Intelligent driving fuel cell vehicle control method integrating multi-dimensional road information spatialization

Abstract

The present invention proposes a control method for an intelligent driving fuel cell vehicle that integrates the spatialization of multi-dimensional information of a driving road, comprising the following steps: obtaining a target vehicle speed output by a decision-making plan, and collecting vehicle status information and battery data information; establishing a time-varying driving environment model through a road recognition convolutional neural network based on VGG16 to estimate the real-time road adhesion coefficient and the optimal slip rate, and calculating the maximum acceleration and deceleration of the current road surface; outputting the acceleration and deceleration required to achieve the target vehicle speed through brain emotion learning, and limited by the current acceleration and deceleration, and obtaining the throttle and brake pedal openings through a throttle and brake calibration table to achieve the target vehicle speed; constructing a multi-dimensional stereo control network suitable for a nine-dimensional state space and a three-dimensional action space, and proposing an energy management strategy based on a deep Q network; the present invention integrates multi-dimensional information, and uses a motor speed fine-tuning strategy and distributes power from a power source, and has good speed tracking accuracy, environmental adaptability and economy.

Description

Technical Field

[0001] This invention relates to the field of intelligent driving technology for vehicles, and in particular to a control method for intelligent driving fuel cell vehicles that integrates multi-dimensional road information spatialization. Background Art

[0002] With the development of information technology and the growing awareness of energy conservation and emission reduction, developing intelligent new energy vehicles is a crucial issue for both the present and the future. New energy vehicles refer to automobiles that use new energy sources that are not fossil fuels and employ new power systems. Intelligent vehicles will transform the transportation industry, potentially saving lives by minimizing traffic accidents, reducing traffic congestion, increasing productivity, and providing convenience for those unable to drive. Intelligent fuel cell vehicles, as products integrating intelligence and new energy, can achieve energy conservation and emission reduction while alleviating driving stress, improving driving safety and mileage limitations, thus meeting future development needs.

[0003] The performance of the motion control module in intelligent driving directly affects whether a vehicle can travel quickly, smoothly, and accurately along the desired trajectory. Therefore, motion control is of great significance to intelligent driving. This control can be functionally divided into lateral motion control and longitudinal motion control. Longitudinal control uses a control strategy to adjust the vehicle's acceleration and deceleration to control the vehicle speed to reach the target value. Regarding longitudinal motion, common control methods (such as PID, LQR, and MPC) suffer from poor adaptability to complex road conditions. How to establish a longitudinal control strategy that is environmentally adaptive, has high speed tracking accuracy, and is simple and feasible is a problem that needs to be solved.

[0004] Furthermore, fuel cell vehicles have two power sources: a lithium-ion battery and a fuel cell. Through energy management strategies, the required power can be rationally allocated between these two power sources, and coordinated control can be achieved by integrating multi-dimensional information such as driving road conditions. This allows for improved vehicle economy while maintaining good power performance. Based on this, the present invention is proposed. Summary of the Invention

[0005] This invention proposes an intelligent driving fuel cell vehicle control method that integrates multi-dimensional road information spatialization, which has good speed tracking accuracy, environmental adaptability, and economy.

[0006] The present invention adopts the following technical solution.

[0007] A control method for intelligent driving fuel cell vehicles that integrates multi-dimensional road information spatialization includes the following steps;

[0008] Step S1: Obtain the target vehicle speed output by the decision planning, and collect vehicle status information and battery data information through sensors;

[0009] Step S2: Using a VGG16-based road recognition convolutional neural network, establish a time-varying driving environment model for road driving, estimate the real-time road adhesion coefficient and the optimal slip ratio during driving, and calculate the maximum allowable acceleration and deceleration of the current road surface.

[0010] Step S3: Through brain emotional learning, output the acceleration and deceleration required to reach the target vehicle speed, and obtain the accelerator and brake pedal openings from the accelerator and brake calibration table based on the current maximum acceleration and deceleration limit of the road surface, so as to reach the target vehicle speed;

[0011] Step S4: Construct a deep Q-network (DQN) containing a multidimensional neural network and propose an energy management strategy for multi-objective control.

[0012] In step S1, the vehicle status information includes: wheel speed, motor speed, and motor torque; the battery data information includes: current, voltage, and battery state of charge.

[0013] In step S2, the time-varying driving environment model is used to improve the environmental adaptability of the vehicle's longitudinal control. The time-varying driving environment model includes: road type, road adhesion coefficient, and optimal slip ratio.

[0014] In step S2, the VGG16-based road recognition convolutional neural network consists of 13 convolutional layers and 3 fully connected layers. When training with images of different road types, different neurons in the Softmax layer correspond to different road types and are used as outputs. The specific process is as follows: the activation function for the convolutional layers of the convolutional neural network is ReLU; RGB format images are compressed to 224×224 and fed into the network, then processed through multiple convolutional layers and max-pooling layers of different sizes. After processing to a suitable size, the data format is adjusted from a three-dimensional convolutional structure to a two-dimensional fully connected structure, and then the features are gradually converted into a predefined number of types. The number of neurons in the Softmax layer is a predefined number of classifications, corresponding to different road types, and is used as the output.

