1. An automobile prediction cruise parameter self-tuning control system fusing driving mode information,

s1, information acquisition;

s2, clustering and identifying the driving modes of the front vehicle;

the method is characterized in that:

s3, designing a prediction cruise controller:

3.1) establishing a vehicle dynamic model facing cruise control, wherein the vehicle dynamic model of the vehicle and the front vehicle:

d(k+1)＝d(k)+(v_l(k)-v_e(k))·T_s

v_e(k+1)＝v_e(k)+u(k)·T_s (1)

T_srepresenting the sampling time, k the time, d (k +1) the longitudinal distance between the host vehicle and the preceding vehicle at the time k +1, d (k) the longitudinal distance between the host vehicle and the preceding vehicle at the time k, v_l(k) Longitudinal speed of the preceding vehicle, v, at time k_e(k) The longitudinal speed v of the vehicle at time k_e(k +1) represents the vehicle longitudinal speed at the time k +1, and u (k) represents the vehicle longitudinal acceleration at the time k;

3.2) establishing an energy consumption model of the vehicle, and establishing an engine oil consumption approximate model by adopting a polynomial fitting method:

m_findicating the instantaneous specific fuel consumption, T, of the engine_eRepresenting engine torque, ω_eIndicating engine speed, p_1～9Representing the fitting parameters; furthermore, T_eAnd ω_eThe following can be calculated by the acceleration and the vehicle speed:

m represents the vehicle mass, r represents the vehicle tire radius, η represents the transmission efficiency, I_gRepresenting transmission ratio of the gearbox, I_finalRepresenting the final gear ratio;

3.3) establishing multi-objective optimization problem description, wherein at the moment k, the multi-objective optimization problem description comprises the following steps:

p denotes the prediction time domain, λ₁、λ₂And λ₃Representing weight parameters in a multi-objective optimization function, where₁The corresponding optimization objective represents security and traceability, d_refIndicating a reference vehicle distance, related to the front vehicle speed; lambda [ alpha ]₂The corresponding optimization target represents the fuel consumption of the vehicle, a complete fuel consumption model is not added into an objective function in order to reduce the calculation burden, and only one part of the fuel consumption model is taken as the optimization target; lambda [ alpha ]₃The corresponding optimization target represents driving comfort, and Deltau represents vehicle accelerationA speed increment; u. of_minAnd u_maxShowing the minimum and maximum values of the acceleration of the vehicle, Deltau_minAnd Δ u_maxMinimum and maximum values representing the acceleration increment of the vehicle, d_minAnd d_maxRepresenting the minimum and maximum longitudinal distance of the two vehicles, v_minAnd v_maxA minimum value and a maximum value representing the vehicle speed of the host vehicle;

s4 parameter setting design based on Bayesian optimization method

4.1) selecting a section of running track data of the preceding vehicle in a certain driving mode from the historical running data processed in the steps S1 and S2 as simulation test data, and specifically explaining the subsequent steps by selecting the running track data of the preceding vehicle in an acceleration running mode;

4.2) randomly selecting a set of weight parameters λ ═ λ₁ λ₂ λ₃]The predictive cruise controller in the step S3 is adopted to execute control simulation to obtain a following vehicle running strategy of the vehicle under the section of the running track of the front vehicle according to V_eThe vehicle speed change sequence of the vehicle under the following vehicle driving strategy is shown, and U represents the acceleration change sequence of the vehicle under the following vehicle driving strategy; the total oil consumption M of the vehicle in the distance can be calculated by the formulas (2) and (3)_fuel

M_fuel＝∑m_f(U,V_e)·T_s (5)

Using negative oil consumption Y to represent M_fuelThe currently selected weight parameter lambda and the negative oil consumption Y form an input/output pair sample { lambda, Y };

4.3) repeating the step 4.2 to obtain a plurality of groups of sample pointsUsing the obtained data as a training sample set, and establishing lambda through a Gaussian process methodⁱAnd YⁱNon-parametric model F, i.e. Y, betweenⁱ＝F(λⁱ) Then the problem of setting the optimal weight parameter

Wherein λ_minDenotes the minimum value of the weight, λ_maxRepresents the maximum value of the weight; the Gaussian process model assumes that there are already sample pointsObeying a multidimensional Gaussian distribution, wherein the correlation degree between any two Y's and Y' can be calculated by the covariance function of the corresponding lambda's and lambda's; covariance functions, also known as kernel functions, have a number of choices, where a Gaussian kernel function is chosen

||·||²Representing Euclidean distance, σ_lRepresenting the characteristic length size as a hyper-parameter; determining sigma for making maximum value of maximum likelihood function of current sample set by conjugate gradient method_lThen, the current sample set can be establishedThe non-parametric model F above; the output of the model F, built by the Gaussian process method, is a Gaussian distribution, denoted by YⁱMeans, σ (λ), representing the Gaussian distribution of the outputⁱ) A variance representing a gaussian distribution of the output; when given a test input λ^testWhen, F (λ)^test) Mean value of (Y)^testAnd variance σ (λ)^test)

WhereinRepresenting Gaussian noise, I_nDenotes an identity matrix of size n, Y_train＝[Y¹ Y² … Yⁿ]^TT denotes transpose, K and K are represented by the following formula:

