Real-time track anomaly detection and prediction correction model for low-altitude traffic control platform
-
摘要: 构建了一种面向低空交通管控平台的航迹异常检测与修正一体化端到端模型，以统计离群、物理包线、形态模式和序列残差4类行为分量融合形成统一异常分数，并结合特征融合模块以及自适应阈值和权重优化实现异常判定；以单向双层长短期记忆网络预测结合注意力机制作为时序骨干，融入可微分物理积分器、自适应噪声估计与卡尔曼更新，获得预测重构序列；设计了可回传损失函数，并采用一致性蒸馏对齐分支输出口径，最终形成端到端的物理感知卡尔曼长短时记忆网络模型。研究结果表明：在异常检测任务中，相较于深度基线模型，本研究模型的固定阈值下的调和均值分数F1、平均精确率-召回率曲线下面积AUPRC和ROC曲线下面积AUROC分别提升了5.95%、4.16%和2.38%；在预测修正任务中，相较于仅滤波方法和仅预测模型，均方根误差分别降低了15.2%和21.7%；在实时部署方面，优选32窗口时，模型计算延迟较最佳F1值对应的64窗口下降了76.9%，而F1值仅下降了0.57%。该模型能够在毫秒级时延约束下兼顾异常检测可靠性与航迹实时纠偏能力，可为低空交通管控平台航迹实时异常检测及预测修正功能的开发提供有效方法支撑。Abstract: An end-to-end integrated model of track anomaly detection and correction for low-altitude traffic control platforms was developed. Four types of behavioral components, such as, statistical outliers, physical envelopes, morphological patterns, and sequence residuals, were fused to form a unified anomaly score, and anomaly identification was achieved by combining a feature fusion module, adaptive thresholding, and weight optimization. A unidirectional two-layer long short-term memory prediction network combined with an attention mechanism was employed as the temporal backbone, and a differentiable physical integrator, adaptive noise estimation, and Kalman update were incorporated to obtain predicted reconstruction sequences. A back-propagatable loss function was designed, and consistency distillation was adopted to align the outputs of different branches, ultimately forming an end-to-end physics-aware Kalman long short-term memory network model. The results show that in the anomaly detection task, compared with deep baseline models, the harmonic mean score (F1), area under the average precision-recall curve (AUPRC), and area under the ROC curve (AUROC) of the developed model under a fixed threshold increase by 5.95%, 4.16%, and 2.38%, respectively. In the prediction correction task, compared with a filtering-only method and a prediction-only model, the root mean square error (RMSE) decreases by 15.2% and 21.7%, respectively. In terms of real-time deployment, when the optimal window size is 32, the computational latency decreases by 76.9% compared with the window size of 64 corresponding to the best F1 value, while the F1 value decreases by only 0.57%. The model can balance the anomaly detection reliability and real-time trajectory correction ability under millisecond-level latency constraints and can provide effective methodological support for the development of real-time trajectory anomaly detection and prediction correction functions of low-altitude traffic control platforms.
-
图 1 低空交通管控平台嵌入航迹实时异常检测及预测修正功能构想
Figure 1. Conceptual of real-time anomaly detection and prediction correction function
图 5 随阈值与权重组合变化下的各指标值
Figure 5. Values of each indicator under the change of threshold and weight combination
图 7 预测修正试验试验结果（轨迹1）
Figure 7. Predicting and correcting experimental results (track 1)
图 8 预测修正试验试验结果（轨迹2）
Figure 8. Predicting and correcting experimental results (track 2)
图 9 预测修正试验试验结果（轨迹3）
Figure 9. Predicting and correcting experimental results (track 3)
表 1 试验环境与关键参数配置
Table 1. Experimental environment and key parameter configuration
配置项 说明 配置项 说明 GPU NVIDIA RTX 3090（24 GB） 内存 16 GB CPU Intel Xeon Gold 6226R（16核，2.9 GHz） CUDA 版本号：11.8 初始学习率 0.001，在验证集性能无提升时按0.5衰减 cuDNN 版本号：v8.9.0 窗口大小 64 PyTorch 版本号：2.0.0 潜在表示维度 64 训练轮数 100 批次大小 128 隐藏层维度 128 
下载: 导出CSV
表 2 原始数据字段
Table 2. Raw data field
