Abstract:Power grid enterprises have a large number of unstructured text data recorded in Chinese, in which contains a lot of im-portant reliability statistical information. However, it is not only inefficient, but also accurate to different people to mine the unstructured text data manually. Therefore, how to excavate the important reliability statistics information in the equipment defect text of power grid enterprises effectively and accurately and intelligently becomes a problem to be solved at present. In this work, we use was used to segment the unstructured text data collected by the whole process of technical supervision was segmented based on the modified Hidden Markov Model(HMM) algorithm, and the structured expression rules of unstructured data was formulated. Natural Language Processing(NLP) algorithms, like Principal Component Analysis(PCA), vector and depth neural network were used to calculated the semantic similarity of words, synonyms and synonyms after dimensionality reduction classification clustering through K Nearest Neighbor algorithm. Relaying on the works, the problems of defects of the text sentence and dividing the digital quantity accurately has been solved, and a data dictionary in state grid operation and maintenance field was created, which has important significance and contribution to the future expansion of technical supervision work.