A BAYESIAN APPROACH TO GROSS ERROR DETECTION IN PROCESS DATA

A new statistical procedure for the detection of gross errors in chemical process data, based on a Bayesian approach, is proposed and tested in the present dissertation.; The statistical tests developed so far for gross error detection were derived from the classical theory of hypothesis testing. The only information used in the classical hypothesis testing is provided by sample data taken at one time instant. If additional information such as the history of occurrence and the order of magnitude of the gross errors are available, the probability of detection of gross errors and their identification may be substantially enhanced. This is done by the Bayesian procedure proposed in the present research.; The Bayesian approach makes use of the prior probabilities of occurrence of the errors and of the current measurements to construct updated probabilities, known as "posteriors". The posteriors serve as basis for the decision criterion concerning the existence and location of the gross errors.; In the present theory we first develop the underlying model and the proposed test for one-time application of the Bayes test. Then, the Bayesian procedure is implemented in a sequential setting, by using a probabilistic model which enables the possibility of updating the prior probabilities of gross error occurrences by exploiting accumulating data collected during the application of the detection scheme. Modifications in the basic model are suggested to take into account unknown magnitudes of gross errors and aging of measuring instruments.; A sensitivity analysis and a performance evaluation of this procedure has been carried out in our work. Since the performance criteria cannot be obtained analytically, computer simulation has been used for evaluation. The performance of the Bayesian detection scheme is compared against that of a classical statistical procedure based on the measurement test. Both detection schemes enable direct identification of the gross errors. The computer results show that the Bayesian procedure performs much better especially when more than one gross error are simultaneously present in the data.