Master-slave cooperative control method, device and storage medium for dynamic following vehicle distance adjustment

Abstract

The present invention relates to a master-slave cooperative control method for dynamic following vehicle distance adjustment, comprising the following steps: step 1) establishing a synchronous control model based on model prediction; step 2) establishing an optimal state estimation model based on Kalman filtering; step 3) determining a Pareto optimization target and solving an optimal control law; step 4) establishing a GRU deep neural network model to calculate the optimal control law; step 5) a leading vehicle and a trailing vehicle obtain a first and a second control law respectively according to the model; step 6) the trailing vehicle sends the second control law to the leading vehicle, and if the leading vehicle determines that the operation requirement is met, the trailing vehicle is allowed to control the interval according to the second control law, otherwise the leading vehicle sends the first control law to the trailing vehicle, and the trailing vehicle is controlled according to the first control law; step 7) judging whether the following process is ended, and if not ended, re-executing step 5)-step 7). Compared with the prior art, the present invention significantly enhances the active safety capability of the vehicle following system, and has the advantages of accurate calculation and strong real-time performance.

Description

Technical Field

[0001] The present invention relates to the field of vehicle following control, and in particular to a master-slave cooperative control method, device and storage medium for dynamic following vehicle distance adjustment. Background Art

[0002] Dynamic following distance control needs to be achieved through the adjustment of vehicle following behavior. Most of the existing technologies focus on the adjustment of the following behavior of the rear vehicle under the condition of changes in the behavior of the front vehicle and its impact on the dynamic following distance. The rear vehicle has the active safety capability to avoid "rear-end collision" with the front vehicle by adjusting its own behavior. In this case, although the behavior of the rear vehicle is constrained by the position, operating status and control strategy of the front vehicle, it is not directly controlled by the front vehicle. Moreover, the front vehicle obtains the road condition information ahead of the rear vehicle, but fails to play its positive role, so it does not have the ability to actively prevent the rear vehicle from rear-ending. As a result, the entire following system has a low level of technical autonomy and intelligence, and a weak active safety capability, so that the control level and quality of the vehicle following behavior and dynamic following distance are greatly limited. Summary of the invention

[0003] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a master-slave collaborative control method, device and storage medium for dynamic following vehicle distance adjustment, so as to realize real-time and accurate vehicle following behavior control and dynamic following vehicle distance adjustment.

[0004] The purpose of the present invention can be achieved by the following technical solutions:

[0005] A master-slave cooperative control method for dynamic following vehicle distance adjustment comprises the following steps:

[0006] Step 1) establishing a synchronous control model of a vehicle following system based on model-predicted vehicle following behavior and real-time continuous adjustment of dynamic following vehicle distance;

[0007] Step 2) establishing an optimal state estimation model based on Kalman filtering to optimize the synchronous control model, and obtaining an optimal estimate of the state variables based on the Kalman filtering algorithm;

[0008] Step 3) Determine the Pareto optimization target and solve the optimal control law of vehicle following behavior and dynamic following distance;

[0009] Step 4) constructing a GRU deep neural network model for vehicle following behavior control, and establishing a nonlinear mapping relationship between the control law and the vehicle following system state and the optimal behavior adjustment process through training;

[0010] Step 5) The leading vehicle and the trailing vehicle input the basic data collected or acquired by communication into their respective GRU deep neural network models, and obtain the first control law and the second control law according to the mapping relationship;

[0011] Step 6) The rear vehicle sends the second control law to the front vehicle, and the front vehicle compares and analyzes the second control law with the first control law. If the safety, efficiency, stability and energy-saving operation requirements are met, the rear vehicle is allowed to control its own behavior according to the second control law. Otherwise, the front vehicle sends the first control law to the rear vehicle, and the rear vehicle controls its own behavior according to the first control law.

[0012] Step 7) determines whether the following process has ended. If the following process has not ended, re-execute steps 5) to 7).

[0013] The step 1) synchronous control model establishment process includes:

[0014] Step 1-1) Establish a continuous-time mathematical model of the vehicle following system:

[0015] Assume that P1(t), v1(t) and a1(t) represent the position, velocity and acceleration of the leading vehicle at time t, respectively, and P2(t), v2(t) and a2(t) represent the position, velocity and acceleration of the trailing vehicle at time t, respectively. u(t)=a1(t)-a2(t), y(t)=d Actual (t) = P1(t) - P2(t), the continuous time mathematical model of the vehicle following system can be established:

[0016]

[0017] in, C =

[01] ;

[0018] Step 1-2) Discretize the continuous-time mathematical model to establish a discrete model of the vehicle following system:

[0019] Let the state vector in the kth sampling period be Control vector u(k) = a1(k) - a2(k) and output y(k) = d Actual (k), the discrete state space model of the vehicle following system is:

