In order to excavate security threats in power grid by making full use of heterogeneous data sources in power information system, this study proposes a multi-source log comprehensive feature extraction method based on Restricted Boltzmann Machine (RBM). Firstly, the RBM neural network is used to normalize coding all kinds of log information. Then, the contrast divergence fast learning method is used to optimize the network weight, and the stochastic gradient rise method is used to maximize the logarithmic likelihood function for the training and learning of the RBM model. The data dimension reduction is realized by processing the normalized coded log information. At the same time, the comprehensive features are obtained, which can effectively solve the problems caused by the heterogeneity of log data. The big data threat early warning monitoring experimental environment was set up in the power information system, and the comprehensive feature extraction and algorithm verification of the security log were carried out. Experimental results show that the proposed method can be applied to all kinds of security analysis, such as clustering analysis, anomaly detection, etc., and it has high accuracy in extracting log features in power information system, which improves the speed and accuracy of network security situation prediction.