The bolt loosening detection method based on acoustic signals is characterized by convenience and high efficiency. However, existing methods for bolt group loosening problems suffer from insufficient detection precision and lack fine-grained descriptions of loosening states. Focusing on typical bolted connection nodes in transmission towers, 649 audio samples corresponding to 48 types of loosening states were collected through tapping. A total of 6 087 augmented samples were obtained by adopting data augmentation techniques such as pitch shifting, time stretching, and noise addition. The time-frequency features of the tapping acoustic signals of bolt groups were extracted based on Mel frequency cepstral coefficients (MFCC). The discriminability of feature channels was enhanced by integrating the squeeze-and-excitation (SE) attention module, and multi-scale time-frequency information was fused using the Inception module. A multi-task convolutional neural network-recurrent neural network (CNN-RNN) architecture was established, and a bolt group loosening state detection method based on MFCC features was proposed. This method achieved accurate multi-dimensional identification of the loosening area, the number of loose bolts, and the loosening positions of bolt groups. Experimental results show that the proposed method achieves average accuracies of 98.13% and 92.20% in the tasks of area identification and loose bolt count recognition, respectively. Meanwhile, the Top-3 coverage rate reaches 91.86% in the loosening position determination task. The research results provide an efficient bolt loosening detection approach for the operation and maintenance of transmission lines.