[0020]

[0021] Among them, the state vector x(k) reflects the real-time behavior and actual following distance of the front and rear vehicles, v1(k) and v2(k) are the speeds of the front and rear vehicles in the kth sampling period, P1(k) and P2(k) are the positions of the front and rear vehicles in the kth sampling period, a1(k) and a2(k) are the accelerations of the front and rear vehicles in the kth sampling period, d Actual (k) is the actual distance between the front and rear vehicles in the kth sampling period, and the state transfer matrix Control Matrix Output matrix C = [0 1], T is the sampling period;

[0022] Step 1-3) Let r(k) = d Safe (k), e(k) = r(k) - y(k), and construct the system state prediction observer:

[0023]

[0024] in, is the system state estimation, d Safe (k) is the dynamic safety distance that the vehicle following system should maintain in the kth sampling period, K is the error compensation matrix of state estimation, and the control vector u(k) is Among them, L is the forecast observation state feedback matrix, and M is the error compensation matrix of the control vector.

[0025] The step 2) establishing an optimal state estimation model based on Kalman filtering includes:

[0026] Establish a vehicle following system state prediction observer:

[0027]

[0028] in, is the system state estimation, u(k) is the control vector, e(k) = r(k) - y(k), r(k) = d Safe (k), d Safe (k) is the dynamic safety distance that the vehicle following system should maintain in the kth sampling period, y(k) is the output of the kth sampling period, w(k) is the noise of the kth sampling period, which conforms to the Gaussian distribution, and the state transfer matrix Control Matrix K is the error compensation matrix of state estimation;

[0029] Establish an estimation model for the optimal state of the vehicle following system:

[0030]

[0031] in, is the state estimate of the k+1th sampling period, P - (k+1) is the estimated value of the covariance matrix in the k+1th sampling period, and P - (k+1) and their respective true values ​​with noise w(k) have uncertainty errors, P(k) represents the covariance matrix of the state estimation in the kth sampling period, and Q(k) is the state transition covariance matrix in the kth sampling period;

[0032] Establishing a correction model for optimal state estimation;

[0033] Let z(k+1)=Fx(k)+v(k), where z(k+1) is the observation value matrix of the system state vector obtained by the measuring instrument or sensor in the k+1th sampling period, F=I (I is the unit matrix), v(k) is the observation noise caused by the external environment in the kth sampling period, and the correction model of the optimal state estimation is:

[0034]

[0035] in, is the Kalman coefficient matrix of the k+1th sampling period, R(k) is the covariance matrix of the observation noise v(k) of the kth sampling period, is the system state estimate after correction in the k+1th sampling period, and P(k+1) is the updated state estimate covariance matrix in the k+1th sampling period.

[0036] The Kalman filter algorithm of step 2) comprises the following steps:

[0037] Step 2-1) Initialize parameters, the parameters include: k = 1, G, H, u(k), K, R(k), F;

[0038] Step 2-2) Determine the covariance matrix P(k) and the state transition covariance matrix Q(k);

[0039] Step 2-3) According to the formula calculate and P - (k+1);

[0040] Step 2-4) According to the formula calculate and P(k+1);

[0041] Step 2-5) determines whether the following process is completed. If the following process is completed, stop the filtering calculation; if the following process is not completed, set k=k+1, update u(k) and R(k), and re-execute steps 2-3) to 2-5).

[0042] The Pareto optimization objective of step 3) is:

[0043]

[0044] Where E(k) is the energy consumption of the following system in the kth sampling period, S is the safe and efficient following steady state, S0 is the initial state of the following system, σ is the control sequence of the following vehicle behavior, Time(S0[σ>S) is the time solution function for the vehicle following system to enter the safe and efficient following steady state S from the initial state S0 through the following vehicle control sequence σ, the optimization target J and the control sequence σ both use the control laws K, L, and M as decision variables, ξ is a small positive number used to limit the adjustment range of the dynamic following distance, and F 1_traction (k), f 1_braking (k), v1(k) and a1(k) represent the traction, braking force, speed and acceleration of the leading vehicle in the kth sampling period, respectively. 2_traction (k), f 2_braking (k), v2(k) and a2(k) represent the traction, braking force, speed and acceleration of the following vehicle in the kth sampling period, respectively. 1_traction_max 、f 1_braking_max 、v 1_max 、a 1_max and a 1_min They represent the maximum traction, maximum braking force, maximum permissible speed, maximum acceleration and minimum acceleration of the vehicle ahead. 2_traction_max 、f 2_braking_max and v 2_max They represent the maximum traction, maximum braking force and maximum running speed of the rear vehicle, respectively. Line_max represents the maximum vehicle speed allowed by the line, J1(k) = |(a1(k+1)-a1(k)) / T| and J2(k) = |(a2(k+1)-a2(k)) / T are the absolute values ​​of the time rate of change of vehicle acceleration before and after the kth sampling period, respectively. 1_min and J 1_max are the minimum and maximum time rates of change of the acceleration of the preceding vehicle, respectively. Comfort and J Comfort are two upper boundaries of vehicle behavior adjustment to satisfy stability; the vehicle following system satisfies the following conditions when it is in a safe and efficient following steady state S:

