GAT-TTN-BFGS多模态融合定位算法

doi:10.12305/j.issn.1001-506X.2026.11.37

摘要/Abstract

摘要：

受动态障碍物与金属反射等因素的影响，室内密集时变多径效应显著加剧，导致室内复杂非视距场景测距误差是视距场景的3~5倍。现有基于深度学习的超宽带（ultra wide band, UWB）定位模型由于对UWB信号特征的表征能力不足，导致定位精度显著下降，尤其是在跨应用场景时性能恶化更为显著。基于此，设计了图注意力-时域卷积-Transformer网络（graph attention-temporal convolution & Transformer network，GAT-TTN），通过多层感知机提取UWB信号特征，利用搭载Inception模块的时序卷积神经网络提取UWB信道特征，采用图注意力网络实现UWB信号特征与信道特征的跨模态融合，通过Transformer编码器增强跨模态适应性，捕捉信号特征变化与信道特征之间的关系，完成测距误差预测。最后建立测距误差方差与位置坐标的联合优化目标函数，基于有限内存布罗伊登-弗莱彻-戈德法布-香奈（有界约束版）（limited memory Broyden Fletcher Goldfarb Shanno-bound constraints, L-BFGS-B）算法求解非线性方程，实现GAT-TTN-BFGS多模态融合位置估计。该方案在4类测试环境中进行了性能验证，工业环境定位精度为0.25 m，两个居民环境与办公环境定位精度分别为0.09 m、0.05 m、0.12 m，定位误差比最小二乘算法分别小2.24 m、0.83 m、0.36 m、0.8 m。

关键词: 超宽带, 信道冲激响应, 非视距传播, 图注意力网络, 有限内存布罗伊登-弗莱彻-戈德法布-香奈（有界约束版）

Abstract:

Influenced by factors such as dynamic obstacles and metal reflections, the dense time-varying multipath effects are significantly intensified, leading to ranging error in indoor complex non-line-of-sight scenarios is 3-5 times that of line-of-sight scenarios. The deep learning-based ultra wide band （UWB） localization models suffer from a significant decrease in localization accuracy due to insufficient representation of UWB signal features, especially in cross application scenarios. Based on this, a graph attention-temporal convolution & Transformer network（GAT-TTN）is proposed. It extracts UWB signal features through multi-layer perceptrons, and extracts UWB channel features using a temporal convolutional neural network equipped with Inception module. The cross modal fusion of UWB signal features and channel features is achieved by using graph attention network. The cross modal adaptability is enhanced through Transformer encoder and the relationship between signal characteristic variations and channel characteristics is captured to accomplish ranging error prediction. Finally, a joint optimization objective function of ranging error variance and position coordinates is established, and the nonlinear equation is solved based on the limited memory Broyden Fletcher Goldfarb Shanno-boundconstraints （L-BFGS-B） algorithm to achieve GAT-TTN-BFGS multimodal fusion position estimation. The performance of the proposed algorithm has been validated in four environments, with a positioning accuracy of 0.25 m in industrial environments and 0.09 m, 0.05 m, and 0.12 m in two residential and office environments, respectively, which is smaller than that of the least squares algorithm by 2.24 m, 0.83 m, 0.36 m, and 0.8 m, respectively.

Key words: ultra wide band (UWB), channel impulse response, non-line-of-sight, graph attention network, limited memory Broyden Fletcher Goldfarb Shanno-bound constraints (L-BFGS-B)

中图分类号:

TN 929.5

刘林, 文健吉, 薛璞钰, 穆再排尔·伊敏托合提. GAT-TTN-BFGS多模态融合定位算法[J]. 系统工程与电子技术, 2026, 48(11): 4080-4091.

Lin Liu, Jianji Wen, Puyu Xue, ·Yimintuoheti Muzaipaier. GAT-TTN-BFGS positioning algorithm based on multi-modal fusion[J]. Systems Engineering and Electronics, 2026, 48(11): 4080-4091.