[0015] In step S2, the time-varying driving environment model estimates the real-time road adhesion coefficient and the optimal slip ratio, and calculates the maximum acceleration and deceleration of the current road surface, as follows:

[0016] The relationship between the adhesion coefficient and slip ratio for different road types is established and expressed by a formula.

[0017]

[0018] Where μ is the road adhesion coefficient, s slip c1, c2, and c3 are the slip ratios, and c1, c2, and c3 are parameters for each road type.

[0019] After identifying different road types, the maximum road adhesion coefficient μ is obtained. max and optimal slip ratio

[0020] During deceleration, the slip ratio is related to vehicle speed and wheel speed as follows:

[0021]

[0022] Among them, v veh1 (t) represents the actual velocity v during deceleration. veh1 w is a function of time t wheel (t) represents the angular velocity of the wheel w wheel r is a function of time t. wheel The radius of the wheel;

[0023] Therefore, the maximum deceleration a under the optimal slip ratio demax Expressed as a formula

[0024]

[0025] v tar1 (t) represents the target velocity v during deceleration. tar1 A function of time t;

[0026] Adhesion rate of drive wheels The relationship with the adhesion coefficient μ is as follows

[0027]

[0028] Among them, F X2 For the tangential reaction force on the ground, F Z2 This is the normal reaction force of the driving wheel;

[0029] During acceleration, adhesion rate Approximately equal to or equal to the coefficient of adhesion μ, in order to maximize the vehicle's acceleration characteristics, this can be solved from vehicle dynamics at the adhesion rate. At that time, the actual speed is v veh2 At this time, the maximum acceleration a inmax for

[0030]

[0031] Among them, v tar2 (t) represents the target velocity v during acceleration. tar2 v is a function of time t. veh2 (t) represents the actual velocity v during acceleration. veh2 A function of time t.

[0032] The brain's emotional learning circuit takes sensory stimulation signal G as input and, under the supervision of emotional cue function EC, ultimately outputs the required acceleration and deceleration values ​​E.

[0033] The brain's emotional learning is composed of the thalamus, sensory cortex, amygdala, and orbitofrontal cortex. The thalamus preprocesses sensory stimulation signals and then transmits them downwards. The sensory cortex receives information from the thalamus. The preprocessed stimulation signals are further analyzed and processed in the sensory cortex and then sent to the orbitofrontal cortex and amygdala, respectively.

[0034] The process of emotional learning in the brain mainly occurs within the amygdala. The orbitofrontal cortex plays a role in enhancing or inhibiting the learning of the amygdala to avoid underlearning or overlearning.

[0035] During the learning process, the strongest sensory stimulation signal is sent directly from the thalamus to the amygdala; each input node in the amygdala simultaneously receives the sensory input signal G, the emotional cue function EC, and the strongest sensory stimulation signal A. th The orbitofrontal cortex receives sensory input signals G and emotional cue functions EC, and regulates its output accordingly; specifically:

[0036] Let the sensory input signal G and the strongest sensory stimulus signal A be... th =max[G]; Each stimulus signal corresponds to node A in the amygdala. i The stimulus signal includes the sensory input signal G, the emotional cue function EC, and the strongest sensory stimulus signal A. th The amygdala output is

[0037] A i =V i ·G i A m+1 =V m+1 ·A th Formula 6, i = 1, 2, ..., m;

[0038] Where m is the number of stimulus signals, V i Let be the weights of each node, and its adjustment law be;

[0039]

[0040]

[0041] Where α is the learning rate, which affects the learning speed of the amygdala; ΔV i monotonicity and G i The symbols must be consistent;

[0042] For stimulation signals, nodes are also located within the orbitofrontal cortex to receive these signals, which include signals from the sensory cortex and the amygdala. Signals from the thalamus are not stimulated, and the output is...

[0043] O i =W i ·G i Formula Nine;

[0044] Sensory cortical signals are sensory input signals, where W i Let be the weights of each node, and its adjustment law be...

[0045] ΔW i =β[G i Formula 10; (E'-EC)]

[0046] Where β is the learning rate, and β > 0, and E' is the output of the amygdala without thalamic signal stimulation, expressed as

[0047]

[0048] ΔW i It can take both positive and negative values, meaning the orbitofrontal cortex can inhibit or enhance the learning process of the amygdala. By eliminating the difference between the emotional cue function EC and the amygdala output E', it makes the amygdala learn towards the desired value; vehicle longitudinal control will control the actual vehicle speed v veh With the target vehicle speed v tar When the error e is used as the input of the sensory stimulus signal G, then the sensory input function is:

[0049] G1=k1e=k1(v veh -v tar ) Formula 12;

[0050] The input signal of the emotional suggestion function EC is divided into two paths, one of which is the target vehicle speed v. tar With actual vehicle speed v veh The error is calculated using one path, while the other path is the output of the brain's emotional learning model controller. Therefore, the emotional cueing function EC is...

