To address the issues of the large number of bolts on transmission towers and the low efficiency of manual looseness detection, this paper proposes a sound-based method for detecting bolt looseness on transmission towers. The method establishes a nonlinear mapping model between Mel-frequency cepstral coefficients (MFCC) acoustic features and the preload state of bolts, revealing the intrinsic relationship between the degree of bolt looseness and the frequency-domain energy distribution; A multi-indicator comprehensive evaluation system was established. Under two sample imbalance ratios (4:1 and 2:1), the detection performance of Random Forest (RF), Extreme Gradient Boosting Tree (XGBoost), Deep Forest (DF), Support Vector Machine (SVM), and Residual Network (ResNet) was systematically compared, and the differences in applicability between traditional machine learning and deep learning models in low-sample-size scenarios were analyzed. Finally, the feasibility of this method was validated through field tests, and key factors influencing acoustic recognition performance were investigated. The results indicate that: in the case of a single loose bolt, the SVM model achieved the highest recognition accuracy (up to 96.1%), with an AUC of 0.918, approaching 1, demonstrating the model’s robustness under imbalanced sample conditions; in scenarios involving multiple loose bolts, the coupled vibrations of multiple bolts enhance acoustic features, enabling MFCC to capture more distinctive characteristics, and the overall recognition accuracy further improves to 98.2%. The installation location of the bolts and the excitation point have no significant impact on detection accuracy, whereas an increase in signal acquisition distance leads to a significant decrease in recognition accuracy. This study provides a novel technical approach for the rapid, automated detection of bolt preload conditions in transmission towers.