[0045]

[0046] in, d Actual (k) is the actual distance between the front and rear vehicles in the kth sampling period, d Safe (k) is the dynamic safety distance that the vehicle following system should maintain in the kth sampling period, v1(k) and v2(k) are the speeds of the front and rear vehicles in the kth sampling period, a1(k) and a2(k) are the accelerations of the front and rear vehicles in the kth sampling period, respectively.

[0047] The process of solving the optimal control law in step 3) includes:

[0048] The Pareto optimal solution set of the control law (K, L, M) is expressed as:

[0049]

[0050] Among them, Ω is the solution space of the control law (K, L, M), J0 is the initial value of the optimization target of the vehicle following system, It means that J is better than J0, that is, J is not greater than J0 in every indicator, and is less than J0 in at least one indicator;

[0051] if Then the control law (K, L, M) obtained by Pareto optimization is acceptable, and J0=J; otherwise, keep J0 unchanged; continue Pareto optimization until the optimization result is satisfactory, and the final control law (K, L, M) is the optimal control law.

[0052] The step 4) comprises the following steps:

[0053] The GRU deep neural network model is trained using the actually measured time series data and the time series data corrected by the Kalman filter as sample data, wherein the time series data includes the speed of the leading vehicle, the speed of the following vehicle, the actual following vehicle distance and the dynamic safety vehicle distance;

[0054] If the optimization goal of the GRU deep neural network is can all achieve Pareto optimality, that is It meets the consistency requirements with minJ, and the control law obtained by the GRU deep neural network If the consistency requirement is met when compared with the control law (K, L, M) that satisfies Pareto optimality, the GRU deep neural network training is successful, and a nonlinear mapping relationship between the control law (K, L, M) and the vehicle following system state and the optimal behavior adjustment process is established; otherwise, the GRU deep neural network training continues.

[0055] In step 6), the rear vehicle controls its own behavior according to the control law, including the following steps:

[0056] Solve the control strategy for vehicle following behavior and dynamic train spacing:

[0057] a2(k)=a1(k)-u(k)

[0058] Among them, u(k) is the control law;

[0059] According to the vehicle mass and force conditions, a2(k) is used to calculate the traction or braking force required for the synchronous control of the vehicle following behavior and the dynamic following distance, so as to realize the real-time synchronous control of the vehicle following behavior and the dynamic following distance.

[0060] A master-slave cooperative control device for dynamic following vehicle distance adjustment includes a memory, a processor, and a program stored in the memory. When the processor executes the program, the method described above is implemented.

[0061] A storage medium stores a vehicle following behavior and a dynamic following distance control program, and the method described above is implemented when the program is executed.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] (1) The present invention combines model prediction with Pareto optimization to achieve synchronous control optimization of following behavior adjustment and dynamic following distance.

[0064] (2) The present invention adopts Kalman filtering to reduce the negative impact of noise and improve the accuracy of vehicle following system state estimation.

[0065] (3) The present invention adopts the GRU deep neural network, which improves the real-time performance of the control law calculation after training, and can achieve the purpose of real-time control of dynamic following vehicle distance.

[0066] (4) In the existing vehicle following control technology, the rear vehicle has the ability to prevent rear-end collision with the front vehicle, but the front vehicle lacks control means over the rear vehicle following behavior and does not have the ability to actively prevent rear-end collision with the rear vehicle. In view of this defect of the existing vehicle following control technology, the present invention adopts a master-slave collaborative control strategy to overcome it, effectively enhancing the active safety capability of the entire following system during the following operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a flow chart of the present invention;

[0068] Figure 2 A discrete model for vehicle following behavior and dynamic following distance synchronization control based on model prediction;

[0069] Figure 3 GRU deep neural network model for vehicle following behavior control;

[0070] Figure 4 A schematic diagram showing the compatibility of the present invention with the existing train control technology. DETAILED DESCRIPTION

[0071] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process of the technical solution of the present invention for dynamic train interval (i.e., dynamic following vehicle distance) adjustment of master-slave collaborative control in the field of rail transit, but the protection scope of the present invention is not limited to the following embodiments.

