A fusion toolbox for sensor data fusion in industrial recycling

Information from different sensors can be fused in various ways. It is often difficult to choose the most suitable method for solving a fusion problem. In a measurement situation, the measured signal is often corrupted by disturbances (noise, etc.). It is, therefore, meaningless to compare crisp values without the corresponding uncertainty intervals. This paper describes a toolbox including nine different fusing methods. All methods are applied on training data, and the most suitable method is then used for solving the real fusion problem. In the example, fusion is performed on data for classification in an industrial recycling operation. The data is from different vision systems and an eddy current system. The fusion methods included in the toolbox are fuzzy logic with triangular and Gaussian shaped membership functions, fuzzy measures with triangular and Gaussian shapes, Bayes' statistics, artificial neural networks, multivariate analysis (PCA), a knowledge-based system, and a neurofuzzy system