Models of steel mass loss by atmospheric corrosion in Colombia using

In order to classify the corrosivity of the different Colombian atmospheres, as part of an extensive research project [1], plates of carbon steel were placed in 21 stations spread along the country electrical infrastructure (transmission lines and substations). There were measured among others at these stations, the time of wetness and deposition of sulfates and chlorides for 12 months, in addition steel plates were taken bimonthly to the laboratory in order to measure the mass loss suffered by these during the time of exposure. The classification of the 21 stations was done in 4 groups, considering the time of moisture, content of chlorides and sulfates, height above sea level and the plates exposure time; these are considered linearly independent variables according to the implemented technique of decomposition unique values (DPS). The criterion used for classification was the similarity of the variables using the Euclidean rule considered in the Kohonen unsupervised neural network. Additionally, models were implemented for the steel mass loss for each one of the groups using feed forward neural networks (RN), defining the above variables as inputs and the mass loss as the output. Besides, the comparison between the RN model for the group 1, with other models using genetic algorithms (GA) and the Simplex method is presented.