CN114881200A - A method for predicting the acceleration of car-following vehicles in foggy environment based on transfer learning and LSTM-NN - Google Patents

Abstract

A following vehicle acceleration prediction method in a foggy environment based on transfer learning and LSTM-NN comprises source domain sample selection and LSTM model training. Selecting source domain samples, including screening basis and quantity selection; the screening basis is that the similarity between the obtained source domain sample and the target domain sample is solved through an improved LCSS algorithm, and the follow-up behavior of a driver is considered to be greatly dependent on the speed difference delta v between a front vehicle and a rear vehicle, the distance h between the front vehicle and the rear vehicle and the acceleration a of the vehicle, so that gamma is selected as a description characteristic for measuring the two samples; the quantity selection is to select different quantities of source domain samples to train the LSTM model according to the similarity of the source domain samples, and select the proper quantity through model performance comparison. The migration sample screening mechanism adopted by the invention can effectively reduce negative migration, so that a following model with more ideal performance is obtained by using fewer following samples in a foggy environment.

Description

A prediction method for the acceleration of car following vehicles in foggy environment based on transfer learning and LSTM-NN

technical field

The invention relates to a method for predicting the acceleration of a car-following vehicle under foggy conditions. The method for calculating the similarity between samples is used to select suitable migration samples, so as to realize the improvement of the performance of the LSTM car-following model under adverse weather conditions through sample migration. It belongs to the field of intelligent transportation.

Background technique

The low visibility in the foggy environment shortens the driver's line of sight, and is prone to nervousness and fatigue. Car-following model is of great significance for understanding and describing the characteristics of traffic flow in foggy environment, and then improving traffic safety in foggy weather.

From the perspective of modeling methods, car following models can be divided into two categories: theory-driven and data-driven. In the current research, most car-following models in bad weather belong to the theory-driven category. The advantage of the theory-driven model is that it can intuitively describe the car-following process with certain variables, but its disadvantage is that it is difficult to accurately describe the driver's driving experience and fuzzy perception characteristics. With the development of artificial intelligence and deep learning, data-driven modeling methods have received extensive attention from researchers. Among them, the artificial neural network method has been proved to be able to better describe the car following behavior of drivers with different characteristics by training data samples. The unique memory ability of long-short-term memory neural network has shown good performance in the modeling of car-following behavior, but the premise is that a large number of training samples are required, which is easy to achieve under normal weather conditions. However, due to the uncertainty and few occurrences of fog, it is relatively difficult to obtain car-following samples in a foggy environment. The small sample size in the foggy environment will lead to poor performance of the LSTM model.

In view of this, the present invention adopts sample migration to increase the number of training samples, and uses the car-following data in normal weather to assist the LSTM neural network to learn the car-following behavior in the foggy environment, thereby improving the performance of the LSTM fog-following model.

SUMMARY OF THE INVENTION

The invention proposes a method for predicting the acceleration of a car-following vehicle in a foggy environment based on migration learning and LSTM-NN, and aims to solve the problem that it is relatively difficult to obtain a car-following sample in the foggy environment. A flow chart of a method for predicting the acceleration of a car-following vehicle under foggy conditions based on transfer learning and LSTM-NN is shown in Figure 1.

The method consists of two parts, source domain sample selection and LSTM model training. The source domain sample selection includes screening basis and quantity selection; the screening basis is the similarity between the source domain sample and the target domain sample obtained by solving the improved LCSS algorithm, and considering that the driver's car following behavior is very large. The degree depends on the speed difference Δv between the front and rear vehicles, the head distance h and the acceleration a of the vehicle, so γ (Equation 1) is selected as the descriptive feature to measure the two samples, where θ is the adjustment coefficient; the number selection is based on The similarity of the source domain samples, select different numbers of source domain samples to train the LSTM model, and select the appropriate number by comparing the model performance. The training of the LSTM model includes a loss function, an optimizer and an iterative termination condition; the loss function selects the mean square error; the optimizer selects the Adam algorithm; the iteration termination condition is that the maximum number of training times is 1000, and the model convergence trend is the target The value is 0.0001.

