Applications of Artificial Intelligence to Embedded Sensor Networks and Their Data

Paul Hanson, NTL LTER

Tony Fountain, SDSC

Yu Hen Hu, UW-Madison, Electrical and Computer Engineering

Deana Pennington, UNM, Biology

Abstract

Recent advances in embedded sensing systems allow us to gather ecosystem data at rates greater than ever before. To take full advantage of embedded sensing systems, we often raise sampling to frequencies adequate for capturing the spatio-temporal dynamics of the ecosystem phenomenon in question. However, the increased sampling frequency creates a number of challenges, including (1) demands on embedded system power that cannot be maintained for prolonged periods, and (2) data densities that cannot be analyzed with adequate rapidity using traditional techniques.

Artificial intelligence (AI) provides many promising techniques that can be applied to the problems identified above. Specifically, algorithms can be used to detect driving events of interest, and then adjust the sampling frequency in response variables to optimize both data collection and sensor system power consumption. Inherent spatio-temporal correlations among sensors in the embedded system allow for a coordinated approach to sampling that reduces the number of sensors that must be active at all times. For analyzing the resulting large data sets, pattern recognition algorithms allow for the detection of phenomena in the data, even when the system is controlled by a complex collection of seemingly random connections and occurrences. AI techniques also can be used for making decisions based on a combination of "expert knowledge" and the data in-hand.

Through this working group, we intend to meet a number of goals. (1) Present a variety of AI techniques, ranging from simple to complex. (2) Describe real-world applications of AI being used in ecosystems analyses. (3) Show how the application of AI can greatly improve the performance of embedded sensor networks. (4) Organize an inter-site working group to develop additional applications of AI at interested LTER member sites.

Workshop Summary

The four presentations consumed approximately 2.5 hrs of the 3 hrs available for the workshop. The speakers covered a diverse set of topics related to AI and ecology, ranging from the fundamentals of AI, to the implementation of AI in embedded sensor networks and in remote sensing. Information dissemination was the primary goal of this portion of the workshop. Informal discussion followed the presentations.

A premise of this workshop was that changing ecological systems, coupled with sensing technologies that provide high data densities, provide analytical challenges not easily addressed using standard methodologies. For example: events of interests occur infrequently; sensing requires agile observation during events, but dormancy modes in between; network operation cost is high; long-term, large scale observations require autonomous operations; and system (sensors, network, power) failure rate can be high. Furthermore, the amount of data is huge.

AI techniques offer a new approach. Intelligent agents are persistent software/hardware systems that perceive, reason, act, and communicate on behalf of human users. They are a realization and embodiment of artificial intelligence methodologies and have these characteristics: autonomous execution; goal seeking; reasoning during action selection; acting for another with authority granted by another; interacting with other agents or humans via dialog or some agent communication language.

A common thread throughout the discussion was that although AI techniques have been commonly used in engineering disciplines for the past couple of decades, they are only at early stages of implementation in ecological disciplines. Tools such as Bayesian Networks, Neural Networks, Support Vector Machines, etc. have application for decision support, pattern recognition, data mining, classification, and optimization in ecological systems. In addition, much domain knowledge (i.e., AI within ecology) has yet to be learned.

This workshop formed the basis of collaboration among scientists from UW-Madison, Center for Limnology, UW-Madison Electrical and Computer Engineering, and the San Diego Super Computer Center. In preparation for this workshop, the collaborators conducted a pilot project during the summer of 2003. The project was formed around the deployment of a network of buoys that sampled limnological and meteorological data at high temporal resolution. These data were analyzed using AI techniques implemented by both UW ECE and SDSC. Some results from this pilot project were presented at the workshop, and will be submitted for publication.

Collaborations fostered through this workshop were partly responsible for further collaborations established at the LTER ASM among scientists from UW, SDSC, NPACI and Taiwan. This consortium proposes to install a limnological buoy in Taiwan. The design of that buoy will be similar to those already deployed in Wisconsin. In addition to the buoy installation, the consortium will provide wireless communication for that buoy to the Internet, and stream and manage data from both sites through a common communications infrastructure. Beyond the technical goals of environmental sensing, communication, and data management, the group will address a suite of ecological questions pertinent to both sites. AI will be a valuable tool in managing and analyzing these data.

Biographies

Presenter: Yu Hen Hu, Professor, University of Wisconsin, ECE

Presentation Title: Intelligent Sensor Network Signal and Information Processing

Yu Hen Hu received BSEE degree from National Taiwan University, Taiwan, ROC in 1976.  He received MS and PHD degree both in electrical engineering from University of Southern California, Los Angeles, CA in 1980 and 1982 respectively.  Currently, he is a professor at the electrical and computer engineering department of the University of Wisconsin - Madison, WI, USA.  Previously, he has been with the electrical engineering department of the Southern Methodist University, Dallas, TX.

Dr.  Hu's research interests include multi-media signal processing, design methodology and implementation of signal processing algorithms and systems, sensor network and distributive signal processing algorithms, and neural network signal processing.  He published more than 200 papers in these areas in technical journal and conferences.  He

edited two books Programmable Digital Signal Processors, and Handbook of Neural Network Signal Processing.

Dr. Hu is a fellow of IEEE.  He served as associate editors for IEEE Transactions on Signal Processing, IEEE Signal Processing letters, Journal of VLSI signal processing, European Journal of Applied Signal Processing.  He served as secretary of IEEE signal processing society, board of governors of IEEE neural network council, chair of IEEE signal processing society, neural network signal processing technical

committee.

Presenter: Paul Hanson, Associate Scientist, University of Wisconsin, Limnology

Presentation Title: The need for artificial intelligence in limnological sampling systems

Paul Hanson received his Ph.D. in Limnology and Marine Science from the University of Wisconsin in 2003. He is a research scientist at the University of Wisconsin, Center for Limnology. His research interests include biogeochemistry, carbon cycling, and ecosystem modeling. Dr. Hanson has been actively involved in the design and development of a buoy sensor network at the North Temperate Lakes LTER field site in northern Wisconsin.

Presenter: Tony Fountain, Directory, San Diego Supercomputer Center, San Diego Knowledge and Information Discovery Laboratory

Presentation Title: Artificial Intelligence and Data Mining for Ecology

Tony Fountain is Director of the San Diego Supercomputer Center Knowledge and Information Discovery Lab (SKIDL). He received his PhD in Computer Science with a dissertation in Machine Learning and Statistical Decision theory. His current research interests include data mining, intelligent systems, and high-performance computational

infrastructure across a broad range of science and engineering domains.

Presenter: Deana Pennington, Research Assistant Professor, University of New Mexico, Biology

Presentation Title: Artificial Intelligence Applications in Remote Sensing

Deana Pennington is Research Assistant Professor at the University of New Mexico, where she is involved in multiple research projects with the LTER Network Office and with Sevilleta LTER.  Her research interests are in Geographic Information Science, Geocomputation, Remote Sensing, and Landscape Ecology.  Deana is currently collaborating with the San Diego Supercomputer Center on computational infrastructure for ecology and remote sensing, including AI approaches to data analysis of massive datasets.