Property:Abstract
From ISLAB/CAISR
Jump to navigationJump to searchThis is a property of type Text.
S
<p>We present a neural network based approach for identifying salient features for classification in neural network committees. Our approach involves neural network training with an augmented cross-entropy error function. The augmented error function forces the neural network to keep low derivatives of the transfer functions of neurons of the network when learning a classification task. Such an approach reduces output sensitivity to the input changes. Feature selection is based on the reaction of the cross-validation data set classification error due to the removal of the individual features. We compared the approach with two other neural network based feature selection methods. The algorithm developed outperformed the methods by achieving a higher classification accuracy on three real world problems tested. ©2002 IEEE</p> +
<p>This paper is concerned with a two stage procedure for designing a sequential SVM committee and selecting features for the committee from multiple feature sets. It is assumed that features of one type comprise one feature set. Selection of both features and hyper-parameters of SVM classifiers comprising the committee is integrated into one learning process based on genetic search. The designing process focuses on feature selection for pair-wise classification implemented by the SVM. In the first stage, a series of pair-wise SVM are designed starting from the original feature sets as well as from sets created by simple random selection from the original ones. Outputs of the SVM are then converted into probabilities and used as inputs to the second stage SVM. When testing the technique in a three-class classification problem of voice data, a statistically significant improvement in classification accuracy was obtained if compared to parallel committees. The number of feature types and features selected for the pair-wise classification are class specific.</p> +
<p>We present a neural network based approach for identifying salient features for classification in feed-forward neural networks. Our approach involves neural network training with an augmented cross-entropy error function. The augmented error function forces the neural network to keep low derivatives of the transfer functions of neurons when learning a classification task. Such an approach reduces output sensitivity to the input changes. Feature selection is based on the reaction of the cross-validation data set classification error due to the removal of the individual features. We compared the approach with five other feature selection methods, each of which banks on different concept. The algorithm developed outperformed the other methods by achieving a higher classification accuracy on all the problems tested.</p> +
<p>To improve recognition results, decisions of multiple neural networks can be aggregated into a committee decision. In contrast to the ordinary approach of utilizing all neural networks available to make a committee decision, we propose creating adaptive committees, which are specific for each input data point. A prediction network is used to identify classification neural networks to be fused for making a committee decision about a given input data point. The jth output value of the prediction network expresses the expectation level that the jth classification neural network will make a correct decision about the class label of a given input data point. The proposed technique is tested in three aggregation schemes, namely majority vote, averaging, and aggregation by the median rule and compared with the ordinary neural networks fusion approach. The effectiveness of the approach is demonstrated on two artificial and three real data sets.</p> +
<p>To improve recognition results, decisions of multiple neural networks can be aggregated into a committee decision. In contrast to the ordinary approach of utilizing all neural networks available to make a committee decision, we propose creating adaptive committees, which are specific for each input data point. A prediction network is used to identify classification neural networks to be fused for making a committee decision about a given input data point. The jth output value of the prediction network expresses the expectation level that the jth classification neural network will make a correct decision about the class label of a given input data point. The effectiveness of the approach is demonstrated on two artificial and three real data sets.</p> +
<p>Aggregating outputs of multiple classifiers into a committee decision is one of the most important techniques for improving classification accuracy. The issue of selecting an optimal subset of relevant features plays also an important role in successful design of a pattern recognition system. In this paper, we present a neural network based approach for identifying salient features for classification in neural network committees. Feature selection is based on two criteria, namely the reaction of the cross-validation data set classification error due to the removal of the individual features and the diversity of neural networks comprising the committee. The algorithm developed removed a large number of features from the original data sets without reducing the classification accuracy of the committees. The accuracy of the committees utilizing the reduced feature sets was higher than those exploiting all the original features.</p> +
