COMPLEX SYSTEMS

Multisensor Fusion - Research Accomplishments

Background

In solving scientific and engineering problems, one is provided with a variety of information such as experimental data, analytical equations and computer simulations. The cause for such variery of information sources can be quite varied and application dependent. In the robotics area, sensors of different types are needed to adequately model the terrain and operate effectivly in it. In the pattern recognition area, a classifier must be designed based on experimental data; in such case these is no single best classifier but there is a wide variety of them each one out performing the other in some part of the state space. In other areas, the analytical equations and computer programs that solve a particular problem (such as embtittlement prediction or estimation of hydrate concentrattion) can be quite varied and provide solutions with different performances. In all these cases, where solutions are based on finite set of data/measuremnts, often there is no single solution that out-performs the others; rather, each has its own strengths and weaknesses.

Our Multisensor and Information Fusion Methods

The area of information fusion deals with the methods to exploit the variety of information components such that the fused information is at least as good as the best or best combination of the components. We developed a number of fusion methods for several classes of problems: (a) multiple classifiers and function estimators can be fused to ensure that the fused method is at least as good as the best estimator, (b) information from multiple sensors measuring the same quantity can be combined to obtain an estimate that is at least as accurate and potentially significantly better than the best sensor, and (c) measurements can be combined with analytical models to produce a system that more closely approximates the underlying process.

Fig. 1: Illustration of fuser.

We developed a number of fusion methods that are guaranteed to be close to the optimal with a high probability based entirely on measurements. If all error distributions are known, then the optimal fuser can be computed in principle, but such approach is not feasible here since we only have measurements. Our performance guarantees are the best of the kind possible when only measurements are known. The methods developed by us include (a) vector space fusers, (b) neural network fusers, (c) nearest neighbor fusers, and (d) Nadaraya-Watson fusers.

We also developed two generic classes of fusers, namely (i) isolation and (ii) projective fusers. The isolation fusers are easier to compute and ensure that the fused system is at least as good as the best predictor - the linear fusers belong to this class, which also includes some non-linear fusers as well. The projective fusers are more powerful in that they are at least as good as the best subset of predictors, but are harder to compute if the error distributions are not known.

We recently developed a nearest neighbor implementation of the projective fuser based entirely on measurements. We mathematically showed that this fuser is a very close approximation to the optimal fuser that can be computed only when all joint error distributions are known. Furthermore, this fuser can be computed with an almost linear-time algorithm.

Applications

We applied our fusion methods to combine multiple sensor measurements collected at various well-sites for the purpose of methane hydrate exploration. Here, several measurements and computational procedures based on measurements are used to obtain the porosity estimate that indicates the presence of methane hydrates. The difficulty is that there are several porosity estimates but with a very limited knowledge about their relative accuracy. We utilized the physical laws that relate the physical parameters to design an isolation fuser whose porosity estimate is an order of magnitude more accurate than the best-known measurement or estimate.

We also developed a fuser for the prediction of embrittlement levels in light water reactors by utilizing a combination of domain models and nonlinear estimators including neural networks and nearest neighbor regressions. This method resulted in about 40% reduction in the uncertainty of prediction compared to the best available model in the literature.

More detailed list of publications can be found at N.S.V. Rao's website.

[Complex Systems] [Current Research Index]

CESAR - Center for Engineering Science Advanced Research Oak Ridge National Laboratory

Disclaimers    -