Abstract: A tensor-based semi-blind joint signal detection and channel estimation algorithm was proposed for multiple input multiple output system. Taking advantage of the uniqueness of tensor decomposition, a tensor decomposition based parallel factor model of the received signal was constructed, and a regular alternating least squares algorithm was designed for joint channel estimation and signal detection. Compared with the traditional method based on pilot channel estimation, the proposed algorithm has higher estimation precision with few pilot sequences; Compared with the alternating least squares algorithm, the proposed algorithm has a faster convergence speed, and dose not suffer from ill-conditional problem when it involves in the pseudo-inverse operation. Moreover, the effect of convergence condition and regular coefficient on the regular alternating least squares algorithm was analyzed in detail.