Abstract: With the concept of optimal classification lines,a pattern recognition method,which uses interface geometric discriminant to generate a pattern classifier,was proposed.Major procedures of the method include:mapping multidimensional inputted characteristic vectors of different pattern classes to a 2-dimensional(2D) discriminant space with a BP neural network which is characterized by its high nonlinear mapping capability,extracting a polygon axis of the polygon which is formed at the interval clearance space among pattern classes,and constructing a decision-making boundary for pattern recognition by extending polygon axes to all discriminating domains.The method was tested in a case study of fault diagnosis for double row tapered roller-bearings used in railway wheels.The result shows that the proposed method can construct decision-making boundaries for different fault patterns on a 2D discriminant space,which provides a condition to operators for intuitive recognition of fault classifications in practice.