Abstract:In view of the shortcomings of inputs redundancy and low fault recognition rate existing in rolling bearings fault diagnosis technology, a method was proposed based on improved neighborhood rough set and s〖BF〗_〖BFQ〗kononen neural network. Since the fault information obtained by sensors mostly was numeric data and the dimension was always very large, the neighborhood rough set was imported, and the forward greedy algorithm was improved which was based on it. The improved method was applied to reduce the original failure data, which reduced the algorithm complexity greatly. The Kohonen neural network was improved, adding an output layer based on the original structure to compel the output classifications to meet the given requirements and making the unsupervised neural network into S〖BF〗_〖BFQ〗Kohonen neural network. Then forward greedy algorithm and improved algorithm were used to divide the original failure data into two parts: one is not reduced, the other is reduced. The two parts were extracted as the inputs of S〖BF〗_〖BFQ〗Kohonen neural network and BP neural network to identify fault state of roller bearing. The test result illustrates that neighborhood rough set can efficiently eliminate duplicate information between attributes, the fault information extracted by improved method can reflect even better the essence of fault state, and S〖BF〗_〖BFQ〗Kohonen neural network has a good ability to identify faults. Combined both approaches for applicaton, this model has a good capability of fault diagnosis.