Abstract: This article studies a LSTM local discharge detection method of the transformer based on ultrasonic signal. First of all, this article systematically analyzes the basic principles of the ultrasonic detection, and explains the collection and pre-processing process of local discharge signals. Secondly, the wavelet transform is used to construct the time-frequency domain eigenvector of the ultrasonic signal, and the LSTM model is optimized to improve the detection effect. The experiment in this article uses the MEMS microphone module of Xunfei Scholars to collect 500 sets of normal signals and local discharge signals, and complete model training and testing in the MATLAB environment. The results show that our method is better than the standard LSTM model in the evaluation indicators such as accuracy, recall and F1 values, and showing significant performance.