A vehicle longitudinal control method based on human-like mechanism

Abstract

The application is suitable for the technical field of automatic driving, and provides a vehicle longitudinal control method based on a human simulation mechanism; on one hand, a unique feature layering input structure can realize a tolerance control mode similar to the control style of a driver, which greatly improves the consistency between the automatic driving control style and the driver control style, and further improves the experience and acceptance of the driver and passenger; on the other hand, according to the proposed driver control mechanism, the overall structure of the network is designed; the special network structure in line with the mechanism can not only improve the consistency between the automatic driving longitudinal control style and the driver, but also provide active explainability based on rules, thereby enhancing the trust of the public in the automatic driving technology.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, and in particular relates to a human-like vehicle longitudinal control method. Background Technology

[0002] Autonomous vehicles are considered a key component of intelligent transportation systems due to their advantages in safety and efficiency. To ensure the safety and comfort of autonomous vehicles during operation, control algorithms must not only fully understand the dynamic characteristics of the deployed vehicles but also incorporate human-like design to reduce passenger unfamiliarity and discomfort with autonomous driving control methods. Considering social acceptance and legal and regulatory limitations, the control algorithms of autonomous vehicles operating on open roads must be interpretable. Therefore, developing interpretable, human-like control algorithms can facilitate the safe and comfortable integration of intelligent vehicles into people's daily lives.

[0003] In response, several institutions have conducted related research. Baidu proposed an industrial-grade vehicle longitudinal dynamics control method, which includes offline modeling and online updating. Offline modeling uses a three-layer feedforward neural network, employing the driver's pedal opening and corresponding speed and acceleration as features to construct accelerator pedal and braking models respectively. Online updating refers to updating the offline learned model in real time by collecting the difference between the expected acceleration and the actual acceleration during autonomous driving mode, to cope with the challenges brought by instantaneous changes. However, the construction of this network model is not related to vehicle control, lacks theoretical analysis, has poor interpretability, and requires a fixed vehicle response delay before use. The University of Texas at Austin proposed a data-driven method to learn complex vehicle inverse dynamics models, but this also implies a high demand for computing power. Jilin University implemented automatic vehicle start-up control based on learning algorithms. Google used reinforcement learning to learn the policy used as a controller during execution. However, most of these are direct applications of machine learning methods, lacking the necessary theoretical explanations related to vehicles, failing to reflect vehicle characteristics, and are difficult to be accepted by society. This lack of interpretability turns algorithms into black boxes, making it difficult to track unexplained and fatal errors, thus reducing the reliability of the algorithms and public trust in autonomous driving. Furthermore, most existing methods are not designed to mimic human control mechanisms; while their control precision is high, their control style is less consistent with humans, leading to poor acceptance by drivers and passengers. Summary of the Invention

[0004] The purpose of this invention is to provide a human-like vehicle longitudinal control method, which aims to solve the problems mentioned in the background art.

[0005] The present invention is implemented as follows: a human-like vehicle longitudinal control method includes the following steps:

[0006] Step 1: Based on the QN cognitive system and hierarchical task structure, construct the driver's control process and mechanism of vehicle dynamics characteristics; the driver's control of vehicle drive includes three steps under the constraints of empirical rules: sensing vehicle information, pedal opening estimation and calculation (including the determination of desired acceleration, the determination of the importance of influencing factors and the determination of pedal opening), and feedback correction.

[0007] Step 2: Construct a human-like vehicle longitudinal control neural network. This network is used to calculate and correct the pedal opening, so that the longitudinal control law can quickly adapt to the new vehicle model and become consistent with the driver's control law. Its input is the desired acceleration sequence over a period of time, the current acceleration, and the current vehicle speed. This network imitates and implements the three parts of step 1: pedal opening estimation and calculation, feedback correction, and general law constraints.

[0008] The neural network comprises three sub-networks: a control response delay estimation layer, a nonlinear parameter estimation layer, and a pedal opening output layer. These three sub-networks correspond to the three sub-processes: the driver determining the desired acceleration, the driver judging the importance of influencing factors, and the driver calculating the accelerator pedal position.

