Modelos de exposición personal a los contaminantes del aire : aplicación de modelos estadísticos clásicos y de red natural

Abstract

El objetivo de este proyecto es el estudio de las aplicaciones de las redes neuronales en el campo de la exposición personal y su comparación con los resultados obtenidos con el método tradicional.

Pare ello hemos utilizado los datos procedentes de dos estudios distintos que se han realizado. Estos estudios son el proyecto Europeo LIFE-MACBETH (1996) y el proyecto Europeo PEOPLE. Estos estudios se realizaron en varias ciudades Europeas pero para este trabajo se han recopilado los datos de la ciudad de Murcia del proyecto LIFE-MACBETH, donde se estudió la exposición personal a Benceno, Tolueno y m, p-Xilenos. Y los de la ciudad de Madrid (2003) donde se midió la concentraciones en aire ambiente y la exposición personal, de los datos que se recopilaron hemos utilizado los correspondientes a Benceno y compuestos aromáticos.

El método clásico utilizado es el de regresión multivariable, que es un método estadístico que establece una relación matemática entre un conjunto de variables independientes y una variable dependiente. El procedimiento utilizado con esta regresión fue el llamado "paso a paso".

Las redes neuronales ensayadas han sido: capa lineal, las dos versiones diseño y entrenamiento; Feed-forward Backprop. Cascade-forward Backprop., la base radial pura, la probabilística y la regresión generalizada y también la opción "fitting tool".

Del proyecto LIFE-MACBETH pudimos concluir que:

• De los métodos aceptables clásicos estadísticos fueron aquellos en donde se relacionó la concentración de exposición personal con los distintos tiempos de exposición a cada ambiente y las concentraciones de los distintos ambientes. Abstract

The aim of this project is to study the applications of neural networks in the field of personal exposure and to compare the obtained results with the traditional method.

For this application we used data from two different studies. These studies are the European LIFE-MACBETH project and the European PEOPLE project. These studies were performed in various European cities, but for this study, we have collected from LIFE-MACBETH project the data of the city of Murcia, where personal exposure to Benzene, Toluene and m, p-Xylenes was studied. And to the PEOPLE project the date of city of Madrid (2003) where concentrations were measured in ambient air and personal exposure. From the data collected, we used those for Benzene and aromatics.

The classic method used is the multivariate regression, which is a statistical method that establishes a mathematical relationship between a set of independent variables and a dependent variable. This regression was made using a called "stepwise” procedure.

The neural networks tested for this project has been the following: Linear layer, both versions design and training; Feed-forward backprop; the Cascade-forward backprop, the pure radial basis, probabilistic and generalized regression and also the option “fitting tool”.

The date of the project LIFE-MACBETH several conclusions are concluded:

• The classical statistical methods acceptable were those in which personal exposure concentration with different times of exposure to each environment and the concentrations of the different environments related.

• Classical statistical study shows that the environments with more influence on personal exposure to the three pollutants are the time spent outdoors and at home (in all the evaluated periods). Also, during periods of weekend and leisure the influence of interiors other than homes is significant.

• In case of neural network models, tested options have been considered acceptable methods for linear layer design and generalized regression in which a neural network of four neurons in parallel was designed.

• In general, the error obtained with neural networks is smaller than those obtained with the conventional method, especially for Toluene and m, p-Xylenes.

From the results with the data of the project PEOPLE several conclusions are concluded too.

• The classic study it is deduced that the environments more influence on personal exposure to the five pollutants are time spent at work and at home; for four of the pollutants also affects the time spent in the car (except for the o-Xylene); for three of them also influences the time spent shopping (except Benzene and m, p-Xylene). For Benzene and m,p-Xylenes other relevant variable is the number of cigarettes smoked.

• With all the contaminants, five neural network simulations and one with the classical statistical multivariate regression method have been made. The simulation for the use of the generalized regression network in two-step option was not acceptable because the percentage of deleted data was very high.

• For all pollutants the Networks option Linear Layer Design (LLD) 1step improves the classical method. Also Generalized Regression (GR) 2 steps shows an improvement over the classical estimates. The GR 1 step serves to refine the data, detecting outliers.