字段 含义 字段 含义 字段 含义 字段 含义 字段 含义 河北省保定市容城县 飞行位置 4370 海拔高度（精确到小数点后2位，乘102后传输） 1 坐标系类型1：WGS-84 2：CGCS2000 3：GLONASS-PZ90 155 航迹角（精确到小数点后1位，乘10后传输） 50 累计飞行时长（s） 23 实时飞行速度（精确到小数点后1位，乘10后传输） 1270 飞行高度（精确到小数点后2位，乘102后传输） 56d6133f-bc26-44bb-a842-ff39dff865d2 Uid生成的字符串 2 违规标志0：正常；1：未注册；
2：未上报；3：未在申请空域内0 是否发送给公安0：不发送1：发送 4104baffof9kl0099901 无具体含义 390217536 当前位置的纬度（精确到小数点后7位，乘107后传输） 1159842275 当前位置经度
（精确到小数点后7位，乘107后传输）39.0217536 当前位置的纬度（精确到小数点后7位） 115.9842275 当前位置的经度（精确到小数点后7位） 914403007954257495 制造商代码 3N3BJ7Q01200H5-20231126-65kG7z2 飞行记录编号 2023-11-26 15：10：19 系统接受时间 20231126151019 时间戳 UAS-DEFAULT 实名登记号（无登记号时以UAS-DEFAULT填充）
下载: 导出CSV
表 3 异常检测对比试验结果
Table 3. Anomaly detection comparison experimental results
模型 精确率 召回率 调和均值 AUPRC AUROC $ \mathrm{T}\mathrm{P}\mathrm{R}@\mathrm{F}\mathrm{P}\mathrm{R}=1\mathrm{\%} $ $ \mathrm{T}\mathrm{P}\mathrm{R}@\mathrm{F}\mathrm{P}\mathrm{R}=0.5\mathrm{\%} $ E2P-AKL 0.903 0.845 0.873 0.902 0.946 0.612 0.492 简化模型 0.881 0.812 0.845 0.878 0.932 0.574 0.456 LSTM-AE 0.836 0.784 0.809 0.852 0.917 0.528 0.402 ConvLSTM 0.842 0.792 0.816 0.858 0.92 0.536 0.408 MSCRED 0.851 0.798 0.824 0.866 0.924 0.542 0.414 Isolation Forest 0.801 0.750 0.775 0.812 0.890 0.480 0.360 One-class SVM 0.792 0.742 0.766 0.804 0.884 0.468 0.348
下载: 导出CSV
表 4 预测修正试验结果
Table 4. Experimental results of the prediction correction function
类型 均方误差 均方根误差 轨迹误差 高度误差 预测模块 0.154 0.392 1.284 0.965 KF模块 0.131 0.362 1.102 0.832 E2P-AKL模型 0.094 0.307 0.846 0.654
下载: 导出CSV
表 5 航迹异常点模型处理性能结果（轨迹1）
Table 5. Performance results of model processing for anomaly points (track 1)
异常类别 异常点数 推理延迟/ms 内存占用峰值/MB 高度异常 59 60 128 转向角异常 102 58 104 纬度异常 105 61 109 经度异常 124 59 115 速度异常 22 57 146 组合异常 318 60 211
下载: 导出CSV
表 6 航迹异常点模型处理性能结果（轨迹2）
Table 6. Performance results of model processing for anomaly points (track 2)
异常类别 异常点数 推理延迟/ms 内存占用峰值/MB 高度异常 54 60 105 转向角异常 99 62 102 纬度异常 101 59 220 经度异常 127 61 155 速度异常 26 59 141 组合异常 303 60 198
下载: 导出CSV
表 7 不同窗口大小试验试验结果
Table 7. Experimental results of different window sizes
窗口大小 调和均值 精确率 召回率 CPU推理延迟/ms GPU推理延迟/ms 内存占用峰值/MB 16 0.862 0.900 0.825 25 12 168 32 0.868 0.910 0.830 60 22 223 64 0.873 0.920 0.830 260 45 568 128 0.858 0.925 0.800 520 90 789 256 0.832 0.935 0.750 1 800 180 1 128
下载: 导出CSV
-
[1] 陶冶, 汤锦辉, 闫震, 等. 融合表征转换与模式回归的航迹插补方法[J]. 航空学报, 2026, 47(1): 272-293.TAO Ye, TANG Jin-hui, YAN Zhen, et al. A trajectory imputation method integrating representation transformation and pattern regression[J]. Acta Aeronautica et Astronautica Sinica, 2026, 47(1): 272-293. [2] DONG X, TIAN Y, DAI L, et al. A new accurate aircraft trajectory prediction in terminal airspace based on spatio-temporal attention mechanism[J]. Aerospace, 2024, 11(9): 718. doi: 10.3390/aerospace11090718 [3] 张鼎开, 王一喆, 王鹏飞, 等. 基于双重时空掩码与双向Mamba状态空间的高速公路交通异常事件检测模型[J]. 交通运输工程学报, 2025, 25(4): 311-327. doi: 10.19818/j.cnki.1671-1637.2025.04.022ZHANG Ding-kai, WANG Yi-zhe, WANG Peng-fei, et al. Expressway traffic anomaly event detection model based on dual spatio-temporal masks and bidirectional Mamba state space modeling[J]. Journal of Traffic and Transportation Engineering, 2025, 25(4): 311-327. doi: 10.19818/j.cnki.1671-1637.2025.04.022 [4] PANG G S, SHEN C H, JIN H D, et al. Deep weakly-supervised anomaly detection[C]//ACM. Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. New York: ACM, 2023: 1795-1807. [5] CHEN B, REN Z, CHENG L. Conformalized time series with semantic features[J]. Advances in Neural Information Processing Systems, 2024, 37: 121449-121474. [6] LEE J, XU C, XIE Y. Kernel-based optimally weighted conformal time-series prediction[C]//ICLR. The Thirteenth International Conference on Learning Representations. Singapore: ICLR, 2025: 1-31. [7] ZENG G X, CHEN S X, WU H, et al. Continuous-discrete extended Kalman filtering based on the neural ordinary differential equations method[J]. Engineering Applications of Artificial Intelligence, 2024, 133: 108611. doi: 10.1016/j.engappai.2024.108611 [8] NI X, REVACH G, SHLEZINGER N. Adaptive KalmanNet: Data-driven Kalman filter with fast adaptation[C]//IEEE. ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). New York: IEEE, 2024: 5970-5974. [9] SHEN S Y, CHEN J C, YU G F, et al. KalmanFormer: Using transformer to model the Kalman Gain in Kalman Filters[J]. Frontiers in Neurorobotics, 2025, 18: 1460255. doi: 10.3389/fnbot.2024.1460255 [10] CHEN X J, LI Y P. An overview of differentiable particle filters for data-adaptive sequential Bayesian inference[EB/OL]. (2023-12-14). https://arxiv.org/abs/2302.09639 [11] WAN Z Y, ZHAO L. DnD filter: Differentiable state estimation for dynamic systems using diffusion models[C]//IEEE. 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). New York: IEEE, 2025: 13851-13858. [12] REIS M J C S, REIS A J D. Edge-based real-time fault detection in UAV systems via B-spline telemetry reconstruction and lightweight hybrid AI[J]. Sensors, 2025, 25(16): 4944. doi: 10.3390/s25164944 [13] AHMAD R, ALKHAMMASH E H. Online adaptive Kalman filtering for real-time anomaly detection in wireless sensor networks[J]. Sensors, 2024, 24(15): 5046. doi: 10.3390/s24155046 [14] POL A A, GOVORKOVA E, GRONROOS S, et al. Knowledge distillation for anomaly detection[EB/OL]. (2023-10-09). https://arxiv.org/abs/2310.06047 .[15] YAO L Y, GAO S B. Dual-student knowledge distillation networks for unsupervised anomaly detection[EB/OL]. (2024-02-01). https://arxiv.org/abs/2402.00448 .[16] KINGMA D P, BA J. Adam: A method for stochastic optimization[EB/OL]. (2014-12-22). https://arxiv.org/abs/1412.6980 .[17] LOSHCHILOV I, HUTTER F. Decoupled weight decay regularization[EB/OL]. (2017-11-14). https://arxiv.org/abs/1711.05101 .[18] 张淼, 何仪娟, 杨博宇, 等. 基于LSTM-KF模型的高速列车群组追踪运行轨迹预测方法[J]. 交通运输工程学报, 2024, 24(3): 296-310. doi: 10.19818/j.cnki.1671-1637.2024.03.021ZHANG Miao, HE Yi-juan, YANG Bo-yu, et al. Trajectory prediction method for high-speed train group tracking operation based on LSTM-KF model[J]. Journal of Traffic and Transportation Engineering, 2024, 24(3): 296-310. doi: 10.19818/j.cnki.1671-1637.2024.03.021 [19] SHRESTHA R, MOHAMMADI M, SINAEI S, et al. Anomaly detection based on LSTM and autoencoders using federated learning in smart electric grid[J]. Journal of Parallel and Distributed Computing, 2024, 193: 104951. doi: 10.1016/j.jpdc.2024.104951 [20] LI J Z, SHI Y, ZHANG T, et al. Radar precipitation nowcasting based on ConvLSTM model in a small watershed in North China[J]. Natural Hazards, 2024, 120(1): 63-85. doi: 10.1007/s11069-023-06193-6 [21] XIE T M, XU Q F, JIANG C X. Anomaly detection for multivariate times series through the multi-scale convolutional recurrent variational autoencoder[J]. Expert Systems with Applications, 2023, 231: 120725. [22] OLIVE X, KRUMMER J, FIGUET B, et al. Filtering techniques for ADS-B trajectory preprocessing[J]. Journal of Open Aviation Science, 2024, 2(2): 7882-7899. -
计量
- 文章访问数: 92
- HTML全文浏览量: 46
- PDF下载量: 10
- 被引次数: 0
