In order to address the challenges of local convergence and robustness degradation in the identification of recursive systematic convolutional (RSC) subcodes of Turbo codes under low signal-to-noise ratio (SNR) conditions, this paper proposes a novel cost function based on an exponential error penalty mechanism, combined with an improved Particle Swarm Optimization (PSO) algorithm for efficient global search. The proposed method applies a nonlinear exponential amplification to the mismatch errors of parity-check equations, which markedly strengthens noise suppression and ensures reliable identification performance even under low-SNR conditions. In addition, an adaptive velocity-position update strategy is incorporated into the PSO framework, allowing particles to maintain strong global exploration in the early search phase and to converge efficiently toward the optimal solution in the later phase, thereby mitigating the risk of stagnation in local optima. Simulation results show that under an SNR of 1.5 dB, the proposed method achieves over 95% identification accuracy for a rate-1/2 RSC code with constraint length 5, using only 2 000 intercepted bits and within 8 iterations. Compared with existing state-of-the-art methods, it achieves a performance gain of approximately 0.5 dB. Additional experiments further confirm that the proposed method exhibits strong adaptability and robustness across different SNR levels and code lengths. Overall, the proposed approach achieves a balance between identification accuracy and computational complexity, making it particularly well-suited for practical applications in low-SNR environments and offering a robust solution for blind Turbo code parameter estimation.