1. Introduction
“One-pedal driving” is one of the many remarkable changes as the paradigm of vehicle platforms switches from an internal combustion engine to electric vehicles (EVs) [
1,
2,
3]. It makes driving with only the accelerator pedal possible by generating regenerative braking torque when the accelerator pedal is released. This can increase driver convenience as there are fewer pedals to shift.
The smart regenerative braking system (SRS) is one of the “one-pedal driving” technologies, which can take advantage of both the improvement of driver convenience and energy efficiency [
4]. It is a type of advanced driver assistance system (ADAS) in that it supports driving using a radar sensor. Unlike general one-pedal driving, it does not always generate regenerative torque when the accelerator pedal is released. However, it does generate adequate regenerative torque according to car-following situations. The amount of regenerative torque is appropriately determined by relative distance, relative velocity, and speed of the ego vehicle.
SRS is an intelligent braking system which controls the friction braking or regenerative braking to meet certain goals such as safety, energy efficiency, and braking performance. There were various studies regarding the intelligent braking system. In Reference [
5], an ADAS system regarding brake assistance based on the driver behavior and situation was proposed. Lin et al. suggested active control of regenerative braking to improve the braking performance of an electric vehicle [
6]. In Reference [
7], a brake-by-wire actuator was designed to shorten the braking distance and time. Similar to Reference [
5], SRS improves the safety of the vehicle by intelligent braking. In addition, the system can enhance driver convenience.
Since SRS requires the deceleration to be used as a reference value in the automatic regenerative torque control, a deceleration planning algorithm is necessary. There are some points to consider when designing a deceleration planning algorithm for SRS. The first essential point is the harmony between the acceleration by the human driver and the deceleration by the SRS. It can be achieved by generating a deceleration profile that is similar to human driving [
8,
9]. A second crucial point is the applicability of SRS to diverse deceleration conditions to maximize the advantages of SRS, such as driver convenience and energy efficiency. A third important point is the usability in vehicles online.
In order to mimic human driving, a machine learning technique is an appropriate method because it can learn the characteristics of human driving. Various researches were conducted to predict acceleration and velocity using an artificial neural network (ANN) [
10,
11,
12,
13,
14,
15,
16]. In particular, Reference [
17] proposed an ANN model to predict acceleration with four inputs: relative distance, relative velocity, velocity, and desired speed. In Reference [
18], a fuzzy rule-based neural network which used relative distance, relative speed, vehicle speed, and actions of the previous time step was designed to capture the vehicle motions. Khodayari et al. proposed a neural network model focusing on human driving behavior in Reference [
16]. Unlike other researches, it used estimated instantaneous reaction delay to capture the realistic driver behavior of moving the foot between the accelerator and the brake pedal. These researches showed an acceptable performance of acceleration prediction. However, they lacked harmonization with the acceleration by human driver when applied to SRS because they could not accurately represent the driver behavior, especially in deceleration scenarios. In addition, these algorithms could only be applied in car-following conditions.
To maximize the advantages of SRS, the deceleration planning algorithm should generate deceleration in not only car-following condition, but other deceleration conditions as well. Intelligent transportation system (ITS) information is actively used in ADAS and energy management systems in various driving situations [
19,
20]. The deceleration planning algorithm can be applied in diverse deceleration situations using ITS information.
A third important point is the usability in vehicles. The deceleration planning algorithm should operate in vehicles online. Accordingly, only the information which is acquired in vehicles on-board can be used as an input of the algorithm.
Some researches related to deceleration planning and speed prediction were conducted. Yeon et al. developed a recurrent neural network (RNN) model to predict vehicle speed with a 10-s prediction horizon [
21]. The proposed algorithm showed better prediction performance than other algorithms that were suggested. However, it had a limitation in that it could only be applied in the specific route used to train the RNN model because the algorithm used the position in the route as an input of the model. Min et al. proposed an RNN model to generate a deceleration profile at braking conditions. The accuracy of the model was improved by using a physical constraint to stop at the specific location. However, the model could only be used in braking conditions at a traffic light.
To overcome these limitations of previous research, this paper proposes a deceleration planning algorithm using classified multi-layer perceptron (MLP) models. The noticeable feature of the proposed algorithm is the classified structure. To improve the planning performance of the model, the deceleration models were developed individually in three different deceleration conditions: car-following, speed bump, and intersection. Each model was trained with the driving data acquired through vehicle experiments. Unlike the previous studies on driver models, the suggested MLP model considers the human reaction delay in deceleration, which results in the generated deceleration mimicking the coasting behavior. In particular, learning of the coasting behavior was adequately performed by appropriately processing the target data. In addition, the reference value of acceleration was used as input. As a result, the vehicle reached the required velocity successfully with the suggested algorithm in the three deceleration conditions. Moreover, the model was applicable in more diverse situations by using ITS information.
