1D ViT-ResNet Method for Magnetic Source Localization of Small Ferromagnetic Targets on Shallow Seabeds
-
摘要: 为克服复杂水下环境下的磁信号采集挑战, 文中研究设计并搭建了一套分体式拖体系统, 搭载磁通门阵列, 以高效采集运动状态下的磁环境噪声和4类典型铁磁小目标的磁异常信号, 并成功构建了相应的实测数据集。为弥补实测数据的局限性并扩充数据多样性, 文中研究基于实测数据特性, 利用COMSOL多物理场仿真软件, 构建了包含4类目标物磁源通过特征曲线的仿真数据集, 为模型的训练提供了数据支撑。最终, 为解决磁源实时检测和定位问题, 文中研究提出了一种创新的1D-ViT检测模型与1D-ResNet定位模型协同的磁源定位方法(简称1D ViT-ResNet磁源定位方法)。通过实测目标信号验证, 该算法实现了约7.0%的定位估计误差均值。与单一模型相比, 双模型方法显著降低了误检率, 在实测目标信号中, 误检率平均降低了11.0个百分点, 有效提升了探测的准确性和可靠性。Abstract: To address the challenges of magnetic signal acquisition in complex underwater environments, this study designed and constructed a modular towed-platform system equipped with a fluxgate magnetometer array. This system efficiently collects magnetic environmental noise and magnetic anomaly signals from four typical small ferromagnetic targets under dynamic conditions, successfully establishing a corresponding real-world measurement dataset . To compensate for the limitations of measured data and enhance data diversity, a simulation dataset was developed using COMSOL Multiphysics software, based on the characteristics of the real-world dataset. This dataset includes the magnetic source signature curves for the four target types, providing robust data support for model training. Ultimately, to enable real-time magnetic source detection and precise localization, this research proposes an innovative collaborative magnetic source localization method. This method integrates a 1D-ViT detection model with a 1D-ResNet localization model (hereinafter referred to as the 1D ViT-ResNet magnetic source localization method). Validated against real-world target signals, the algorithm achieved an average localization estimation error of approximately 7.0%. Compared to single-model approaches, this dual-model strategy significantly reduced the false detection rate, with an average reduction of 11.0 percentage points observed in real-world target signals, thereby substantially enhancing detection accuracy and system robustness .
-
图 1 组合式水下拖曳平台系统示意图
Figure 1. Schematic diagram of the modular underwater towed platform system
图 11 4类目标物磁源通过特性曲线检测结果
Figure 11. Detection results of magnetic source passage characteristic curves for four target types
图 13 检测模型和定位模型扫描整个片段
Figure 13. Detection model and positioning model scan the whole segment
表 1 磁通门参数
Table 1. Parameters of fluxgate magnetometer
序号 技术条目 参数 1 通道数 3 2 输入电压 ±10 V 3 AD采样 24 bit 4 采样率 200 Hz 5 1 Hz处频域噪声 ≤0.3 μVrms/√Hz 6 通信方式 内记 7 供电电压 24 V 8 功耗 ≤3 W 9 尺寸 ≤270 mm×56 mm×56 mm 
下载: 导出CSV
表 2 COMSOL磁异常仿真参数设置
Table 2. Parameter settings for the COMSOL magnetic anomaly simulation
