A 10,000-word article on automobile control in the intelligent era

3.3 Gradual application of advanced control theories and methods

The continuous improvement of the computing power of automotive chips has brought opportunities for the application of advanced control theories and methods, machine learning and deep learning in real vehicles. For information about machine learning and deep learning, please refer to the literature [123−124]. Due to space considerations, this article will not elaborate on them. The following will take advanced control methods such as model predictive control and nonlinear control as examples to briefly summarize their latest progress in automotive control applications.

Intelligent vehicle control and decision-making problems are usually multi-objective and constrained. Model predictive control (MPC) [125] was first applied to industrial process control and is capable of handling multi-variable, multi-objective optimization problems with constraints. Since its first application in engine idle speed control in the 1990s [126], a large number of literature on the application of MPC in vehicle control has emerged. Currently, MPC has gradually become an important method for solving key control problems such as automotive powertrain [14, 127-128], energy management [129-130], emissions [131], and vehicle stability [133]. General Motors (GM) GM Motor spent five years developing an MPC-based turbocharged engine management system[132]. The system will be gradually applied to GM's mass-produced models after 2018, achieving a breakthrough in the mass production of predictive control applications in automobiles worldwide. In the context of smart cities and connected transportation, by introducing future traffic preview information, MPC has broader application prospects in vehicle predictive energy saving[130] and predictive safety[134]. With the increasing complexity of vehicle systems, which contain a large amount of uncertainty or dynamics that cannot be modeled mechanically, learning-based MPC has gradually become a hot topic in automotive control[135]. In addition, embedded optimization algorithm research[136] will also accelerate the application of MPC in future intelligent automotive control.

The automobile control system involves many tracking control problems of underlying actuators. Most tracking control systems have the characteristics of strong nonlinearity and high coupling, which makes it difficult for traditional PID control to meet the control requirements. Model-based nonlinear control methods such as feedback linearization [137] and backstepping [10] can greatly improve the tracking performance of the control system and have good robustness. Although they require relatively complex formula derivation and simplification, they still have great application potential. To solve the tracking control problem in the actual engineering of automobile control, the reference [138] proposed a novel "three-step method" nonlinear control system design method. The obtained control law is presented in the form of "steady-state control-feedforward control-error feedback control", which has the advantages of clear design ideas, clear physical meaning, and simple engineering implementation structure. This method was first applied to the rail pressure control of direct injection gasoline engines, and the experimental bench test was completed with the support of FAW Technology Center [139]. Subsequently, this method was extended to some high-order and non-affine systems, providing solutions to problems such as transmission shift control [140], vehicle stability control [141], and fault diagnosis [142].

In the future, advanced control theories and methods will penetrate into all aspects of automotive control to empower intelligence, such as the application of game theory in human-machine co-driving[143] and energy distribution of electric vehicles[144], and the application of logic control[145] in vehicle scheduling and fault diagnosis.

3.4 New functions of intelligent automobile control systems

In the context of the development of smart transportation and vehicle intelligence, cars can obtain more external information, and advanced control methods represented by model predictive control and new artificial intelligence technologies represented by deep learning and reinforcement learning will receive widespread attention. As mentioned earlier, with the gradual maturity of cloud computing and the significant improvement of vehicle edge computing capabilities, it has become possible to implement complex advanced algorithms online in real vehicle control systems. The development of these technologies not only enhances the upper-level driving decision-making and planning capabilities of automobiles, but also creates conditions for the derivative development of new systems/new functions for automobile steering, braking, and driving.

3.4.1 Driving Decision-making and Planning in Complex Scenarios on Open Roads

Compared with simple working conditions, driving decisions in complex environments on open roads require analysis and modeling of complex information, and making optimal decisions with the support of massive information. The control context is shown in Figure 4. Vehicle and environment modeling, behavior prediction of other traffic participants, and uncertainty analysis are the difficulties in driving decisions on open roads. In recent years, deep learning cognitive capabilities[201] and the optimization framework of reinforcement learning have brought new ideas for autonomous decision-making in complex environments on open roads. End-to-end autonomous driving deep learning technology can achieve direct control from images to vehicle direction by training on collected driving image sets[110]. NVIDIA further proposed an explainable neural network system for end-to-end solutions[111], and applied it to lane estimation[112], lane keeping, etc.[113].

