Adaptive Rectification of Data from Nonlinear Dynamic Processes via Recursive Neural Networks

点击翻译按钮获取中文摘要

A problem in many areas of science and engineering is that of data rectification. For example, in a plant a wide variety of measurements are made of the process variables for purposes such as cost accounting, process control, statistical quality control, and performance evaluation. Not all of the possible variables are measured, some of the measurements are defective in the sense that gross errors creep into the measurements because of instrument degradation (or human error), noise exists, and unmeasured disturbances enter the processes. The objective of data rectification is to provide the best estimates of the process variables that satisfy the process model assuming the process model is an accurate one - which may or not be the case. If flawed information is used for process control or cost accounting, process performance can be degraded, perhaps substantially. Data rectification involves not only reduction of noise but removing of other types of contaminants in the data, particularly gross errors. A problem with some existing techniques for data rectification in dynamic processes is their inability to detect gross errors. Often a process model of a complex nonlinear process may be inadequate to represent the dynamics of a process. Model mismatch prevents correct detection of disturbances and outliers, and rectified variables that satisfy an erroneous model will include systematic errors. The use of artificial neural networks (ANN) for data rectification can ameliorate this problem because the mapping they accomplish can easily be of higher quality than that yielded by models based on physical principles or regression analysis. At the same time, an ANN could serve as a filter to damp out random noise and gross errors. This research involves the investigation of the possible uses of recursive ANN to implicitly model dynamic chemical process equipment with the goal of data rectification. The PI plans to generate deterministic simulations of simple and complex nonlinear dynamic processes (so that the true values of the process variables are known), and to these values add various kinds of random noise and gross errors to generate simulated measurements akin to real process data. Scaled simulated data will become input to various kinds of recursive ANN, and the outputs of the nets will be the rectified values of the variables. The effect of: (1) the magnitude and frequency of the gross errors, (2) the degree of auto and cross correlation in the noise, (3) model mismatch between the net and the true model, (4) missing measurements, (5) the sampling rate and delay in the measurements, and (6) the appropriate types of recursive nets will be evaluated.***//