[0072] The present invention relates to a master-slave cooperative control method for dynamic train interval adjustment, which is technically compatible with the existing train control technology. Figure 4 As shown. Safe (t), d Actual (t) are the dynamic safety distance and actual distance of the leading and trailing vehicles at time t, P1(t) and P2(t) are the positions of the leading and trailing vehicles at time t, v1(t) and v2(t) are the velocities of the leading and trailing vehicles at time t, v e1 (P1(t)) and v e2 (P2(t)) are the expected speeds of the leading and trailing cars at time t, automatically generated according to the expected speed curve; a1(t) and a2(t) are the accelerations of the leading and trailing cars at time t; P1(t0) and P2(t0) are the line positions of the leading and trailing cars at the initial time t0; v1(t0) and v2(t0) are the speeds of the leading and trailing cars at the initial time t0; m1(t) and m2(t) are the masses of the leading and trailing cars at time t. Train dismantling, marshaling operations or passenger boarding and alighting operations may cause train mass F1(t) and F2(t) are the traction forces (F1(t)>0, F2(t)>0) or braking forces (F1(t)<0, F2(t)<0) of the leading and trailing vehicles at time t, respectively; f1(t) and f2(t) are the running resistances (>0, which refer to the forces that are opposite to the running direction of the train and are not actively applied by the train for deceleration, caused by the influence of external conditions such as friction, impact, vibration, and the plane and section of the line during the running of the train, and ∫ represents the time integral.

[0073] When the train's autonomous intelligence and active safety capabilities are low, a combination of "train station-to-station behavior control based on expected speed curve tracking" and "train timetable" can be adopted to achieve train following operation on section lines. However, the adjustment accuracy and real-time performance of the dynamic train interval are greatly limited, and it is often only possible to complete "extensive" active control of the following distance by adjusting the train interval time at the stop station; "target distance speed curve" is mainly used for safety protection during the train following operation. Once the actual train interval is equal to or less than the safe distance specified by the target distance speed curve, the following car must use normal braking or emergency braking to slow down and increase the dynamic train interval to improve the safety of following operation; "mobile authorization" gives constraints on the real-time and continuous adjustment of the train following behavior. The following car must adjust its own behavior within the range allowed by the mobile authorization to ensure its own driving safety during the following operation.

[0074] A master-slave coordinated control method for dynamic train interval adjustment comprises the following steps:

[0075] Step 1) Establish a synchronous control model of a train following system based on model-predicted train following behavior and real-time continuous adjustment of dynamic train interval.

[0076] Step 1-1) Establish a continuous-time mathematical model of the train following system:

[0077] Assume that P1(t), v1(t) and a1(t) represent the position, velocity and acceleration of the leading vehicle at time t, respectively, and P2(t), v2(t) and a2(t) represent the position, velocity and acceleration of the trailing vehicle at time t, respectively. u(t)=a1(t)-a2(t), y(t)=d Actual (t) = P1(t) - P2(t), a continuous-time mathematical model of the train following system can be established:

[0078]

[0079] in, C =

[01] ;

[0080] Step 1-2) Discretize the continuous-time mathematical model and establish a discrete model of the train following system:

[0081] Let the state vector in the kth sampling period be Control vector u(k) = a1(k) - a2(k) and output y(k) = d Actual (k), the discrete state space model of train following operation is:

[0082]

[0083] Among them, the state vector x(k) reflects the real-time behavior and actual following distance of the front and rear trains, v1(k) and v2(k) are the speeds of the front and rear trains in the kth sampling period, P1(k) and P2(k) are the positions of the front and rear trains in the kth sampling period, a1(k) and a2(k) are the accelerations of the front and rear trains in the kth sampling period, and d Actual (k) is the actual distance between the front and rear trains in the kth sampling period, and the state transfer matrix Control Matrix Output matrix C = [0 1], T is the sampling period;

[0084] Step 1-3) Let r(k) = d Safe (k), e(k) = r(k) - y(k), and construct the system state prediction observer:

[0085]

[0086] in, is the system state estimation, d Safe (k) is the dynamic safe vehicle distance that the train following system should maintain in the kth sampling period, K is the error compensation matrix of state estimation, and the control vector u(k) is Among them, L is the forecast observation state feedback matrix, and M is the error compensation matrix of the control vector.

[0087] Discrete model of train following behavior and dynamic train interval synchronization control based on model prediction, see Figure 2 As shown in the figure, the dynamic safe vehicle distance is taken as the tracking target, and the actual vehicle distance and speed difference are taken as the state variables of the following system, thus creating favorable conditions for the synchronous control and optimization of the train following behavior and the dynamic train interval.

[0088] Step 2) Establishing an optimal state estimation model based on Kalman filtering to optimize the synchronous control model, and obtaining an optimal estimate of the state variables based on the Kalman filtering algorithm.

