Abstract: Model-driven spectrum sensing algorithm is difficult to accurately simulate the complex communication environment, so the accuracy of spectrum sensing is not high. Meanwhile, due to the high algorithm complexity, the real-time performance of the system is also low. To solve the above problems, this paper combines deep learning with spectrum sensing and proposes a spectrum sensing model based on Transformer under Alpha noise. The observation data was collected and processed by fractional lower order moments (FLOM), and the data was sent to the spectral convolution layer for local feature crude extraction, and finally the final decision was made by Transformer network with multi-head-attention mechanism module. The simulation results show that the proposed algorithm can enhance the correlation of signals, strengthen the parallel computing capability of the network, and shows superior spectrum sensing performance even under the condition of low generalized signal-to-noise ratio (GSNR).