A wireless sensor node (WSN) system, capable of sensing animal motion and transmitting motion data wirelessly, is an effective and efficient way to monitor pigs' activity. However, the raw sensor data sampling and transmission consumes lots of power such that WSNs' battery have to be frequently charged or replaced. The proposed work solves this issue through WSN edge computing solution, in which a Random Forest Classifier (RFC) is trained and implemented into WSNs. The implementation of RFC on WSNs does not save power, but the RFC predicts animal behavior such that WSNs can adaptively adjust the data sampling frequency to reduce power consumption. In addition, WSNs can transmit less data by sending RFC predictions instead of raw sensor data to save power. The proposed RFC classifies common animal activities: eating, drinking, laying, standing, and walking with a F-1 score of 93%. The WSN power consumption is reduced by 25% with edge computing intelligence, compare to WSN power that samples and transmits raw sensor data periodically at 10 Hz. / Master of Science / A wireless sensor node (WSN) system that detects animal movement and wirelessly transmits this data is a valuable tool for monitoring pigs' activity. However, the process of sampling and transmitting raw sensor data consumes a significant amount of power, leading to frequent recharging or replacement of WSN batteries. To address this issue, our proposed solution integrates edge computing into WSNs, utilizing a Random Forest Classifier (RFC). The RFC is trained and deployed within the WSNs to predict animal behavior, allowing for adaptive adjustment of data sampling frequency to reduce power consumption. Additionally, by transmitting RFC predictions instead of raw sensor data, WSNs can conserve power by transmitting less data. Our RFC can accurately classify common animal activities, such as eating, drinking, laying, standing, and walking, achieving an F-1 score of 93%. With the integration of edge computing intelligence, WSN power consumption is reduced by 25% compared to traditional WSNs that periodically sample and transmit raw sensor data at 10 Hz.

Industrial Control Systems (ICSes), networked through conventional IT infrastructures, are vulnerable to attacks originating from network channels. Perimeter security techniques such as access control and firewalls have had limited success in mitigating such attacks due to the frequent updates required by standard computing platforms, third-party hardware and embedded process controllers. The high level of human-machine interaction also aids in circumventing perimeter defenses, making an ICS susceptible to attacks such as reprogramming of embedded controllers. The Stuxnet and Aurora attacks have demonstrated the vulnerabilities of ICS security and proved that these systems can be stealthily compromised. We present several run-time methods for preemptive intrusion detection in industrial control systems to enhance ICS security against reconfiguration and network attacks. A run-time prediction using a linear model of the physical plant and a neural-network based classifier trigger mechanism are proposed for preemptive detection of an attack. A standalone, safety preserving, optimal backup controller is implemented to ensure plant safety in case of an attack. The intrusion detection mechanism and the backup controller are instantiated in configurable hardware, making them invisible to operating software and ensuring their integrity in the presence of malicious software. Hardware implementation of our approach on an inverted pendulum system illustrates the performance of both techniques in the presence of reconfiguration and network attacks. / Master of Science

El aprendizaje autom�atico es un �area de investigaci�on que proporciona algoritmos y t�ecnicas que son capaces de aprender autom�aticamente a partir de experiencias pasadas. Estas t�ecnicas son esenciales en el �area de descubrimiento de conocimiento de bases de datos (KDD), cuya fase principal es t�ÿpicamente conocida como miner�ÿa de datos. El proceso de KDD se puede ver como el aprendizaje de un modelo a partir de datos anteriores (generaci�on del modelo) y la aplicaci�on de este modelo a nuevos datos (utilizaci�on del modelo). La fase de utilizaci�on del modelo es muy importante, porque los usuarios y, muy especialmente, las organizaciones toman las decisiones dependiendo del resultado de los modelos. Por lo general, cada modelo se aprende de forma independiente, intentando obtener el mejor resultado (local). Sin embargo, cuando varios modelos se usan conjuntamente, algunos de ellos pueden depender los unos de los otros (por ejemplo, las salidas de un modelo pueden ser las entradas de otro) y aparecen restricciones. En este escenario, la mejor decisi�on local para cada problema tratado individualmente podr�ÿa no dar el mejor resultado global, o el resultado obtenido podr�ÿa no ser v�alido si no cumple las restricciones del problema. El �area de administraci�on de la relaci�on con los clientes (CRM) ha dado origen a problemas reales donde la miner�ÿa de datos y la optimizaci�on (global) deben ser usadas conjuntamente. Por ejemplo, los problemas de prescripci�on de productos tratan de distinguir u ordenar los productos que ser�an ofrecidos a cada cliente (o sim�etricamente, elegir los clientes a los que se les deber�ÿa de ofrecer los productos). Estas �areas (KDD, CRM) carecen de herramientas para tener una visi�on m�as completa de los problemas y una mejor integraci�on de los modelos de acuerdo a sus interdependencias y las restricciones globales y locales. La aplicaci�on cl�asica de miner�ÿa de datos a problemas de prescripci�on de productos, por lo general, ha / Bella Sanjuán, A. (2012). Model Integration in Data Mining: From Local to Global Decisions [Tesis doctoral]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/16964