参数名称 参数设置 海水电导率/(S/m) 4 海水相对磁导率 1 海水相对介电常数 80 目标物(铁)电导率/(S/m) 1.12×107 目标物(铁)相对磁导率 200 目标物(铁)相对介电常数 1 地磁场总强度/(nT) 56348 地磁倾角设置/(°) 50 地磁偏角设置/(°) 2 稳态求解器 PARDISO 直接求解器 相对容差 1×10−5 物理场控制网格单元规模 常规
下载: 导出CSV
表 3 1D-CNN网络结构与主要参数
Table 3. Architecture and main parameters of the 1D-CNN network
层数 层类型 核长 步长 填充 输出形状 1 Conv1d 7 2 3 [64, 32, 150] 2 Batch
Norm1d[64, 32, 150] 3 ReLU [64, 32, 150] 4 MaxPool1d 2 2 0 [64, 32, 75] 5 Conv1d 5 1 2 [64, 64, 75] 6 Batch
Norm1d[64, 64, 75] 7 ReLU [64, 64, 75] 8 MaxPool1d 2 2 0 [64, 64, 37] 9 Conv1d 3 1 1 [64, 64, 37] 10 Batch
Norm1d[64, 64, 37] 11 ReLU [64, 64, 37] 12 Adaptive
AvgPool[64, 64, 1] 13 Sequential [64, 2] 14 Dropout [64, 64] 15 Linear [64, 2]
下载: 导出CSV
表 4 1D-ResNet网络结构与主要参数
Table 4. Architecture and main parameters of the 1D-ResNet network
层数 层类型 核长 步长 填充 输出形状 1 Conv1d 5 2 2 [64, 8, 150] 2 Batch
Norm1d[64, 8, 150] 3 ReLU [64, 8, 150] 4 MaxPool1d 3 2 1 [64, 8, 75] 5 BasicBlock
1D_1[64, 16, 38] 6 Conv1d 3 2 1 [64, 8, 38] 7 Conv1d 1 1 [64, 16, 38] 8 Batch
Norm1d[64, 16, 38] 9 ReLU [64, 16, 38] 10 Conv1d 3 1 1 [64, 16, 38] 11 Conv1d 1 1 [64, 16, 38] 12 Batch
Norm1d[64, 16, 38] 13 Sequential [64, 16, 38] 14 Conv1d 1 2 [64, 16, 38] 15 Batch
Norm1d[64, 16, 38] 16 ReLU [64, 16, 38] 17 BasicBlock
1D_2[64, 32, 19] 18 BasicBlock
1D_3[64, 64, 10] 19 AdaptiveAvg
Pool1d[64, 64, 1] 20 Dropout [64, 64] 21 Linear [64, 2]
下载: 导出CSV
表 5 1D-ViT网络结构与主要参数
Table 5. Architecture and main parameters of the 1D-ViT network
层数 层类型 输出形状 1 Conv1d_编码层 [64, 20, 10] 2 Dropout [64, 11, 20] 3 Transformer
EncoderLayer_1[64, 11, 20] 4 Multi-head Attention [64, 11, 20] 5 Dropout [64, 11, 20] 6 LayerNorm [64, 11, 20] 7 Linear [64, 11, 40] 8 Dropout [64, 11, 40] 9 Linear [64, 11, 20] 10 Dropout [64, 11, 20] 11 LayerNorm [64, 11, 20] 12 TransformerEncoderLayer_2 [64, 11, 20] 13 LayerNorm [64, 20] 14 Linear [64, 2]
下载: 导出CSV
表 6 海试数据集参数
Table 6. Parameters of the sea trial dataset
参数 数值 总时长/h 19.7 航速/kn 2~3 水深/m 10~20 海况平均波高/m 0.4~0.5 磁通门采样频率/Hz 200 姿态传感器采样频率/Hz 10
下载: 导出CSV
表 7 仿真磁源通过特性曲线参数
Table 7. Parameters of simulated magnetic source passing characteristic curves
参数 数值 球型峰峰值均值/nT 16.30 球柱型峰峰值均值/nT 15.50 管道型峰峰值均值/nT 12.50 球柱二型峰峰值均值/nT 10.10 采样间距/m 0.12 滤波器(S-G)平滑度 1.00 随机缩放系数 0.10
下载: 导出CSV
表 8 模型验证损失值
Table 8. Model validation loss value
任务 模型 信噪比 −1 dB 1 dB 3 dB 5 dB 7 dB 检测 1D-CNN 0.434 0.386 0.332 0.283 0.239 1D-ResNet 0.433 0.385 0.318 0.301 0.247 1D-ViT 0.393 0.307 0.322 0.292 0.257 定位 1D-CNN 1.838 1.732 1.287 0.963 1.087 1D-ResNet 1.743 1.670 1.020 0.998 0.867 1D-ViT 1.758 1.832 1.311 1.212 1.099
下载: 导出CSV
表 9 模型验证准确率
Table 9. Model validation accuracy
