Distributed pattern recognition schemes for wireless sensor networks
2017-02-28T23:59:40Z (GMT) by
Smart systems are increasingly in use in daily life applications, replacing old-fashioned processes and procedures as a result of technological evolution. However, these systems can be limited in their resources capacity. Wireless Sensor Networks (WSNs) are considered to be one form of such resource-constrained smart systems. One of the main goals of WSNs is to sense physical activities so as to detect events in an area of interest. Adaptive and machine learning techniques have been proposed and implemented to work in conjunction with WSNs to serve a number of applications, such as physical activities detection, network security threats detection, artificial intelligence applications and decision making support. Pattern recognition is one of the most useful machine learning techniques that can perform event detection for WSNs. However, the nature of WSNs poses extreme challenges for the implementation of these learning techniques so that they can serve the goals of different types of applications. Such networks have limited resources available for performing learning operations. Additionally, WSNs are of a dynamic nature in terms of network deployment and the appearance of activities in the field of interest. Global events can also span very large regions requiring vast quantities of data exchange and processing in order to detect such events. These challenges become critical when detection time limits are required by applications such as mission critical and online applications. The aim of this research project is to propose pattern recognition schemes that are capable of addressing the limitations associated with resource-constrained networks such as WSNs. The research first investigates the existing learning techniques for WSNs and their limitations. Then the research proposes novel collaborative in-network global pattern recognition based event detection schemes that are light-weight, scalable and suit resource-constrained networks such as WSNs well. The proposed schemes address the limitations and challenges for WSNs to provide reasonable detection capabilities for mission critical, online, and decision making applications. The proposed schemes adopt the distributed and parallel recognition mechanisms of Graph Neuron (GN) in order to minimise recognition computations and communications and thus will lead to maintaining low levels of limited resources consumption. The distributed network structure of the proposed schemes will result in loosely coupled connectivity between a network’s nodes and avoid iterative learning. Hence, the proposed schemes will perform recognition operations in a single learning cycle of predictable duration, which will provide online learning capabilities that can support mission critical applications. In addition to minimal resources and time requirements, the distributed structure of the schemes will sustain large-scale networks in performing pattern recognition operations. To deal with a WSN’s dynamic nature and limited prior knowledge of events, a pattern transformation invariant scheme is proposed in this research. The proposed scheme implements a weighting mechanism that searches the edges and boundaries of patterns and replaces traditional local information storing. This mechanism allows the scheme to identify dynamic and continuous changes in patterns. Consequently, the scheme will be capable of performing recognition operations in dynamic environments and will also provide a high level of detection accuracy using a minimal amount of available information about patterns. Required protocols for performing scheme operations are also presented and discussed. Theoretical and experimental analysis and evaluation of the presented schemes is conducted in the research. The evaluation includes time complexity, recognition accuracy, communicational and computational overhead, energy consumption and lifetime analysis. The scheme’s performance is also compared with existing recognition schemes. This shows that the scheme is capable of minimising computational and communicational overheads in resource-constrained networks, enabling those networks to perform efficient recognition activities for patterns that involve transformations within a single learning cycle while maintaining a high level of scalability and accuracy. The results show that the scheme’s time complexity is proportional to the square root of the pattern size which allows the network to scale up to adopt large patterns. It is shown that a network that implements mica 2 motes and requires 3.0625 milliseconds to send a single message can perform recognition operations within a single learning cycle duration ranging between 126.4 and 323.1 milliseconds for 10,000- and 40,000-node network settings respectively. The results also show that energy requirements can be decreased up to 89.66 per cent by using the proposed schemes in comparison to other recognition techniques. In terms of efficiency, theoretical and experimental analyses show that the proposed schemes are capable of dealing with transformed patterns with a high level of accuracy. The analyses show that the scheme is able to detect translated patterns, rotated or even flipped patterns with a rotational angle of up to 23 degrees, and dilated patterns with a dilation level of up to 26 per cent. The results show that the proposed schemes have features that will be best suited for implementing pattern recognition applications on resource-constrained networks such as WSNs. The research also discusses the use of the proposed pattern-recognition-based schemes in different machine learning and artificial intelligence (AI) applications. This aims to explore new research opportunities that can lead to enhancing existing schemes’ performance by involving the proposed schemes in different technological disciplines and models. Two disciplines are presented as examples in this context: optimisation and classification. In the first example, a new model that involves the proposed schemes in the process of optimisation techniques, in this case genetic algorithms (GA), is presented. The proposed model enhances the performance of traditional GA in terms of speed and accuracy. The second example proposes a classification model using the pattern-recognition-based proposed schemes. The proposed model shows a high level of classification capabilities compared with other well-known existing schemes.