天津职业技术师范大学学报

针对汽车加工车间内人工语音信号中噪声难以有效去除的问题，提出一种基于冠豪猪优化（CPO）算法的变分模态分解（VMD）与小波阈值（WT）相结合的CPO-VMD-WT去噪方法。该方法使用CPO对VMD模态分解数K与惩罚因子α进行优化，实现了带噪语音信号的最优分解；利用Pearson相关系数筛选法，保留本征模态函数（IMF）分量中的有效分量；对IMF有效分量进行WT去噪处理，去除部分噪声后重构得到最终去噪信号。实验结果表明，与WT、经验模态分解（EMD）结合WT(EMD-WT)以及VMD结合WT(VMD-WT)去噪方法相比，经CPO-VMD-WT去噪后的信号信噪比（8.950 0 dB）最高，均方根误差（0.0344）最低，与原始信号的相关系数（0.9400）最高，CPO-VMD-WT去噪效果最优。

To address the challenge of effectively removing noise from human speech signals in automotive manufacturing workshops, this study introduces a speech signal denoising method, termed CPO-VMD-WT, combining Crown Porcupine Optimization(CPO), Variational Mode Decomposition(VMD) and Wavelet Thresholding(WT). This method uses CPO to optimize the number of VMD decomposition modes(K) and the penalty factor(α), thereby achieving an optimal decomposition of noisy speech signals. After decomposition, the Pearson correlation coefficient is employed to select and retain the effective Intrinsic Mode Function(IMF) components. These components are subsequently denoised via WT, followed by reconstruction of the final clean speech signal. Experimental results demonstrate that the proposed CPO-VMD-WT method outperforms traditional WT, Empirical Mode Decomposition combined with WT(EMD-WT), and VMD combined with WT(VMD-WT). Specifically, it achieves the highest signal-to-noise ratio(8.950 0 dB), the lowest root mean square error(0.034 4), and the highest correlation coefficient with the original signal(0.940 0), demonstrating superior denoising capability and robustness under complex acoustic conditions typical of industrial environments.