[0009] In a further technical solution, step 1 includes the following specific steps:

[0010] Step 1.1: Sensing vehicle information, which is the process by which the driver obtains vehicle speed and acceleration data and performs noise reduction.

[0011] Step 1.2, Estimation and Calculation of Accelerator Pedal Opening:

[0012] This process specifically includes three steps: determining the desired acceleration, determining the importance of influencing factors, and determining the pedal position. Among these, the desired acceleration *a*... exp It is the specific value that the driver expects the vehicle to reach at a certain time t0. The process by which the driver determines the expected acceleration is represented by equation (1):

[0013] a exp =E m～p(t) [a(m)] (1)

[0014] In the formula, different accelerations a correspond to different driver acceptance m, and driver acceptance m follows a distribution function p(t) with respect to time t.

[0015] After determining the desired acceleration, the driver calculates the accelerator pedal position. The driver's calculation of the accelerator pedal position is only related to three influencing factors: desired acceleration, current vehicle speed, and current acceleration. The driver first judges the importance of these three factors based on the currently perceived vehicle information, which is the determination of the importance of the influencing factors. Then, the driver comprehensively considers the importance of the three factors and finally calculates the value of the accelerator pedal position, which is the determination of the pedal position. In the above two steps, the estimation of the importance of the influencing factors is obtained through the operation of a human neural network. The human neural network is represented as a multivalued nonlinear function φ(·), whose input is desired acceleration, current acceleration, and current vehicle speed. Then, the driver's calculation process of the accelerator pedal position can be described as shown in equation (2):

[0016]

[0017] In the formula, p represents the pedal position as a percentage; a exp a now v now These represent the desired acceleration, current acceleration, and current vehicle speed, respectively; φ(a exp ,a now ,v now ) [i] Let k be the i-th output of the multivalued function, representing the driver's emphasis on the i-th influencing factor. This will be abbreviated as k below. i .

[0018] Step 1.3, Feedback Correction:

[0019] After estimating the pedal position, the driver conducts multiple trials and adjusts the estimation of the vehicle's longitudinal driving characteristics based on the results to improve control performance. Mathematically, this is reflected in the adjustment and final determination of the internal weights of the multivalued function φ(·).

[0020] Step 1.4, General Constraints: Steps 1.1 to 1.3 cover the portion of the driver's control of vehicle dynamics specific to a particular vehicle model; general rules, also known as empirical rules, are analyzed and extracted from the perspective of three influencing factors, yielding the following rules:

[0021]

[0022] In summary, the driver's control of vehicle drive involves three steps under the constraints of empirical rules: sensing vehicle information, estimating and calculating pedal opening (including determining the desired acceleration and pedal opening), and feedback correction.

[0023] In a further technical solution, in step 2, the error function of the network is as follows:

[0024]

[0025] In the formula, p is the accelerator pedal position calculated by the model, p r The pedal opening position given by the driver, and p and p r All figures are expressed as percentages.

[0026] A further technical solution involves designing a feature-layered input structure in step 2, deploying the control response delay estimation layer before some nodes of the main network. This enables the model to simulate human-like autonomous estimation of control response delay and tolerance control. The feature-layered input structure inputs the features required by the network at different levels. It is deployed before the nonlinear parameter estimation layer and the pedal opening output layer. The features input into the main network (nonlinear parameter estimation layer + pedal opening output layer) are divided into two parts. The current acceleration and current vehicle speed are directly input into the nonlinear parameter estimation layer and the pedal opening output layer for training. The other part is input into the control response delay estimation layer, and after pre-learning and preprocessing in the control response delay estimation layer, it is then fed into the nonlinear parameter estimation layer and the pedal opening output layer.

[0027] The input to the control response delay estimation layer is the desired acceleration sequence, and the structure is a fully connected feedforward network to reproduce the process by which the driver determines the desired acceleration. The process of solving for the mathematical expectation of acceleration according to the acceptance distribution function p(t) is simplified to a weighted calculation of different acceleration values, and the corresponding weights are learned through a fully connected layer. The acceptance is denoted as m. i The output of this layer is then obtained as follows:

[0028]

[0029] In the formula, a i Let m be the acceleration value at time i. i The value of acceleration corresponds to the level of acceptance, where n represents the state that the vehicle needs to reach within a time period of n seconds, where n is 1.25 seconds.

