针对当前局部放电信号去噪方法普适性不足的问题，提出一种以变分模态分解（VMD）为基础，通过自适应参数寻优，并结合离散小波变换（DWT）进行去噪的方法。建立了以“软阈值”结合“硬阈值”的方式进行峭度筛选并以信号分量能量差最小为目标函数的VMD参数优化模型，实现了基于北方苍鹰优化算法（NGO）的自适应寻优，达成了去噪方法的自适应目标。通过对构造模拟数据和实测数据的双方面测试验证了方法的有效性。其中模拟数据囊括了四种典型波形表达式的局部放电脉冲以及由白噪声和两个窄带干扰组成的强干扰。经本方法的处理实现了信噪比从-18.46dB到12.54dB的提升。实测数据在具有较高工程现场噪声环境类似性的场景中经自研的局部放电超声信号采集装置获取，能够较好反映工程现场干扰和硬件电路噪声干扰。局部放电由真型缺陷产生，避免了标准放电模型与实际绝缘缺陷的偏差。获取的数据经本方法的处理实现了38.19dB的信噪比提升和9.88dB的噪声抑制比，并获得了明显特征，且与该缺陷典型局部放电信号特征相符的波形。通过与DWT、EMD-DWT和PSO-VMD-DWT算法处理效果的对比验证了本方法各个环节的有效性、必要性和性能优势。

To address the issue of inadequate adaptivity of conventional denoising methods in partial discharge (PD) monitoring engineering practice, a Variational Mode Decomposition (VMD)-based adaptive denoising method was proposed. The approach incorporates an optimization module to determine the optimal number of intrinsic mode functions (IMFs) and penalty factor, followed by a Discrete Wavelet Transform (DWT) module. The objective function of the optimization model was proposed as minimizing the energy difference between the original signal and the reconstructed signal, which is the combination of IMFs that passed a kurtosis-based screening strategy using both hard and soft threshold methods. The Northern Goshawk Optimization (NGO) algorithm was applied for parameter optimization, successfully achieve the purpose of high adaptivity. The effectiveness of the proposed method was validated using both simulated data and field data from high-voltage tests. Simulated data comprised PD signals generated from all the four well accepted pulse waveform expressions superimposed with strong interference (white noise and 2 different narrowband interference) that fully submerged the PD signals. Processing with the proposed method increased the signal-to-noise ratio (SNR) from -18.46 dB to 12.54 dB. Field data were collected by a self-developed ultrasonic acquisition device in on-site high-voltage tests, exhibiting inherent interference from engineering environments and hardware circuitry. PD signals originated from a test sample engineered to simulate actual switchgear defect, thereby providing a more accurate representation of practical engineering scenarios compared to conventional standard PD models. Processing with the proposed method achieved an increasement of 38.19dB in SNR and a noise rejection ratio (NRR) of 9.88 dB, as well as PD waveform with clearly identifiable features highly consistent with the typical PD features of the defect sample. Comparative analyses were performed with DWT, EMD-DWT, and PSO-VMD-DWT algorithms, as the effectiveness and necessity of each module of the proposed method as well as the overall superiority were validated.