Application of neural networks to fluorescent diagnostics of organic pollution in natural waters

The use of a neural net in a sea water pollutant rapid diagnosis system is described. The neural net classifies a sea water pollutant on the basis of a total luminescent spectroscopy (TLS) spectrum and is insensitive to the dissolved organic matter (DOM) spectrum variations. The gradual complication of task during learning is used to reach the minimal decision threshold value. The net gives adequate answers to presentation of a mixture of pollutants spectra, or spectra of unknown substances. The three-step determination of pollutant concentration comprises classification of a pollutant by the basic net, its identification by an auxiliary net, and concentration determination by a linear neural net with a typical accuracy of 0.05 ppm. It is shown that the use of a net with two hidden layers for classification of TLS-spectra of low resolution allows one to achieve classification thresholds close to those of standard TLS-spectra