基于半监督极限学习机的隧道内车辆RSSI定位方法

为了提高公路隧道突发事件的判别效率，实现道路交通状态全天候监测，以智能公路上泛在无线传感网络为基础，研究了基于信号强度指示值(RSSI)的网联车辆定位问题；考虑到隧道内车辆的连续运动特性，提出了一种带有局部线性嵌入(LLE)算法的半监督极限学习机(SSELM)实现RSSI指纹定位；离线阶段利用LLE对少量已标记位置的RSSI样本和大量无标记样本进行降维处理，辨识表征目标位置信息的高维数据对应的低维流形，再基于改进的半监督学习拟合降维后的RSSI与位置的映射关系；在线阶段将实时采集的RSSI数据进行流形降维后，输入校准好的SSELM中估计目标位置；采用无迹卡尔曼滤波平滑估计位置。试验结果表明：相比于已有半监督学习算法，提出的方法在不同车辆行驶速度和部署间距下均能取得较优的定位性能；当已标记数据占比(减少了50%~90%)、未标记数据数量(0~1 000个)和检测器部署间距(10~25 m)等关键指标变化后，本文方法的定位性能仍然保持最佳，其平均误差最低为3.09 m；计算复杂度上，当已标记数据为30%，即仅采集96个参考点样本时，其平均定位误差为3.8 m，训练时间低至8.7 s。可见，带有局部线性嵌入算法的半监督极限学习机在稀疏或密集传感器部署环境中，对不同行驶速度的车辆均能提供理想的定位性能，且训练时间短、样本依赖性低，是进行隧道内网联车辆辅助定位的一种有效方法。

RSSI positioning method of vehicles in tunnels based on semi-supervised extreme learning machine

To improve the identification efficiency of highway tunnel emergencies and to realize the full-time monitoring of road traffic conditions, the problem of positioning connected automated vehicles based on the received signal strength indicator (RSSI) was studied based on a ubiquitous wireless sensor network on an intelligent road. Considering the continuous motion characteristics of vehicles in a tunnel, a semi-supervised extreme learning machine (SSELM) with a locally linear embedding (LLE) algorithm was proposed to achieve the RSSI fingerprint positioning. In the offline phase, the dimension reductions for a few RSSI sample datasets with their vehicle positions marked and for a mass of unmarked ones were conducted by using the LLE, and the low-dimensional manifolds corresponding to their high-dimensional data, which represented the target's location information, were recognized. The mapping relationship between the RSSI data and vehicle positions was fitted based on the SSELM. In the online phase, real-time collected RSSI data after manifold dimensionality reduction were put into the calibrated SSELM to estimate the positions of vehicles. The estimated position was smoothed using an unscented Kalman filter (UKF). Analysis result shows that compared with the existing semi-supervised learning algorithms, the proposed method can achieve better positioning performance regardless of the vehicle travel speed and deployed distances. Under a change in key variables, such as the proportion of marked data (reduced by 50%-90%), number of unmarked data (0-1 000), and deployed sensor distance (10-25 m), the proposed method still has the best positioning performance with a minimum average error of 3.09 m. In terms of computational complexity, when the marked data comprise 30% of the dataset (only 96 reference points), the average positioning error is 3.8 m and the training time reduces to 8.7 s. Therefore, the proposed SSELM with LLE algorithm can provide promising positioning performance for vehicles with different driving speeds in an environment with sparsely or densely deployed sensors. In addition, it has a shorter training time and lower dependence on sample size, which makes it an effective method for the auxiliary positioning of connected automated vehicles in tunnels. 2 tabs, 11 figs, 34 refs.