A distributed reinforcement-learning system is designed and implemented on a mobile robot for the study of complex task decomposition in real robot learning environments. The Distributed Q-learning Classifier System (DQLCS) is evolved from the standard Learning Classifier System (LCS) proposed by J.H. Holland. Two of the limitations of the standard LCS are its monolithic nature and its complex apportionment of credit scheme, the bucket brigade algorithm (BBA). The DQLCS addresses both of these problems as well as the inherent difficulties faced by learning systems operating in real environments. We introduce Q-learning as the apportionment of credit component of the DQLCS, and we develop a distributed learning architecture to facilitate complex task decomposition. Based upon dynamic programming, the Q-learning update equation is derived and its advantages over the complex BBA are discussed. The distributed architecture is implemented to provide for faster learning by allowing the system to effectively decrease the size of the problem space it must explore. Holistic and monolithic shaping approaches are used to distribute reward among the learning modules of the DQLCS in a variety of real robot learning experiments. The results of these experiments support the DQLCS as a useful reinforcement learning paradigm and suggest future areas of study in distributed learning systems. / Master of Science

Yes / Multimodal biometric systems have been widely applied in many real-world applications due to its ability to deal with a number of significant limitations of unimodal biometric systems, including sensitivity to noise, population coverage, intra-class variability, non-universality, and vulnerability to spoofing. In this paper, an efficient and real-time multimodal biometric system is proposed based on building deep learning representations for images of both the right and left irises of a person, and fusing the results obtained using a ranking-level fusion method. The trained deep learning system proposed is called IrisConvNet whose architecture is based on a combination of Convolutional Neural Network (CNN) and Softmax classifier to extract discriminative features from the input image without any domain knowledge where the input image represents the localized iris region and then classify it into one of N classes. In this work, a discriminative CNN training scheme based on a combination of back-propagation algorithm and mini-batch AdaGrad optimization method is proposed for weights updating and learning rate adaptation, respectively. In addition, other training strategies (e.g., dropout method, data augmentation) are also proposed in order to evaluate different CNN architectures. The performance of the proposed system is tested on three public datasets collected under different conditions: SDUMLA-HMT, CASIA-Iris- V3 Interval and IITD iris databases. The results obtained from the proposed system outperform other state-of-the-art of approaches (e.g., Wavelet transform, Scattering transform, Local Binary Pattern and PCA) by achieving a Rank-1 identification rate of 100% on all the employed databases and a recognition time less than one second per person.

The purpose of this project was to expand the applications of time series prediction and action recognition for use with motion capture data and football plays. Both the motion capture data and football play trajectories were represented in the form of multidimensional time series. Each point of interest on the human body or football players path, was represented in two or three time series, one for each dimension of motion recorded in the data. By formulating a phase space reconstruction of the data, the remainder of each input time series was predicted utilizing kernel regression. This process was applied to the first portion of a play. Utilizing features from the theory of chaotic systems and specialized geometric features, the specific type of motion for the motion capture data or the type of play for the football data was identified by using the features with various classifiers. The chaotic features used included the maximum Lyapunov exponent, the correlation integral, and the correlation dimension. The variance and mean were also utilized in conjunction with the chaotic features. The geometric features used were the minimum, maximum, mean, and median of the x, y, and z axis time series, as well as various angles and measures of the trajectory as a whole. The accuracy of the features and classifiers was compared and combinations of features were analyzed. The novelty of this work lies in the method to recognize types of actions from a prediction made from only a short, initial portion of an action utilizing various sets of features and classifiers.