[0051]

[0052] Where k1, k2, k3, k4, and k5 are weight adjustment coefficients;

[0053] In summary, the actual output speed u1 of the brain's emotion learning model is

[0054] u1 = k1e(V1 - W1) Formula Fourteen;

[0055] Therefore, the actual output acceleration E1 is

[0056]

[0057] During acceleration, the actual acceleration E1 must be less than or equal to the maximum acceleration; during deceleration, due to the maximum deceleration a... demax The value is negative, meaning that the actual acceleration E1 must be greater than or equal to the maximum deceleration, so that optimal longitudinal control can be achieved under any road conditions.

[0058] The limited acceleration aact is

[0059]

[0060] In step S3, the throttle and brake calibration table is based on the vehicle dynamics model, and is a set of all accelerations, decelerations and vehicle speeds corresponding to each throttle or brake pedal opening.

[0061] The method for the vehicle to reach the target speed is: to reduce the limited acceleration a act Input the values ​​into the accelerator and brake calibration table to obtain the corresponding accelerator or brake pedal opening, so that the vehicle can reach the target speed.

[0062] In step S4, the dimension of the multidimensional neural network DQN is affected by the action space A. Each dimension receives the same action value and state value as input. The optimal control action π for each dimension is determined based on maximizing the expected value of the cumulative reward in different dimensions. * ;

[0063] The cumulative reward for each dimension in the future is defined as a different action value Q. π (s,a), and using the action value Q π The process of evaluating the quality of action a(t) using (s,a) is as follows:

[0064]

[0065] Where s is the state value, a is the action value, N is the total time length, r is the reward, and γ is the discount factor; the discount factor reflects the importance of future rewards.

[0066] If the optimal action a is chosen in state s(t) * If (t), then the maximum action value, which is also the optimal action value Q, will be generated. * (s,a), expressed by the formula as

[0067]

[0068] Ultimately, the optimal control action π in this dimension * Expressed as

[0069]

[0070] Where s0 is the initial state.

[0071] The energy management strategy of the multi-objective control is as follows: by using the neural network of DQN, the synchronous learning of multiple actions and the comprehensive control of the hybrid power system are realized; in the strategy, the optimal actions of different dimensions are not independent, but under the influence of the reward factor R, the parameters of the neural network are continuously updated to achieve simultaneous fitting of the optimal actions of each dimension, so as to improve the economy of intelligent driving fuel cell vehicles.

[0072] In multidimensional neural networks, the number of neurons in the input and output layers is related to the state space S and action space A, respectively. Furthermore, the hidden space, which serves as a key module, no longer uses a single-planar network but a three-dimensional structure with the same dimension as the number of control actions.

[0073] The state space S is defined as having nine dimensions, expressed by the formula:

[0074] S = {soc, v} veh ,a act ,ω motor ,P li ,P fuel ,θ angle Road, M people Formula 20;

[0075] Where SOC stands for State of Charge, and V... veh It is the actual speed, a act It is acceleration, ω motor It is the motor speed, P li It refers to the power output of lithium-ion power batteries, P. fuel It is the fuel cell power, θ angle It refers to the road gradient, and Road is the road type, M. people It is the total mass of passengers;

[0076] The action space A is defined as three-dimensional, expressed by the formula as follows:

[0077]

[0078] Where, Δω motor ΔPli represents the change in motor speed, while ΔP represents the change in lithium-ion battery power. fuel It refers to the change in fuel cell power; and the fine-tuning of the motor speed must be verified. The range of this action cannot be arbitrarily selected. When compared with other ranges, this range can obtain the maximum cumulative reward, which is the basis of the definition.

[0079] The above actions are continuous variables applicable to deep deterministic policy gradient DDPG. Using the changes as control actions can ensure better continuity; specifically...

[0080] Discretize all three actions described above, representing them as a reward factor R, defined as follows:

[0081]

[0082] Where k'1, k'2, k'3, k'4, and k'5 are weighting coefficients, and soc tar For the target battery state of charge, ω ref This is the reference speed for the motor. For the hydrogen consumption of fuel cells, i fuel For fuel cell current, u fuel For fuel cell voltage, i li For lithium-ion power battery current, u li For the lithium-ion power battery voltage, η li For the efficiency of lithium-ion power batteries; ω ref The motor speed, which reflects the optimal acceleration or deceleration during acceleration or deceleration, is obtained from wheel speed, coefficient of adhesion, or optimal slip ratio.

[0083] During the DQN training process, the weight coefficients are continuously adjusted, and after the reward r of each optimization term can converge to an approximate order of magnitude, appropriate fine-tuning is performed according to the importance.

[0084] All variables in the state space S are divided into three types: environment state (θ) angle Road, M people ), vehicle status (soc, v) veh 、a act ) and the state of the control object (ω) motor 、P li 、P fuel );

[0085] The multi-objective control function reflects the optimal mapping relationship between the nine-dimensional state space and the three-dimensional action space, and is fitted by a multi-dimensional neural network, described by the formula:

[0086] [Δω motor ,ΔP li ,ΔP fuel ] = f(soc,v veh ,a act ,ω motor ,P li ,P fuel ,θ angle Road, M people Formula 23;

[0087] In the energy management strategy, the change in motor speed is defined as the primary control object. The motor speed is fine-tuned to maintain the lowest coefficient of adhesion or optimal slip ratio during acceleration or braking. Then, the power of the fuel cell and the lithium-ion battery are rationally adjusted to minimize energy consumption, improve economy, and achieve multi-objective control of the hybrid power system.