[0089] First, considering the influence of noise, a train following system state prediction observer that conforms to the actual operating environment is established:

[0090]

[0091] in, is the system state estimation, u(k) is the control vector, e(k) = r(k) - y(k), r(k) = d Safe (k), d Safe(k) is the dynamic safe distance that the train following system should maintain in the kth sampling period, y(k) is the output of the kth sampling period, w(k) is the noise of the kth sampling period, which conforms to the Gaussian distribution, and the state transfer matrix Control Matrix K is the error compensation matrix of state estimation;

[0092] Then, the optimal state estimation model of the train following system is established:

[0093]

[0094] in, is the state estimate of the k+1th sampling period, P - (k+1) is the estimated value of the covariance matrix in the k+1th sampling period, and P - (k+1) and their respective true values ​​with noise w(k) have uncertainty errors, P(k) represents the covariance matrix of the state estimation in the kth sampling period, and Q(k) is the state transition covariance matrix in the kth sampling period;

[0095] Finally, a correction model for optimal state estimation is established;

[0096] Let z(k+1)=Fx(k)+v(k), where z(k+1) is the observation value matrix of the system state vector obtained by the measuring instrument or sensor in the k+1th sampling period, F=I (I is the unit matrix), v(k) is the observation noise caused by the external environment in the kth sampling period, and the correction model of the optimal state estimation is:

[0097]

[0098] in, is the Kalman coefficient matrix of the k+1th sampling period, R(k) is the covariance matrix of the observation noise v(k) of the kth sampling period, is the system state estimate after correction in the k+1th sampling period, and P(k+1) is the updated state estimate covariance matrix in the k+1th sampling period.

[0099] According to the above-mentioned Kalman filter model for optimal state estimation of the train following system, a Kalman filter algorithm for achieving optimal estimation of state variables is established, which specifically includes the following steps:

[0100] Step 2-1) Initialize parameters, the parameters include: k = 1, G, H, u(k), K, R(k), F;

[0101] Step 2-2) Determine the covariance matrix P(k) and the state transition covariance matrix Q(k);

[0102] Step 2-3) According to the formula calculate and P - (k+1);

[0103] Step 2-4) According to the formula calculate and P(k+1);

[0104] Step 2-5) determines whether the following process is completed. If the following process is completed, stop the filtering calculation; if the following process is not completed, set k=k+1, update u(k) and R(k), and re-execute steps 2-3) to 2-5).

[0105] Step 3) Determine the Pareto optimization objective and solve the optimal control law for train following behavior and dynamic train interval.

[0106] The Pareto optimization objective is:

[0107]

[0108] Where E(k) is the energy consumption of the following system in the kth sampling period, S is the safe and efficient following steady state, S0 is the initial state of the following system, σ is the control sequence of the following behavior of the following vehicle, Time(S0[σ>S) is the time solution function for the train following system to enter the safe and efficient following steady state S from the initial state S0 through the following vehicle control sequence σ, the optimization target J and the control sequence σ are both based on the control laws K, L, and M as decision variables, ξ is a small positive number used to reasonably limit the adjustment range of the dynamic train interval, and F 1_traction (k), f 1_braking (k), v1(k) and a1(k) represent the traction, braking force, speed and acceleration of the leading vehicle in the kth sampling period, respectively. 2_traction (k), f 2_braking (k), v2(k) and a2(k) represent the traction, braking force, speed and acceleration of the following vehicle in the kth sampling period respectively (the traction and braking force of the train are not presented at the same time), F 1_traction_max 、f 1_braking_max 、v 1_max 、a 1_max and a 1_min They represent the maximum traction, maximum braking force, maximum permissible speed, maximum acceleration and minimum acceleration of the vehicle ahead. 2_traction_max 、f 2_braking_max and v 2_max They represent the maximum traction, maximum braking force and maximum running speed of the rear vehicle, respectively. Line_maxrepresents the maximum train running speed allowed by the line, J1(k) = |(a1(k+1)-a1(k)) / T| and J2(k) = |(a2(k+1)-a2(k)) / T| are the absolute values ​​of the time rate of change of train acceleration before and after the kth sampling period, respectively. 1_min and J 1_max are the minimum and maximum time rates of change of the acceleration of the preceding vehicle, respectively. Comfort and J Comfort The two upper bounds for the train behavior to satisfy stationarity are adjusted respectively.

