An Applied Research On Information Fusion And Ensemble Learning For Spectral Analysis Of Water Quality

According to the analysis principles, the existing analysis methods of the comprehensive organic contaminant indexes of water can be divided into two categories: chemical analysis method and optical analysis method. Compared with chemical analysis method, the optical analysis method which is based on optical principles has some notable advantages, such as, high analysis speed, no chemical reagent pollution, easy operation and maintenance with low cost. Therefore, the optical analysis method is more suitable for the application on automatic monitoring system for environment pollution and represents the development direction of the "Green" environment monitoring. However, this analysis has one disadvantage, that is, the analysis accuracy is relatively low at the present stage. This disadvantage is becoming the "bottlenecks" to the development of this method, and obstructs its application in the environment monitoring system. This thesis deals with the issues on the optical analysis method of the comprehensive organic contaminant indexes of water. Specifically, it focuses on how to improve the accuracy of the method via adopting information fusion and ensemble learning techniques. The following contents have been presented.1. After collecting a group of representative surface water samples, a detail analysis of the optical characters of the primary ingredients of these samples has been carried out through two aspects: the analysis object and the analysis method. Subsequently, two widely adopted spectral analysis method, i.e. ultraviolet absorption spectrum and the three-dimensional fluorescence spectrum, for water quality have been compared and their advantages as well as disadvantages have been identified.2. Considering the interferential signals and noises in the water samples, wepresent a features extraction method, Improved ICA with Reference (IICA-R) method, as an effective approach to reduce noise. Prior knowledge extracted from the comprehensive organic contaminant indexes of water has been incorporated into traditional ICA, thereafter a two-level architecture has been introduced to effectively extract more effective features. Computational results of the surface water samples show that IICA-R is valid.3. To address the low accuracy problem of the existing single-spectral water quality analysis methods, a novel method based on multi-spectral information fusion for water quality analysis has been presented. After analyzing the characteristics of spectral fusion, a specific fusion strategy has been presented which combines feature extraction based information fusion and support vector machine that possesses good generalization capability, and the problem of information cover is avoided.4. Multi-spectral information fusion requires optimal selection of models involved. The thesis addresses this critical problem in an ensemble learning approach. A boosting method based on multi-spectral modeling, Boosting -LSSVM, has been proposed. A stopping criterion of iterations has been derived ensuring model precision by optimizing the parameters of the models automatically.5. The above proposed methods have been applied to the analysis on collected surface water samples, experimental results confirm the validity of the methods proposed in the thesis.