다변량 통계분석을 이용한 산업폐수 처리공정의 통합 모니터링

Abstract

As the size of collected database which should be managed in modern industrial wastewater plants becomes larger and complexity of the involved processes is increased, an efficient and structured approach for monitoring and analysis of process behaviors is being highly demanded. Multivariate statistical data analysis can provide wide range of methodologies to extract information from large amount of data and application of multivariate statistical analysis to the wastewater treatment processes seems very promising. In this study, we tried to identify challenges and difficulties encountered in the implementation of multivariate statistical process monitoring methods into industrial wastewater treatment processes and constructed reliable multivariate statistical process monitoring methodology utilizing various kinds of process measurements. To achieve these objectives, both real-time fault detection algorithms and multivariate calibration methods were employed together to provide complementary information about chemical and physical states of the process. In order to develop advanced process monitoring and fault detection algorithms, multi-scale and adaptive PLS algorithms were studied. A multi-scale process monitoring scheme assumes that the most processes have multi-scale nature due to the contributions of the events occurring at different locations and with different localizations both in time and frequency and these issues are highly important, especially, in the wastewater treatment processes. On the other hand, an adaptive/recursive process monitoring scheme can provides efficient ways to update the model by to incorporate newly available measurements into current model structures. Therefore, both properties were complementary to each other and should be met when we want to implement data-based fault detection and monitoring system to the wastewater treatment plant. In order to monitor key process components, systematic and statistics-oriented methodologies were studied to construct reliable calibration models for spectrometry-based chemical measurements. To achieve this sub-objective, we sequentially optimized the individual steps involved in the multivariate calibration for improving generalization ability of the calibrated model. These steps included selection of regression algorithm, data pretreatment, feature selection, etc. During the step-by step optimization procedures, wavelet packet transform and multi-objective genetic algorithm were employed as powerful tools to improve accuracy and robustness of the multivariate regression model. Finally, the integrated monitoring modules, which combine all the developed methods in a unified platform, were developed as a software package and implemented into the field plant, realizing the practical advantages of the proposed scheme. When considering that most industrial wastewater treatment processes, by their inherent nature, are not precisely defined because the underlying physical and chemical states cannot be fully described, it could be expected that both approaches provided complementary information about chemical and physical states of the system and integrating them were crucial to develop the reliable monitoring system.