Fuzzified neural network for similar/dissimilar sensor fusion

We explore the robustness of a sensor fusion system as a function of failed sensors. Neural networks are applied to classify data from a sensor suite. Two dissimilar sensor types are used to produce three spectral patterns (in red, green, and blue wavelength regions) per sensor location (three sensor locations were used). The main goal of this effort is to improve the sensor fusion confidence level by introducing several realizations of a neural network. Each specific neural network realization is activated upon a specific sensor failure configuration during the recognition stage. In such a case, the number of NN realization is equal to the number of failed sensor combinations. To reduce the number of NN realizations, fuzzification of the NN weights is proposed. An experimental demonstration of the proposed concept is also included.