A self-adaptive phase recovery algorithm based on K-means clustering is proposed to solve the problem of self-adaptive phase recovery of unknown probability shaping factor signals. Through theoretical analysis and numerical simulation, the feasibility of constellation point modulus radius repositioning based on K-means clustering is verified. Then, the K-means clustering is combined with the feedforward carrier phase recovery algorithm to solve the problem of relative amplification of constellation points after normalization of probabilistic shaped signals, and the adaptive recovery of signal phase of unknown probability shaping factor is realized. The 16 quadrature amplitude modulation (QAM) and 64QAM signals at different optical signal-to-noise ratio (OSNR) and laser linewidth are simulated. The results show that the proposed algorithm can be used not only for phase recovery of signals with unknown probability shaping factor, but also for phase recovery of uniform QAM signals. Because the proposed algorithm considers the influence of noise in the repositioning of the constellation point modulus, higher precision phase compensation can be realized under the same number of test phases. The algorithm improves the tolerance of OSNR by about 1 dB for uniform QAM signals.