<p>We present a neural network based approach for identifying salient features for classification in neural network committees. Our approach involves neural network training with an augmented cross-entropy error function. The augmented error function forces the neural network to keep low derivatives of the transfer functions of neurons of the network when learning a classification task. Feature selection is based on two criteria, namely the reaction of the cross-validation data set classification error due to the removal of the individual features and the diversity of neural networks comprising the committee. The algorithm developed removed a large number of features from the original data sets without reducing the classification accuracy of the committees. By contrast, the accuracy of the committees utilizing the reduced feature sets was higher than those exploiting all the original features. © Springer-Verlag Berlin Heidelberg 2003.</p> +
<p>The aim of the variable selection is threefold: to reduce model complexity, to promote diversity of committee networks, and to find a trade-off between the accuracy and diversity of the networks. To achieve the goal, the steps of neural network training, aggregation, and elimination of irrelevant input variables are integrated based on the negative correlation learning (1) error function. Experimental tests performed on three real world problems have shown that statistically significant improvements in classification performance can be achieved from neural network committees trained according to the technique proposed.</p> +
<p>The ability to diagnose deviations and predict faults effectively is an important task in various industrial domains for minimizing costs and productivity loss and also conserving environmental resources. However, the majority of the efforts for diagnostics are still carried out by human experts in a time-consuming and expensive manner. Automated data-driven solutions are needed for continuous monitoring of complex systems over time. On the other hand, domain expertise plays a significant role in developing, evaluating, and improving diagnostics and monitoring functions. Therefore, automatically derived solutions must be able to interact with domain experts by taking advantage of available a priori knowledge and by incorporating their feedback into the learning process.</p><p>This thesis and appended papers tackle the problem of generating a real-world self-monitoring system for continuous monitoring of machines and operations by developing algorithms that can learn data streams and their relations over time and detect anomalies using joint-human machine learning. Throughout this thesis, we have described a number of different approaches, each designed for the needs of a self-monitoring system, and have composed these methods into a coherent framework. More specifically, we presented a two-layer meta-framework, in which the first layer was concerned with learning appropriate data representations and detectinganomalies in an unsupervised fashion, and the second layer aimed at interactively exploiting available expert knowledge in a joint human-machine learning fashion.</p><p>Furthermore, district heating has been the focus of this thesis as the application domain with the goal of automatically detecting faults and anomalies by comparing heat demands among different groups of customers. We applied and enriched different methods on this domain, which then contributed to the development and improvement of the meta-framework. The contributions that result from the studies included in this work can be summarized into four categories: (1) exploring different data representations that are suitable for the self-monitoring task based on data characteristics and domain knowledge, (2) discovering patterns and groups in data that describe normal behavior of the monitored system/systems, (3) implementing methods to successfully discriminate anomalies from the normal behavior, and (4) incorporating domain knowledge and expert feedback into self-monitoring.</p>
<p>An approach for intelligent monitoring of mobile cyberphysical systems is described, based on consensus among distributed self-organised agents. Its usefulness is experimentally demonstrated over a long-time case study in an example domain: a fleet of city buses. The proposed solution combines several techniques, allowing for life-long learning under computational and communication constraints. The presented work is a step towards autonomous knowledge discovery in a domain where data volumes are increasing, the complexity of systems is growing, and dedicating human experts to build fault detection and diagnostic models for all possible faults is not economically viable. The embedded, self-organised agents operate on-board the cyberphysical systems, modelling their states and communicating them wirelessly to a back-office application. Those models are subsequently compared against each other to find systems which deviate from the consensus. In this way the group (e.g. a fleet of vehicles) is used to provide a standard, or to describe normal behaviour, together with its expected variability under particular operating conditions. The intention is to detect faults without the need for human experts to anticipate them beforehand. This can be used to build up a knowledge base that accumulates over the life-time of the systems. The approach is demonstrated using data collected during regular operation of a city bus fleet over the period of almost four years. © 2017 The Author(s)</p> +
