Elastic Sensor Networks: Towards Attention-Based Information Management in Large-Scale Sensor Networks

Lead Research Organisation: Imperial College London
Department Name: Dept of Computing


This project, which will be co-funded by the Institute of Security Science and Technology at Imperial College London, aims to develop a novel theoretical framework and associated computational model for information management in Large-scale Sensor Networks (LSSN). Applications of such networks are drawing wide attention from both academia and industry ranging from home monitoring to industry sensing, including environment and habitat monitoring, security, traffic control and health care. A key challenge in managing such networks is that of avoiding information overload as the amount of information monitored increases. A second key challenge is how to effectively maximize the value of collected information under resource and real-time constraints. Addressing both challenges requires developing effective and efficient methods for organizing information collection and information processing that focus on analyzing only information relevant to the user needs.Our key hypothesis in this proposal is that an analogy between information processing by humans, in particular their well-evolved human attention mechanism, and information processing in sensor networks would lead to the development of novel and highly effective information management strategies for LSSNs. This analogy would enable us to exploit effectively the relationship between local and global information, avoid information overload in the application and also minimize unnecessary resource consumption (processing and communication) in the network. Our interest in developing and using an attention-like mechanism in sensor networks is driven by the fact that it could be mapped easily to a concise and robust Bayesian formulation. Such a formulation would enable us and other researchers to reason about the correctness of the approach and also to reason about its extensions and potential improvements beyond this project.Our work in this project thus focuses on addressing a number of key challenges both at the theoretical and practical levels, including the extension and application a standard Bayesian probabilistic framework to the LSSN setting, developing the foundations for an elastic resource allocation model for such networks and supporting a decentralized approach for our implementation that scales to large scale networks implementations.In addition to developing the theoretical foundations, our work will also include developing functional prototypes of a distributed LSSN information management system using both simulations and real sensor hardware. The evaluation of our methods will proceed using case studies from two application areas: multi-modality security monitoring and urban pollution monitoring. The evaluation will be conducted in close collaboration with end users in the Institute of Security Science and Technology (ISST) and the Cenre of Transportation Studies (CTS) at Imperial College London as well as with collaborators in three international institutions (Rutgers University, Harvard University and Monash University). The evaluation will be based on real and simulated data sets to compare the efficacy and efficiency of the proposed approach against traditional and competing methods.

Planned Impact

This project aims to achieve a number of high impact benefits beyond the project consortium: Academic Impact: The theoretical foundations of our research aim to make a significant academic contribution to the analysis of the large-scale and distributed data sets in general and not only in LSSN applications. It extends naturally to the management and analysis of traditional healthcare and bioinformatics data and even financial data. The same methods can also provide a basis for a more general problem of managing complex dynamic behaviours and emergent phenomena under resource constraints, including the analysis of Large-scale Complex IT Systems which is regarded as urgently needed by the EPSRC. To achieve these academic benefits, the research outputs will be published in high impact academic journals (e.g. IEEE TKDE and Sensors) and top academic conference (e.g. VLDB, ACM KDD, SIGMOD and SenSys, Mobicom and IPSN). End Users of Large-scale Sensor monitoring Applications: The application focus of our research is on two areas (Security and Pollution Monitoring). End users include researchers and practitioners. The proposed methods generic and can be applied to similar applications of LSSN including environment and industrial sensing, habitat monitoring, security, traffic control and health care. The resulting attention-based mechanisms provide new ways of managing and analysing the sensor networks used in thesedomains, and will in the long term translate into enabling them to address more complex applications and problems. Potential stakeholders include environmental organisations, city planners, law enforcement agencies, defence companies, health care institutions and scientific researchers in these domains. Through collaboration with ISST, CTS and the international partners, we will seek engagement with a wide user base through targetted information dissemination workshop to facilitate knowledge and technology transfer, and to collect feedback on the applicability of our results in other areas. Industrial outreach: The successful development and demonstration of our framework and methods would, in the long term, be of direct relevance to commercial players developing and using sensors and sensor management platforms for various applications and industries. The success of our methods, and their adoption by such palyers would increase their competitiveness. Within the UK, these players include large telecommunication operators, companies engaged in security and defence applications, specialized hardware and network management companies and software companies. The interest of UK-based commercial players in sensor networks can be seen in their participation in a large number of research projects for sensor network applications funded by the DTI over the past few years. The results of this project would provide these companies with the theoretical foundation and algorithms for developing novel data management architectures and systems. Based on a strong track record of industrial collaboration of the Discovery Sciences Group and also the collaboration programme of ISST, we will follow various routes for disseminating our results to such industries including seeking their collaboration on future research projects to identify and address more research challenges and research opportunities. Our primary route to engaging such players will be through a focused workshop series. Software Development: The successful output of this research project would include a number of software prototypes that can be used directly by non-profit end users and/or licensed to industrial partners. The Discovery Sciences Research Group has a strong track record in the development of commercialization of technology, and will explore the best routes for making the software outputs available for use by a wide user based through consultation with both ISST and the Imperial College consulting arm ICON and commercialization office, Imperial Innova