4.4) selecting the next set of λ using the desired gain acquisition function^newSo that Y is^new＝F(λ^new) Obtaining a new maximum value; the acquisition function generally assumes that when the acquisition function takes a maximum value, the target model can also take the maximum value; desired gain acquisition function

F(λ⁺) Represents the known maximum taken from the current model F, phi (-) represents the cumulative distribution function of the standard gaussian distribution, phi (-) represents the probability density function of the standard gaussian distribution; by finding out EI (lambda)^new) Lambda to obtain a maximum value^newThe next set of lambda is obtained^newSo that Y is^new＝F(λ^new) Obtaining a new maximum value;

4.5) obtaining a new set of weights λⁿ⁺¹Then, the total fuel consumption of the vehicle corresponding to the new weight is obtained in the step 4.2, and the weight lambda is usedⁿ⁺¹And negative oil consumption Yⁿ+¹Adding the new sample point into the existing sample point, and using the new sample data setUpdating the non-parametric model F;

4.6) repeating the step 4.4 and the step 4.5, and obtaining a group of optimization weight parameters aiming at the multi-objective optimization function when the front vehicle is in the acceleration driving mode after the maximum iteration step number is reached;

4.7) selecting other preceding vehicle running track data segments of the acceleration running mode in the step 4.1, and repeating the steps 4.2 to 4.6 to obtain other optimization weight parameters of multiple groups of multi-target optimization functions;

4.8) averaging all the obtained multiple groups of weight parameters, and taking the average as the optimal weight parameter of the multi-objective optimization function in the PCC system of the host vehicle under the accelerating driving mode of the host vehicle;

4.9) based on the steps 4.1-4.8, the optimal weight parameters of the multi-objective optimization function in the PCC system of the vehicle can be obtained similarly when the front vehicle is in the constant-speed driving mode and the deceleration driving mode; when the vehicle runs in subsequent following cruising, the driving mode of the front vehicle is identified and classified on line at each sampling time based on the speed information of the front vehicle, and the weight parameters of the multi-objective optimization function in the PCC system are adjusted to be the optimal weight parameters in the corresponding driving mode by fusing the identification result of the current driving mode.

Background

The cruise control system can replace a driver to complete longitudinal following running, and is widely applied to an automobile auxiliary driving system. At the present stage, with the development of technologies such as mobile internet, big data and the like, the intelligent internet automobile can acquire surrounding traffic information in real time, and by utilizing the information, the cruising and car-following strategies are optimized under the energization of an advanced control method, so that the fuel economy, comfort and safety of the automobile can be further improved.

Patent CN105857309A discloses a vehicle Adaptive Cruise Control (ACC) method considering multiple targets. The method adopts a layered control strategy, obtains the expected longitudinal acceleration through upper-layer control, and then realizes the tracking of the expected longitudinal acceleration through lower-layer control so as to meet the driving targets of safety, following performance, comfort, fuel economy and the like. However, the invention does not consider the real-time online adjustment of the weight of multiple optimization targets, and each optimization target still has the possibility of further improvement.

Patent CN107117170A discloses a real-time Predictive Cruise Control (PCC) system based on economical driving. The control system obtains optimal control input by solving a constrained multi-objective optimization problem based on the running state information and traffic speed limit information of the front vehicle and the self vehicle, and achieves the aims of reducing fuel consumption, improving comfort and the like. However, the method does not relate to setting of multiple optimization target weights, and the fuel economy still has a space for further excavation.

Patent CN406154836A discloses an online dynamic particle swarm PID parameter self-tuning optimization method. The self-tuning method has a strong self-adaptive adjusting function and strong stability, and can effectively overcome the defects of the existing PID tuning technology. However, the method does not relate to vehicle intelligent control and is not applied to a cruise control system of the vehicle.

Patent CN111505936A discloses an automatic safety tuning method based on gaussian process PID control parameters. The method can manually set the objective function, so that the finally calibrated PID parameter can meet the initially set ideal requirement. However, the method does not relate to the self-tuning problem of the weight parameters in the multi-objective optimization, and the method is not applied to the cruise control system of the vehicle.

In a cruising and following driving scene, a PCC system is generally expected to take the fuel economy of the vehicle as a main optimization target on the premise of ensuring the safety, and secondary indexes such as driving comfort, tracking capability and the like are added on the basis to establish a multi-target optimization function. These control objectives are usually balanced by a large number of trial and error human choices of a fixed set of weight parameters in order to hopefully maximize the energy saving space of the own vehicle, but currently some challenges remain in PCC systems: firstly, a single set of fixed weight parameters is likely to generate a negative control strategy when a front vehicle is in some specific driving modes, so that the fuel economy of the vehicle is reduced, therefore, the weight parameters in a multi-objective optimization function according to different driving modes of the front vehicle are generally required to be adjusted in time to achieve an optimal control strategy and an energy-saving effect, and how to identify the driving mode of the front vehicle on line based on rich multi-dimensional internet information in the following driving process is a problem to be solved urgently at present; secondly, a large number of trial and error are usually needed for manually adjusting the weight parameters in the multi-objective function according to the identification result of the driving mode of the front vehicle, and the finally selected weighted values are often not optimal values, so how to quickly and accurately set the optimal weight parameters aiming at different driving modes of the front vehicle, the energy-saving effect of the self vehicle can be further improved by a system control strategy, and the method is another problem to be solved urgently.

