航磁非线性干扰的LSTM与BP神经网络复合架构补偿方法

Aeromagnetic Nonlinear Interference Compensation Method Based on Hybrid LSTM and BP Neural Network Architecture

摘要: 针对传统Tolles-Lawson(T-L)模型在航空磁法勘探中难以有效补偿非线性磁干扰的局限性，本文提出一种基于LSTM-BP组合神经网络的航空磁干扰补偿方法。该方法深度融合长短期记忆网络(LSTM)对时序动态特征的处理能力与误差反向传播网络(BP)对静态非线性关系的强拟合优势，构建了一种兼具时序建模与非线性补偿能力的混合模型。利用LSTM网络提取由日变磁场及飞行姿态变化引起的动态干扰特征，同时引入BP网络对LSTM难以充分表征的残余非线性干扰成分进行补偿优化，从而实现对复杂磁扰动的精准建模与抑制。基于T-L模型生成仿真数据，并分别采用LSTM-BP、GRU-BP、LSTM及BP神经网络进行补偿性能对比。实验结果表明，本文所提LSTM-BP模型使干扰磁场峰峰值降至478.44 nT，改善比单一神经网络模型提升近3~7倍，达到了25.2。研究验证了LSTM-BP组合结构在动态与静态非线性特征建模上的有效互补性，为高精度航空磁测数据补偿提供了新的技术途径。

Abstract: The traditional Tolles-Lawson (T-L) model faces limitations in effectively compensating for nonlinear magnetic interference in aeromagnetic surveys. To address this challenge, this paper proposes a novel magnetic compensation method based on a combined LSTM-BP neural network. This approach deeply integrates the capability of the Long Short-Term Memory (LSTM) network to process temporal dynamic features with the strong nonlinear fitting ability of the Back Propagation (BP) neural network, constructing a hybrid model capable of both temporal modeling and nonlinear compensation. The LSTM network is utilized to extract dynamic interference features caused by diurnal variations in the magnetic field and changes in flight attitude, while the BP network is introduced to compensate for and optimize the residual nonlinear interference components that are difficult for the LSTM network to fully characterize, thereby achieving accurate modeling and suppression of complex magnetic disturbances. Simulation data were generated based on the T-L model, and compensation performance was compared using the LSTM-BP, GRU-BP, LSTM, and BP neural networks, respectively. Experimental results show that the proposed LSTM-BP model reduces the peak-to-peak value of the interference magnetic field to 478.44 nT, and the improvement ratio is nearly 3 to 7 times higher than that of a single neural network model, reaching 25.2. The study verifies the effective complementarity of the LSTM-BP combined structure in modeling both dynamic and static nonlinear features, providing a new technical approach for high-precision aerial magnetic survey data compensation.