Because there are three models which are specialized in the three deceleration conditions, the planning algorithm should select the MLP model to be used. For this, a state recognition algorithm was designed. Using data such as accelerator pedal position, velocity, relative distance, and distance to speed bump, it recognizes the necessity of deceleration and the cause of it. Using the cause of deceleration, the adequate model which has specialized inputs in each deceleration condition is selected and used in planning. This results in acceptable performance in using the suggested algorithm in the SRS.
The proposed algorithm was validated through driving simulations using driving data. Then, the safety of the proposed algorithm was evaluated. In addition, the similarity to human driving was analyzed. Moreover, the planning results of the proposed algorithm were compared to the results with integrated structure.
The rest of the paper is organized as follows:
Section 2 illustrates an overview of the entire algorithm.
Section 3 describes the state recognition algorithm which determines the driving state and the cause of deceleration, called the “deceleration condition”. The set of input and hidden layers of the MLP model, and the hyper-parameter optimization methods are described in
Section 4. In
Section 5, the training of the MLP models and the data used in the training are described. In
Section 6, the simulation results of the suggested algorithm are shown and compared to the algorithm of integrated structure.
Section 7 discusses the results and concludes the paper.
2. System Overview
2.1. Description of Deceleration Conditions
To minimize the number of braking instances by a driver, the deceleration planning algorithm should be able to generate deceleration in diverse situations. In this research, three deceleration situations were selected: car-following, speed bump, and intersection. Car-following was chosen because car-following is most common in both urban and highway driving. Speed bump and intersection were selected because they are major deceleration causes in urban driving. With these three deceleration conditions, the algorithm can be applied in most deceleration situations. In urban driving, traffic lights represent a major deceleration situation. However, its signal cannot be acquired in the target vehicle, which was used in the vehicle experiments on-board; thus, the condition is excluded in this research because the proposed algorithm was designed considering its usability in vehicles online. Details of each deceleration condition are given below.
The car-following condition is a driving situation where the leading car gets close to the ego vehicle. In this situation, the relative velocity is negative and the relative distance decreases for a few seconds.
The second situation is a speed bump, where the human pushes the brake pedal to pass through it smoothly. The location of the speed bump and ego vehicle in the route are acquired from the navigation device in the vehicle used for the experiment in this research. Distance to speed bump was calculated using the locations and used to check the reason for deceleration.
The third condition is an intersection. At an intersection, there are three options of driving according to the path: right turn, left turn, and straight. In right-driving countries, the movement of the vehicle is decided by the traffic signal for a left turn and going straight. To design the deceleration model in the two situations, the signal of the traffic light is required. However, it cannot be acquired from the current navigation device in the vehicle. Therefore, only the right turn was included in the intersection condition.
2.2. Algorithm Overview
As mentioned in
Section 1, the suggested algorithm has a classified structure. Compared to the integrated structure, the number of inputs in each MLP model is reduced, and the size of the model decreases. In addition, better planning performance can be anticipated, because each model is trained with the deceleration profile which is matched to each deceleration situations.
Because of the classified structure, the model used in the deceleration planning should be selected. In addition, the vehicle motion should be monitored because automatic regeneration can operate only when the driver does not push the pedals. To monitor the vehicle motion and determine the current cause of deceleration, a state recognition algorithm was designed.
Figure 1 shows the overall structure of the entire algorithm. Firstly, the state recognition algorithm determines the driving state which refers to the vehicle motion based on the driver’s behavior of pushing the pedal. Then, it determines the cause of deceleration, called the “deceleration condition”. After the deceleration condition is selected, the deceleration model generate the deceleration of the next time step using the trained MLP model suitable for the deceleration condition.
7. Conclusions
This study suggested a deceleration planning algorithm which consisted of classified MLP models. The MLP models were trained with the human driving data acquired from vehicle experiments. In addition, diverse hyper-parameters were selected by hand-tuning and a grid-search algorithm. The validation results in each deceleration condition are summarized as follows:
The best model in car-following showed a validation RMSE of and a total RMSE of ;
The best model in speed bump showed a validation RMSE of and a total RMSE of ;
The best model in intersection showed a validation RMSE of and a total RMSE of .
The driving simulation was conducted using trained models with the best hyper-parameters. Through the driving simulation, it was validated that the vehicle motion with SRS based on the proposed algorithm is safe in car-following conditions and satisfies the minimum velocity at speed bumps and intersections. In addition, the deceleration planning results showed the coasting behavior in the initial part of the deceleration by using the “coasting time” and the “reference acceleration” as inputs for the MLP models. This can reduce the driver discomfort in using the SRS and improve the acceptability of the SRS. The proposed algorithm was compared with other deceleration planning algorithms with an integrated MLP model. The results showed that the planning algorithm with a classified structure has more similarity with human driving than the integrated structure.
In the future, this research will be applied to the SRS in EVs via integration with a regenerative torque controller. Furthermore, the deceleration conditions will be extended by using more types of ITS information such as curvature, traffic lights, and speed limits.