<p>Operators of fleets of vehicles desire the best possible availability and usage of their vehicles. This means the preference is that maintenance of a vehicle is scheduled with as long intervals as possible. However, it is then important to be able to detect if a component in a specific vehicle is not functioning properly earlier than expected (due to e.g. manufacturing variations). This paper proposes a telematic based fault detection scheme for enabling fault detection for diagnostics by using a population of vehicles. The basic idea is that it is possible to create low-dimensional representations of a sub-system or component in a vehicle, where the representation (or model parameters) of a vehicle can be monitored for changes compared to the model parameters observed in a fleet of vehicles. If a model in a vehicle is found to deviate compared to a group of models from a fleet of vehicles, then the vehicle is judged to need diagnostics for that component (assuming the deviation in the model cannot be attributed to e.g. a different driver behavior). The representation should be low-dimensional so it is possible to have it transferred over a limited wireless communication channel to a communications center where the comparison is made. The algorithm is shown to be able to detect leakage on simulated data from a cooling system, work is currently in progress for detecting other types of faults.</p> +
<p>A telematic based system for enabling automatic fault detection of a population of vehicles is proposed. To avoid sending huge amounts of data over the telematics gateway, the idea is to use low-dimensional representations of sensor values in sub-systems in a vehicle. These low-dimensional representations are then compared between similar systems in a fleet. If a representation in a vehicle is found to deviate from the group of systems in the fleet, then the vehicle is labeled for diagnostics for that subsystem. The idea is demonstrated on the engine coolant system and it is shown how this self-organizing approach can detect varying levels of clogged radiator.</p> +
<p>This thesis and appended papers present the process of tacking the problem of environment modeling for autonomous agent. More specifically, the focus of the work has been semantic mapping of warehouses. A semantic map for such purpose is expected to be layout-like and support semantics of both open spaces and infrastructure of the environment. The representation of the semantic map is required to be understandable by all involved agents (humans, AGVs and WMS.) And the process of semantic mapping is desired to lean toward full-autonomy, with minimum input requirement from human user. To that end, we studied the problem of semantic annotation over two kinds of spatial map from different modalities. We identified properties, structure, and challenges of the problem. And we have developed representations and accompanied methods, while meeting the set criteria. The overall objective of the work is “to develop and construct a layer of abstraction (models and/or decomposition) for structuring and facilitate access to salient information in the sensory data. This layer of abstraction connects high level concepts to low-level sensory pattern.” Relying on modeling and decomposition of sensory data, we present our work on abstract representation for two modalities (laser scanner and camera) in three appended papers. Feasibility and the performance of the proposed methods are evaluated over data from real warehouse. The thesis conclude with summarizing the presented technical details, and drawing the outline for future work.</p> +
<p>Place categorisation; i.e., learning to group perception data into categories based on appearance; typically uses supervised learning and either visual or 2D range data. This paper shows place categorisation from 3D data without any training phase. We show that, by leveraging the NDT histogram descriptor to compactly encode 3D point cloud appearance, in combination with standard clustering techniques, it is possible to classify public indoor data sets with accuracy comparable to, and sometimes better than, previous supervised training methods. We also demonstrate the effectiveness of this approach to outdoor data, with an added benefit of being able to hierarchically categorise places into sub-categories based on a user-selected threshold. This technique relieves users of providing relevant training data, and only requires them to adjust the sensitivity to the number of place categories, and provide a semantic label to each category after the process is completed. © 2017 IEEE.</p> +
Semi-Supervised Semantic Labeling of Adaptive Cell Decomposition Maps in Well-Structured Environments +
<p>We present a semi-supervised approach for semantic mapping, by introducing human knowledge after unsupervised place categorization has been combined with an adaptive cell decomposition of an occupancy map. Place categorization is based on clustering features extracted from raycasting in the occupancy map. The cell decomposition is provided by work we published previously, which is effective for the maps that could be abstracted by straight lines. Compared to related methods, our approach obviates the need for a low-level link between human knowledge and the perception and mapping sub-system, or the onerous preparation of training data for supervised learning. Application scenarios include intelligent warehouse robots which need a heightened awareness in order to operate with a higher degree of autonomy and flexibility, and integrate more fully with inventory management systems. The approach is shown to be robust and flexible with respect to different types of environments and sensor setups. © 2015 IEEE</p> +
Sensor Based Adaptive Metric-Topological Cell Decomposition Method for Semantic Annotation of Structured Environments +
