1999 年 65 巻 633 号 p. 1946-1953
When using computer for pattern recognition, excellent featrue parameters are necessary, by which patterns can be precisely distinguished. Currently, there is not acceptable method for extracting optimum feature parameter. For overcoming this difficulty, this paper proposes a new method called "self-reorganization of feature parameters" in frequency domain by genetic programing. When patterns cannot be precisely recognized by using conventional feature parameters which are also called "primitive feature parameters" in this paper, the primitive feature parameters can be automatically reorganized into a new one by using genetic programming. The new feature parameter can be used to distinguish the patterns sensitively. For raising the accuracy of the recognitions, the noise contained in the pattern signal can be canceled by the method called "standard spectrum ratio" proposed in this paper. The distinction index has been defined by statistical theory to evaluate the goodness of a feature parameter, and it can also be used for the fitness in the genetic programing operations. By using the method to many practices, the optimum feature parameter can be quickly discovered, and the distinction rate of the optimum feature parameter is raised much higher than that of any primitive feature parameter. Examples of failure diagnosis for a gear equipment are shown to verify the efficiency of this method.