Disclosure of Invention

The invention aims to identify the driving modes of a vehicle by means of data mining, further optimize and set cruise control multi-target weights under different driving modes by using a Bayesian optimization method, further improve the fuel economy of the vehicle and greatly reduce the setting workload of the multi-target optimization control weight.

The method comprises the following steps:

s1, information acquisition;

s2, clustering and identifying the driving modes of the front vehicle;

s3, designing a prediction cruise controller:

3.1) establishing a vehicle dynamic model facing cruise control, wherein the vehicle dynamic model of the vehicle and the front vehicle:

d(k+1)＝d(k)+(v_l(k)-v_e(k))·T_s

v_e(k+1)＝v_e(k)+u(k)·T_s (1)

T_srepresenting the sampling time, k the time, d (k +1) the longitudinal distance between the host vehicle and the preceding vehicle at the time of k +1, d (k) the longitudinal distance between the host vehicle and the preceding vehicle at the time of k, v_l(k) Longitudinal speed of the preceding vehicle, v, at time k_e(k) The longitudinal speed v of the vehicle at time k_e(k +1) represents the vehicle longitudinal speed at the time k +1, and u (k) represents the vehicle longitudinal acceleration at the time k;

3.2) establishing an energy consumption model of the vehicle, and establishing an engine oil consumption approximate model by adopting a polynomial fitting method:

m_findicating the instantaneous specific fuel consumption, T, of the engine_eRepresenting engine torque, ω_eIndicating engine speed, p_1～9Representing fitting parameters; furthermore, T_eAnd ω_eThe following can be calculated by the acceleration and the vehicle speed:

m represents the vehicle mass, r represents the vehicle tire radius, η represents the transmission efficiency, I_gRepresenting transmission ratio of the gearbox, I_finalRepresenting the final gear ratio;

3.3) establishing multi-objective optimization problem description, wherein at the moment k, the multi-objective optimization problem description comprises the following steps:

p denotes the prediction time domain, λ₁、λ₂And λ₃Representing weight parameters in a multi-objective optimization function, where₁The corresponding optimization target represents security and traceability, d_refIndicating a reference vehicle distance, related to the front vehicle speed; lambda [ alpha ]₂The corresponding optimization target represents the fuel consumption of the vehicle, a complete fuel consumption model is not added into an objective function in order to reduce the calculation burden, and only one part of the fuel consumption model is taken as the optimization target; lambda [ alpha ]₃The corresponding optimization target represents driving comfort, and delta u represents the acceleration increment of the vehicle; u. of_minAnd u_maxShowing the minimum and maximum values of the acceleration of the vehicle, Deltau_minAnd Δ u_maxMinimum and maximum values representing the acceleration increment of the vehicle, d_minAnd d_maxRepresenting the minimum and maximum longitudinal distance of the two vehicles, v_minAnd v_maxA minimum value and a maximum value representing the vehicle speed of the host vehicle;

s4 parameter setting design based on Bayesian optimization method

4.1) selecting a section of travel track data of the front vehicle in a certain type of driving mode from the historical travel data processed in the steps S1 and S2 as simulation test data, and selecting the section of travel track data of the front vehicle in an acceleration travel mode to specifically explain the subsequent steps;

4.2) randomly selecting a set of weight parameters λ ═ λ₁ λ₂ λ₃]The method is used for setting a multi-objective optimization function, and the control simulation is executed by adopting the forecast cruise controller in the step S3 to obtain the following vehicle running of the vehicle under the section of the running track of the front vehicleDriving strategy in V_eThe vehicle speed change sequence of the vehicle under the following vehicle driving strategy is shown, and U represents the acceleration change sequence of the vehicle under the following vehicle driving strategy; the total oil consumption M of the vehicle in the distance can be calculated by the formulas (2) and (3)_fuel

M_fuel＝∑m_f(U,V_e)·T_s (5)

Using negative oil consumption Y to represent M_fuelThe currently selected weight parameter lambda and the negative oil consumption Y form an input/output pair sample { lambda, Y };

4.3) repeating the step 4.2 to obtain a plurality of groups of sample pointsUsing the obtained data as a training sample set, and establishing lambda through a Gaussian process methodⁱAnd YⁱNon-parametric model F, i.e. Y, betweenⁱ＝F(λⁱ) Then the problem of setting the optimal weight parameter

Wherein λ_minDenotes the minimum value of the weight, λ_maxRepresents the maximum value of the weight; the Gaussian process model assumes that there are already sample pointsObeying a multidimensional Gaussian distribution, wherein the correlation degree between any two Y's and Y' can be calculated by the covariance function of the corresponding lambda's and lambda's; covariance functions, also known as kernel functions, have a number of choices, where a Gaussian kernel function is chosen

||·||²Representing Euclidean distance, σ_lRepresenting the characteristic length size as a hyper-parameter; tong (Chinese character of 'tong')Over-conjugate gradient method for obtaining sigma which enables maximum value of maximum likelihood function of current sample set_lThen, the current sample set can be establishedThe non-parametric model F above; the output of the model F, built by the Gaussian process method, is a Gaussian distribution, denoted by YⁱMeans, σ (λ), representing the Gaussian distribution of the outputⁱ) A variance representing a gaussian distribution of the output; when given a test input λ^testWhen, F (λ)^test) Mean value of (Y)^testAnd variance σ (λ)^test)