A method for predicting the acceleration of a car-following vehicle in a foggy environment based on transfer learning and LSTM-NN. The implementation steps of the method include the following:

Step 1, collect car-following samples under normal weather (source domain) and foggy environment (target domain);

Step 2: First, determine the descriptive feature γ that measures the two samples, and convert the multi-dimensional time series of the samples in the source domain and the target domain into a one-dimensional time series. Descriptive feature γ is calculated according to formula 1: where Δv is the speed difference between the vehicle in front and the rear of the sample, h is the distance between the fronts of the vehicle, a is the acceleration of the vehicle, and θ is the adjustment coefficient (take 0.5);

Step 3: Calculate the similarity between the source domain and the target domain samples. The similarity described in step 3 is determined according to the following method: first, one sample is selected from the source domain and the target domain, and each element of the description feature of the target domain sample is placed in a rectangular range with a length of δ=4 and a width of ε. Search the common elements in the source domain samples, calculate the longest common subsequence between the target domain samples and the source domain samples, divide it by the maximum sequence length to obtain the similarity S; then, calculate the source domain samples {x ₁ according to the above process ,x ₂ ,...,x _n } each sample x _i and the sample {y ₁ ,y ₂ ,...,y _m } in the target domain are similar to S _i =max{s ₁ ,s ₂ ,. ..,s _m }, and finally obtain the similarity S={S ₁ ,S ₂ ,...,S _n } of all source domain samples;

ε=0.6×σ(2)

Among them: σ is the minimum value of the standard deviation of the description feature γ of the selected two samples.

个试验组；Step 4: According to the calculated similarity, different numbers of source domain samples and target domain samples are used to form different experimental groups to train the LSTM together to obtain different acceleration prediction models of car following vehicles under fog. The samples of different experimental groups are determined according to the following methods: first, the migration samples are arranged in descending order according to the value of similarity Si, then the number of migration samples starts from 0 and increases according to the step size B (take 50), until the number of samples _n so far, constitute

a test group;

Step 5: By comparing the model performance indicators (MSE, RMSE, MAE) obtained by different experimental groups, the model with the best performance is selected as the acceleration prediction model of the car-following vehicle in the foggy environment.

Compared with the prior art, the present invention has the following advantages:

(1) Compared with the car-following model driven by theory, the present invention adopts a data-driven modeling method, and can better describe the car-following behavior of drivers with different characteristics by training the LSTM model.

(2) The migration sample screening mechanism adopted in the present invention can effectively reduce the negative migration and improve the performance of the car following model.

(3) For practical applications, a car-following model with more ideal performance can be obtained by using fewer car-following samples under foggy conditions.

Description of drawings

Fig. 1 is the flow chart of the acceleration prediction method of the following vehicle under foggy conditions;

Figure 2 shows the structure of the transfer learning-LSTM fog-following model.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

The LSTM neural network consists of an input layer, several hidden layers and an output layer. In this paper, the current speed v _f (t) of the following vehicle, the acceleration a _f (t), the head distance Δx _f (t), the speed of the preceding vehicle v _p (t) is used as the input variable, and the acceleration of the following vehicle at the next moment is used as the output variable. The time taken by the LSTM neural network model is 3s, the frequency of the collected driving behavior data is 30Hz, and the input matrix length of each time step is 4×90. It is determined that the hidden layer is two layers, and the number of neurons is 40 and 50. In order to prevent the model from overfitting, Dropout is set in the hidden layer, so that a certain neuron in the layer of neurons can be used with a certain probability. (0.2) Temporarily stop working. The structure of the transfer learning-LSTM fog-following model is shown in Figure 2.

The flowchart of the intelligent vehicle intersection parking method according to the present invention is shown in FIG. 1, and specifically includes the following steps:

Step 1, collect car-following samples under normal weather and foggy conditions;

Step 2: First, determine the descriptive feature γ that measures the two samples, and convert the multi-dimensional time series of the samples in the source domain and the target domain into a one-dimensional time series. Descriptive feature γ is calculated according to formula 1: where Δv is the speed difference between the vehicle in front and the rear of the sample, h is the distance between the fronts of the vehicle, a is the acceleration of the vehicle, and θ is the adjustment coefficient, which is taken as 0.5;

