Recognizing and modeling towers in three dimensions using 3D point cloud data is crucial for digitizing power lines and ensuring lightning protection. However, tower point cloud data lacks clear structure, which makes it harder to work with more general point cloud data. This poses a significant challenge for developing algorithms that can accurately recognize tower points. In this paper, we introduce a 3D point cloud shape recognition algorithm based on self-supervised learning. Through a series of experiments, we demonstrate the algorithm's accuracy and stability in classifying power transmission towers. Our results show that the algorithm performs well even with limited labeled data, particularly in scenarios where data samples are sparse. This capability is advantageous for overcoming challenges associated with understanding point clouds, thus making the algorithm more accessible for practical implementation and providing a robust solution for real-world applications in relevant fields.