WhereinRepresenting Gaussian noise, I_nDenotes an identity matrix of size n, Y_train＝[Y¹ Y² … Yⁿ]^TT denotes transpose, k_*And K is represented by the following formula:

4.4) selecting the next set of λ using the desired gain acquisition function^newSo that Y is^new＝F(λ^new) Obtaining a new maximum value; the acquisition function generally assumes that when the acquisition function takes a maximum value, the target model can also take the maximum value; desired gain acquisition function

F(λ⁺) Representing data taken from the current model FA known maximum value, Φ (-) represents a cumulative distribution function of the standard gaussian distribution, Φ (-) represents a probability density function of the standard gaussian distribution; by finding out EI (lambda)^new) Lambda to obtain a maximum value^newThe next set of lambda is obtained^newSo that Y is^new＝F(λ^new) Obtaining a new maximum value;

4.5) obtaining a new set of weights λⁿ⁺¹Then, the total fuel consumption of the vehicle corresponding to the new weight is obtained in the step 4.2, and the weight lambda is usedⁿ⁺¹And negative oil consumption Yⁿ⁺¹Adding the new sample point into the existing sample point, and using the new sample data setUpdating the non-parametric model F;

4.6) repeating the step 4.4 and the step 4.5, and obtaining a group of optimized weight parameters aiming at the multi-objective optimization function when the front vehicle is in the accelerating running mode after the maximum iteration step number is reached;

4.7) selecting other preceding vehicle running track data segments of the acceleration running mode in the step 4.1, and repeating the steps 4.2 to 4.6 to obtain other optimization weight parameters of multiple groups of multi-target optimization functions;

4.8) averaging all the obtained multiple groups of weight parameters, and taking the average as the optimal weight parameter of the multi-objective optimization function in the PCC system of the host vehicle under the accelerating driving mode of the host vehicle;

4.9) based on the steps 4.1-4.8, the optimal weight parameters of the multi-objective optimization function in the PCC system of the vehicle can be obtained similarly when the front vehicle is in the constant-speed driving mode and the deceleration driving mode; when the vehicle runs in subsequent following cruising, the driving mode of the front vehicle is identified and classified on line at each sampling time based on the speed information of the front vehicle, and the weight parameters of the multi-target optimization function in the PCC system are adjusted to be the optimal weight parameters in the corresponding driving mode by fusing the identification result of the current driving mode.

The method adopts a clustering method to cluster and identify the driving mode of the front vehicle based on rich multi-source heterogeneous network connection information, and provides data support for adjusting each weight parameter in a multi-objective optimization function; based on the historical driving state data of the preceding vehicle, the driving modes of the preceding vehicle are identified and clustered, the weight parameters of the multi-objective optimization function are respectively adjusted according to different categories by using a Bayesian optimization method, a group of optimal weight parameters can be rapidly and accurately solved, a large number of trial and error required when the weights are manually selected is effectively avoided, and the manpower for adjusting the weight parameters is greatly reduced. In addition, the weight set by the Bayes optimization method is often better than the weight selected artificially, so that the oil consumption of the vehicle during cruising and following is further reduced, the fuel economy is effectively improved, and the tracking performance and the driving comfort are improved; based on the set optimal weight parameters corresponding to different driving mode categories of the front vehicle, the weight parameters of the multi-objective optimization function in the vehicle prediction cruise controller can be adjusted in real time according to the driving mode identification result of the front vehicle when the vehicle runs in subsequent cruising and following modes, so that the vehicle can achieve the aim of further improving the fuel economy on the premise of not greatly changing a control method and a control structure. Aiming at the problem of distinguishing the driving mode of the front vehicle during cruising and following driving, the invention provides a method for identifying the driving mode of the front vehicle based on rich multi-source heterogeneous network connection information by adopting a clustering algorithm, and provides effective data support for adjusting and predicting the weight parameter of a multi-target optimization function in a cruising controller. The invention provides a weight parameter optimization self-tuning method based on a Bayesian optimization method aiming at the problem of multi-objective optimization function weight parameter selection in different driving modes of a front vehicle, so that the optimal weight parameters corresponding to the different driving modes of the front vehicle can be quickly and accurately obtained.

Drawings

FIG. 1 is a process flow diagram of the present invention;

FIG. 2 is a schematic diagram of the control system of the present invention;

FIG. 3 is an overall block diagram of the control system of the present invention;

FIG. 4 is a schematic diagram of the results of the preceding vehicle driving pattern clustering;

FIG. 5 is a view showing the vehicle speed trajectory data of the preceding vehicle and the vehicle speed trajectory maps obtained by the above two methods, respectively;

FIG. 6 is a diagram showing the variation of the difference between the longitudinal vehicle distance of two vehicles and the reference vehicle distance obtained by the above two methods;

FIG. 7 is a graph showing the change in the acceleration increment of the vehicle obtained by the above two methods;

fig. 8 is a graph showing the results of the total fuel consumption change of the vehicle under the control strategies obtained by the above two methods.

