The parameters of transformer is complex, and many state quantities are difficult to collect in daily operation inspection, which increases the difficulty of transformer comprehensive state evaluation. The core of state evaluation is the processing and mining of multi-source heterogeneous state parameter data. Because the transformer data collected at present are often low dimensional and nonlinear. The existing data processing methods cannot perform data cleaning and filtering well. In order to solve this problem, the paper uses KPCA algorithm to map high-dimensional space to eliminate invalid information in massive data through data cleaning and filtering, and extract effective information from it to further improve the accuracy of state evaluation.