[0030] A further technical solution involves the following step: In step 2, the inputs to the nonlinear parameter estimation layer are the desired acceleration, the current acceleration, and the current vehicle speed. The desired acceleration is provided by the control response delay estimation layer, and the output is an estimate of the weights of the three influencing factors, namely k1, k2, and k3 in formula (3). The hidden layer uses a Gaussian radial basis function, defined as follows:

[0031]

[0032] In the formula, x p Given the input feature vector, c iThe coordinates of the center points of the radial basis functions are given, and the distance between them is measured using the 2-norm. The output layer uses the Rectified Linear Function (ReLU) as the activation function. The overall structure and principle of this layer are the same as those of the standard RBF network, but its output should satisfy the range constraints described by formula (3) to conform to the general laws of driver drive control.

[0033] In a further technical solution, in step 2, the pedal opening output layer is used to mimic the process by which the driver obtains the pedal opening through weighted calculation, and finally calculates and determines the accelerator pedal position (percentage). Its input is the weights k1, k2, and k3 of three influencing factors, which are also the outputs of the nonlinear parameter estimation layer. In order to reproduce the calculation process of formula (2) and make the network structure more consistent with the vehicle's delay response characteristics and obtain clearer rule-based interpretability, it adopts a special fixed weight design, the weights of which are the three influencing factors a exp a now and v now This means that its weights are only related to the current input and do not change with training. In other words, the structural design of the pedal opening output layer takes into account the influence of the current vehicle state on future motion, reflecting an adaptation to the delay response characteristics.

[0034] This invention provides a human-like vehicle longitudinal control method. On one hand, through a unique feature-layered input structure, it can achieve a tolerance control mode similar to the driver's control style, which greatly improves the consistency between the autonomous driving control style and the driver's control style, thereby improving the experience and acceptance of passengers. On the other hand, based on the proposed driver control mechanism, the overall network structure is designed. This special network structure that conforms to the mechanism can not only improve the consistency between the autonomous driving longitudinal control style and the driver, but also provide rule-based active interpretability, thereby enhancing public trust in autonomous driving technology. Attached Figure Description

[0035] Figure 1 A flowchart illustrating the driver control mechanism;

[0036] Figure 2 A flowchart illustrating a human-like vehicle longitudinal control method provided in an embodiment of the present invention;

[0037] Figure 3 This is a diagram of a neural network structure for longitudinal control of a vehicle, based on human-like mechanisms. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0039] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0040] An embodiment of the present invention provides a human-like vehicle longitudinal control method, comprising the following steps:

[0041] Step 1: Based on the QN cognitive system and hierarchical task structure, construct the driver's control process and mechanism of vehicle dynamics characteristics; the driver's control of vehicle drive includes three steps under the constraints of empirical rules: sensing vehicle information, pedal opening estimation and calculation (including the determination of desired acceleration, the determination of the importance of influencing factors and the determination of pedal opening), and feedback correction.

[0042] Step 2: Construct a human-like vehicle longitudinal control neural network. This network is used to calculate and correct the pedal opening, so that the longitudinal control law can quickly adapt to the new vehicle model and become consistent with the driver's control law. Its input is the desired acceleration sequence over a period of time, the current acceleration, and the current vehicle speed. This network imitates and implements the three parts of pedal opening estimation and calculation, feedback correction, and general law constraints in Step 1.

[0043] The neural network comprises three sub-networks: a control response delay estimation layer, a nonlinear parameter estimation layer, and a pedal opening output layer. These three sub-networks correspond to the three sub-processes: the driver determining the desired acceleration, the driver judging the importance of influencing factors, and the driver calculating the accelerator pedal position.

[0044] like Figure 1 As shown, in a preferred embodiment of the present invention, the driver's adaptation and control of the vehicle's longitudinal dynamics is a process combining empirical rules with specific instances, that is, an online adjustment process of offline experience. Empirical rules are general principles followed by the driver when performing dynamic control; they are offline rules that do not change with vehicle model. Specific instances refer to the specific controlled vehicle. The process of combining the two involves, under the guidance of offline empirical rules, the driver making online adaptive adjustments based on feedback obtained from the human cognitive system and human interaction with the outside world, for a specific vehicle, ultimately achieving an accurate estimate and control of its dynamic performance.

