大地电磁测深是一种通过观测天然电磁场获取地下电性结构的勘探方法，较易受到噪声干扰。脉冲类噪声是大地电磁工作中的常见噪声，其幅值高、频带宽，会对数据质量产生较大影响。

为了压制脉冲类噪声，以插补思想为基础，提出了基于时间序列双向循环插补模型(Bidirectional Recurrent Imputation for Time Series，BRITS)的大地电磁脉冲类噪声处理方法。首先，将噪声干扰段删除，此时大地电磁时间序列可视为待插补的缺失序列，而后利用该缺失序列构建训练集，对BRITS模型进行插补训练，训练完成后对缺失序列进行插补，即可得到去噪结果。通过仿真及实测含噪声数据处理，并与经验模态分解(Empirical mode decomposition，EMD)阈值方法进行了对比。

结果表明：BRITS方法对仿真噪声数据处理后与原始数据的归一化互相关系数可达0.999以上，信噪比可达29 dB以上，EMD阈值方法处理前后相关系数为0.778，信噪比为3.09 dB；在实测数据处理中，BRITS方法有效恢复了噪声干扰数据，相比EMD阈值方法，其阻抗奈奎斯特图更接近天然大地电磁信号特征。通过不同训练样本试验得出：对4分量大地电磁数据而言，数据中至少需包含两道正常分量，单个含噪分量中噪声占比不大于20%，且噪声连续干扰长度不超过10个采样点，此时，BRITS方法去噪后数据的相关系数在0.96以上，可以保证一定的去噪精度。

Magnetotelluric sounding is an exploration method to obtain the underground electrical structure by observing the natural electromagnetic field, which is easily disturbed by noise. Impulse noise, frequently occurring in magnetotelluric sounding, generally exhibits high amplitude and wide frequency bands, producing significant impacts on the data quality.

To suppress such noise, this study proposed a method based on a bidirectional recurrent imputation for time series (BRITS) model. First, the data segments with noise interference were deleted. Second, for the magnetotelluric time series with missing data to be imputed, training sets were constructed for imputation training of the BRITS model. Third, imputation was conducted to supplement the missing data, yielding the denoising results. Last, the proposed method was applied to process the simulated and measured data with noise, and the application results were compared with the results derived using the empirical mode decomposition (EMD) threshold method.

The results of this study are as follows: (1) Relative to the original data, the simulated data with noise, after being processed using the BRITS method, manifested normalized cross-coefficient reaching up to 0.999 and signal-to-noise ratios of over 29 dB. In contrast, the simulated noise data, after being treated using the EMD threshold method displayed cross-coefficient of 0.778 and signal-to-noise of 3.09 dB; (2) In the processing of the measured data, the BRITS method effectively restored the data with noise interference, with the obtained Nyquist diagram closer to the characteristics of natural magnetotelluric signals compared to the EMD threshold method; (3) As indicated by the test results of different training samples, in the case where four-component magnetotelluric data contain at least two normal components, with the proportion of noise in a single unnormal component not exceeding 20% and the continuous noise interference length of 10 sampling points or less, the data denoised using the BRITS method can yield cross-coefficient exceeding 0.96, thus ensuring certain denoising accuracy.