Step 3: Calculate the similarity of car-following samples under normal weather and foggy conditions. The similarity described in step 3 is determined according to the following method: first, one sample is selected from the source domain and the target domain, and each element of the description feature of the target domain sample is placed in a rectangular range with a length of δ=4 and a width of ε. Search the common elements in the source domain samples, calculate the longest common subsequence between the target domain samples and the source domain samples, divide it by the maximum sequence length to obtain the similarity S; then, calculate the source domain samples {x ₁ according to the above process ,x ₂ ,...,x _n } each sample x _i and the sample {y ₁ ,y ₂ ,...,y _m } in the target domain are similar to S _i =max{s ₁ ,s ₂ ,. ..,s _m }, and finally obtain the similarity S={S ₁ ,S ₂ ,...,S _n } of all source domain samples;

ε=0.6×σ(2)

Among them: σ is the minimum value of the standard deviation of the description feature γ of the selected two samples.

Step 4: According to the calculated similarity, different numbers of samples from the source domain and samples from the target domain are used to form different experimental groups to train the LSTM collaboratively, so as to obtain the acceleration prediction model of the car-following vehicle in different foggy environments. The samples of different experimental groups are determined according to the following method: first, the migration samples are arranged in descending order according to the value of similarity _Si , and then the number of migration samples starts from 0 and increases according to the step size of 50, until the number of samples is 296, which constitutes 7 a test group;

Step 5: By comparing the model performance indicators (MSE, RMSE, MAE) obtained by different experimental groups, the model with the best performance is selected as the acceleration prediction model of the car-following vehicle in the foggy environment.

By comparing the effects of the method for predicting the acceleration of a car following vehicle under foggy conditions based on transfer learning and LSTM-NN, it can be seen that the LSTM model trained by migrating 150 source domain samples to the target domain has the highest accuracy , the sample transfer improves the performance of the model.

Claims (3)

1. A following vehicle acceleration prediction method based on transfer learning and LSTM-NN is characterized by comprising the following steps:

collecting follow-up samples in normal weather, namely a source domain and a foggy environment, namely a target domain;

determining description characteristics gamma for measuring two samples, and converting a multi-dimensional time sequence of the source domain samples and the target domain samples into a one-dimensional time sequence; the description characteristic gamma is calculated according to formula 1: wherein, Deltav is the speed difference between the front vehicle and the rear vehicle of the sample, h is the distance between the vehicle heads, a is the acceleration of the vehicle, and theta is an adjustment coefficient, and 0.5 is taken;

step three, calculating the similarity of the source domain and the target domain samples;

performing collaborative training on the LSTM by using different experimental groups formed by different numbers of source domain samples and target domain samples according to the calculated similarity to obtain different following vehicle acceleration prediction models in foggy days;

and step five, comparing the performance indexes of the models obtained by different experimental groups, and taking the model with the optimal performance as a following vehicle acceleration prediction model in the foggy weather environment.

2. The method for predicting the acceleration of the following vehicle based on the transfer learning and LSTM-NN as claimed in claim 1, wherein the similarity in the third step is determined according to the following method: firstly, selecting a sample from a source domain and a target domain, searching common elements in a source domain sample in a rectangular range with the length delta being 4 and the width being epsilon for each element of description characteristics of the target domain sample, calculating to obtain the longest common subsequence of the target domain sample and the source domain sample, and dividing the longest common subsequence by the length of the maximum sequence to obtain the similarity S; then, the source domain samples { x ] are computed according to the above process ₁ ,x ₂ ,...,x _n Every sample x in _i Samples with target Domain y ₁ ,y ₂ ,...,y _m S similarity _i ＝max{s ₁ ,s ₂ ,...,s _m Get all the source domain samples finally(iii) similarity of S ═ S ₁ ,S ₂ ,...,S _n }；

ε＝0.6×σ (2)

Wherein: σ is the minimum of the standard deviation of the two samples selected to describe the characteristic γ.

3. The method for predicting the acceleration of the following vehicle based on the transfer learning and LSTM-NN as claimed in claim 1, wherein the samples of the different experimental groups in step four are determined according to the following method: firstly, the migration samples are determined according to the similarity S _i Is arranged from large to small, then the number of shifted samples is increased from 0 and by the step size B until the number of samples n

For each test group, 50 were taken for B.

Images