[0088] Compared with existing technologies, the method of the present invention integrates multi-dimensional information such as driving road conditions, and uses motor speed fine-tuning strategy and power source power control strategy. It combines brain emotion learning, road recognition and multi-objective control of power system based on deep reinforcement learning in time-varying environments, and has good speed tracking accuracy, environmental adaptability and economy. It can improve economy, speed tracking accuracy and environmental adaptability of longitudinal control strategy.

[0089] Compared with existing technologies, the intelligent driving fuel cell vehicle control method that integrates multi-dimensional spatial information of the driving road described in this invention has the following advantages:

[0090] (1) This invention utilizes the “self-learning function” of the brain’s emotional learning to adjust control parameters and improve the vehicle’s speed following ability.

[0091] (2) The acceleration and deceleration capabilities of the vehicle of the present invention are limited by the driving road environment. Therefore, a time-varying driving environment model is established and integrated into the longitudinal control of the vehicle to improve the environmental adaptability and stability of the vehicle.

[0092] (3) This invention takes the motor speed fine-tuning strategy to maintain the lowest adhesion coefficient or the best slip ratio during acceleration or braking, the lithium-ion power battery power control strategy and the fuel cell power control strategy as multi-objective optimization objects, and proposes a multi-objective energy management strategy based on DQN to improve the economy of vehicle driving. Attached Figure Description

[0093] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0094] Attachment Figure 1 This is a schematic diagram of the intelligent driving fuel cell vehicle control method of the present invention, which integrates multi-dimensional spatial information of the driving road.

[0095] Attachment Figure 2 This is a schematic diagram of the intelligent driving fuel cell vehicle control method of the present invention, which integrates multi-dimensional information spatialization of driving roads. Detailed Implementation

[0096] As shown in the figure, the intelligent driving fuel cell vehicle control method that integrates multi-dimensional road information spatialization includes the following steps;

[0097] Step S1: Obtain the target vehicle speed output by the decision planning, and collect vehicle status information and battery data information through sensors;

[0098] Step S2: Using a VGG16-based road recognition convolutional neural network, establish a time-varying driving environment model for road driving, estimate the real-time road adhesion coefficient and the optimal slip ratio during driving, and calculate the maximum allowable acceleration and deceleration of the current road surface.

[0099] Step S3: Through brain emotional learning, output the acceleration and deceleration required to reach the target vehicle speed, and obtain the accelerator and brake pedal openings from the accelerator and brake calibration table based on the current maximum acceleration and deceleration limit of the road surface, so as to reach the target vehicle speed;

[0100] Step S4: Construct a deep Q-network (DQN) containing a multidimensional neural network and propose an energy management strategy for multi-objective control.

[0101] In step S1, the vehicle status information includes: wheel speed, motor speed, and motor torque; the battery data information includes: current, voltage, and battery state of charge.

[0102] In step S2, the time-varying driving environment model is used to improve the environmental adaptability of the vehicle's longitudinal control. The time-varying driving environment model includes: road type, road adhesion coefficient, and optimal slip ratio.

[0103] In step S2, the VGG16-based road recognition convolutional neural network consists of 13 convolutional layers and 3 fully connected layers. When training with images of different road types, different neurons in the Softmax layer correspond to different road types and are used as outputs. The specific process is as follows: the activation function for the convolutional layers of the convolutional neural network is ReLU; RGB format images are compressed to 224×224 and fed into the network, then processed through multiple convolutional layers and max-pooling layers of different sizes. After processing to a suitable size, the data format is adjusted from a three-dimensional convolutional structure to a two-dimensional fully connected structure, and then the features are gradually converted into a predefined number of types. The number of neurons in the Softmax layer is a predefined number of classifications, corresponding to different road types, and is used as the output.

[0104] In step S2, the time-varying driving environment model estimates the real-time road adhesion coefficient and the optimal slip ratio, and calculates the maximum acceleration and deceleration of the current road surface, as follows:

[0105] The relationship between the adhesion coefficient and slip ratio for different road types is established and expressed by a formula.

[0106]

[0107] Where μ is the road adhesion coefficient, sslip c1, c2, and c3 are the slip ratios, and c1, c2, and c3 are parameters for each road type.

[0108] After identifying different road types, the maximum road adhesion coefficient μ is obtained. max and optimal slip ratio

[0109] During deceleration, the slip ratio is related to vehicle speed and wheel speed as follows:

[0110]

[0111] Among them, v veh1 (t) represents the actual velocity v during deceleration. veh1 w is a function of time t wheel (t) represents the angular velocity of the wheel w wheel r is a function of time t. wheel The radius of the wheel;

[0112] Therefore, the maximum deceleration a under the optimal slip ratio demax Expressed as a formula

[0113]

[0114] v tar1 (t) represents the target velocity v during deceleration. tar1 A function of time t;

[0115] Adhesion rate of drive wheels The relationship with the adhesion coefficient μ is as follows

[0116]

[0117] Among them, F X2 For the tangential reaction force on the ground, F Z2 This is the normal reaction force of the driving wheel;

[0118] During acceleration, adhesion rate Approximately equal to or equal to the coefficient of adhesion μ, in order to maximize the vehicle's acceleration characteristics, this can be solved from vehicle dynamics at the adhesion rate. At that time, the actual speed is v veh2 At this time, the maximum acceleration a inmax for

[0119]

[0120] Among them, v tar2 (t) represents the target velocity v during acceleration. tar2 v is a function of time t. veh2 (t) represents the actual velocity v during acceleration. veh2 A function of time t.

