SENSORS: Multisensor Information Fusion for Biological Sensor Networks

Massachusetts Institute Of Technology, Cambridge MA

Investigators

Abstract

Detection of biological attack is a critical element in civilian biodefense. Various approaches to such detection have been proposed, including periodic sample analysis, and use of biological sensors, e.g., fluorescence-based. Sample collection methods can provide the needed specificity and sensitivity, however they are time-consuming and costly. Faster-acting (e.g., fluorescence) biological sensors offer lower specificity and sensitivity. Consequently, detection of bioattacks using traditional sensing methods is constrained by serious tradeoffs involving significant attack-to-detection delays, high false-alarm rates, and excessive costs. Therefore, new paradigms for bioattack detection are of paramount importance, especially for such large-scale civilian biodefense applications as protection of urban transport systems (e.g., subways), airports and other public facilities. This proposal concentrates on a novel paradigm for bioattack detection, based on multisensor information fusion. Objectives and methods. Detection of biological attack is in essence a recognition problem where the attack/no-attack decision is based on the sensor information. The proposed research concentrates on methods for fusion of information from multiple sensors within a network, to boost performance beyond the levels offered by any sensor separately. Multisensor data fusion should be considered as a key enabling technology for biological sensor networks. While topics such as delivery of data from multiple sensors, associated network protocols and databases, have received attention, including other network application domains, methods for robust combining of multisensor information present within such networks have not been adequately addressed. Multisensor information fusion is a critical part of the recognition process, i.e., as the task of intelligent aggregation of information within and as part of the reasoning process to achieve an emergent behavior conclusions regarding the phenomena to be recognized (in this case the biological attack occurrences) that would be unattainable from either single sources or from simple information superposition approaches. The multisensor information fusion techniques are thus a crucial enabling element that will allow exploiting to the fullest extent the information present in biological sensor networks. The proposed research addresses cardinal problems of information fusion, in particular information uncertainty and disparity, in the setting of biodefense-oriented automatic decision-making. Novel techniques for biodefense multisensor information fusion, based on uncertainty calculi (e.g., Dempster-Shafer theory) and advanced pattern recognition (e.g., Vapnik-Chervonenkis theory), will be investigated. The effort will use biosensor data collected previously by MIT Lincoln Laboratory (MIT/LL) as well as data that will be collected within the proposed effort. Intellectual merit. Offering novel techniques for exploitation of multisensor information in the decisionmaking process, the proposed research will contribute to advancing state of the art in both the bioattack detection domain and in the domain of multisensor-based automatic recognition in general. While biodefense sensing will be the focus of this research, the multisensor information fusion techniques developed are expected to be applicable to a range of other sensing applications. Moreover, the developed information fusion techniques will not be constrained to a specific sensor suite. Therefore, their value will not be diminished by development of more robust sensors. Rather, these methods will be able to take advantage of the increased performance of future more robust sensors. (In fact, a likely byproduct of the proposed effort will be the insight into mutual value of various information sources, valuable for improving future sensor designs.) Thus, in summary, the multisensor information fusion techniques investigated within this proposal offer the prospect to attain performance levels beyond any nonideal single sensor constituents, and thus will likely become a vital enabling element of future biodefense systems. The proposal authors' prior research in biodefense, sensor information fusion, and pattern recognition, form a solid basis for the proposed effort. In addition, MIT/LL team is in a unique position to undertake this research, because of unique access to biodefense sensor data stemming from other MIT/LL biodefense programs and from the development of MIT/LL's own biological-agent sensors. Broader impacts. Biological weapons are a serious threat to the civilian population. Small quantities of biological agents are sufficient to cause significant casualties, and attacks can be difficult to detect at the time they occur. This research will contribute to the knowledge needed to develop such detection capabilities. Early detection of an attack through deployment of sensor networks can save innumerable lives and therefore has an enormous value to society. In addition, it is expected that the participation of MIT/LL will benefit the NSF 03-512 program and its other academic participants since MIT/LL is familiar with the broader issues and needs surrounding biological defense.