Over the past years Distributed Sensor Networks (DSN) have been closing the gap between theory and application in real-life scenarios, thereby gaining prominence as the key enabling technology for addressing significant societal challenges. DSN research conducted at SPL emphasizes on both theoretical, as well as practical aspects related to distributed signal and data processing, networking over limited resources, and context retrieval, while serving the design paradigms of IoT and CPS.
Distributed Signal Acquisition, Representation, and Recovery
Distributed signal acquisition and representation
- Investigate coherent processing issues for wireless sensor network systems to enable their stable and robust operation in a wide range of adverse environmental conditions.
- Develop beamforming methods to identify dominant sources, while also providing high rejection of interference and noise.
- Design optimal maximum-likelihood (ML) methods and suboptimal, computationally efficient, distributed techniques using sub-arrays yielding cross-bearing information, for accurate source localization.
- Develop efficient algorithms for network self-organization and dynamic configuration of the needed sensor nodes to perform complex beamforming operations.
-Negative Signals using Symmetric Alpha-Stable Distributions, in Proc. 18th European Signal Processing Conference (EUSIPCO '10), Aalborg, Denmark, August 23-27, 2010.
Distributed Data Sampling & Compression
Energy-efficient acquisition and communication of measurements is a critical aspect of WSNs, directly determining the lifetime and usability of the network infrastructure. We focus on three situations where measurement matrices are extremely under-sampled, i.e. contain a large number of zero-valued entries:
A. While frequent sampling offers high quality monitoring of the underlying processes, increasing the frequency of sampling may have very dramatic effect on the lifetime of the network, due to the relationship between measurement acquisition and energy consumption. If communications with other nodes is also necessary, then the impact of sampling rate on network lifetime is even more pronounced.
B. A second scenario that entails a large number of missing measurements is attributed to measurements encoded in packets that lost due to communications failures. This scenario occurs in industrial environments, where heavy machinery has a detrimental impact on link quality and in multi-hop networks where congestion and duty cycling can also lead to dropped packets. A significant number of missing measurements that can have a dramatic impact on subsequent tasks, such as detection of unusual events or clustering of the measurements.
C. A third scenario is related to the temporal sampling frequency of a WSN. Either by design, or due to network clocks de-synchronization, each sensor may end-up sampling the underlying field at a different time instance. We argue that such a limitation can actually be used in our advantage and achieve temporal super-resolution.
To achieve strict performance requirements and overcome the limitations of challenging operating environments, we exploit redundancies in data collected over time, which are modeled by low rank measurement matrices. In SPL, we explore the cutting edge signal processing and learning paradigms of Compressed Sensing (CS) and Matrix Completion (MC).
Related Projects: Hydrobionets
Savvaki, Sofia; Tsagkatakis, Grigorios; Panousopoulou, Athanasia; Tsakalides, Panagiotis, Application of Matrix Completion on Water Treatment Data, Proceedings of the 1st ACM International Workshop on Cyber-Physical Systems for Smart Water Networks, Article #3.
Tzagkarakis, George; Tsagkatakis, Grigorios; Alonso, Daniel; Celada, Eugenio; Asensio, Cesar; Panousopoulou, Athanasia; Tsakalides, Panagiotis; Beferull-Lozano, Baltasar, Signal and Data Processing Techniques for Industrial Cyber-Physical Systems, Rawat, Danda; Rodrigues, Joel; Stojmenovic, Ivan (Ed.): Cyber Physical Systems: From Theory to Practice, CRC Press, USA, 2015.
Autonomic Networking for Distributed Sensor Networks
Modern sensing applications have been facing up to the challenge of moving from episodic, periodic sampling to pervasive paradigms, relying on the resilient, long-term and unattended operation of Distributed Sensor Networks. This necessitates the transition from conventional distributed sensor/actuator schemes, which act as transparent gateways between complex physical spaces and sophisticated decision makers, to more dynamic networked structures that seamlessly integrate within the environment and emerge as the conditions change. From a network perspective, the problem at hand is how we can exploit the inherent redundancy of heterogeneous data (e.g., sensing information, network indices and metrics at different layers of the protocol stack) that are flowing within the network, in order to realize self-organizing, self-optimized, and self-healing network topologies. Towards this direction, our research agenda moves along the following lines:
A. Synthesis, implementation, and evaluation on realistic scenarios and platforms of distributed and localized algorithms for optimizing the energy conservation at large- and massive-scale networks of distributed sensors, with limited power and computational efficiency.
Our approach exploits the concept of Delaunay triangulations in order to establish a scalable framework for solving problems associated to topology control in polynomial time and a localized fashion that, against current state-of-art, eliminates the necessity of additional network traffic.