[0121] The brain's emotional learning circuit takes sensory stimulation signal G as input and, under the supervision of emotional cue function EC, ultimately outputs the required acceleration and deceleration values ​​E.

[0122] The brain's emotional learning is composed of the thalamus, sensory cortex, amygdala, and orbitofrontal cortex. The thalamus preprocesses sensory stimulation signals and then transmits them downwards. The sensory cortex receives information from the thalamus. The preprocessed stimulation signals are further analyzed and processed in the sensory cortex and then sent to the orbitofrontal cortex and amygdala, respectively.

[0123] The process of emotional learning in the brain mainly occurs within the amygdala. The orbitofrontal cortex plays a role in enhancing or inhibiting the learning of the amygdala to avoid underlearning or overlearning.

[0124] During the learning process, the strongest sensory stimulation signal is sent directly from the thalamus to the amygdala; each input node in the amygdala simultaneously receives the sensory input signal G, the emotional cue function EC, and the strongest sensory stimulation signal A. th The orbitofrontal cortex receives sensory input signals G and emotional cue functions EC, and regulates its output accordingly; specifically:

[0125] Let the sensory input signal G and the strongest sensory stimulus signal A be... th =max[G]; Each stimulus signal corresponds to node A in the amygdala. i The stimulus signal includes the sensory input signal G, the emotional cue function EC, and the strongest sensory stimulus signal A. th The amygdala output is

[0126] A i =V i ·G i A m+1 =V m+1 ·A th Formula 6, i = 1, 2, ..., m;

[0127] Where m is the number of stimulus signals, V i Let be the weights of each node, and its adjustment law be;

[0128]

[0129]

[0130] Where α is the learning rate, which affects the learning speed of the amygdala; ΔV i monotonicity and G i The symbols must be consistent;

[0131] For stimulation signals, nodes are also located within the orbitofrontal cortex to receive these signals, which include signals from the sensory cortex and the amygdala. Signals from the thalamus are not stimulated, and the output is...

[0132] O i =W i ·G i Formula Nine;

[0133] Sensory cortical signals are sensory input signals, where W i Let be the weights of each node, and its adjustment law be...

[0134] ΔW i =β[G i Formula 10; (E'-EC)]

[0135] Where β is the learning rate, and β > 0, and E' is the output of the amygdala without thalamic signal stimulation, expressed as

[0136]

[0137] ΔW i It can take both positive and negative values, meaning the orbitofrontal cortex can inhibit or enhance the learning process of the amygdala. By eliminating the difference between the emotional cue function EC and the amygdala output E', it makes the amygdala learn towards the desired value; vehicle longitudinal control will control the actual vehicle speed v veh With the target vehicle speed v tar When the error e is used as the input of the sensory stimulus signal G, then the sensory input function is:

[0138] G1=k1e=k1(v veh -v tar ) Formula 12;

[0139] The input signal of the emotional suggestion function EC is divided into two paths, one of which is the target vehicle speed v. tar With actual vehicle speed v veh The error is calculated using one path, while the other path is the output of the brain's emotional learning model controller. Therefore, the emotional cueing function EC is...

[0140]

[0141] Where k1, k2, k3, k4, and k5 are weight adjustment coefficients;

[0142] In summary, the actual output speed u1 of the brain's emotion learning model is

[0143] u1 = k1e(V1 - W1) Formula Fourteen;

[0144] Therefore, the actual output acceleration E1 is

[0145]

[0146] During acceleration, the actual acceleration E1 must be less than or equal to the maximum acceleration; during deceleration, due to the maximum deceleration a... demax The value is negative, meaning that the actual acceleration E1 must be greater than or equal to the maximum deceleration, so that optimal longitudinal control can be achieved under any road conditions.

[0147] The limited acceleration aact is

[0148]

[0149] In step S3, the throttle and brake calibration table is based on the vehicle dynamics model, and is a set of all accelerations, decelerations and vehicle speeds corresponding to each throttle or brake pedal opening.

[0150] The method for the vehicle to reach the target speed is: to reduce the limited acceleration a act Input the values ​​into the accelerator and brake calibration table to obtain the corresponding accelerator or brake pedal opening, so that the vehicle can reach the target speed.

[0151] In step S4, the dimension of the multidimensional neural network DQN is affected by the action space A. Each dimension receives the same action value and state value as input. The optimal control action π for each dimension is determined based on maximizing the expected value of the cumulative reward in different dimensions. * ;

[0152] The cumulative reward for each dimension in the future is defined as a different action value Q. π (s,a), and using the action value Q π The process of evaluating the quality of action a(t) using (s,a) is as follows:

[0153]

[0154] Where s is the state value, a is the action value, N is the total time length, r is the reward, and γ is the discount factor; the discount factor reflects the importance of future rewards.

