Based on Φ-OTDR, distributed fiber optic sensing technology can be utilized to achieve online, real-time health monitoring of power transmission lines by analyzing the vibration characteristics of optical fibers within OPGW. By utilizing Φ-OTDR, vibration signals under three operating conditions—no icing, level 1 icing, and level 2 icing—were collected, and the temporal, frequency, and time-frequency domain features, along with their corresponding statistical properties, of both phase and amplitude signals were thoroughly explored. To enhance the accuracy of icing condition identification, this paper proposes an optimal feature subset selection algorithm based on LCPSO-AdaBoost-MCG. This algorithm employs the classification error rate of the AdaBoost-MCG as the fitness function and iterates with the LCPSO to calculate the optimal feature subset. The AdaBoost ensemble incorporates four weak classifiers: Simple Cartesian Network (SCN), K-Nearest Neighbors (KNN), Probabilistic Neural Network (PNN), and Support Vector Machine (SVM), forming a heterogeneous strong classifier. By leveraging the strengths of each weak classifier, the model′s generalization performance and recognition accuracy are improved. Field data validation demonstrates that the proposed method achieves a 98.7% accuracy in identifying icing levels. Using the optimal feature subset identified in this study, an icing level warning feature library can be established, offering a valuable reference for intelligent transmission line inspection.