Plant monitoring and diagnosis using input-training neural networks
- Northwestern Univ., Evanston, IL (United States). Dept. of Chemical Engineering
Analysis of process measurements is important in monitoring, diagnosis and control of power plants. Advances in instrumentation and computational resources now allow access to large volumes of data obtained from the plant. Dimensionality reduction is a way of summarizing, with a few latent variables, information carried by a large number of observed variables. Given an m x n matrix representing m patterns of measurements made on n variables, reduction of data dimensionality aims to map each pattern to a lower-dimensional one, containing only f latent variables, which is able to reproduce the original pattern with minimum distortion through a demapping model. Thus, the original data matrix is mapped to a much smaller matrix of dimension m x f (f {much_lt} n). The dimensionality reduction is useful when there exist correlations among the observed variables; the mapping and demapping models should capture these relationships among the observed variables by relating to the smaller number of latent variables. In this paper, the authors present a new method for reducing data dimensionality using neural networks as nonlinear models between observed variables and latent variables. With this method, only one single-hidden-layer network is needed for dimensionality reduction of a given data set, avoiding the training difficulties of autoassociative networks. The authors demonstrate the use of IT-nets for data-dimensionality reduction, missing-measurement replacement, and gross error detection.
- OSTI ID:
- 376155
- Report Number(s):
- CONF-960426--
- Country of Publication:
- United States
- Language:
- English
Similar Records
A non-linear dimension reduction methodology for generating data-driven stochastic input models