Industrial sites are typically subject to complex noise environments, where mixed interference from non-Gaussian noise and unknown but bounded noise severely degrades the performance of conventional replay attack detection techniques. To address this challenge, a dynamic threshold detection algorithm based on state estimation is proposed. A novel state estimation is developed, leveraging ellipsoid theory and unbiased finite impulse response filtering, which accounts for both the amplitude constraint of unknown but bounded and the higher-order statistical properties of non-Gaussian noise in linear systems. It solves the problem of estimation performance degradation caused by prior statistical model mismatch in the dual noise environment. By employing real-time system information, the dynamic threshold detection strategy is designed to overcome the defect of insufficient sensitivity of the fixed threshold detector in the context of replay attacks. Utilizing the system's residuals, a detection function is formulated to recognize potential replay attacks occurring within the system. The simulation results indicate that the proposed approach provides a more reliable technical means for replay attack detection of such noisy systems, while exhibiting superior adaptability.