In order to solve the problem that the accuracy of word segmentation in power defect descriptions is not good and the single neural network model has its own shortcomings , a determination method of defect grades in electrical equipment based on combination neural network optimized by attention mechanism is proposed in this paper. The distributed character granularity vector is used for representation of power defect descriptions. The local features and sequence features of power defect descriptions are extracted by using the convolutional recurrent neural network which is composed by convolutional neural network and bidirectional long short-term memory network. The attention mechanism is used to assign weights of the semantic features obtained by the combination neural network, so as to reduce the loss of key features and further enhance the influence of key information on the lassification results. Taking 110 000 defect description data of Yunnan Power Grid Company from 2014 to 2019 as experimental objects, the Acc, MF? and WF? values of the method proposed in this paper are 0.927 5, 0. 9112 and 0. 927 5, which illusrates that the proposed method is effective and feasible in the determination of the power defect grades, and provides help for itellgent operation of power grid.