Wearable activity classifiers, so far, have been able to perform well with simple activities, strictly-scripted activities, and application-specific activities. In addition, current classification systems suffer from using impractical tight-fitting sensor networks, or only use one loose-fitting sensor node that cannot capture much movement information (e.g., smartphone sensors and wrist-worn sensors). These classifiers either do not address the bigger picture of making activity recognition more practical and being able to recognize more complex and naturalistic activities, or try to address this issue but still perform poorly on many fronts. This dissertation works toward having practical, in-the-wild, and reusable wearable activity classifiers by taking several steps that include the four following main contributions. The dissertation starts by quantifying users' needs and expectations from wearable activity classifiers to set a framework for designing ideal wearable activity classifiers. Data collected from user studies and interviews is gathered and analyzed, then several conclusions are made to set a framework of essential characteristics that ideal wearable activity classification systems should have. Afterwards, this dissertation introduces a group of datasets that can be used to benchmark different types of activity classifiers and can accommodate for a variety of goals. These datasets help comparing different algorithms in activity classification to assess their performance under various circumstances and with different types of activities. The third main contribution consists of developing a technique that can classify complex activities with wide variations. Testing this technique shows that it is able to accurately classify eight complex daily-life activities with wide variations at an accuracy rate of 93.33%, significantly outperforming the state-of-the-art. This technique is a step forward toward classifying real-life natural activities performed in an environment that allows for wide variations within the activity. Finally, this dissertation introduces a method that can be used on top of any activity classifier that allows access to its matching scores in order to improve its classification accuracy. Testing this method shows that it improves classification results by 11.86% and outperforms the state-of-the-art, therefore taking a step forward toward having reusable activity classification techniques that can be used across users, sensor domains, garments, and applications. / Ph. D. / Wearable activity classifiers are wearable systems that can recognize human activities. These systems are needed in many applications. Nowadays, they are mainly used for fitness purposes – e.g., Fitbits and Apple Watches – and in gaming consoles – e.g., Microsoft Kinect. However, these systems are still far from being ideal. They still miss many characteristics that would make them practical and usable for different purposes, such as in medical applications, industrial applications, and other types of applications where recognizing human activities can be useful. This dissertation works toward having practical wearable activity classifiers that can be reused for different purposes in real-life scenarios. Four contributions are introduced in this dissertation. The dissertation starts by quantifying users’ needs and expectations from wearable activity classifiers and sets a framework for designing them. Afterward, this dissertation introduces a group of datasets that can be used to benchmark and compare different types of activity classifiers. The third main contribution consists of a technique that enables activity classifiers to recognize complex activities having a wide range of variations within each activity. Finally, this dissertation introduces a method that can be used to improve the recognition accuracy of activity classifiers.

Thesis (MSc)--Stellenbosch University, 2012. / ENGLISH ABSTRACT: This research considered the automated remote sensing-based classification of wetland extent within the Nuwejaars and Heuningnes River systems on the Agulhas Plain. The classification process was based on meaningful image objects created through image segmentation rather than on single pixels. An expert system classifier was compared to a nearest-neighbour supervised classifier, and one multispectral (SPOT 5) image (dry season) and two C-band, VV-polarisation synthetic aperture radar (SAR: ERS-2) images (dry and wet season) were used separately and in combination. Classifications were performed within two subset areas. Final classes identified were Permanent waterbody, Other wetland and Non-wetland. Statistical accuracy assessment was performed. Validation data was derived from a combination of high-resolution aerial photographs, the SPOT 5 image, high-resolution imagery on Google Earth and observations during a field visit. Wetland extent was defined as the total extent of wetland-specific vegetation, unvegetated seasonal pans and waterbodies. More detailed classes were originally envisaged, but available validation data was not considered adequate for assessing their accuracy with any confidence. The supervised classifier was found to be more accurate overall than the developed expert system. The difference between the two was however not always significant. The two SAR images alone did not contain sufficient information for the accurate classification of Agulhas wetlands’ extent, with recorded overall accuracies not exceeding 65% regardless of the classifier used. The SPOT image alone achieved accuracies higher than 80%; this was considered a good result. In comparison, combining the SAR and SPOT data did not improve the classification accuracy. The potential of the expert system to be applied with little modification to images acquired over other areas or over the same area in other years should be further investigated. However, several reservations are noted in this regard. Future research could potentially improve the results obtained from supervised classification by augmenting it with expert system rules to identify more complicated classes. KEYWORDS ERS-2, SPOT 5, SAR, wetlands, expert system classifier, nearest-neighbour supervised classifier / AFRIKAANSE OPSOMMING: Hierdie navorsing het die geoutomatiseerde afstandswaarneminggebaseerde klassifikasie van vleilandomvang binne die Nuwejaars- en Heuningnesrivier stelsels op die Agulhasvlakte ondersoek. Die klassifikasieproses was gebaseer op betekenisvolle beeldobjekte geskep deur middel van beeldsegmentasie eerder as op enkele beeldelemente. ‘n Deskundige stelsel klassifiseerder is vergelyk met ‘n naaste-naburige gerigte klassifiseerder. Een multispektrale (SPOT 5) beeld vir die droë seisoen, sowel as twee C-band, VV-polarisasie sintetiese diafragma radar (SAR, ERS2) beelde (vir die droë en nat seisoene) is afsonderlik en in kombinasie gebruik. Klassifikasies is uitgevoer binne twee sub-areas in die beelde. Finale klasse wat geïdentifiseer is was Permanente waterliggaam, Ander vleiland en Nie-vleiland. Statistiese akkuraatheidsassessering is uitgevoer. Verwysingsdata is geskep vanuit ‘n kombinasie van hoë- resolusie lugfoto’s, die SPOT 5 beeld, hoë-resolusie beelde op Google Earth en waarnemings tydens ‘n besoek aan die studiegebied. Vleiland omvang is gedefinieer as die totale omvang van vleiland-spesifieke plantegroei, onbegroeide seisoenale panne en waterliggame. Die gerigte klassifiseerder blyk om oor die algemeen meer akkuraat as die ontwikkelde deskundige stelsel te wees. Die verskil was egter nie altyd beduidend nie. Die twee SAR beelde alleen het nie genoegsame inligting bevat vir die akkurate klassifikasie van Agulhas-vleilande se omvang nie, met behaalde algehele akkuraatheidsvlakke wat nie 65% oorskry het nie, ongeag van die klassifiseerder. Die SPOT-beeld alleenlik het algehele akkuraathede van meer as 80% behaal; wat as ‘n goeie resultaat beskou kan word. In vergelyking hiermee kon die kombinering van SAR- en SPOT-data nie ‘n verbetering teweeg bring nie. Die potensiaal van die deskundige stelsel om met min aanpassing op beelde van ander gebiede of van dieselfde gebied in ander jare toegepas te word, verg verdere ondersoek. Verskeie voorbehoude word egter in hierdie verband gemeld. Toekomstige navorsing kan potensieel die resultate van gerigte klassifikasie verbeter deur dit aan te vul met deskundige stelsel reëls vir die klassifikasie van meer komplekse klasse. TREFWOORDE ERS-2, SPOT 5, SAR, vleilande, deskundige stelsel klassifiseerder, naaste-naburige gerigte klassifiseerder.

