OVERVIEW

Objectives – Methods – Accomplishment

Advances in microelectronics, array processing, and wireless networking, have motivated the analysis and design of low-cost integrated sensing, computing, and communicating nodes capable of performing various demanding collaborative space-time processing tasks. In this work, we considered two methods using coherent acoustic sensor array processing and localization on distributed wireless sensor networks. In recent years, in the first method, we and others have proposed various algorithms based on time-delay estimation using all the sensor data followed by least-squares smoothing of these estimates to perform localization and enhanced signal beamforming for detection and identification. This method of performing coherent processing algorithms depends crucially on fine-grain synchronization of the signals obtained from all the sensors in the array. In practical wireless sensor systems, fine-grain synchronization is a challenging task, but has been partially solved by using the Reference-Broadcast Synchronization scheme of J. Elson. In the second method, we use subarrays, where sensors in each subarray are closely separated and can be connected by wires. Thus, the time-synchronization issue among the sensors is no longer an issue. Within each subarray, we were able to use a maximum likelihood method (denoted as the Approximate Maximum Likelihood (AML) to estimate the direction-of-arrival (DOA) of the source at each subarray. Then least-squares smoothing of bearing crossings of these DOAs from different subarrays can provide accurate source localization. Recently, we have also developed effective algorithms based on a novel “virtual array model” for beamforming applications in controlled reverberant environments. These algorithms are capable of performing source localization under multipath reflection conditions not possible using conventional free-space beamforming algorithms. Preliminary simulations and limited field measurements indicate the “virtual array model” is useful. All three of these algorithms have been implemented and tested on a real-time sensor platform using iPAQs. Numerous journal and conference publications as well as an invited tutorial paper on beamforming have been reported.

ACCOMPLISHMENTS

From 2004 and 2005

The use of sensor networks for monitoring has been proposed in recent years. The location of each sensor node must be determined before performing any useful monitoring. For large number of networks, it is not possible to have GPS capability on each sensor node. However, some anchor nodes whose locations are assumed to be known. Several centralized algorithms have been proposed using convex optimization to estimate location of sensor nodes given ranging information, i.e., pair-wise distance measurements between sensor-to-sensor nodes and sensor-to-anchor nodes. However, as the size of the network increases, it is not desirable to have a centralized algorithm because of costs due to computational complexity and communication energy of transmitting information to a fusion center. Recently, we have proposed new distributed algorithms for sensor localization based on the Gauss-Newton method. We assume that each sensor can only communicate to nodes within its radio range, i.e., only local communication is allowed. Based on pair-wise distance information between nodes and estimated locations of its neighbors, we minimize the local cost function and then sent its estimated location to all its neighbors. The algorithm terminates when all the sensor nodes reach a given value of local cost function. It is shown that the proposed algorithms guarantee a non-increasing value of global cost function by proper choice of step length.Analysis and simulation of the proposed scheme have shown it is capable of working in realistic conditions.

FUTURE DIRECTIONS

From 2004 and 2005

We expect to have completed various sequential and parallel processing of the above distributed algorithm in the coming months. Then we plan to further develop and test methods for performing node localization in sensor networks with measured data. We are attempting to contact other researchers who may have such measured data.

From 2003

In area (B), we propose to continue studying the reverberant array processing topic and initiate two new topics. In the "virtual array model," we need to enumerate all the possible reflecting multipath rays and determine those rays that are mutually consistent to localize the source. We found different formulations of the same problem can have different numerical least-squares implications for real-time measured data. We need to find adaptive versions of our array processing algorithms capable of handling moderate practical uncertainties in the model. The first new proposed topic deals with modeling the use of advanced optimization techniques for sensor network applications. In the last ten years or so, a new class of convex optimization method based on semi-definite programming (SDP) method (which includes and supersedes the classical linear programming and the quadratic programming optimization methods) has revolutionalized the possibility of performing optimization for many practical real-time system problems. When an original optimum solution is not in a convex space, many locally optimum solutions may exist and it is not clear how to determine the true global optimum solution from these locally optimum solutions. Upon proper relaxing of the constraints (e.g., relaxing the solution from a subset of discrete values to a subset of the reals; etc.) on many of these problems, these problems become convex problems. SDP is a modern numerical iterative method that can solve convex optimization problems involving a small number of iterations (i.e., 3-5 iterations) in almost all cases. S. Boyd of Stanford Univ. and L. Vandenberge of UCLA have been the two early leading researchers and advocates of SDP. They and other leading researchers (Y. Ye of Stanford Univ. and Z.Q. Luo of Univ. of Minnesota) all have proposed using SDP methodology to address various sensor network optimization problems. Unfortunately, most of the problems they have posed are not very realistic from the physical and engineering points of view. We have initiated a study on these problems (with one graduate student and some advice from L. Vandenberghe) and propose to use SDP methodology to tackle sensor subset selection problem, optimum robust beamforming strategy, optimum combined detection, estimation, and tracking problem, etc. in various sensor network settings. We believe having had extensive prior experiences in working with practical aspects of sensor systems, the models we pose and want to solve using SDP will be more realistic than those posed by other excellent optimization researchers but who have not worked before in sensor network. While the early part of this proposed research will be performed on a centralized processor, the eventual goal is to perform the processing in a distributed manner. The second new topic we propose to conduct is to find new and useful event detection algorithms to determine the start of an event of interest (e.g., an earthquake) from sensor waveforms. We have had some discussions with the seismic processing group and obtained some limited sets of data to formulate and test these algorithms. We propose to start considering various advanced adaptive and robust detection and estimation algorithms to tackle this problem.

PEOPLE

FACULTY

K. Yao and D. Estrin

GRADUATE STUDENT

B.H. Cheng

RESEARCH ENGINEERS

Dr. R. E. Hudson
Dr. F. Lorenzelli