Detailed Description

In the invention, a cruise car-following scene is considered, and the system firstly adopts a clustering algorithm to identify the driving mode of the front car based on rich multi-source heterogeneous network connection information; then, based on the historical driving state information of the front vehicle and the corresponding driving mode information, automatically setting each weight parameter in the multi-objective optimization function by using a Bayesian optimization method to obtain an optimal weight parameter which enables the energy consumption of the vehicle to be the least under a certain driving mode of the front vehicle; and finally, based on the current driving mode identification result of the front vehicle, the set weight parameters are used for adjusting each weight parameter of the multi-objective optimization function in the PCC system of the vehicle, so that the aim of further improving the fuel economy of the vehicle when the vehicle is driven by following cruising is fulfilled.

The invention is realized by the following technical scheme:

1. information acquisition: collecting historical speed, historical position and other driving state information of a front vehicle and historical state information of a road traffic environment;

2. clustering and identifying the driving modes of the front vehicle: classifying the driving modes of the front vehicles by adopting a clustering method according to the collected historical driving state information of the front vehicles and the historical state information of the road traffic environment;

3. the design of the predictive cruise controller comprises the following processes:

3.1) establishing a PCC-oriented vehicle dynamics model;

3.2) establishing an energy consumption model of the vehicle;

3.3) selecting the energy consumption of the vehicle in a prediction time domain as a main optimization target, and simultaneously adding tracking performance and comfort as a secondary optimization target; determining constraint conditions of a multi-objective optimization problem, wherein the constraint conditions comprise vehicle speed constraint, vehicle acceleration variation constraint, longitudinal distance constraint between two vehicles and the like;

3.4) solving the constrained multi-objective optimization problem of the rolling time domain by using a nonlinear programming solver;

4. the parameter setting design based on the Bayesian optimization method comprises the following processes:

4.1) selecting a section of driving track data of the front vehicle in a certain driving mode from the historical driving data processed in the step 1 and the step 2 as simulation test data;

4.2) randomly selecting a group of weight parameters for setting each weight parameter of a multi-objective optimization function in the predictive cruise controller, obtaining a cruise and follow vehicle driving control strategy of the vehicle aiming at the driving track of the front vehicle in the step 4.1 through control simulation, and calculating the total energy consumption of the vehicle in the distance by adopting an energy consumption model;

4.3) repeating the step 4.2 to obtain a plurality of groups of weight parameters and the total energy consumption of the vehicle under the corresponding control simulation, representing the negative number of the total energy consumption of the vehicle by negative energy consumption, constructing the weight parameters and the corresponding negative energy consumption as input and output pair samples, and establishing a non-parametric fitting model with the input as the weight parameters and the output as the negative energy consumption by a Gaussian process method;

4.4) selecting a new group of weight parameters which can enable the negative energy consumption to be larger on the non-parametric model obtained in the step 4.3 through the acquisition function, and executing control simulation by using the new group of weight parameters to obtain the corresponding total energy consumption of the vehicle, so as to obtain a new group of weight parameters and input/output pair samples of the negative energy consumption;

4.5) updating the non-parametric model in the step 4.3 by using the newly obtained sample in the step 4.4 and all original samples;

4.6) repeating the steps 4.4 and 4.5 until the maximum iteration times is reached, and finally obtaining a group of optimization weight parameters aiming at the multi-objective optimization function in the PCC system of the vehicle in the driving mode of the front vehicle;

4.7) selecting other multi-section running track data under the driving mode of the preceding vehicle, repeating the steps 4.2-4.6, and obtaining a plurality of groups of optimization weight parameters aiming at the multi-target optimization function in the PCC system of the vehicle under the driving mode of the preceding vehicle;

4.8) averaging all the obtained multiple groups of optimized weight parameters, and taking the averaged optimized weight parameters as the optimal weight parameters of the multi-objective optimization function in the PCC system of the vehicle in the driving mode of the front vehicle;

4.9) similarly obtaining the optimal weight parameters of the multi-target optimization function in the PCC system of the vehicle corresponding to the front vehicle in other driving mode types through the steps 4.1-4.8, and adjusting the weight parameters to the optimal weight parameters in the corresponding driving mode according to the identification result of the driving mode of the current front vehicle when the vehicle runs in the subsequent cruising and following modes;

5. performance evaluation: giving a section of complete and long-time running track data of a front vehicle; according to the step 4.9, the cruising and vehicle following driving track of the vehicle under the Bayesian optimization setting weight can be obtained; and comparing the running track with the running track obtained under the single set of fixed weight selected manually, and analyzing and evaluating the improvement degree of each optimization target.

