Analysis and rectification of data from dynamic chemical processes via artificial neural networks

Modern chemical plants produce and store a tremendous amount of measurement data collected in the course of operation. The process measurements are typically tagged, time stamped, and saved in a computer data base until an engineer, a researcher, or a business manager requires the historical information for the purpose of adjusting the process operating conditions, analyzing the results of process changes, modeling the behavior of the process, or analyzing the economic return or business strategy associated with the process.; This dissertation was concerned with investigating the potential of recurrent neural networks as tools for analyzing and/or utilizing the measurements generated by dynamic chemical processes.; Data rectification. The raw measurements that are stored in a data historian are often corrupted by various random and non-random errors. The measurements might contain only a small component of random noise associated with the irreproducibility of the measuring device, or the measurements might be corrupted by non-random errors such as biases caused by faulty instrument calibration or installation, or large non-random errors associated with human recording of measurement values, computer failures, power surges, or other problems. The success of applications that depend on the use of historical process data makes the removal or significant reduction of measurement errors important.; The field of data rectification is concerned with the removal of errors from process measurements by statistical analysis, pattern recognition, filtering, and/ or reconciliation with a process model. Recurrent neural networks have been proposed as useful tools for removing the errors present in measurements taken from dynamic chemical processes, and have been employed as one-step ahead nonlinear predictive filters for removing noise from process data. In this dissertation, a new technique for employing recurrent neural networks for data rectification is proposed, implemented, and analyzed. The proposed techniques are based on Basis Pursuit De-noising and are demonstrated to perform well on the input and output variables from several nonlinear processes.; Dynamic process modeling. Dynamic process modeling should certainly be included under the general heading of data analysis. In fact, the ultimate goal of collecting and storing measurements is often the development of accurate process models for predicting the future behavior of the plant based on historical data.; Recurrent neural networks offer the potential of accurately modeling a wide variety of nonlinear dynamic processes. In this work, the utility of recurrent neural networks was investigated for modeling a complex industrial polymerization process using historical data collected from the actual operating unit. Recurrent neural network models were compared to a variety of models obtained via other techniques, and some conclusions were drawn regarding the types of modeling applications in which recurrent neural network models are likely to enjoy the most success or to encounter the most difficulties.; The objective of the polymerization process modeling was to produce an effective model for predicting the end-use property of the polymer product, not to validate a preconceived preference for a particular modeling technique. Therefore, recurrent neural networks were analyzed as objectively as possible with regard to the selection of the best modeling technique for the industrial polymerization reactor.