PhD
Bio-inspired Algorithms for Data Fusion in Hazardous Threat Detection
Investigated the use of sensor networks for timely and efficient detection of chemical, biological and radiological leaks. Developed bio-inspired techniques for detecting such threats in a timely manner, while saving energy to prolong sensor lifetime. These new methods will have both civilian and military applications.
Thesis Abstract
Bio-inspired systems focus on the design, development and understanding of systems composed of algorithms that mimic the behaviours or processes of biological entities. In order to achieve a better abstraction, these bio-inspired algorithms can be constructed on a multi-agent platform which comprises multiple interaction between autonomous agents. These components are equipped with cognitive and processing abilities and have access to information extracted from the environment. In our endeavour to explore the viability of using bio-inspired algorithms in chemical, biological, radiological and nuclear environments, we carried out four research studies.
First, we present a numerical verification of a simple population and physics-based epidemiological model for dynamic collaboration in a network of chemical sensors. The modelling approach is based on the known analogy between the information spread in a sensor network and the propagation of epidemics across a population. In this framework, we verify the derived analytical expressions, which relate the parameters to the network (e.g., number of sensors, their density, sensing time, etc.), with parameters of the external challenge (e.g., the chemical pollutant) and the environment (e.g., turbulence). Using numerical simulations of wireless sensor networks with random, line and circle topologies, we show that simulated and analytical results agree.
Secondly, we apply the epidemiology based protocol to a wireless chemical sensor network in a chemical environment with spatial characteristics. The chemical tracers dispersed by turbulent motion in the environment display rather complex and even chaotic properties. Meanwhile, chemical tracer detecting sensors with air sampling to consume significant energy, hazardous chemical releases are rare events which will not require continuity in detection. If all sensors in a wireless chemical sensor network (WCSN) are left in the active state continuously, it would result in significant power consumption. Therefore, dynamic sensor activation is crucial for the longevity of WCSNs. Moreover, the statistical characteristics of chemical tracers to be detected (temporal and spatial correlations, etc.) and placement of chemical sensors can also become the key parameters that influence the WCSN design and performance. In this research study, we investigate the effect of the spatial correlation of a chemical tracer field, and also the effect of network topology, on the performance of a WCSN that employs an epidemiology-based dynamic sensor activation protocol. We present
a simulation framework that comprises models of the spatially correlated tracer field, individual chemical sensor nodes, and the sensor network. After validating this simulation framework against an analytical model, we perform simulation experiments to evaluate the effect of spatial correlation and network topology on selected performance metrics: response time, level of sensor activation, and network scalability. Our simulations show that the spatial correlation of the chemical tracer field has a detrimental effect on the performance of a WCSN that uses an epidemiological activation protocol. These results also suggest that a WCSN with random network topology has poor performance compared to one with a regular grid topology in this application.
Thirdly, we apply a gossip based protocol to the wireless chemical sensor network
to overcome the detrimental effect of the epidemiological protocol. In this study, we investigate the performance of a variant of epidemiology based protocols, the gossip based sensor activation protocol of a WCSN in a chemical tracer field. The simulation framework with the gossip protocol is validated against an analytical model. We then perform simulation experiments to evaluate the performance of the gossip-based sensor activation protocol on selected performance metrics: the sensor activation and chemical tracer detection. We show by simulations that, by varying the communication radii of sensors; we can achieve better energy conservation while maintaining better performance of a WCSN with a gossip-based activation protocol which was the drawbacks of the epidemiological protocol in a turbulent environment.
Fourthly, we explore the possibility of localising a radiological source using bio inspired genetic algorithms. We consider localisation of point sources of gamma radiation using dose rate measurements. As bio-inspired genetic algorithms, we use binary and continuous genetic algorithms (GA). They are used to implement maximum likelihood estimation (MLE) of the position and strength of point radiation sources. MLE was achieved by minimising the objective function which computes the negative log likelihood. Real experimental data collected during a field trial was used to test and verify the performance of the algorithms. The performance of the GA based implementation was compared to an implantation that used gradient descent optimisation. Source parameters estimated by the algorithms were also compared to the theoretical bounds obtained using a Cramer-Rao bound (CRB) analysis, which quantifies the accuracy with which it is possible to localise the source and estimate its strength. All three mplementations localised a single-point source well, nearly approaching the CRB. Reasonable position estimates were achieved for two and three source cases, but the source strength estimates were found to have much larger root mean square (RMS) errors than that predicted by CRB. While the GA-based implementations
took longer to converge compared to the gradient descent algorithm, they encountered fewer divergent runs than the latter algorithm. Also we examine
the data collection geometries that influence the topography of objective function
surfaces in radiological source localisation. It is shown that data gathered along the circumference of a circle around a point a radiation source has an associated mirror image source of different strength that exists outside the circle. It appears that data acquired along an irregular path generated using the random walk algorithm to eliminate the image sources making source parameter estimation easier.
Given these considerations, it is evident that Bio-inspired algorithms are viable solutions for CBRN data fusion.
Inspiration from ancient texts:
Research Inspired by Ancient Scripts on sensing and quoted in my PhD Thesis,
In Buddhist text, sensing and fusion is described as Rupa, Vedana, samjna & vijnana. (Further Ref: Wikipedia)
What is an ideal Sensing outcome?
In Sri Lankan text ‘Subashita’, “Though we have hundred of non-worthy children, it will be worthless, but if we have a single child with intellect and good human qualities, it is the greatest blessing. In analogy, though we have hundreds of sensors, if we can have selected quality sensors, that will give us the most concurrent and reliable outcomes.
Further References: eThesis, Publications
Applications of my studies
Hazard reporting to the nearest centre Propagation of a message across the environment
Social structures and communication
Towards Data Science
Mathematics:
- Engineering Mathematics
- Optimisation
- Statistics & Probability
Business
- Management Studies (One Year Study)
- Financial Mathematics
- Actuarial Mathematics
- Share Market Analytics (Bachelors case study)
Information Systems
- Bachelors Special Honours, Masters & PhD in Computing & Information Systems
Artificial Intelligence, Machine Learning - Accounting Information Systems
- Information & Cyber Security
PhD requirements: allreports[Champake Mendis]-[15 March 2012]
Masters in Engineering Science
Optimised Sink Trajectories for Sensor Networks
Abstract
The performance of the Wireless Sensor Network (WSN) depends on the location of the sink node in the sensor field, assuming the sensor nodes to be relatively stationary. Most researchers have used static sink nodes and some have used multiple sink nodes. In this research study, I propose a mobile sink node for efficient data collection while maintaining the efficient performance of the network. I introduce a technique to obtain the optimum path of the sink node in a network of stationery sensor nodes, considering practical constraints such as the limitation in the sink movement. The proposed evolutionary computer-based simulator, PSO-SIMSENS, is an integrated system of particle swarm optimisation and a sensor network simulator, with suitable fitness functions derived from the SIMSENS traffic parameters. This system can be used in manufacturing applications and in disaster monitoring. Based on the results obtained from simulations, it is deduced that the proposed system achieves an efficient WSN with maximum field coverage while the sink node is mobile.
Champ's Blog 