The present invention is described in detail below:

1. and (5) information acquisition. The method comprises the steps of collecting historical speed, position and other driving state information of a front vehicle, removing noise in the speed data of the front vehicle by using a Gaussian filtering method, calculating acceleration change of the front vehicle, and providing necessary data support for front vehicle driving mode clustering. The adopted running track data is obtained by the mass production type SUV of a certain enterprise(Hubei)Economic(Wuhan)Collected in the road working condition of city. It should be noted that, in the present embodiment, the subsequent driving pattern clustering and identification are performed only by using the acceleration as a characteristic, and subsequently, more kinds of multi-source internet information can be used as a characteristic to further improve the accuracy of clustering and identification.

2. And clustering and identifying the driving modes of the front vehicle. The specific scheme is based on a K-means clustering method, and the acceleration data after collection and processing is used for updating a clustering center, and the method mainly comprises the following steps: initializing three clustering centers; calculating Euclidean distances from the sample data to the centers of the clusters; classifying the sample data into a class with the minimum distance from the clustering center; after all sample data are classified, calculating the average value of each class of data, and taking the average value as a new clustering center; and repeating the steps until the maximum iteration times is reached to obtain the optimal clustering center. According to the specific scheme, the driving modes of the front vehicle are divided into three categories, as shown in fig. 2, which respectively represent that the front vehicle is in a constant speed driving mode, an acceleration driving mode and a deceleration driving mode. The specific scheme is divided into three categories which are only used for illustration, and the categories can be further refined according to actual identification requirements.

3. The design of the prediction cruise controller comprises the following specific steps:

3.1) establishing a vehicle dynamic model facing the cruise control. When the vehicle is cruising and driving, the vehicle dynamics model of the vehicle and the front vehicle can be described by the following formula:

T_srepresenting the sampling time, k the time, d (k +1) the longitudinal distance between the host vehicle and the preceding vehicle at the time of k +1, d (k) the longitudinal distance between the host vehicle and the preceding vehicle at the time of k, v_l(k) Longitudinal speed of the preceding vehicle, v, at time k_e(k) The longitudinal speed v of the vehicle at time k_e(k +1) represents the vehicle longitudinal speed at the time k +1, and u (k) represents the vehicle longitudinal acceleration at the time k.

3.2) establishing an energy consumption model of the vehicle. The specific method takes the fuel consumption of the vehicle as an energy consumption index, and in order to avoid building a complex engine oil consumption model from the mechanism, a polynomial fitting method is adopted to build an engine oil consumption approximate model, which can be described by the following formula:

m_findicating the instantaneous specific fuel consumption, T, of the engine_eRepresenting engine torque, ω_eIndicating engine speed, p_1～9Express fitParameters. Furthermore, T_eAnd ω_eCan be calculated from the acceleration and the vehicle speed, and can be described by the following formula:

m represents the vehicle mass, r represents the vehicle tire radius, η represents the transmission efficiency, I_gRepresenting transmission ratio of the gearbox, I_finalRepresenting the final gear ratio.

3.3) establishing multi-objective optimization problem description. At time k, the multiobjective optimization problem description may be represented by:

p denotes the prediction time domain, λ₁、λ₂And λ₃Representing weight parameters in a multi-objective optimization function, where₁The corresponding optimization target represents security and traceability, d_refIndicating a reference vehicle distance, related to the front vehicle speed; lambda [ alpha ]₂The corresponding optimization target represents the fuel consumption of the vehicle, a complete fuel consumption model is not added into an objective function in order to reduce the calculation burden, and only one part of the fuel consumption model is taken as the optimization target; lambda [ alpha ]₃The corresponding optimization target represents driving comfort, and Δ u represents the vehicle acceleration increment. u. of_minAnd u_maxShowing the minimum and maximum values of the acceleration of the vehicle, Deltau_minAnd Δ u_maxMinimum and maximum values representing the acceleration increment of the vehicle, d_minAnd d_maxRepresenting the minimum and maximum longitudinal distance of the two vehicles, v_minAnd v_maxThe minimum value and the maximum value of the vehicle speed of the host vehicle are indicated.

And 3.4) solving a multi-objective optimization problem. In the specific scheme, an fmincon nonlinear programming solver in MATLAB software is adopted to solve the multi-objective optimization problem, and in order to accelerate the problem solving speed, a commercial solver can be adopted in the follow-up process, such as: cplex, Gurobi, etc.

4. The parameter setting design based on the Bayesian optimization method comprises the following specific steps:

4.1) selecting a section of driving track data of the front vehicle in a certain driving mode from the historical driving data processed in the steps 1 and 2 as simulation test data. Here, the following steps will be specifically described by taking the historical vehicle speed data of the preceding vehicle shown in fig. 4 as an example, and selecting a piece of travel track data of the preceding vehicle in the acceleration travel mode.

4.2) randomly selecting a set of weight parameters λ ═ λ₁ λ₂ λ₃]The method is used for setting a multi-objective optimization function, and executing control simulation by adopting the prediction cruise controller in the step 3 to obtain a following vehicle running strategy under the section of running track of the vehicle in front, and the following vehicle running strategy is expressed by V_eThe following driving strategy is a following driving strategy of the vehicle, and the following driving strategy is a following driving strategy of the vehicle. The total oil consumption M of the vehicle in the distance can be calculated by the formulas (2) and (3)_fuelCan be described by the following formula:

M_fuel＝∑m_f(U,V_e)·T_s, (5)

using negative oil consumption Y to represent M_fuelThe currently selected weight parameter lambda and the negative fuel consumption Y form an input/output pair sample { lambda, Y }.