[0109] The purpose of train following behavior adjustment is to establish a safe and efficient train following steady state, rather than being in an endless behavior adjustment process. When the train following system is in a safe and efficient train following steady state, the following conditions are met:

[0110]

[0111] in, d Actual (k) is the actual distance between the front and rear trains in the kth sampling period, d Safe (k) is the dynamic safe vehicle distance that the train following system should maintain in the kth sampling period, v1(k) and v2(k) are the speeds of the leading and trailing trains in the kth sampling period, a1(k) and a2(k) are the accelerations of the leading and trailing trains in the kth sampling period, respectively.

[0112] During the dynamic train interval adjustment process, Actual (k)-d Safe (k) The adjustment range is relatively large. Once a safe and efficient car-following steady state is established, d Actual (k)-d Safe The value of (k) is small, slightly greater than 0, to meet the safety and efficiency of train following operation. When v1(k) = v2(k) = 0, the leading and trailing trains will be in a stopped state, and the state described at this time can be regarded as a special safe and efficient following steady state.

[0113] The Pareto optimal solution set of the control law (K, L, M) is expressed as:

[0114]

[0115] Among them, Ω is the solution space of the control law (K, L, M), J0 is the initial value of the optimization target of the train following system, It means that J is better than J0, that is, J is not greater than J0 in every indicator, and is less than J0 in at least one indicator;

[0116] The process of solving the optimal control law includes:

[0117] if Then the control law (K, L, M) obtained by Pareto optimization is acceptable, and J0=J; otherwise, keep J0 unchanged; continue Pareto optimization until the optimization result is satisfactory, and the final control law (K, L, M) is the optimal control law.

[0118] Step 4) Construct a GRU deep neural network model for vehicle following behavior control, and establish a nonlinear mapping relationship between the control law and the vehicle following system state and the optimal behavior adjustment process through training.

[0119] If the time taken to solve the control law (K, L, M) by Pareto optimization is too long, it will not be able to meet the real-time control requirements of train following behavior and dynamic train interval continuous adjustment. Deep neural networks such as GRU can be used to solve the control law (K, L, M) that satisfies Pareto optimality for train following behavior and dynamic train interval real-time continuous adjustment. The constructed GRU deep neural network model is shown in Figure 3 As shown, N is a positive integer, the leading vehicle speed, the trailing vehicle speed, the actual following distance and the dynamic safety distance are all variables in the form of time series, and the input data formed should be able to cover all states of the train following system and its transition process.

[0120] The GRU deep neural network model is trained using the actually measured time series data and the time series data corrected by the Kalman filter as sample data, wherein the time series data includes the speed of the leading vehicle, the speed of the following vehicle, the actual following vehicle distance and the dynamic safety vehicle distance;

[0121] If the optimization goal of the GRU deep neural network is can all achieve Pareto optimality, that is It meets the consistency requirements with minJ, and the control law obtained by the GRU deep neural network If the consistency requirement is met when compared with the control law (K, L, M) that satisfies Pareto optimality, the GRU deep neural network training is successful, and a nonlinear mapping relationship between the control law (K, L, M) and the various states of the train following system and the optimal behavior adjustment process is established; otherwise, the GRU deep neural network training continues.

[0122] After the model training is completed, the GRU deep neural network can be used to solve the control law (K, L, M) in real time, thereby realizing master-slave collaborative control with real-time and precise adjustment of dynamic train intervals.

[0123] Step 5) The leading vehicle and the trailing vehicle input the basic data collected or acquired by communication into their respective GRU deep neural network models, and obtain the first control law and the second control law respectively according to the mapping relationship.

[0124] Step 6) The rear vehicle sends the second control law to the front vehicle, and the front vehicle compares and analyzes the second control law with the first control law. If the safety, efficiency, stability and energy-saving operation requirements are met, the rear vehicle is allowed to control its own behavior according to the second control law. Otherwise, the front vehicle sends the first control law to the rear vehicle, and the rear vehicle controls its own behavior according to the first control law.

[0125] The optimal control law is used for real-time control of train following behavior and dynamic train interval, and includes the following steps:

[0126] First, the control strategy of train following behavior and dynamic train interval is solved, and the calculation formula is:

[0127] a2(k)=a1(k)-u(k)

[0128] Then, according to the train mass and force conditions, the above formula is used to calculate the traction or braking force required for the synchronous control of the train following behavior and the dynamic train interval, so as to realize the real-time synchronous control of the train following behavior and the dynamic train interval.

[0129] Step 7) determines whether the following process has ended. If the following process has not ended, re-execute steps 5) to 7).

[0130] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program code.