任务 模型 信噪比 −1 dB 1 dB 3 dB 5 dB 7 dB 检测 1D-CNN 0.782 0.812 0.871 0.913 0.925 1D-ResNet 0.805 0.827 0.868 0.887 0.915 1D-ViT 0.842 0.889 0.907 0.904 0.921 定位 1D-CNN 0.381 0.352 0.502 0.695 0.632 1D-ResNet 0.386 0.408 0.428 0.653 0.705 1D-ViT 0.345 0.319 0.493 0.652 0.623
下载: 导出CSV
表 10 模型验证ROC-AUC值
Table 10. Model validation ROC-AUC values
任务 模型 信噪比 −1 dB 1 dB 3 dB 5 dB 7 dB 检测 1D-CNN 0.877 0.907 0.941 0.946 0.978 1D-ResNet 0.890 0.948 0.951 0.955 0.966 1D-ViT 0.931 0.952 0.962 0.978 0.974 定位 1D-CNN 0.734 0.752 0.830 0.901 0.867 1D-ResNet 0.748 0.764 0.845 0.893 0.908 1D-ViT 0.720 0.705 0.816 0.890 0.866
下载: 导出CSV
表 11 模型验证PR-AUC值
Table 11. Model validation PR-AUC values
任务 模型 信噪比 −1 dB 1 dB 3 dB 5 dB 7 dB 检测 1D-CNN 0.875 0.906 0.940 0.945 0.979 1D-ResNet 0.892 0.949 0.952 0.956 0.965 1D-ViT 0.932 0.952 0.962 0.978 0.975 定位 1D-CNN 0.394 0.466 0.596 0.775 0.726 1D-ResNet 0.438 0.502 0.627 0.785 0.797 1D-ViT 0.374 0.409 0.577 0.763 0.706
下载: 导出CSV
表 12 模型验证Loss值波动系数
Table 12. Coefficient of variation of model validation loss
任务 模型 信噪比 −1 dB 1 dB 3 dB 5 dB 7 dB 检测 1D-CNN 0.109 0.161 0.143 0.155 0.128 1D-ResNet 0.118 0.137 0.125 0.160 0.093 1D-ViT 0.039 0.025 0.036 0.040 0.023 定位 1D-CNN 0.015 0.081 0.062 0.029 0.032 1D-ResNet 0.033 0.028 0.030 0.036 0.030 1D-ViT 0.005 0.008 0.009 0.011 0.012
下载: 导出CSV
表 13 模型单帧推理时长
Table 13. Model inference time per frame
任务 模型 参数量 单帧推理时长/ms 检测 1D-CNN 23586 1.16±0.39 1D-ResNet 12154 2.83±0.41 1D-ViT 8382 2.28±1.08 定位 1D-CNN 23846 1.15±0.35 1D-ResNet 12414 2.82±0.37 1D-ViT 8466 2.12±0.33
下载: 导出CSV
表 14 模型检测误差估计
Table 14. Model detection error estimation
目标物 陆地标定距离/m 模型定位距离/m 估计误差/％ 管道型 4.5 3.7 17.8 球柱型 4.4 4.0 9.1 球柱型第二次 4.7 4.5 4.3 球柱二型 2.9 3.0 3.4 球柱二型第二次 4.4 4.5 2.3 球型 4.1 4.3 4.9
下载: 导出CSV
表 15 模型误检率
Table 15. Model false detection rate
目标物 检测误
检率/%定位误
检率/%联合误
检率/%降低的
误检率/%管道型 25 80.0 0 25 球柱型 0 60 0 0 球柱型第二次通过 0 33.3 0 0 球柱二型 84.6 74.5 66.7 7.8 球柱二型第二次通过 33.3 33.3 0 33.3 球型 0 89.4 0 0
下载: 导出CSV
-
[1] Tian B, Wu Y, Hong H. A review of research on magnetic detection methods for underwater target [J]. 2024 3rd International Conference on Artificial Intelligence and Computer Information Technology (AICIT), 2024: 1-6. [2] 李佳, 邱伟, 邹劭芬. 基于改进LM-GN算法的磁性目标定位方法研究[J]. 数字海洋与水下攻防, 2023, 6(5): 552-561. doi: 10.19838/j.issn.2096-5753.2023.05.004Li J, Qiu W, Zou S F. Research on magnetic target localization method based on improved LM-GN algorithm[J]. Digital Ocean & Underwater Warfare, 2023, 6(5): 552-561. doi: 10.19838/j.issn.2096-5753.2023.05.004 [3] 迟铖, 王丹, 于振涛, 等. 磁梯度张量不变量约束条件下的两点定位方法[J]. 