小样本下基于迁移学习与LSTM的雾天高速公路车辆跟驰模型

由于在现实生活中能够采集到的不同雾天等级的高速公路车辆跟驰样本有限，导致雾天跟驰模型精度不佳，为此在长短时记忆神经网络（long short-term memory,LSTM）跟驰模型的基础上，采用迁移学习（transfer learning,TL）方法来提升雾天跟驰模型的性能。利用驾驶模拟实验平台搭建高速公路雾天与正常天气2种实验场景进行驾驶模拟实验，获得296组正常天气下（源域）的跟驰样本与100组雾天下（目标域）的跟驰样本。提出了基于最长公共子序列（longest common sequence solution,LCSS）的迁移样本选择方法，从源域中选出100个样本迁移至目标域中，通过扩大训练样本提升LSTM从源域、目标域特征到目标域输出的端对端泛化学习能力，得到雾天高速公路车辆跟驰模型。为对比所提样本迁移方法对LSTM模型的效用，将LSTM-TL模型与训练样本全部来源于源域的LSTM-S模型和训练样本全部来源于目标域的LSTM-T模型进行对比，LSTM-TL模型的均方误差、均方根误差和平均绝对误差比LSTM-S模型分别减小47.5%、27.7%和46.5%，比LSTM-T模型...

A Car-following Model for Expressway under Foggy Weather Based on Transfer Learning and LSTM with Small-sample

Due to the fact that it is difficult to collect car-following samples at different fog levels and the samples that can be collected are limited, and the accuracy of car-following models is generally poor under the condition of foggy weather. A transfer learning (TL) approach is used to improve the performance of a car-following model under the condition of foggy weather based on the long short-term memory (LSTM) neural network technique. A driving simulator is used to set up two types of experimental scenes (normal and foggy weather) for driving experiments on an expressway. Driving behavior data from 296 groups of car-following samples under the condition of normal weather (source domain), and 100 groups of car-following samples under the condition of foggy weather (source domain) is collected. A selection method for transfer samples is proposed based on the longest common sequence solution (LCSS). 100 samples are selected from the source domain and transferred to the target domain. The end-to-end generalization learning capability of the LSTM from features of both source and target domains to output of target domain is improved by expanding the training samples to develop a car-following model for expressway under the condition of foggy weather. To compare the utility of the proposed method in improving the LSTM model, the LSTM-TL model is compared with the LSTM-S model with all training samples from the source domain, and the LSTM-T model with all training samples from the target domain. The mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE) of the LSTM-TL model is 47.5%, 27.7%, and 46.5% less than the LSTM-S model respectively; while 31.1%, 17.0%, and 29.9% less than the LSTM-T model. To compare the performance of different models when only 100 groups of samples from the target domain are available, the LSTM-TL model is compared with three models, Gipps, IDM, and BP. The MSE, RMSE, and MAE of the LSTM-TL model is 18.5%, 8.0%, and 25.9% less than the Gipps model respectively, which performs best among the three models. Study results also show that the LSTM-S model has poor prediction accuracy when directly applied to the prediction of the target domain, and the use of sample transfer can significantly improve its accuracy. The LCSS method is effective for sample screening from the source domain, and the LSTM-TL model trained by transferring 100 samples from the source domain to the target domain has the highest accuracy. In case of a small sample, the Gipps model with fewer parameters has a better prediction accuracy than the LSTM-T or LSTM-S models. However, the LSTM-TL model still achieves the highest accuracy among all of the above models, due to the fact that the transfer learning can transfer useful knowledge from source domain samples to the target domain.