Abstract: Accurate and rapid identification of abnormal events in the power system is the key to ensure the safe and stable operation of the power grid. Aiming at the shortcomings of traditional threshold method and shallow machine learning model in complex timing signal feature extraction, transient anomaly capture and multi-type fault discrimination accuracy, this paper proposes a new recognition framework based on multimodal feature fusion deep learning. Firstly, a mixed feature engineering method is designed to simultaneously extract the time-domain statistical features of power signals, frequency- domain features based on wavelet transform, and gramian angular summation field.The generated image features are encoded to fully mine the multi-scale information of the signal. Furthermore, a two-branch hybrid deep learning model is constructed, in which the feature engineering branch uses the gated recurrent unit to capture the long time series dependencies, and the image processing branch uses the convolutional neural network to extract the local spatial pattern. Finally, high-dimensional feature fusion and classification are realized through feature splicing and fully connected layers. The experimental results based on the IEEE 39 system simulation data demonstrate that this method exhibits good computational speed, robustness, and potential for engineering applications.