图/表 30

表1

图1

图2

图3

图4

图5

图6

图7

表2

表3

表4

图8

图9

图10

图11

表5

图12

图13

图14

图15

表6

图16

图17

图18

图19

表7

图20

图21

图22

图23

参考文献 29

1	Rana M, Mittal V. Wearable sensors for real-time kinematics analysis in sports: a review[J]. IEEE Sensors Journal, 2021, 21 (2): 1187. doi: 10.1109/JSEN.2020.3019016
2	Li X X, Huang J X, Li X, et al. Review of PPP-RTK: achievements, challenges, and opportunities[J]. Satellite Navigation, 2022, 3 (3): 16.
3	Qi L, Liu Y, Yu Y, et al. Current status and future trends of meterlevel indoor positioning technology: a review[J]. Remote Sensing, 2024, 16 (2): 398. doi: 10.3390/rs16020398
4	Gao Y K, Postolache O, Yang Y S, et al. UWB system and algorithms for indoor positioning[C]//Telecoms Conference, 2021.
5	Tan S, Yang J. Commodity WiFi sensing in ten years: status, challenges, and opportunities[J]. IEEE Internet of Things Journal, 2022, 9 (18): 17832. doi: 10.1109/JIOT.2022.3164569
6	Han K, Yu S M, Kim S L, et al. Exploiting user mobility for WiFi RTT positioning: a geometric approach[J]. IEEE Internet of Things Journal, 2021, 8 (19): 14589. doi: 10.1109/JIOT.2021.3070367
7	Xiao K, Wang Y T, Wang Q, et al. Research on indoor localization algorithm based on multi-fusion bluetooth AOA using two-dimensional DOA estimation[C]//International Conference on Network Simulation and Evaluation, 2023: 121.
8	Ji M X, Ren G H, Zhang H J, et al. Collaborative positioning for emergency rescuers based on INS, GPS and ZigBee[J]. Physica Scripta, 2024, 99 (6): 065530. doi: 10.1088/1402-4896/ad38e1
9	Gezici S, Poor H V. Position estimation via ultra-wide-band signals[J]. Proceedings of the IEEE, 2009, 97 (2): 386. doi: 10.1109/JPROC.2008.2008840
10	Marano S, Gifford W M, Wymeersch H, et al. NLOS identification and mitigation for localization based on UWB experimental data[J]. IEEE Journal on Selected Areas in Communications, 2010, 28 (7): 1026. doi: 10.1109/JSAC.2010.100907
11	Wymeersch H. A machine learning approach to ranging error mitigation for UWB localization[J]. IEEE Trans. on Communications, 2012, 60 (6): 1719. doi: 10.1109/TCOMM.2012.042712.110035
12	Nguyen T V, Jeong Y, Shin H, et al. Machine learning for wideband localization[J]. IEEE Journal on Selected Areas in Communications, 2015, 33 (7): 1357. doi: 10.1109/JSAC.2015.2430191
13	Li W F, Yang A N, Zhou J L, et al. NLOS occlusion recognition method to improve UWB spatial sensing[J]. IEEE Internet of Things Journal, 2024, 11 (19): 31715. doi: 10.1109/JIOT.2024.3419796
14	袁驰, 黄波. UWB 定位中基于 XGBoost 的 NLOS 识别方法[J]. 移动通信, 2023, 47 (5): 83. doi: 10.3969/j.issn.1006-1010.20220509-0003
15	Bregar K, Mohorcic M. Improving indoor localization using convolutional neural networks on computationally restricted devices[J]. IEEE Access, 2018, 6, 17429. doi: 10.1109/access.2018.2817800
16	Angarano S, Mazzia V, Salvetti F, et al. Robust ultra-wide-band range error mitigation with deep learning at the edge[J]. Engineering Applications of Artificial Intelligence, 2021, 102 (4): 10428. doi: 10.1016/j.engappai.2021.104278
17	刘万青, 魏国, 高春峰, 等. 基于深度学习的 UWB NLOS 传播影响抑制技术研究[J]. 红外与激光工程, 2023, 52 (12): 285. doi: 10.3788/IRLA20230183
18	Deng B W, Xu T W, Yan M D. UWB NLOS identification and mitigation based on Gramian angular field and parallel deep learning model[J]. IEEE Sensors Journal, 2023, 23 (22): 28513. doi: 10.1109/JSEN.2023.3323564
19	Nkrow R E, Silva B, Boshoff D, et al. Transfer learning-based NLOS identification for UWB in dynamic obstructed settings[J]. IEEE Trans. on Industrial Informatics, 2024, 20 (3): 4839. doi: 10.1109/TII.2023.3329655
20	苏毛弟, 王建博, 齐辉, 等. 斜井高频电磁波传输特性的研究[J]. 工矿自动化, 2009, 35 (5): 1.
21	Byrd R H, Lu P, Nocedal J, et al. A limited memory algorithm for bound constrained optimization[J]. SIAM Journal on Scientific Computing, 1995, 16 (5): 1190. doi: 10.1137/0916069
22	Zhu C, Byrd R H, Lu P, et al. Algorithm 778: L-BFGS-B: fortran subroutines for large-scale bound-constrained optimization[J]. ACM Transactions on Mathematical Software, 1997, 23 (4): 550. doi: 10.1145/279232.279236
23	Ding C Y, Sun S L, Zhao J. MST-GAT: a multimodal spatial-temporal graph attention network for time series anomaly detection[J]. Information Fusion, 2023, 89, 527. doi: 10.1016/j.inffus.2022.08.011
24	Mo X Y, Xing Y, Lv C. Heterogeneous edge-enhanced graph attention network for multi-agent trajectory prediction[EB/OL]. [2024-08-08]. https://doi.org/10.48550/arXiv.2106.07161.
25	Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[C]// 31st Annual Conference on Neural Information Processing Systems, 2017: 5999.
26	Bai S, Kolter J Z, Koltun V. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling[EB/OL]. [2024-08-08]. https: //arxiv.org/abs/1803.01271.
27	Szegedy C, Liu W, Jia Y, et al. Going deeper with convolutions[C]// IEEE Conference on Computer Vision and Pattern Recognition, 2015.
28	Bregar K. Indoor UWB positioning and position tracking data set[J]. Scientific Data, 2023, 10 (1): 744. doi: 10.1038/s41597-023-02639-5
29	Kirmaz A, Sahin T, Michalopoulos D S, et al. ToA and TDoA estimation using artificial neural networks for high-accuracy ranging[J]. IEEE Journal on Selected Areas in Communications, 2023, 41 (12): 3816. doi: 10.1109/JSAC.2023.3322803