10 25 50
Description The key findings through the Elastic Sensor Network project include the following aspects:
1. Cognitive sensing methodology: We developed a cognitive sensing methodology and computational framework during the project. Intelligent agents including human are able to achieve equilibrium with the dynamic and fluctuate environment with limited cognitive computation resource. Such ability drives us to design a sensing system to fit dynamic environment. In this cognitive sensing framework, based on the principle of free-energy and surprise minimisation , perception is viewed as modelling actions to minimise the KL divergence between (subjective) recognition density and (objective approximate) distribution, while sensing is a connected action to reduce the space between such KL divergence with real KL divergence (that is between recognition density and real distribution). In order to conduct the sensing and modelling in such way, we addressed two aspects of issues: how to adapt the model (and even the model space) so that it fits the constantly changes of sensing target; and how to reduce the dimension of sensing, to fit the resource constrain. For the first issue, we used non-parametric approaches including dictionary based compressive sensing and deep learning to deal with the change of model space, and a decay moving window for training data to simulate 'memory'. For the second issue, we showed although both random and surprise minimisation sensor selection can achieve the goal, the later has better performance.

2. WikiSensing platform: Based on the methodology in Compressive information measurement and collaborative sensing used in Elastic Sensor Network project, We designed and implemented a research platform to support urban and ubiquitous crowd-sensing, turning sensor big data to data product. Data sources, analysis models, and sensing data products are all user-contributed and fused in the platform. Functions like model training and composition are provided as platform support. Also it provides a sensor data and application marketplace as well as an economic ecosystem to encourage user contribution.

3. Others: We applied the methodology some real case applications. We proposed a pollution monitoring system, which employs only a small number of sensors to monitor pollutant distributions over a region of interest. Compressive sensing technique is applied to implement our signal construction framework so as that a high resolution pollutant distribution can be accurately reconstructed with under-sampling measurements (less than 2%).
Exploitation Route 1. The cognitive sensing framework could be used in a wide range of sensing and modeling applications, including environmental monitoring, etc. The compressive sensing based / deep network based modeling methods are relatively general, and can be adopted to those sensing target with evolving and dynamic behaviour. The sensor placement method could be used in selecting optimzed sensor placement in various applications.
2. The WikiSensing system is public accessible and open to all researchers, application developers and end users, for them to store, manage, and model their sensor data. The WikiHealth platform upon it is also open and can be used in various healthcare related projects.
Sectors Digital/Communication/Information Technologies (including Software),Environment,Healthcare

URL http://wikisensing.org/
Description The output of the Elastic Sensor Network project (algorithms, computational frameworks, tools, etc.) has been contributed to the following non-academic impacts: 1. WikiSensing platform: the sensor data management and modelling methods developed in the Elastic Sensor Network project had been further developed to a cloud-based sensor data management and modelling system called WikiSensing. It's a cutting-edge user oriented and crowdsourcing based system for large scale of sensor data. The system has been used in many practical projects and public events including UpLondon Hackthon in 2013. 2. WikiHealth system: The similar approach was adopted to devleop the healthcare sensor data platform named WikiHealth. WikiHealth will be a social platform for data-driven and context-specific discovery of citizen communities in the areas of health, fitness and well-being. WikiHealth will harness personalised and location-based information for health-related knowledge discovery available to individual users at a massive scale. WikiHealth is based on the concept of the wiki (or mass collaboration), according to which aggregated user-contributed content is collectively curated to produce unprecedented knowledge in a multi-perspective and socially engaging way. The system has been used in several practical healthcare related projects. 3. Pay by data platform: During the Elastic Sensor Network project, we found current model of mobile marketplace suffers the risk of violating user's data privacy, because only the price of application is explicitly indicated without clear agreement on usage of data, and the granularity of data access authentication is not enough to protect users privacy. Targeting at this problem, we proposed a new approach of data pricing model, where the data usage of the application is explicitly shown, and controlled by an authentication service. Such model and system protect users from the abuse of their data, and show good application potential in various application marketplaces. This PBD model has been demonstrated in several public events including Digital Economy Conference UK and Imperial Festival, to increase the public awareness of mobile privacy issue.
First Year Of Impact 2013
Sector Digital/Communication/Information Technologies (including Software),Healthcare
Impact Types Societal,Economic