地磁时变观测数据中高压直流输电干扰事件多尺度表示及识别方法

Multi-scale representation and identification of High Voltage Direct Current interference events in time-varying geomagnetic observation data

高效准确地识别地磁时变观测数据中受高压直流输电干扰的波形对于提高地磁时变观测数据质量具有重要意义.然而,由于高压直流输电干扰事件持续时间长短不一、干扰程度变化多样,给识别任务带来巨大困难.为了能自动识别长短不同的高压直流输电干扰事件,本文提出一种高压输电干扰事件的多尺度表示及识别方法:利用小波技术具有多尺度的特性,卷积神经网络具有自动特征提取的特性,将二者结合,设计了一个多输入卷积神经网络模型来识别地磁中的高压直流输电干扰事件.首先使用离散小波技术将地磁时变观测样本进行多尺度分解,得到原始样本的多尺度表示,再将分解后的多尺度地磁时变观测样本分别输入到含有多个输入分支的卷积神经网络中,每个分支分别自动提取不同尺度的特征,然后将多个尺度的特征融合,并加入注意力机制来自适应计算每个尺度特征的权重,对多尺度特征进行加权处理,再采用全连接层和SoftMax层进行分类,本文将该模型命名为CBAM-MCNN.在中国地震前兆台网中心提供的高压直流输电干扰样本上进行试验,并将本文所提出模型的识别效果与现有的全卷积网络、残差神经网络、多输入卷积神经网络、IICM-HVDCT-CNN-LSTM进行了对比,在5271条测试样本集上,本文所提出的CBAM-MCNN模型识别准确率达到了97.14%,F_1值达到了97.12%,远远高于其他4种对比模型.

Efficient and accurate identification of geomagnetic waveform disturbances caused by High Voltage Direct Current (HVDC) transmission is crucial to improve the quality of time-varying geomagnetic observation data. However, identifying these disturbances is challenging due to their varying duration and degree of disturbance. The present study introduces a novel approach for the automatic detection of HVDC transmission disturbance events of varying durations. A multi-scale representation and identification method is proposed, incorporating wavelet technology with multi-scale properties and convolutional neural network (CNN) known for their ability to extract features automatically. By integrating these techniques, a multi-input CNN model is devised to effectively identify HVDC transmission disturbance events in time-varying geomagnetic observation data. Firstly, discrete wavelet technique is used to decompose multi-scale geomagnetic samples into a multi-scale representation. Next, these decomposed samples are input into a convolutional neural network with multiple input branches, where each branch automatically extracts features at different scales. The features are then combined and the attention mechanism is added to calculate the weights of each scale feature adaptively. Then the fully connected layer and SoftMax layer are used for classification, and the proposed model is named CBAM-MCNN. This paper conducted experiments on HVDC transmission disturbance samples provided by the China Earthquake Precursor Network Center. The accuracy of the CBAM-MCNN model reached 97.14% on 5271 test samples, which is much higher than the existing Full Convolutional Neural Networks, Residual Network, Multi-input Convolutional Neural Network, and IICM-HVDCT-CNN-LSTM models.