参数	说明
RANGE	标签到某个锚点的测量距离
RSS_FP	第一路径峰值的RSS
FP_INDEX	CIR序列第一路径峰值的采样点索引值
FP_POINT1	第一路径峰值所在索引位置的信号幅度值
FP_POINT2	第一路径峰值索引+2处的信号幅度值
FP_POINT3	第一路径峰值索引+3处的信号幅度值
STDEV_NOISE	噪声的标准差
MAX_NOISE	测量到的最大噪声幅值
CIR_POWER	CIR的总功率

网络结构	参数类型	参数值
Conv0	卷积核	1×1
	输入	157×8
	输出	157×8
膨胀因果Conv1-1	卷积核	5×1
	输入	157×8
	输出	157×8
	膨胀率	2^N
	填充	（卷积核大小−1）×膨胀率
膨胀因果Conv1-2 &Conv2	卷积核	3×1
	输入	157×8
	输出	157×8
	膨胀率	2^N
	填充	（卷积核大小−1）×膨胀率
MLP1	输入	10
	输出	157
	隐藏层神经元	128
	隐藏层数	2
MPL2	输入	157×10
MPL2	输出	157×10
GATConv	输入	157×10
	节点	10
	输出	157×10
Tranformer 编码器	层数	2
	输入	157×10
	自注意力头数	2
	前馈神经网络维度	32

参数类型	参数值
学习率	0.001
优化器	AdamW
批量大小	128
迭代次数	600
早停门限值	45
丢失率	0.2/0.1
损失函数	MSE

学习器参数配置	参数值
忍耐度	30
学习率下降比例	0.5
学习器类型	ReduceLROnPlateau

环境	模型	CDF=67%	CDF=90%
0	Inception-TCN-Transformer	0.09	0.16
	Inception-TCN-GAT	0.08	0.16
	Inception-TCN-Transformer-GAT	0.07	0.14
	GAT-TTN	0.05	0.09
1	Inception-TCN-Transformer	0.05	0.10
	Inception-TCN-GAT	0.04	0.08
	Inception-TCN-Transformer-GAT	0.05	0.08
	GAT-TTN	0.03	0.06
2	Inception-TCN-Transformer	0.14	0.29
	Inception-TCN-GAT	0.11	0.24
	Inception-TCN-Transformer-GAT	0.13	0.28
	GAT-TTN	0.07	0.15
3	Inception-TCN-Transformer	0.14	0.25
	Inception-TCN-GAT	0.12	0.22
	Inception-TCN-Transformer-GAT	0.10	0.21
	GAT-TTN	0.07	0.12