In this thesis have been proposed new detection schemes based on the use of nonlinear prediction, in order to apply them to detect fires in large areas. These schemes have been verified with real data obtained by means of an infrared sensor located in a mountainous area in the vicinity of Alcoi, from the data have been built three blocks of information (A, B and C). The aim, in particular, focuses on determining the ability to resolve a noise between two situations: fire alarms and unwanted, which implies the detection based on a vector (vector of decision). To validate the proposed schemes we have carried out different simulations that model the two types of signals: growing fires and different unwanted alarms (occasional effects) that we can find in a real situation. From a basic prediction, detection, characterized by two main parts: the first corresponds to the stage of prediction (linear type) and the second stage of detection (of the type adapted to filter the signal subspace), we have verified the feasibility of such a scheme. Then the thesis presents improvements over the original scheme. The first, the interest of including non-linearity in the predictor design, mainly because the data do not show a Gaussian distribution. In the thesis presents the structure as a Wiener scheme to construct the nonlinear predictor. The second is the need to include the so-called growth detector to distinguish between increasing trends (possible fire) and decreasing (possible unwanted alarms). The design of the Wiener structure (characterized by a linear plus a nonlinear part) we have developed two techniques for characterizing non-linearity: the direct method and the polynomial approximation. The linear part is designed to apply .........