Many real-world problems like hand-written digit recognition or semantic scene classification are treated as multiclass or multi-label classification prob-lems. Solutions to these problems using support vector machines (SVMs) are well studied in literature. In this work, we focus on building sparse max-margin classifiers for multiclass and multi-label classification. Sparse representation of the resulting classifier is important both from efficient training and fast inference viewpoints. This is true especially when the training and test set sizes are large.Very few of the existing multiclass and multi-label classification algorithms have given importance to controlling the sparsity of the designed classifiers directly. Further, these algorithms were not found to be scalable. Motivated by this, we propose new formulations for sparse multiclass and multi-label classifier design and also give efficient algorithms to solve them. The formulation for sparse multi-label classification also incorporates the prior knowledge of label correlations. In both the cases, the classification model is designed using a common set of basis vectors across all the classes. These basis vectors are greedily added to an initially empty model, to approximate the target function. The sparsity of the classifier can be controlled by a user defined parameter, dmax which indicates the max-imum number of common basis vectors. The computational complexity of these algorithms for multiclass and multi-label classifier designisO(lk2d2 max), Where l is the number of training set examples and k is the number of classes. The inference time for the proposed multiclass and multi-label classifiers is O(kdmax). Numerical experiments on various real-world benchmark datasets demonstrate that the proposed algorithms result in sparse classifiers that require lesser number of basis vectors than required by state-of-the-art algorithms, to attain the same generalization performance. Very small value of dmax results in significant reduction in inference time. Thus, the proposed algorithms provide useful alternatives to the existing algorithms for sparse multiclass and multi-label classifier design.

This thesis describes the design and development of a system for detection and recognition of license plates. The work is divided into three basic parts: licence plates detection, finding of character positions and optical character recognition. To fullfill the goal of this work, a new dataset was taken. It contains 2814 license plates used for training classifiers and 2620 plates to evaluate the success rate of the system. Cascade Classifier was used to train detector of licence plates, which has success rate up to 97.8 %. After that, pozitions of individual characters were searched in detected pozitions of licence plates. If there was no character found, detected pozition was not the licence plate. Success rate of licence plates detection with all the characters found is up to 88.5 %. Character recognition is performed by SVM classifier. The system detects successfully with no errors up to 97.7 % of all licence plates.