Fig.4 The technology of decision -making and planning in open road with complex scenario

In response to environmental uncertainty, the dynamic Bayesian network in machine learning can establish a probability distribution model of complex states and is a typical solution for identifying and predicting traffic environment uncertainty [114]. For example, reference [115] uses Bayesian change points to predict the potential behavior of other traffic participants. By evaluating the results of the interaction between the vehicle and other vehicles, control strategies such as slowing down or accelerating can be implemented. Under the premise of effectively utilizing complex information, the core of autonomous driving decision-making is how to make the best decision in an interactive environment, which is where reinforcement learning comes in handy: for example, reference [116] studies the optimal control strategy based on the AC algorithm for the right turn speed control problem under weak signal conditions; reference [117] proposes a throttle control method based on reinforcement learning for the following vehicle problem, and It claims that its control strategy is close to the optimal control solution. Similarly, reference [118] proposed a vehicle longitudinal optimization control scheme based on parameterized reinforcement learning. In response to the uncertainty problem caused by the difference between the simulated environment and the actual environment, reference [119] proposed a new reinforcement learning framework based on image semantic segmentation network. In response to numerical problems in practical applications, reference [120] adopted an inverse reinforcement learning method to learn high-value strategies in the human driving process. It was verified in simulation that this method can learn human driving skills, such as accelerating to force surrounding vehicles to slow down. In order to reduce dependence on data and consider safety, reference [121] proposed an autonomous learning process, which can perform self-improvement learning by checking classification, clustering, autonomous learning, online learning and optimization models using only a small amount of data.

In fact, at present, there are still problems with driving decision-making and planning based on deep learning [122], such as data dependence. Some studies have attempted to propose new learning frameworks and strategies [121]. In the future, a development trend in the research of autonomous driving control of automobiles is to continuously improve the ability of machine learning in processing, transforming and optimizing decisions on complex information, break through the autonomous decision-making technology in complex scenes on open roads, and improve the ability to handle emergencies.

3.4.2 Intelligent Energy Saving of Connected Vehicles

Intelligent energy-saving technology is based on traditional energy-saving technology. It uses the network information of human-vehicle, vehicle-vehicle, and vehicle-road communication to comprehensively optimize the vehicle's driving state, thereby further improving the energy efficiency of the entire vehicle. As shown in Figure 5, the vehicle uses navigation, high-precision maps, and traffic environment prediction information to comprehensively consider the impact of future traffic conditions on the vehicle's driving economy. It uses the optimization idea of ​​the rolling time domain to predictively coordinate and optimize the vehicle's driving, braking, and transmission system states, ultimately achieving a reduction in the vehicle's driving energy [146-149]. Related studies have shown that using network information to reasonably plan the speed of vehicles can increase road traffic rates by more than 10%. With the continuous increase in the penetration rate of networked vehicles, it can eventually increase by 50% to 90% [150]. In addition, through the coordinated optimization of vehicle driving strategies and traffic, the overall traffic energy consumption can be reduced by 15% to 20% [151-152]. If the optimization matching of the vehicle powertrain system is further considered, the energy consumption can be reduced to 30%. Considering the huge energy-saving potential that connected technology brings to automobiles, the U.S. Department of Energy's Advanced Research Projects Agency-Energy (ARPA-E) spent $32 million in 2017 to fund 10 research teams composed of top American universities and automotive R&D departments to carry out a three-year research and development project code-named "NEXTCAR". The project aims to study the energy-saving technology of the next generation of intelligent connected vehicles from different angles, hoping to achieve a 20% improvement in fuel economy [153]. In this context, intelligent energy-saving technology has become one of the hot research directions in the field of automotive intelligence.