Abstract: The research on blind demodulation of frequency-shift keying signals has significant implications for spectrum monitoring, electronic countermeasures, and other related fields. However, the existing two-stage architecture method is constrained by accumulated parameter estimation errors, insufficient noise resistance capability, and high demodulation loss. Aiming at this problem, a novel blind demodulation method based on real-time symbol rate estimation is proposed in this paper. The approach initially establishes a mathematical model for demodulation and extracts instantaneous frequency sequences. Subsequently, it incorporates frequency distribution entropy and centroid position features to quantitatively describe the obtained demodulation results. By minimizing these proposed features, an integrated design that combines high-precision symbol rate estimation with low bit error rate non-cooperative demodulation is achieved. simulation experiments demonstrate that proposed algorithm exhibits a maximum demodulation loss of 0.5 dB and achieves a symbol rate estimation accuracy above 92% at 0 dB signal-to-noise ratio (SNR), surpassing existing algorithms significantly while demonstrating excellent engineering practicality.