China is building a new type of power system mainly based on new energy. The large-scale access of new energy, the complex structure of AC-DC hybrid grid, and the application of power electronic devices make the grid faults show more complex modes. Fast and accurate diagnosis of grid faults is a necessary condition to guarantee the reliable operation of power grid. In this paper, based on the theory of equipment operation state diagnosis algorithm of multimodal data fusion analysis of power system, we add the bidirectional long and short-term memory network with attention mechanism, respectively, through Word2Vec and Fast Fourier Transform, extract text and audio data features of power industry, and fuse the extracted multimodal data features. The XGBoost decision tree algorithm is used to achieve the training objectives such as data prediction and pattern recognition. From the extracted time-domain plots, it can be seen that the amplitude of the windings fluctuates between -0.9~0.6m·s-2 before loosening, and the vibration signals after loosening are between -1~1 m·s-2, with obvious amplitude variations when a fault occurs. The error of the prediction curve when a fault occurs suddenly becomes larger, and the error values of the two are 60.491 and 45.469, respectively, and the prediction model proposed in this paper has a high monitoring accuracy for normal operation of the power system.
