ABSTRACT
      In the last years electronic tongues have become an excellent alternative to traditional analysis methods to control processes and products, among others in the food and agriculture field. These are systems which through electrochemical techniques, such as the potentiometry or voltammetry together with multivariable analysis tools, are able to classify samples and quantify their physicochemical parameters. Their performance is based on the use of sensors of cross-sensitivity, allowing to measure samples in which there are interferences between the different compounds integrating them.
      Currently most of the methods used to determine the physicochemical properties are destructive. The design of non destructive measure systems is a challenge. The new analytic techniques must not only preserve the integrity of the samples analyzed but also have a low cost and a simple performance, non dependant on qualified workforce.
      To carry out the analysis of the data it is common to use non supervised techniques of recognition of patterns, like Principal Components Analysis (PCA). But frequently it is convenient to realize supervised analysis, where the categories of the samples are predefined, being the aim to check if it is possible to obtain a new measurement system. One of the most used methods to carry out a classification of the samples with supervised techniques is the artificial neuronal nets (ANN).
      There are several types of neuronal networks; one of the most well-known is the so-called    Perceptron multilayer. The training of this red consists of fixing the weights of every neuron. This type of neuronal net has checked its usefulness in several applications with electronic tongues, but it has also shown its limitations, which are principally determined by the complexity of the mathematic algorithm that implements it, the high computing cost when having a big size, and the opacity of the internal calculus to determine the weight of the neurons. Those nets using the so-called adaptive resonance technique (ART) arose in order to try to supplement some of the problems of the Perceptron nets.
      In order to carry out the analysis of the structure of Perceptron multilayer and Fuzzy Artmap neuronal networks and their application to a microcontroller (µC) data from different measurement equipment have been used to realize the training of the RNA and subsequently implement it in a µC.
      In the first type of experiences, data come from a device designed to calculate the coefficients of extinction of the solar light in the water columns with five different wave lengths (three for the visible range, UV and IR). The perceptron RNA is implemented as a routine of additional software, being trained in a computer by means of MATLAB   and using real data of the photo-sensor unit, allowing to determine the called depth of the Secchi disc.
      In the second type of experiences, polyester films have been used as substrates obtaining humidity sensors. The sensor obtained by this means presents a non-lineal answer. The perceptron RNA implemented in the µC allows a lineal answer of the sensor.
      In the third type of experiences, data come from the potentiometric measurements carried out with different types of drinking waters that have different concentrations and types of salts. Measurements of four types of honey have also been made. The Fuzzy Artmap  RNA, implemented in the µC,  is able to realize a classification of the samples of water or honey.
      In the fourth type of experiences data obtained by pulse voltammetry have been used to realize a classification (Fuzzy Artmap) and prediction (three perceptron multilayer networks) of the concentration of glyphosate.
      The measurement systems used in the different experiences have been developed ‘ad hoc’, from an ‘electronic eye’ in the case of Secchi, up to ‘electronic tongues’: potentiometric or voltammetric.
      To sum up, the previous selection of variables results in a reduction of the amount of memory necessary in the µC and a significant increase in the corresponding results of classification.