Abstract: With the rapid growth of power grid load, the heavy overload operation of transformer has become a serious problem restricting the safe and stable operation of power grid. The traditional monitoring methods based on parameters such as electrical quantity and temperature have a certain lag. Therefore, this study proposes an abnormal identification method of transformer heavy overload state based on voiceprint feature and deep learning. A high fidelity acoustic sensor is installed on the transformer body to collect the vibration and noise signals during the operation of the transformer. Using the improved Mel frequency cepstrum coefficient and power spectral density combined with the multi-dimensional voiceprint feature extraction method, the voiceprint feature signal with large discrimination is obtained. One dimensional convolutional neural network (A-1D-CNN) classifier optimized by attention mechanism is used to recognize and classify normal heavy load and abnormal heavy load conditions. The results show that the recognition accuracy of abnormal winding looseness can reach 97.8%, which is higher than other models. This research provides a new noninvasive intelligent diagnosis method for transformer condition based maintenance.