A carregar...
Publicação
Exploratory and chemometric analysis of stormwater runoff using UV spectroscopy
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 5.12 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Contexto: Este trabalho é um projeto de dissertação de mestrado para examinar a viabilidade do uso de um espectrofotómetro UV, sonda espectrofotómetro spectro::lyser 15 mm, para monitorar águas pluviais.
Objetivo: Verificar a viabilidade do spectro::lyser para monitorar águas pluviais em tempo real; a correlacionar a carência química de oxigénio (CQO) e o carbono orgânico total (COT) com informações espectrais usando algoritmos de regressão; a segregar as informações espectrais como contaminadas ou não contaminadas usando algoritmos de classificação.
Método: Foi realizada uma avaliação do risco para determinar as substâncias que provavelmente contaminam as águas pluviais. Com base nessas informações, foram preparadas amostras sintéticas representativas do escoamento de águas pluviais. Modelos de regressão e classificação foram construídos com os dados de duas fontes: amostras sintéticas e amostras de redes de drenagem pluvial. A regressão foi realizada em dois conjuntos de dados: o comprimento de onda original e as razões de comprimento de onda. Da mesma forma, para a classificação, foram definidos dois conjuntos de dados: definido com base no valor limite de emissão do CQO e definido com classes definidas como 1 para água e 0 para espectros diferentes de água. Para examinar o desempenho do modelo, os dados obtidos das medições contínuas tempo real foram usados como um conjunto de teste. Como alguns contaminantes não absorvem na gama UV, a análise de regressão e classificação foi repetida no conjunto de dados sem os contaminantes não absorventes para examinar o desempenho na remoção.
Resultados: Os modelos de regressão sofrem de problemas de sobreajuste indicando a complexidade do problema e a necessidade de tratar um elevado número de dados na fase de implementação. O desempenho para a classificação baseada no limite de descarga do CQO foi insatisfatório para todos os modelos. Na classificação baseada na diferenciação de água ou não água, todos os modelos tiveram bom desempenho, com regressão logística melhor que os demais. Ao utilizar o conjunto de dados sem dados de contaminantes não absorventes, observou-se uma melhoria no desempenho na classificação e regressão.
Conclusão: Independentemente do tipo de estratégia da metodologia escolhida, a limitação está no fundamento teórico e seletividade da espectroscopia UV. Como algumas substâncias químicas não absorvem luz na gama UV, isso afeta o desempenho dos algoritmos preditivos à medida que são testados usando os dados obtidos do espectrofotômetro UV. Assim, se a Hovione decidir implementar o espectro::lyser como ferramenta de monitorização, tem de ser complementado por um analisador COT ou qualquer outro sensor sensível a contaminantes não absorventes de luz UV.
Background: This work is a master's thesis project to examine the feasibility of using a UV spectrophotometer, spectro::lyser 15 mm spectrophotometer probe, to monitor stormwater runoff. Objective: To check the viability of spectro::lyser for real-time stormwater monitoring; by correlating COD (Chemical Oxygen Demand) or TOC (Total Organic Carbon) with spectral information using regression algorithms; by segregating the spectral information as contaminated or uncontaminated water using classification algorithms. Method: A risk assessment was performed to determine the substances likely contaminating the stormwater runoff. Based on that information, synthetic stormwater runoff samples were prepared. Regression and classification models were then built by merging the data from two sources: synthetic samples and samples from stormwater drainage networks. Regression was performed in two datasets: the original wavelength and the wavelength ratios. Similarly, for the classification, two datasets were defined: set based on the COD discharge limit and set with labels defined as 1 for water and 0 for spectra different from water. In order to examine the performance of the model, data obtained from the continuous online measurements were used as a test set. As some contaminants do not absorb in the UV range, the regression and classification analysis were repeated on the dataset without the non-absorbing contaminants to examine the performance on removal. Results: The regression models suffer from overfitting issues indicating the complexity of the problem and the need for large training data. The performance for the classification task based on the COD discharge limit was unsatisfactory for all the models. In the classification task based on the differentiation of water or non-water, all the models performed well, with Logistic regression performing better than others. On using the dataset without data of non-absorbing contaminants, an improvement in the performance was observed in the case of both classification analysis and regression analysis. Conclusion: Irrespective of the type of modeling strategy chosen, the limitation lies in the selectivity of UV spectroscopy. Since some chemical substances do not absorb light in the UV range, it affects the performance of predictive algorithms as they are trained using the data obtained from the UV spectrophotometer. Therefore, if Hovione decides to implement the spectro::lyser as a monitoring tool, it has to be complemented by a TOC analyzer or any other sensor sensitive to UV light non-absorbing contaminants.
Background: This work is a master's thesis project to examine the feasibility of using a UV spectrophotometer, spectro::lyser 15 mm spectrophotometer probe, to monitor stormwater runoff. Objective: To check the viability of spectro::lyser for real-time stormwater monitoring; by correlating COD (Chemical Oxygen Demand) or TOC (Total Organic Carbon) with spectral information using regression algorithms; by segregating the spectral information as contaminated or uncontaminated water using classification algorithms. Method: A risk assessment was performed to determine the substances likely contaminating the stormwater runoff. Based on that information, synthetic stormwater runoff samples were prepared. Regression and classification models were then built by merging the data from two sources: synthetic samples and samples from stormwater drainage networks. Regression was performed in two datasets: the original wavelength and the wavelength ratios. Similarly, for the classification, two datasets were defined: set based on the COD discharge limit and set with labels defined as 1 for water and 0 for spectra different from water. In order to examine the performance of the model, data obtained from the continuous online measurements were used as a test set. As some contaminants do not absorb in the UV range, the regression and classification analysis were repeated on the dataset without the non-absorbing contaminants to examine the performance on removal. Results: The regression models suffer from overfitting issues indicating the complexity of the problem and the need for large training data. The performance for the classification task based on the COD discharge limit was unsatisfactory for all the models. In the classification task based on the differentiation of water or non-water, all the models performed well, with Logistic regression performing better than others. On using the dataset without data of non-absorbing contaminants, an improvement in the performance was observed in the case of both classification analysis and regression analysis. Conclusion: Irrespective of the type of modeling strategy chosen, the limitation lies in the selectivity of UV spectroscopy. Since some chemical substances do not absorb light in the UV range, it affects the performance of predictive algorithms as they are trained using the data obtained from the UV spectrophotometer. Therefore, if Hovione decides to implement the spectro::lyser as a monitoring tool, it has to be complemented by a TOC analyzer or any other sensor sensitive to UV light non-absorbing contaminants.
Descrição
Tese de mestrado, Engenharia Farmacêutica, 2022, Universidade de Lisboa, Faculdade de Farmácia.
Palavras-chave
Wastewater monitoring COD Regression Classification UV spectroscopy Teses de mestrado - 2022