Abstract: To construct an accurate three-dimensional digital model of substations for quantitative evaluation of UAV inspection blind spots, the instance segmentation of equipment point clouds is of critical importance. However, the complex geometric structures, significant noise interference, and severe occlusion commonly present in substation scenes pose great challenges to the feature representation capability and robustness of segmentation algorithms. To address these issues, we propose an innovative point cloud segmentation algorithm based on the Fractional Fourier Transform, namely FrFT-PointNet. The core of the proposed method is a time-frequency adaptive perception mechanism. This mechanism, on one hand, employs a trainable multi-order fractional Fourier convolution layer, enabling the network to adaptively identify the optimal time-frequency analysis domain for varying local structures, thereby synergistically capturing both spatial geometric details and fractional domain energy features within a unified framework. On the other hand, it incorporates a designed fractional Fourier amplitude pooling layer, which utilizes energy-based statistical aggregation to obtain a more invariant and robust global representation against noise and occlusion. This mechanism fundamentally enhances the model's ability to perceive complex boundaries and improves its overall feature representation capability. Experimental results show that FrFT-PointNet significantly outperforms existing mainstream methods, achieving a mean Intersection over Union of 79.38% and an overall accuracy of 86.92%. Compared with the PointNet++ algorithm, these metrics represent improvements of 4.75% and 6.45%, respectively.