Spectrum sensing is a crucial technology for cognitive radios and cognitive wireless sensing networks. In order to improve spectrum utilization and avoid interference to primary users, it is necessary to detect whether the spectrum is occupied accurately. This paper proposes a sequence-to-sequence model based on WaveNet structure for spectrum sensing as a practical solution and obtaining the occupied time location. Compared with the traditional Convolutional Neuron Network, the model proposed in this paper can be based on the signal data, reducing the dimension of input signal, and alleviating the computational burden. Furthermore, the model considers the data sequence dependence on time to obtain a comprehensive judgment, and achieves the classification of the corresponding sampling point data on each time step to realize spectrum sensing and time location. Based on the data-sequence analysis, researchers can develop more efficient wireless sensor management strategies. The model alleviates gradient-vanishing and gradient-exploding problems in longtime dependence or long time-series data that generated by high sampling data. Reducing the computational cost and energy consumption of wireless sensor networks is another novel feature of the proposed model. Compared with RNN models, the proposed model reduces the number of model parameters on a large scale. At the same time, the model can achieve the parallel signal processing and energy-saving optimization without extra parameters.