Feature Detection for Model Assessment in State Estimation
Abstract
The presence of deterministic features in a residual sequence is oftentimes an indication of a modeling error in the state estimation process. Here, three methods of detecting and extracting the features of jump and/or drift in a predicted residual sequence are developed. Two of the methods are traditional ones. The first is a multiple hypothesis, generalized, likelihood ratio test that results in a chi-squared (XSQ) test statistic. The second is a similar, but computationally more efficient, intuitively derived test resembling a modified Neyman-Pearson (MNP) test. The third method is a nontraditional one; it uses a backpropagation artificial neural network trained to emulate the MNP test. Monte Carlo experimental results show that the XSQ and MNP give essentially identical results, while the ANN -- although apparently outperforming the XSQ in feature detection -- does so at the expense of a higher feature misclassification probability, which is an undesirable effect. Overall, the ANN is judged as a feasible approach to feature detection, and improved performance is expected with better network training.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 15, 1991
- Accession Number
- ADA252755
Entities
People
- D. J. Ferkinhoff
- J. G. Baylog
- K. F. Gong
- S. C. Nardone
Organizations
- Naval Underwater Systems Center