Claims

1. A master-slave cooperative control method for dynamic following vehicle distance adjustment, characterized in that: The following steps are involved: Step 1) establishing a synchronous control model of a vehicle following system based on model-predicted vehicle following behavior and real-time continuous adjustment of dynamic following vehicle distance; Step 2) establishing an optimal state estimation model based on Kalman filtering to optimize the synchronous control model, and obtaining an optimal estimate of the state variables based on the Kalman filtering algorithm; Step 3) Determine the Pareto optimization target and solve the optimal control law of vehicle following behavior and dynamic following distance; Step 4) constructing a GRU deep neural network model for vehicle following behavior control, and establishing a nonlinear mapping relationship between the control law and the vehicle following system state and the optimal behavior adjustment process through training; Step 5) The leading vehicle and the trailing vehicle input the basic data collected or acquired by communication into their respective GRU deep neural network models, and obtain the first control law and the second control law according to the mapping relationship; Step 6) The rear vehicle sends the second control law to the front vehicle, and the front vehicle compares and analyzes the second control law with the first control law. If the safety, efficiency, stability and energy-saving operation requirements are met, the rear vehicle is allowed to control its own behavior according to the second control law. Otherwise, the front vehicle sends the first control law to the rear vehicle, and the rear vehicle controls its own behavior according to the first control law. Step 7) determining whether the car-following process has ended, if the car-following process has not ended, re-execute steps 5) to 7); The step 1) synchronous control model establishment process includes: Step 1-1) Establish a continuous-time mathematical model of the vehicle following system: Assume that P1(t), v1(t) and a1(t) represent the position, velocity and acceleration of the leading vehicle at time t, respectively; P2(t), v2(t) and a2(t) represent the position, velocity and acceleration of the trailing vehicle at time t, respectively. u(t)=a1(t)-a2(t), y(t)=d Actual (t) = P1(t) - P2(t), the continuous time mathematical model of the vehicle following system can be established: in, C = [01]; Step 1-2) Discretize the continuous-time mathematical model to establish a discrete model of the vehicle following system: Let the state vector in the kth sampling period be Control vector u(k) = a1(k) - a2(k) and output y(k) = d Actual (k), the discrete state space model of the vehicle following system is: Among them, the state vector x(k) reflects the real-time behavior and actual following distance of the front and rear vehicles, v1(k) and v2(k) are the speeds of the front and rear vehicles in the kth sampling period, P1(k) and P2(k) are the positions of the front and rear vehicles in the kth sampling period, a1(k) and a2(k) are the accelerations of the front and rear vehicles in the kth sampling period, d Actual (k) is the actual distance between the front and rear vehicles in the kth sampling period, and the state transfer matrix Control Matrix Output matrix C = [01], T is the sampling period; Step 1-3) Let r(k) = d Safe (k), e(k) = r(k) - y(k), and construct the system state prediction observer: in, is the system state estimation, d Safe (k) is the dynamic safety distance that the vehicle following system should maintain in the kth sampling period, K is the error compensation matrix of state estimation, and the control vector u(k) is Among them, L is the forecast observation state feedback matrix, and M is the error compensation matrix of the control vector; The step 2) establishing an optimal state estimation model based on Kalman filtering includes: Establish a vehicle following system state prediction observer: in, is the system state estimation, u(k) is the control vector, e(k) = r(k) - y(k), r(k) = d Safe (k), d Safe (k) is the dynamic safety distance that the vehicle following system should maintain in the kth sampling period, y(k) is the output of the kth sampling period, w(k) is the noise of the kth sampling period, which conforms to the Gaussian distribution, and the state transfer matrix Control Matrix K is the error compensation matrix of state estimation; Establish an estimation model for the optimal state of the vehicle following system: in, is the state estimate of the k+1th sampling period, P - (k+1) is the estimated value of the covariance matrix in the k+1th sampling period, and P - (k+1) and their respective true values ​​with noise w(k) have uncertainty errors, P(k) represents the covariance matrix of the state estimation in the kth sampling period, and Q(k) is the state transition covariance matrix in the kth sampling period; Establishing a correction model for optimal state estimation; Let z(k+1)=Fx(k)+v(k), where z(k+1) is the observation value matrix of the system state vector obtained by the measuring instrument or sensor in the k+1th sampling period, F=I, I is the unit matrix, and v(k) is the observation noise caused by the external environment in the kth sampling period. The correction model of the optimal state estimation is: in, is the Kalman coefficient matrix of the k+1th sampling period, R(k) is the covariance matrix of the observation noise v(k) of the kth sampling period, is the system state estimate after correction in the k+1th sampling period, and P(k+1) is the updated state estimate covariance matrix in the k+1th sampling period; The Kalman filter algorithm of step 2) comprises the following steps: Step 2-1) Initialize parameters, the parameters include: k = 1, G, H, u(k), K, R(k), F; Step 