4.3) repeating the step 4.2 to obtain a plurality of groups of sample pointsUsing the obtained data as a training sample set, and establishing lambda through a Gaussian process methodⁱAnd YⁱNon-parametric model F, i.e. Y, betweenⁱ＝F(λⁱ) Then, the tuning problem of the optimal weight parameter can be described by the following optimization problem:

wherein λ_minDenotes the minimum value of the weight, λ_maxRepresenting the maximum value of the weight. The Gaussian process model assumes that there are already sample pointsObey a multidimensional gaussian distribution, and the degree of correlation between any two Y's and Y "s therein can be calculated from the corresponding λ's and λ" s by means of a covariance function.

Covariance functions, also known as kernel functions, have a number of choices, where a gaussian kernel function is chosen, which can be described by the following equation:

||·||²representing Euclidean distance, σ_lThe characteristic length size is expressed as a hyper-parameter. Determining sigma for making maximum value of maximum likelihood function of current sample set by conjugate gradient method_lThen, the current sample set can be establishedThe non-parametric model F above.

The output of the model F, built by the Gaussian process method, is a Gaussian distribution, denoted by YⁱMeans, σ (λ), representing the Gaussian distribution of the outputⁱ) Representing the variance of the output gaussian distribution. When given a test input λ^testWhen, F (λ)^test) Mean value of (Y)^testAnd variance σ (λ)^test) This can be found by the following equation:

whereinRepresenting Gaussian noise, I_nDenotes an identity matrix of size n, Y_train＝[Y¹ Y² … Yⁿ]^TT denotes transpose, k_*And K is represented by the following formula:

4.4) selecting the next set of λ using the desired gain acquisition function^newSo that Y is^new＝F(λ^new) A new maximum value is taken. The acquisition function generally assumes that when the acquisition function takes a maximum value, the object model also takes a maximum value. The desired gain acquisition function may be described by:

F(λ⁺) Represents the known maximum taken from the current model F, phi (-) represents the cumulative distribution function of the standard gaussian distribution, phi (-) represents the probability density function of the standard gaussian distribution. By finding out EI (lambda)^new) Lambda to obtain a maximum value^newThe next set of lambda is obtained^newSo that Y is^new＝F(λ^new) A new maximum value is taken.

4.5) obtaining a new set of weights λⁿ⁺¹Then, the total fuel consumption of the vehicle corresponding to the new weight is obtained in the step 4.2, and the weight lambda is usedⁿ⁺¹And negative oil consumption Yⁿ+¹Adding the new sample point into the existing sample point, and using the new sample data setThe non-parametric model F is updated.

4.6) repeating the step 4.4 and the step 4.5, and obtaining a group of optimization weight parameters aiming at the multi-objective optimization function when the front vehicle is in the acceleration driving mode after the maximum iteration step number is reached.

4.7) selecting other preceding vehicle running track data segments of the acceleration running mode in the step 4.1, and repeating the steps 4.2 to 4.6 to obtain other optimization weight parameters of the multi-group multi-objective optimization function.

4.8) averaging all the obtained multiple groups of weight parameters, and taking the average as the optimal weight parameter of the multi-objective optimization function in the PCC system of the vehicle when the front vehicle is in the accelerating driving mode.

4.9) based on the steps 4.1-4.8, the optimal weight parameters of the multi-objective optimization function in the PCC system of the vehicle can be obtained similarly when the front vehicle is in the constant-speed driving mode and the deceleration driving mode; when the vehicle runs in subsequent following cruising, the driving mode of the front vehicle is identified and classified on line at each sampling time based on the speed information of the front vehicle, and the weight parameters of the multi-target optimization function in the PCC system are adjusted to be the optimal weight parameters in the corresponding driving mode by fusing the identification result of the current driving mode.

5. And (6) performance evaluation. Taking the previous vehicle speed trajectory data shown in fig. 5 as an example, the vehicle starts to perform cruising and vehicle following driving actions from the 2 nd second by using the prediction cruise controller in the step 3, identifies the driving mode of the previous vehicle in real time at each sampling time, and adjusts the multi-target optimization function weight parameter in the prediction cruise controller into the optimal parameter in the corresponding driving mode by combining the driving mode identification result; finally, the driving strategy of the vehicle under the Bayesian optimization setting weight can be obtained, the driving strategy is compared with the driving strategy obtained under the single set of fixed weight selected manually, and the results of the graphs in FIGS. 6, 7 and 8 show that the specific scheme can further improve the tracking performance, the comfort and the fuel economy of the vehicle.