<p>A fundamental ingredient for semantic labeling is a reliable method for determining and representing the relevant spatial features of an environment. We address this challenge for planar metric-topological maps based on occupancy grids. Our method detects arbitrary dominant orientations in the presence of significant clutter, fits corresponding line features with tunable resolution, and extracts topological information by polygonal cell decomposition. Real-world case studies taken from the target application domain (autonomous forklift trucks in warehouses) demonstrate the performance and robustness of our method, while results from a preliminary algorithm to extract corridors, and junctions, demonstrate its expressiveness. Contribution of this work starts with the formulation of metric-topological surveying of environment, and a generic n-direction planar representation accompanied with a general method for extracting it from occupancy map. The implementation also includes some semantic labels specific to warehouse like environments. © 2014 IEEE.</p> +
<p>Machine learning methods are increasingly being used to solve real-world problems in the society. Often, the complexity of the methods are well hidden for users. However, integrating machine learning methods in real-world applications is not a straightforward process and requires knowledge both about the methods and domain knowledge of the problem. Two such domains are colour print quality assessment and anomaly detection in smart homes, which are currently driven by manual monitoring of complex situations. The goal of the presented work is to develop methods, algorithms and tools to facilitate monitoring and understanding of the complex situations which arise in colour print quality assessment and anomaly detection for smart homes. The proposed approach builds on the use and adaption of supervised and unsupervised machine learning methods.</p><p>Novel algorithms for computing objective measures of print quality in production are proposed in this work. Objective measures are also modelled to study how paper and press parameters influence print quality. Moreover, a study on how print quality is perceived by humans is presented and experiments aiming to understand how subjective assessments of print quality relate to objective measurements are explained. The obtained results show that the objective measures reflect important aspects of print quality, these measures are also modelled with reasonable accuracy using paper and press parameters. The models of objective measures are shown to reveal relationships consistent to known print quality phenomena.</p><p>In the second part of this thesis the application area of anomaly detection in smart homes is explored. A method for modelling human behaviour patterns is proposed. The model is used in order to detect deviating behaviour patterns using contextual information from both time and space. The proposed behaviour pattern model is tested using simulated data and is shown to be suitable given four types of scenarios.</p><p>The thesis shows that parts of offset lithographic printing, which traditionally is a human-centered process, can be automated by the introduction of image processing and machine learning methods. Moreover, it is concluded that in order to facilitate robust and accurate anomaly detection in smart homes, a holistic approach which makes use of several contextual aspects is required.</p>
<p>Development, testing and validation of algorithms for smart home applications are often complex, expensive and tedious processes. Research on simulation of resident activity patterns in Smart Homes is an active research area and facilitates development of algorithms of smart home applications. However, the simulation of passive infrared (PIR) sensors is often used in a static fashion by generating equidistant events while an intended occupant is within sensor proximity. This paper suggests the combination of avatar-based control and probabilistic sampling in order to increase realism of the simulated data. The number of PIR events during a time interval is assumed to be Poisson distributed and this assumption is used in the simulation of Smart Home data. Results suggest that the proposed approach increase realism of simulated data, however results also indicate that improvements could be achieved using the geometric distribution as a model for the number of PIR events during a time interval. © IEEE 2015</p> +
<p>We present a soft computing techniques based option for assessing the quality of colour prints. The values of several print distortion attributes are evaluated by employing data clustering, support vector regression, and image analysis procedures and then aggregated into an overall print quality measure using fuzzy integration. The experimental investigations performed have shown that the print quality evaluations provided by the measure correlate well with the print quality rankings obtained from the experts. The developed tools are successfully used in a printing shop for routine print quality control.</p> +
<p>This paper presents four schemes for soft fusion of the outputs of multiple classifiers. In the first three approaches, the weights assigned to the classifiers or groups of them are data dependent. The first approach involves the calculation of fuzzy integrals. The second scheme performs weighted averaging with data-dependent weights. The third approach performs linear combination of the outputs of classifiers via the BADD defuzzification strategy. In the last scheme, the outputs of multiple classifiers are combined using Zimmermann's compensatory operator. An empirical evaluation using widely accessible data sets substantiates the validity of the approaches with data-dependent weights, compared to various existing combination schemes of multiple classifiers.</p> +