B. Synthesis, implementation, and evaluation of unsupervised feature-level fusion algorithms that exploit the inherited redundancy of network information available at: (a) across different sides of the network and (b) across different layers of a fully functional protocol stack, ranging from the Physical to the Transport and Application layers. Regardless of the type of deployment (e.g., monitoring an industrial process or the human activity), our approach entails techniques for decentralized and in-network feature selection of the dominant characteristics with respect to inferring user-defined classes, whilst promoting real-time characteristics at computationally constrained devices.
Phivou, Phivos; Panousopoulou, Athanasia; Tsakalides, Panagiotis, On Realizing Distributed Topology Control in Low-power IoT Platforms, in Proc. IEEE 2nd World Forum on Internet of Things (WF-IoT) – Enabling Internet Evolution, Milan, Italy, Dec. 14-16, 2015.
Panousopoulou, Athanasia; Mikel Azkune; Tsakalides, Panagiotis, Feature Selection for Performance Characterization in Multi-hop Wireless Sensor Networks, Ad Hoc Networks, 49 , pp. 70 - 89, 2016, ISSN: 1570-8705
Body Sensor Networks and Human Activity Recognition
Latest trends in wearable-based Human Activity Recognition are challenged by the need to transit from off-line, centralized classification of a-priori known activities towards on-line processing and learning architectures, which are dictated by the need to analyse and interpret complex activities while in data capture. To respond to this challenge, our research considers the combination of modern signal processing techniques and Body Sensor Networks architectures for:
A. Coping with the operating imperfections of the underlying sensing infrastructure, which lead to substantial losses of missing values. In a nutshell, we synthesize classification frameworks that consider the presence ofmissing values during runtime and propose the use of imputation methods (matrix completion, tensor completion) for reconstructing inertial sensing data streams. Our methodology additionally consider the effect of the reconstruction technique on the classification accuracy of landmark classifiers with respect to the persentage of missing values. Ongoing investigations highlight the efficacy of signal processing techniques to effectively exploit the inherent correlations between heterogeneous sensing data for achieving optimal classification performance even for the case of extremely sub-sampled datasets.
B. The automatic calculation of the data attributes that are considered to have sufficient information for inferring the labels of different classes, thereby reducing the feature space prior applying any classification technique. We design and develop on-line feature selection architectures for human activity recognition based on multi-modal and dense Body Sensor Networks for performing feature selection in parallel with data acquisition, and implement it on modern mobile devices. We emphasize on analytical results on the dynamic FSA performance with respect to limited windows of sensing streams, and the variation of human activities, sampled at different parts of the human body.
Related material: FORTH-TRACE dataset and FORTH-TRACE Library.
Savvaki, Sofia; Grigorios, Tsagkatakis; Panousopoulou, Athanasia; Tsakalides, Panagiotis, Matrix and Tensor Completion on a Human Activity Classification Framework, submitted, IEEE Journal of Biomedical and Health Informatics, 2017.
Savvaki, Sofia; Grigorios, Tsagkatakis; Panousopoulou, Athanasia; Tsakalides, Panagiotis, Effects of matrix completion on the classification of undersampled human activity data streams, in Proc. 2016 24th European Signal Processing Conference (EUSIPCO), Aug 2016, pp. 2010–2014, 2016.
Karagiannaki, Katerina; Panousopoulou, Athanasia; Tsakalides, Panagiotis, An Online Feature Selection Architecture For Human Activity Recognition, accepted for publication, 42nd IEEE International Conference on Acoustic, Speech and Signal Processing (ICASSP 2017), 2017.
Karagiannaki, Katerina; Panousopoulou, Athanasia; Tsakalides, Panagiotis, A Benchmark Study on Feature Selection for Human Activity Recognition, in Proceedings of the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing: Adjunct (Ubicomp 2016), 2016, pp. 105-108.
Data Management and Context Retrieval
Collaborative detection, classification, and tracking
- Develop novel algorithmic solutions for data management and information fusion in unstructured and uncertain environments.
- Design and implement effective methods for self-organization of heterogeneous sensors and information fusion for intelligent learning and decision making.
- Develop distributed pattern matching algorithms that are robust to uncertainties, and can be adapted to the observed data for the early detection of abnormal behavior in industrial infrastructures.
Related Projects: Hydrobionets.
Tzagkarakis, George; Seliniotaki, Alexandra; Christophides, Vassilis; Tsakalides, Panagiotis, "Uncertainty-Aware Sensor Data Management and Early Warning for Monitoring Industrial Infrastructures," International Journal of Monitoring and Surveillance Technologies Research, Vol. 2, No. 4, pp. 1-24, October-December 2014 (doi:10.4018/IJMSTR.2014100101).
Seliniotaki, Alexandra; Tzagkarakis, George; Christophides, Vassilis; Tsakalides, Panagiotis, Stream Correlation Monitoring for Uncertainty-Aware Data Processing Systems, in Proc. 5th International Conference on Information, Intelligence, Systems and Applications (IISA '14), Chania Crete, Greece, July 7-9, 2014.