基于弱监督学习的大地电磁噪声检测和分类

Magnetotelluric Noise Detection and Classification Based on Weakly Supervised Learning

摘要: 在大地电磁法中，提高数据质量一直是大地电磁信号采集与处理的核心问题。本文提出使用Anomaly Transformer异常检测模型对大地电磁数据进行噪声识别和CNN-BiLSTM噪声分类的弱监督学习联合算法，首先，将处理好的数据集放入Anomaly Transformer异常检测模型训练，采用两个并行的分支计算，在先验关联分支输出为先验关联矩阵，在序列关联分支输出序列关联矩阵，之后用KL散度计算关联差异，对测试集数据进行无监督异常点识别并标记，然后将标记好的测试集数据切分好放入CNN-BiLSTM模型进行训练，利用CNN提取信号的时域和频域特征，并将其融合，最后利用BiLSTM网络学习信号的序列特征，完成噪声分类任务。实验结果表明：在噪声检测部分，与LSTM对比Anomaly Transformer无监督异常检测模型在噪声识别方面表现出明显的优势，在噪声分类部分，CNN-BiLSTM与CNN、LSTM、BiLSTM、CNN-LSTM对比该模型仍具有卓越的噪声分类诊断性能、出色的泛化能力和快速的计算速度。这一结果充分验证了该模型在弱监督大地电磁噪声识别和分类任务上的有效性，为后续信号去噪、反演和地质解释提供了有力保障和重要的实际应用价值。

Abstract: In magnetotelluric (MT) methods, improving data quality has consistently been a central challenge in signal acquisition and processing. This paper proposes a joint algorithm based on weakly supervised learning, integrating the Anomaly Transformer for noise detection and a CNN-BiLSTM network for noise classification in MT data. First, the preprocessed dataset is input into the Anomaly Transformer model for training. The model employs two parallel branches: the prior association branch outputs a prior association matrix, while the series association branch produces a sequence association matrix. The Kullback-Leibler (KL) divergence is then used to compute the difference between the two, enabling unsupervised identification and labeling of anomalies in the test dataset. The labeled test data are then segmented and fed into the CNN-BiLSTM model. The CNN is responsible for extracting and fusing time-domain and frequency-domain features of the signal, while the BiLSTM captures the sequential characteristics to complete the noise classification task. Experimental results show that the Anomaly Transformer significantly outperforms traditional LSTM models in noise detection. In the classification phase, the CNN-BiLSTM model also demonstrates superior performance, better generalization, and faster computation compared with CNN, LSTM, BiLSTM, and CNN-LSTM models. These results fully validate the effectiveness of the proposed model for weakly supervised MT noise detection and classification, offering strong support and practical value for subsequent denoising, inversion, and geological interpretation.