GIS mechanical defects seriously threaten the safety of power grid. It is of great significance to deeply understand its characteristics and defect diagnosis methods to ensure the safety of power system. The traditional threshold method is limited by fixed parameters of feature extraction method, which is difficult to deal with complex vibration characteristics, and the deep learning model lacks physical interpretation. To overcome these limitations, this study uses a STFT primary function to construct a frequency band adaptive convolution kernel, and proposes a neural network combined with a convolutional layer embedded with time-frequency transform, which dynamically decouples the time-frequency energy of the vibration signal through convolution, and feed the obtained time-frequency physical information into a simplified 1-D CNN model to diagnose the defect. The method improves the time-frequency feature extraction ability and the physical interpretability of the model through the frequency band attention of the convolutional layer. The result demonstrates that the defect classification accuracy of the proposed method exceeded 98% in mechanical defect classification task, which provides a reliable and interpretable method for GIS mechanical defect diagnosis.