Power grid control alarm recognition is an important aspect of realizing smart grid dispatching. In order to improve the accuracy of power grid control alarm recognition, in view of the huge amount of grid data, the difficulty of extracting effective information, and the poor ability of traditional knowledge base knowledge migration, a power grid control online alarm recognition method based on BERT-DSA-CNN and knowledge base is proposed. First, using natural language processing-deep learning text data mining architecture, after the steps of word segmentation and removal of stop words, the BERT model is used to obtain the word vector of the power grid control warning information. Then the word vector is input into the CNN deep learning model for training, and the DSA mechanism is introduced according to the characteristics of the power grid warning information. Finally, an online warning recognition scheme for power grid regulation is proposed, one which combines the deep learning model and the traditional knowledge base. Through the analysis of a large number of calculation examples, it is concluded that this method has higher accuracy and effectiveness than Word2vec, traditional CNN, traditional knowledge base, offline learning and other methods, and has better recognition ability for different types of faults, providing a basis for engineering application. This work is supported by the National Key Research and Development Program of China (No. 2018YFB2100103).