2-2) Determine the covariance matrix P(k) and the state transition covariance matrix Q(k); Step 2-3) According to the formula calculate and P - (k+1); Step 2-4) According to the formula calculate and P(k+1); Step 2-5) determines whether the following process is completed. If the following process is completed, stop the filtering calculation; if the following process is not completed, set k = k + 1, update u(k), R(k), and re-execute steps 2-3) to 2-5); The Pareto optimization objective of step 3) is: Where E(k) is the energy consumption of the following system in the kth sampling period, S is the safe and efficient following steady state, S0 is the initial state of the following system, σ is the control sequence of the following vehicle behavior, Time(S0[σ>S) is the time solution function for the vehicle following system to enter the safe and efficient following steady state S from the initial state S0 through the following vehicle control sequence σ, the optimization target J and the control sequence σ both use the control laws K, L, and M as decision variables, ξ is a small positive number used to limit the adjustment range of the dynamic following distance, and F 1_traction (k), f 1_braking (k), v1(k) and a1(k) represent the traction, braking force, speed and acceleration of the leading vehicle in the kth sampling period, respectively. 2_traction (k), f 2_braking (k), v2(k) and a2(k) represent the traction, braking force, speed and acceleration of the following vehicle in the kth sampling period, respectively. 1_traction_max 、f 1_braking_max 、v 1_max 、a 1_max and a 1_min They represent the maximum traction, maximum braking force, maximum permissible speed, maximum acceleration and minimum acceleration of the vehicle ahead. 2_traction_max 、f 2_braking_max and v 2_max They represent the maximum traction, maximum braking force and maximum running speed of the rear vehicle, respectively. Line_max represents the maximum vehicle speed allowed by the line, J1(k) = |(a1(k+1)-a1(k)) / T| and J2(k) = |(a2(k+1)-a2(k)) / T| are the absolute values ​​of the time rate of change of vehicle acceleration before and after the kth sampling period, respectively. 1_min and J 1_max are the minimum and maximum time rates of change of the acceleration of the preceding vehicle, respectively. Comfort and J Comfort are two upper boundaries of vehicle behavior adjustment to satisfy stability; the vehicle following system satisfies the following conditions when it is in a safe and efficient following steady state S: in, d Actual (k) is the actual distance between the front and rear vehicles in the kth sampling period, d Safe (k) is the dynamic safety distance that the vehicle following system should maintain in the kth sampling period, v1(k) and v2(k) are the speeds of the leading and trailing vehicles in the kth sampling period, a1(k) and a2(k) are the accelerations of the leading and trailing vehicles in the kth sampling period, respectively; The process of solving the optimal control law in step 3) includes: The Pareto optimal solution set of the control law (K, L, M) is expressed as: Φ={(K,L,M)∈Ω|J<J0} Where Ω is the solution space of the control law (K, L, M), J0 is the initial value of the optimization target of the vehicle following system, and J < J0 means that J is better than J0, that is, J is not greater than J0 in every indicator and is less than J0 in at least one indicator; If J < J0, the control law (K, L, M) obtained by Pareto optimization is acceptable, and J0 = J; otherwise, keep J0 unchanged; continue Pareto optimization until the optimization result is satisfactory, and the final control law (K, L, M) is the optimal control law; The step 4) comprises the following steps: The GRU deep neural network model is trained using the actually measured time series data and the time series data corrected by the Kalman filter as sample data, wherein the time series data includes the speed of the leading vehicle, the speed of the following vehicle, the actual following vehicle distance and the dynamic safety vehicle distance; If the optimization goal of the GRU deep neural network is can all achieve Pareto optimality, that is It meets the consistency requirements with minJ, and the control law obtained by the GRU deep neural network If the consistency requirement is met when compared with the control law (K, L, M) that satisfies Pareto optimality, the GRU deep neural network training is successful, and a nonlinear mapping relationship between the control law (K, L, M) and the vehicle following system state and the optimal behavior adjustment process is established; otherwise, the GRU deep neural network training continues.

2. A master-slave cooperative control method for dynamic following vehicle distance adjustment according to claim 1, characterized in that: In step 6), the rear vehicle controls its own behavior according to the control law, including the following steps: Solve the control strategy for vehicle following behavior and dynamic train spacing: a2(k)=a1(k)-u(k) Among them, u(k) is the control law; According to the vehicle mass and force conditions, a2(k) is used to calculate the traction or braking force required for the synchronous control of the vehicle following behavior and the dynamic following distance, so as to realize the real-time synchronous control of the vehicle following behavior and the dynamic following distance.

3. A master-slave cooperative control device for dynamic following vehicle distance adjustment, comprising a memory, a processor, and a program stored in the memory, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 2 is implemented.

4. A storage medium storing a vehicle following behavior and a dynamic following distance control program, characterized in that: When the program is executed, the method according to any one of claims 1 to 2 is implemented.

Images