FUN Papers Analysis of operating data for evaluation, diagnosis and control of batch operations Bhavik R. Bakshi”, Georg Lecher, Gregory Stephanopoulos and George Stephanopoulos Department of Chemical Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139, USA Received 3 January 1994; revised 13 June 1994 Correct interpretation of measured process variables is essential for the evaluation, diagnosis and control of batch operations. The extraction, though, of the ‘pivotal’ temporal features from operating data is not a trivial task for two reasons: the localization in time of the operating features, and the multi-scale character of operating trends. This paper introduces an integrated systematic approach, based on the wavelet decomposition of batch records of operating variables, which accomplishes the following goals: (1) It extracts the temporal features of the process trends at various time-scales. (2) It generalizes the multi-scale description of operating variables, by identifying the common features among many records of these operating variables. (3) It establishes relationships among the temporal patterns of operating variables, thus leading to pattern-based diagnosis and control of batch operations. Data from an industrial batch fermen- tor are used to illustrate the ideas of the proposed approach and its value for evaluating batch operations. zyxwvutsrqponmlkjih Keywords: fault diagnosis; supervisory control; pattern recognition Batch manufacturing of specialty organic chemicals, bio- technological products and materials has been a fast growing segment of the chemical processing industry. Batch plants allow the cost-effective production of a broad array of traditional or newly developed products and enable rapid response to changing market con- ditions. Thus, the same pieces of equipment may be used for a variety of different processing operations (e.g. a vessel could be used for the dissolving of solid raw mater- ials, the carrying out of a chemical reaction, the extrac- tion of by-products and the crystallization of a product) under different operating conditions, and the manufac- turing of a variety of products. Such multi-purpose, multi-product definition of a batch plant poses certain special challenges and requires the development of special techniques for the monitor- ing, analysis and control of batch operations, which are quite distinct from their counterparts employed in con- tinuous processes’. Specifically, batch operations are essentially dynamic operations over a broad range of operating conditions. Consequently, conventional diag- nosis and control techniques for continuous processes which rely on the stationarity of a nominal steady state are not applicable. Furthermore, the physico-chemical phenomena occurring in a batch process have to be *Present address: Department of Chemical Engineering, Ohio State University, Columbus, OH 43210, USA 0959-1524/94/04/017%16 0 1994 Butterworth-Heinemann Ltd understood and modelled over a wide range of operating conditions and not at a single steady state. For example, in the course of batch fermentation the micro-organisms go through the phases of growth, production and decay, all of which are driven by distinct mechanisms that must be understood and modelled individually. As a result, batch operations are not usually as well understood and modelled as their continuous counterparts. In addition, the short lifetime of certain batch-manufactured pro- ducts does not justify the expenditure of resources for the modelling tasks required. Given all the above limitations, the development of process models and their use in the deployment of pro- cedures for the analysis, evaluation, diagnosis and control of batch operations is a very challenging task. It tends to rely more on data-driven generation of empirical models and the ensuing development of methodologies which are based on such models. Thus, the correct moni- toring, analysis and interpretation of batch operating data is more pivotal for the efficient execution of batch operations than for continuous processes. Measurement of batch operating conditions yields records of data which reflect the underlying events that influence the value of process variables, i.e. the dynamic physico-chemical phenomena. In addition, the values of measured variables reflect the contributions from a var- iety of external factors, such as: sensor noise, disturb- ances due to the variation in the quality of raw materials, J. Proc. Cont. 1994, Volume 4, Number 4 179