[0155] If the optimal action a is chosen in state s(t) * If (t), then the maximum action value, which is also the optimal action value Q, will be generated. * (s,a), expressed by the formula as

[0156]

[0157] Ultimately, the optimal control action π in this dimension * Expressed as

[0158]

[0159] Where s0 is the initial state.

[0160] The energy management strategy of the multi-objective control is as follows: by using the neural network of DQN, the synchronous learning of multiple actions and the comprehensive control of the hybrid power system are realized; in the strategy, the optimal actions of different dimensions are not independent, but under the influence of the reward factor R, the parameters of the neural network are continuously updated to achieve simultaneous fitting of the optimal actions of each dimension, so as to improve the economy of intelligent driving fuel cell vehicles.

[0161] In multidimensional neural networks, the number of neurons in the input and output layers is related to the state space S and action space A, respectively. Furthermore, the hidden space, which serves as a key module, no longer uses a single-planar network but a three-dimensional structure with the same dimension as the number of control actions.

[0162] The state space S is defined as having nine dimensions, expressed by the formula:

[0163] S = {soc, v} veh ,a act ,ω motor ,P li ,P fuel ,θ angle Road, M people Formula 20;

[0164] Where SOC stands for State of Charge, and V... veh It is the actual speed, a act It is acceleration, ω motor It is the motor speed, P li It refers to the power output of lithium-ion power batteries, P. fuel It is the fuel cell power, θ angle It refers to the road gradient, and Road is the road type, M. people It is the total mass of passengers;

[0165] The action space A is defined as three-dimensional, expressed by the formula as follows:

[0166]

[0167] Where, Δω motor It is the change in motor speed, ΔP li It refers to the change in the power of the lithium-ion battery, ΔP fuel It refers to the change in fuel cell power; and the fine-tuning of the motor speed must be verified. The range of this action cannot be arbitrarily selected. When compared with other ranges, this range can obtain the maximum cumulative reward, which is the basis of the definition.

[0168] The above actions are continuous variables applicable to deep deterministic policy gradient DDPG. Using the changes as control actions can ensure better continuity; specifically...

[0169] Discretize all three actions described above, representing them as a reward factor R, defined as follows:

[0170]

[0171] Where k'1, k'2, k'3, k'4, and k'5 are weighting coefficients, and soc tar For the target battery state of charge, ω ref This is the reference speed for the motor. For the hydrogen consumption of fuel cells, i fuel For fuel cell current, u fuel For fuel cell voltage, i li For lithium-ion power battery current, u li For the lithium-ion power battery voltage, η li For the efficiency of lithium-ion power batteries; ω ref The motor speed, which reflects the optimal acceleration or deceleration during acceleration or deceleration, is obtained from wheel speed, coefficient of adhesion, or optimal slip ratio.

[0172] During the DQN training process, the weight coefficients are continuously adjusted, and after the reward r of each optimization term can converge to an approximate order of magnitude, appropriate fine-tuning is performed according to the importance.

[0173] All variables in the state space S are divided into three types: environment state (θ) angle Road, M people ), vehicle status (soc, v) veh 、a act ) and the state of the control object (ω) motor 、P li 、P fuel );

[0174] The multi-objective control function reflects the optimal mapping relationship between the nine-dimensional state space and the three-dimensional action space, and is fitted by a multi-dimensional neural network, described by the formula:

[0175] [Δω motor ,ΔP li ,ΔP fuel ] = f(soc,v veh ,a act ,ω motor ,P li ,P fuel ,θ angle Road, M people Formula 23;

[0176] In the energy management strategy, the change in motor speed is defined as the primary control object. The motor speed is fine-tuned to maintain the lowest coefficient of adhesion or optimal slip ratio during acceleration or braking. Then, the power of the fuel cell and the lithium-ion battery are rationally adjusted to minimize energy consumption, improve economy, and achieve multi-objective control of the hybrid power system.