(II) FIG. 2 is a schematic diagram of a networked automobile prediction cruise parameter self-tuning control system integrating driving pattern identification information, in the diagram, firstly, a host vehicle collects historical driving state information of a preceding vehicle and historical state information such as road traffic environment and the like in the cruising and following process, and the driving patterns of the preceding vehicle are identified and clustered by using the historical driving state information of the preceding vehicle and the historical state information of the road traffic environment and the like, so that the driving pattern information of the historical driving state of the preceding vehicle can be obtained; then, setting various weight parameters of the multi-objective optimization function in the prediction cruise controller based on a Bayesian optimization method by combining the information to respectively obtain optimal weight parameters corresponding to different driving modes of the front vehicle, and storing the weight parameters in the prediction cruise controller to facilitate subsequent parameter adjustment; and finally, at the current moment, the vehicle adjusts each weight parameter in the multi-target optimization function into the optimal weight parameter in the corresponding driving mode according to the driving mode identification result, and the optimal cruising and following vehicle driving strategy is obtained by solving the multi-target optimization problem.

(III) FIG. 3 is an overall block diagram of the online automobile prediction cruise parameter self-tuning control system integrating driving pattern identification information, firstly, a host vehicle collects information such as historical driving state information of a preceding vehicle and historical state of road traffic environment in the process of cruising and following vehicle driving, and the information such as historical driving state information of the preceding vehicle and historical state of road traffic environment is used for identifying and clustering the driving pattern of the preceding vehicle, so that the driving pattern information of the historical driving state of the preceding vehicle can be obtained; then, setting various weight parameters of a multi-objective optimization function in the prediction cruise controller based on a Bayesian optimization method in combination with the information to respectively obtain optimal weight parameters corresponding to different driving modes of the front vehicle, and storing the weight parameters in the prediction cruise controller to facilitate subsequent parameter adjustment; and finally, at the current moment, the vehicle adjusts each weight parameter in the multi-target optimization function into the optimal weight parameter in the corresponding driving mode according to the recognition result of the driving mode of the previous vehicle, and the optimal cruising and following driving strategy is obtained by solving the multi-target optimization problem.

(IV) FIG. 4 is a schematic diagram of the clustering results of the driving patterns of the front vehicles

The figure shows the driving mode classification result of the running track data of the front vehicle in a certain long time, wherein the type 1 shows that the front vehicle is in a uniform speed driving mode, the type 2 shows that the front vehicle is in an acceleration driving mode, and the type 3 shows that the front vehicle is in a deceleration driving mode.

(V) FIGS. 5, 6, 7, and 8 are schematic diagrams showing performance evaluation results

And giving a section of complete and long-time running track data of the front vehicle, and finishing the cruising and following running of the front vehicle by the vehicle through the predictive cruising controller. Therefore, the driving track of the vehicle under the single set of fixed weight parameters and the driving track of the vehicle under the Bayesian optimization weight parameters can be obtained, and the improvement degrees of three optimization targets, namely the tracking capability, the comfort and the fuel economy of the vehicle, can be analyzed and evaluated according to the two driving tracks.

Fig. 5 shows the vehicle speed trajectory data of the preceding vehicle and the vehicle speed trajectory maps obtained by the above two methods. The black solid line represents the vehicle speed track of the front vehicle, the gray dotted line represents the vehicle speed track obtained under the single set of artificially fixed weight parameters, and the black dotted line represents the vehicle speed track obtained under the Bayesian optimization weight parameters.

Fig. 6 is a diagram showing the variation of the difference between the longitudinal vehicle distance of the two vehicles and the reference vehicle distance obtained by the above two methods. The black solid line represents a difference change curve of the longitudinal vehicle distance and the reference vehicle distance obtained under the artificially set single group of fixed weight parameters, and the black dotted line represents a difference change curve of the longitudinal vehicle distance and the reference vehicle distance obtained under the Bayesian optimization weight parameters. It can be calculated from the results shown in the figure that the average absolute error of the solid line is 2.4848m, and the average absolute error of the dashed line is 1.9264 m. Therefore, the error between the longitudinal vehicle distance and the reference vehicle distance obtained under the Bayesian optimization weight parameter is smaller, which means that the vehicle has stronger tracking capability and higher safety.

Fig. 7 shows the change in the acceleration increment of the vehicle obtained by each of the above two methods. The black solid line represents the acceleration increment change curve of the vehicle obtained under the single set of artificial fixed weight parameters, and the gray dotted line represents the acceleration increment change curve of the vehicle obtained under the Bayesian optimization weight parameters. Generally, the smaller the absolute value of the acceleration increment, the higher the driving comfort, and it can be calculated from the results shown in the figure that the absolute value of the solid line is 0.0327m/s on average²The absolute value of the dotted line averages 0.0291m/s². Therefore, the driving comfort of the vehicle can be improved by the control strategy obtained under the Bayesian optimization weight parameter.

Fig. 8 is a graph showing the results of the total fuel consumption change of the vehicle under the control strategies obtained by the above two methods. The solid line represents the total oil consumption change curve of the vehicle when the driving mode of the front vehicle is not identified and the multi-objective optimization function weight parameters in the PCC system only use a single set of artificially set fixed weight parameters; and the dotted line represents the driving mode of the vehicle before online identification clustering, and the multi-target optimization function weight parameter in the PCC system uses the total oil consumption change curve of the vehicle of the optimal weight parameter set by the Bayesian optimization method under the corresponding driving mode. The inside of the rectangular frame represents a partially enlarged view. As can be seen from the figure, the method provided by the invention can effectively reduce the fuel consumption of the vehicle by 2.13%, and improve the fuel economy.