Method and apparatus for predicting object properties and events using similarity-based information retrieval and modeling
Inventors
Birdwell, J. Douglas • Wang, Tse-Wei • Icove, David J. • Horn, Sally P. • Horn, Roger • Rader, Mark S. • Stansberry, Dale V.
Assignees
University of Tennessee Research Foundation • United States Department of the Army
Publication Number
US-8775428-B2
Publication Date
2014-07-08
Expiration Date
2030-06-25
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Abstract
Method and apparatus for predicting properties of a target object comprise application of a search manager for analyzing parameters of a plurality of databases for a plurality of objects, the databases comprising an electrical, electromagnetic, acoustic and thermal spectral database (ESD), a micro-body assemblage database (MAD) and a database of image data whereby the databases store data objects containing identifying features, source information and information on site properties and context including time and frequency varying data. The method comprises application of multivariate statistical analysis and principal component analysis in combination with content-based image retrieval for providing two-dimensional attributes of three dimensional objects, for example, via preferential image segmentation using a tree of shapes and to predict further properties of objects by means of k-means clustering and related methods. By way of example, an evidence tree display showing a target object linked by a pathway to a predicted property comprises a similarity value, a speculation value and a model-based rank.
Core Innovation
The invention provides a method and apparatus for predicting properties of a target object by applying a search manager to analyze parameters in multiple databases including an electrical, electromagnetic, acoustic and thermal spectral database (ESD), a micro-body assemblage database (MAD), and a database of image data (CBIR). These databases store data objects with identifying features, source information, and site properties and context including time and frequency varying data. Multivariate statistical analysis and principal component analysis (PCA) in combination with content-based image retrieval provide two-dimensional attributes of three-dimensional objects via preferential image segmentation using a tree of shapes. Clustering methods such as k-means facilitate prediction of further properties.
The problem addressed involves improving identification and characterization of objects by utilizing similarity-based information retrieval and modeling across diversified data sources. Whereas existing systems may focus on exact or schema-driven matching, the invention aims to efficiently locate and cluster nearest neighbor objects with similar properties in complex multi-dimensional databases that incorporate spectral, micro-body assemblage, and image data. Challenges include integrating disparate modalities, handling multi-dimensional and time/frequency varying data, and predicting properties for novel objects with no exact matches.
Claims Coverage
The patent includes a set of independent claims primarily directed to a computer-implemented method and apparatus for dynamic indexing of spectral data and generating evidence trees for property prediction based on similarity measures, modeling, and associated displays.
Dynamic indexing of spectral data using multi-dimensional attribute vectors
The method stores spectral data represented by multi-dimensional attribute vectors comprising electromagnetic, acoustic, optic, isotopic, mass spectra, or thermal data. It defines a root node in a spectral database tree with similarity tests that separate data into clusters in a three-dimensional subspace and determines reference evidence related to location, manufacturer, or event for a target object.
Similarity measures using various distance metrics
The method applies Euclidean distance, cosine distance metric, or squared chord distance to determine similarity between spectral data samples, enabling effective clustering and indexing.
Graphical display of objects and associations in evidence trees
The invention displays the target object centrally with symbols linked by lines or curves to other objects forming pathways annotated with similarity values, speculation levels, and model-based ranks balancing similarity and speculation.
Recursive clustering using support vector machines within spectral database trees
Recursive application of similarity tests using support vector machines to partition spectral data into multiple clusters by defining indexing surfaces in the three-dimensional subspace.
Integration of multiple object associations and metadata in graphical forms
Displaying associations such as containment or operational data relating objects, grouping associated objects in concentric circles, and labeling objects. Forensic measurement data (spectral or thermal signatures) are represented as distinct graphical elements within layered circular graphs.
Prediction of worn or failed components in machines or processes using multivariate statistical analysis and modeling
Using query inputs with spectral and attribute data to search multiple databases via tree-structured indices, retrieving similar objects, and applying models including optimization methods to predict physical status such as component wear or failure.
The claims cover a comprehensive computer-implemented method and apparatus that dynamically indexes diverse spectral and associative data using multi-dimensional clustering and distance metrics, provides graphical evidence tree displays linking objects and their properties, and applies model-based prediction for object characterization including forensic and machine health applications.
Stated Advantages
Enables rapid identification of objects with similar attributes from large multi-dimensional databases through efficient similarity-based search.
Integrates multi-modal data including spectral, micro-body assemblage, and image data to improve prediction accuracy of object properties and origins.
Supports dynamic updating and incorporation of new data to refine models and cluster memberships, enhancing inference precision over time.
Provides a graphical representation of evidentiary data as evidence trees linking objects with similarity measures, speculation levels, and ranks to support forensic and other analyses.
Utilizes multivariate statistical analysis, PCA, k-means clustering, and support vector machines to improve clustering and indexing, reducing search complexity.
Allows modeling of time and frequency varying data for predictive maintenance and identification of worn or failed machine components.
Documented Applications
Forensic analyses including identification and attribution of objects using spectral signatures, micro-body assemblages (e.g., pollen, charcoal, diatoms, foraminifera), and image data.
Source identification and characterization of fire events and residual materials based on electromagnetic and acoustic emission spectra.
Micro-particle and micro-body analysis such as automated recognition and classification of pollen grains, diatoms, and charcoal particles via content-based image retrieval and preferential image segmentation.
Monitoring and identification of livestock and intruders using thermal and electromagnetic signatures.
Detection and differentiation of vehicle movements and fire events using microwave and acoustic spectral data.
Financial data mining, portfolio analysis, and discovering patterns in criminal activities or communications networks using similarity-based clustering and modeling.
Disease modeling and drug discovery using measurable and inferred characteristics of biological objects such as individuals or microorganisms.
Predictive maintenance and fault detection in machines and industrial processes by comparing time series or spectral data with historical data to identify worn or failed components.
Interested in licensing this patent?