Claims

1. An intelligent driving fuel cell vehicle control method integrating multi-dimensional road information spatialization, characterized in that: The steps include: Step S1, obtaining the target vehicle speed output by the decision planning, and collecting vehicle status information and battery data information through sensors; Step S2: Establish a time-varying driving environment model for road driving through a road recognition convolutional neural network based on VGG16, estimate the real-time road adhesion coefficient and the optimal slip rate during driving, and calculate the maximum acceleration and deceleration allowed by the current road surface; Step S3: output the acceleration and deceleration required to reach the target vehicle speed through brain emotion learning, and obtain the accelerator and brake pedal openings through the accelerator and brake calibration table according to the maximum acceleration and deceleration limit of the current road surface to reach the target vehicle speed; Step S4, constructing a deep Q network DQN including a multi-dimensional neural network, and proposing an energy management strategy for multi-objective control; The brain emotion learning loop is input by the sensory stimulation signal G, and under the supervision of the emotion suggestion function EC, it finally outputs the required acceleration and deceleration value E; The brain emotional learning, whose component modules include thalamus, sensory cortex, amygdala and orbitofrontal cortex; The thalamus pre-processes the sensory stimulus signal and transmits it downward. The sensory cortex receives the information from the thalamus. The pre-processed stimulus signal is further analyzed and processed in the sensory cortex and then sent to the orbitofrontal cortex and amygdala respectively. The process of emotional learning in the brain mainly occurs in the amygdala, and the orbitofrontal cortex enhances or inhibits the learning of the amygdala to avoid under-learning or over-learning. During the learning process, the strongest sensory stimulation signal is sent directly from the thalamus to the amygdala; each input node in the amygdala simultaneously receives the sensory input signal G, the emotional cue function EC and the strongest sensory stimulation signal A th ; The orbitofrontal cortex receives sensory input signal G and emotional cue function EC to adjust the output; Specifically: Let sensory input signal G, the strongest sensory stimulation signal A th =max[G]; Each stimulus signal corresponds to node A in the amygdala. i , the stimulation signal includes the sensory input signal G, the emotional cue function EC and the strongest sensory stimulation signal A th , the amygdala output is A i = V i · G i , A m+1 = V m+1 · A th , i = 1, 2, ..., m Equation Six; Where m is the number of stimulus signals, V i is the weight of each node, and its adjustment law is; Among them, α is the learning rate, which affects the learning speed of the amygdala; ΔV i The monotonicity and G i The signs remain consistent; For the stimulation signal, a node is also provided in the orbitofrontal cortex to receive the stimulation signal. The stimulation signal is the sensory cortex signal and the signal from the amygdala, but the signal of the thalamus is not stimulated. The output is O i =W i ·G i Formula 9; Sensory cortical signals are sensory input signals, where W i is the weight of each node, and its adjustment law is ΔW i = β[G i ·(E' - EC)] Equation Ten; Where β is the learning rate, and β>0, E' is the output of the amygdala without thalamic signal stimulation, expressed as ΔW i It can take positive or negative values, that is, the orbitofrontal cortex can inhibit or enhance the learning process of the amygdala, and by eliminating the difference between the emotional cue function EC and the amygdala output E', the amygdala can learn towards the expected value; The vehicle longitudinal control changes the actual vehicle speed v veh and the target speed v tar When the error e is used as the input of the sensory stimulation signal G, the sensory input function is G1=k1e=k1(v veh -v tar ) Formula 12; The input signal of the emotion cue function EC is divided into two paths, one of which is the target vehicle speed v tar The actual vehicle speed v veh The other is the output of the brain emotion learning model controller, then the emotion suggestion function EC is Formula XIII; Among them, k1, k2, k3, k4, and k5 are weight adjustment coefficients; In summary, the actual output speed u1 of the brain emotion learning model is u1=k1e(V1-W1) Formula 14; Therefore, the actual output acceleration E1 is During acceleration, the actual acceleration E1 must be less than or equal to the maximum acceleration; during deceleration, due to the maximum deceleration a demax It is a negative value, that is, the actual acceleration E1 must be greater than or equal to the maximum deceleration, so that the longitudinal control with the best performance can be achieved under any road conditions; The limited acceleration a act for In step S3, the accelerator and brake calibration table is based on the vehicle dynamics model, and each accelerator or brake pedal opening corresponds to all sets of acceleration, deceleration and vehicle speed; The method for the vehicle to reach the target speed is: act Input into the accelerator and brake calibration table to obtain the corresponding accelerator or brake pedal opening to make the vehicle reach the target speed; In step S4, the dimension of the multidimensional neural network DQN is affected by the action space A. The same action value and state value are input into each dimension, and the optimal control action π of the dimension is determined according to the expected value of maximizing the cumulative reward in different dimensions. * ; The future cumulative reward in each dimension is defined as the different action values ​​Q π (s,a), and use the action value Q π (s,a) evaluates the quality of action a(t). The specific process is as follows: Among them, s is the state value, a is the action value, N is the total time length, r is the reward, and γ is the discount factor; the discount factor reflects the importance of future rewards; If the optimal action a is chosen in state s(t) * (t), then the maximum action value will be generated, which is also the optimal action value Q * (s,a), expressed as Finally, the optimal control action π in this dimension * Expressed as Among them, s0 is the initial state; The energy management strategy of the multi-objective control is specifically: using the DQN neural network to achieve synchronous learning of multiple actions and comprehensive control of the hybrid power system; in the strategy, the optimal actions of different dimensions are not independent, but under the influence of the reward factor R, the parameters of the neural network are continuously updated to achieve simultaneous fitting of the optimal actions of each dimension, so as to improve the economy of the intelligent driving fuel cell vehicle; In a multidimensional neural network, the number of neurons in the input layer and the output layer are related to the state space S and the action space A respectively; and the hidden space as a key module no longer uses a single-plane network, but a three-dimensional structure with the same dimension as the number of control actions; The state space S is defined as nine dimensions, expressed as S={soc,v veh ,a act ,ω motor ,P li ,P fuel ,θ angle ,Road,M people } Formula 20; Where soc is the battery state of charge, v veh is the actual speed, a act is the acceleration, ω motor is the motor speed, P li is the lithium-ion battery power, P fuel is the fuel cell power, θ angle is the road slope, Road is the road type, M people is the total mass of passengers; The action space A is defined as three-dimensional, expressed as Among them, Δω motor is the change in motor speed, ΔP li is the change in power of the lithium-ion battery, ΔP fuel is the change in fuel cell power; and the fine-tuning of the motor speed must be verified, and the range of this action cannot be arbitrarily selected. This range can obtain the maximum cumulative reward when compared with other ranges, which is the basis of the definition; The above three actions of change in motor speed, change in lithium-ion battery power, and change in fuel cell power are continuous variables suitable for deep deterministic policy gradient DDPG. Taking the change as the control action can ensure better continuity. Specifically, the above three actions of change in motor speed, change in lithium-ion battery power, and change in fuel cell power are all discretized and expressed as a reward factor R, which is defined as Among them, k'1, k'2, k'3, k'4, k'5 are weight coefficients, soc tar is the target battery state of charge, ω ref is the motor reference speed, is the hydrogen consumption of the fuel cell, i fuel is the fuel cell current, u fuel is the fuel cell voltage, i li is the lithium-ion power battery current, u li is the lithium-ion power battery voltage, η li is the efficiency of lithium-ion power battery; ref Reflects the motor speed that maintains the best acceleration or deceleration during acceleration or deceleration, obtained from the wheel speed, adhesion coefficient or optimal slip ratio; During the DQN training process, the weight coefficients are continuously adjusted, and after the reward r of each optimization item can converge to an approximate order of magnitude, appropriate fine-tuning is performed according to the importance; All variables in the state space S are divided into three types: environment state (θ angle 、Road、M people ), vehicle status (soc, v veh 、a act ) and the state of the controlled object (ω motor , P li , P fuel ); The multi-objective control function reflects the optimal mapping relationship between the nine-dimensional state space and the three-dimensional action space, and is fitted by a multi-dimensional neural network, which can be described by the formula: [Δω motor , ΔP li , ΔP fuel = f(soc, v veh , a act , ω motor , P li , P fuel , θ angle , Road, M people ) Formula 23; In the energy management strategy, the change in motor speed is defined as the first control object; the motor speed is fine-tuned to maintain the lowest adhesion coefficient or the best slip rate during acceleration or braking, and then the fuel cell power and lithium-ion power battery power are reasonably changed to minimize energy consumption, improve economy, and achieve multi-objective control of the hybrid system.

