| Water eutrophication is one of the main water pollution problem,facing the worldtoday.Since the excess nitrogen is the key reason of eutrophication,detection of the totalnitrogen concentration is important for understanding and study of the water environmentquality. Taking into account the characteristics of the natural water environment,afterproceeding an oxidation process on the water sample,this thesis combines the waveletdenoising and the support vector regression to establish the prediction model for water test.The main contributions of this thesis are summarized as follows:1Considering the oxidizing digestion method used in the National Standard requires toconsume chemical oxidation reagents, resulted in secondary pollution on the environment,aset of advanced oxidation method by combination of ozone, ultraviolet,high voltage staticdischarge and other oxidation technologies is proposed in this thesis.2To avoid the disadvantages of the soft threshold denoising method and the hardthreshold denoising method,the spatial correlation based wavelet denoising method isintroduced into the soft threshold denoising method,and an improved thresholding waveletdenoising method is proposed in this thesis; Furthermore, the proposed method has beencompared with the optimized spatial correlation wavelet method is optimized, and the twomethods are compared in terms of the denoising effect.3Considering the high dimensional characteristic of the quality spectral data thesemi-supervised affinity propagation algorithm, successive projections algorithm, partial leastsquares algorithm and uninformative variable elimination algorithm are applied for spectralband selection, then each sub-model is established by using the support vector machine. Toimprove the precision of the single modeling method,the sub-models are further combined topredict the water quality.4Since the traditional water quality detection method based on continuous spectrumanalysis is often complex and can introduce unknown substances to interfere the spectralinformation an ultraviolet absorption spectrum analysis based local linear embedded(LLE)-support vector regression modeling(SVR) method. is proposed to predict the totalnitrogen content in the water. The basic idea is to use the LLE algorithm for nonlineardimensionality reduction of the ultraviolet absorption spectrum data, and then use the SVRmethod for modeling. The experimental results show that: the proposed method is effective toreduce the model complexity,and improves the prediction accuracy.5The same as other learning algorithms, the performance of the support vectorregression machine depends on its parameters,different parameter can result in differentgeneralization capability.To relieve this situation,the hybrid particle swarm optimizationalgorithm is proposed to optimize the mixed kernel function parameters of the water quality prediction model. The proposed method can automatically select the kernel parametervariables of the mixed kernel function.Furthermore,the experimental results show that,thismethod can significantly improve the prediction accuracy. |