水下无人系统学报, 2023, 31(4): 582-587.Chi C, Wang D, Yu Z T, et al. Two-point positioning method with magnetic gradient tensor invariant constraints[J]. Journal of Unmanned Undersea Systems, 2023, 31(4): 582-587. [4] Wang M J, Liang X F, Wang H D, et al. The magnetic array study of effective detection and location for submarine pipeline[C]//Proceedings of the Twenty-ninth (2019) International Ocean and Polar Engineering Conference. ISOPE, 2019: 1504. [5] Zeng F, Zhang X, Liu J, et al. Magnetic gradient tensor positioning method implemented on an autonomous underwater vehicle platform[J]. Journal of Marine Science and Engineering, 2023, 11(10): 1909. doi: 10.3390/jmse11101909 [6] Tang W, Huang G, Li G, et al. Eigenvector constraint-based method for eliminating dead zone in magnetic target localization[J]. Remote Sensing, 2023, 15(20): 4959. doi: 10.3390/rs15204959 [7] Liu G, Zhang Y, Liu W. Structural design and parameter optimization of magnetic gradient tensor measurement system[J]. Sensors, 2024, 24(13): 4083. doi: 10.3390/s24134083 [8] Wu X, Huang S, Li M, et al. Vector magnetic anomaly detection via an attention mechanism deep-learning model[J]. Applied Sciences, 2021, 11(23): 11533. doi: 10.3390/app112311533 [9] Hui R, Liang X, Zuo C, et al. 2D CNN-based multi-feature fusion detection method for the magnetic anomaly generated by submarine wake[J]. Journal of Ocean Engineering and Science, 2025, 10(6): 1139-1154. doi: 10.1016/j.joes.2023.11.001 [10] Wang S, Zhang X, Zhao Y, et al. Self-supervised marine noise learning with sparse autoencoder network for generative target magnetic anomaly detection[J]. Remote Sensing, 2024, 16(17): 3263. doi: 10.3390/rs16173263 [11] Zhang K, You X, Liu X, et al. Inversion of target magnetic moments based on scalar magnetic anomaly signals[J]. Electronics, 2023, 12(24): 4900. doi: 10.3390/electronics12244900 [12] Cheng S, Wang J, Wang J, et al. Application of polynomial chaos expansion in sensitivity analysis of towed cable parameters of the underwater towing system[J]. Journal of Ocean Engineering and Science, 2025, 10(4): 367-385. doi: 10.1016/j.joes.2023.09.001 [13] He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2016: 770-778. [14] Dosovitskiy A, Beyer L, Kolesnikov A, et al. An image is worth 16x16 words: Transformers for image recognition at scale[C]//International Conference on Learning Representations. 2021. [15] Lecun Y, Bottou L, Bengio Y, et al. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 1998, 86(11): 2278-2324. doi: 10.1109/5.726791 [16] Savitzky A, Golay M J E. Smoothing and differentiation of data by simplified least squares procedures[J]. Analytical Chemistry, 1964, 36(8): 1627-1639. doi: 10.1021/ac60214a047 -
下载:
点击查看大图
计量
- 文章访问数: 22
- HTML全文浏览量: 12
- PDF下载量: 14
- 被引次数: 0