2. The intelligent driving fuel cell vehicle control method integrating multi-dimensional road information spatialization according to claim 1 is characterized in that: In step S1, the vehicle status information includes: wheel speed, motor speed, and motor torque; the battery data information includes: current, voltage, and battery state of charge.

3. The intelligent driving fuel cell vehicle control method integrating multi-dimensional road information spatialization according to claim 1 is characterized in that: In step S2, the time-varying driving environment model is used to improve the environmental adaptability of the longitudinal control of the vehicle, and the time-varying driving environment model includes: road type, road adhesion coefficient and optimal slip ratio.

4. The intelligent driving fuel cell vehicle control method integrating multi-dimensional road information spatialization according to claim 1 is characterized by: In the step S2, the road recognition convolutional neural network based on VGG16 is composed of 13 convolutional layers and 3 fully connected layers. When concentrated training is performed on different types of road images, different neurons in the Softmax layer correspond to different road types and are used as outputs. The specific process is as follows: the activation function applicable to the convolutional layer of the convolutional neural network is ReLU; the RGB format image is compressed to 224×224 and fed to the network, and then passes through multiple convolutional layers and maximum pooling layers of different sizes; after processing to a suitable size, the format of the data is adjusted from a three-dimensional convolutional structure to a two-dimensional fully connected structure, and then the features are gradually converted into a predefined number of types; the number of neurons in the Softmax layer is a predefined classification number, corresponding to different road types, and is used as output.

5. The intelligent driving fuel cell vehicle control method integrating multi-dimensional road information spatialization according to claim 4 is characterized in that: In step S2, the time-varying driving environment model estimates the real-time road adhesion coefficient and the optimal slip rate, and calculates the maximum acceleration and deceleration of the current road surface, as follows: The relationship between the adhesion coefficient and slip rate of different road types is established and expressed as; Where μ is the road adhesion coefficient, s slip is the slip rate, c1, c2, c3 are the parameters of each road type; After identifying different road types, the maximum road adhesion coefficient μ is obtained max And the best slip ratio During deceleration, the relationship between slip rate, vehicle speed and wheel speed is: Among them, v veh1 (t) is the actual speed v during deceleration veh1 As a function of time t, w wheel (t) is the wheel angular velocity w wheel As a function of time t, r wheel is the wheel radius; Therefore, the maximum deceleration a at the optimal slip ratio is demax The formula is v tar1 (t) is the target speed v during deceleration tar1 Function of time t; Adhesion rate of driving wheel The relationship between the adhesion coefficient μ and Among them, F X2 is the ground tangential reaction force, F Z2 is the normal reaction force of the driving wheel; During the acceleration process, the adhesion rate Approximate or equal to the adhesion coefficient μ, in order to maximize the acceleration characteristics of the car, the adhesion rate can be solved by the vehicle dynamics. When the actual speed is v veh2 ; At this time, the maximum acceleration a inmax for Among them, v tar2 (t) is the target speed v during acceleration tar2 As a function of time t, v veh2 (t) is the actual speed v during acceleration veh2 as a function of time t.

Images