâce à l'utilisation de l'apprentissage profond. Celui-ci permet d'entraîner des réseaux de neurones profonds afin d'automatiser une tâche habituellement réalisée par l'homme. Certaines approches d'apprentissage profond ont permis d'atteindre de meilleurs résultats lors la résolution de problèmes complexes. L'objectif principal de ce travail est donc d'utiliser l'apprentissage profond afin d'automatiser la reconnaissance d'espèces animales dans des vidéos de pièges photographiques installés dans les forêts tropicales d'Afrique centrale. Pour ce faire, trois jeux de données ont été constitués et 22 classes ont été définies. Différentes architectures de modèles ont ensuite été testées. Ces dernières sont composées de réseaux de neurones convolutifs (ResNet à deux dimensions et à trois dimensions) et de réseaux de neurones récurrents (mémoire convolutive à long court terme (ConvLSTM) et mémoire à long court terme (LSTM)). Ce travail aborde également la comparaison de différents outils qui ont été développés afin de classifier automatiquement des données de pièges photographiques. Les meilleurs modèles entraînés ont atteint, sur un jeu de données de test, une exactitude globale de 67,93 % pour la classification multi-espèces et de 84,89 % pour la classification binaire (animal / arrière-plan). Ces modèles ont mieux performés que les autres outils testés, pour la classification multi-espèces mais pas pour la classification binaire. Enfin, les modèles développés pourraient être utilisés sous certaines conditions dans le but d'aider à l'analyse des données de pièges photographiques. Les résultats obtenus sont prometteurs.

Listing 1 - 3 of 3
Sort by