[0045] The human process of handling driving tasks, including adapting to new vehicles, can be divided into three parts: information perception, decision-making, and correction based on results. This method introduces empirical rules as constraints and proposes a three-stage process for drivers to control vehicle actuation: perceiving vehicle information, estimating and calculating accelerator pedal opening, and feedback correction. These will be described in detail below.

[0046] Step 1.1, Sensing vehicle information:

[0047] Perceiving vehicle information means acquiring vehicle dynamics data related to acceleration strategies. Limited by human perception, drivers can only vaguely sense speed and acceleration information, and due to attentional limitations, they subconsciously ignore noise information unrelated to driving. Therefore, perceiving vehicle information is the process by which the driver acquires vehicle speed and acceleration data and then denoises it.

[0048] Step 1.2, Estimation and Calculation of Accelerator Pedal Opening (Tolerance Control Characteristics):

[0049] The estimation and calculation of accelerator pedal opening primarily reflects the driver's drive control mechanism and their understanding of vehicle characteristics. This process involves the driver determining the desired pedal opening based on the desired acceleration and perceived information, specifically including three steps: determining the desired acceleration, determining the importance of influencing factors, and determining the pedal position. The process of determining the desired acceleration reflects the driver's tolerance control characteristics.

[0050] Expected acceleration a exp This refers to the specific value the driver expects the vehicle to reach at a given time t0. However, it is difficult for humans to precisely control a vehicle to reach an exact acceleration value at a specific point in time. Therefore, when the vehicle's acceleration reaches a value near time t0... exp When the value is around 100°, the driver will accept it with a lower degree of acceptance, which is the specific manifestation of the tolerance control characteristic in this process. The process by which the driver determines the desired acceleration is expressed by equation (1):

[0051] a exp =E m～p(t) [a(m)] (1)

[0052] In the formula, different accelerations a correspond to different driver acceptance m, and driver acceptance m follows a distribution function p(t) with respect to time t.

[0053] Furthermore, the process by which the driver determines the desired acceleration includes their estimation of the vehicle control response delay. In the above formula, point a has the highest acceptance. iThis represents the state the driver believes the vehicle is most likely to reach, and the corresponding time represents the driver's estimate of the control response delay. Instead of directly using the point with the highest acceptance as the desired acceleration input, the driver comprehensively considers and tracks various possible future scenarios. This is precisely the embodiment and application of tolerance control mode in vehicle longitudinal control.

[0054] After determining the desired acceleration, the driver calculates the accelerator pedal position. The driver's calculation of the accelerator pedal position is only related to three influencing factors: desired acceleration, current vehicle speed, and current acceleration. The driver first judges the importance of these three factors based on the currently perceived vehicle information. Then, the driver comprehensively considers the importance of the three factors and finally calculates the value of the accelerator pedal position. In the above two steps, the estimation of the weights is obtained through the operation of a complex human neural network. If the human neural network is represented as a multivalued nonlinear function φ(·), and its input is desired acceleration, current acceleration, and current vehicle speed, then the driver's calculation process of the accelerator pedal position can be described as shown in equation (2):

[0055]

[0056] In the formula, p represents the pedal position as a percentage; a exp a now v now These represent the desired acceleration, current acceleration, and current vehicle speed, respectively; φ(a exp ,a now ,v now ) [i] Let k be the i-th output of the multivalued function, representing the driver's emphasis on the i-th influencing factor. This will be abbreviated as k below. i .

[0057] Step 1.3, Feedback Correction:

[0058] After estimating the pedal position, the driver continuously conducts trials and adjusts the estimation of the vehicle's longitudinal driving characteristics based on the results to achieve more comfortable control—this is the feedback correction process. In biology, this process manifests as the formation of conditioned reflexes; that is, humans adjust the connections between neurons through continuous experimentation and feedback, strengthening connections that lead to expected results and weakening connections that lead to erroneous results. This ultimately forms a sophisticated human neural network for controlling vehicle dynamics, i.e., the formation of corresponding conditioned reflexes, which mathematically manifests as the adjustment and final determination of the weights within the multivalued function φ(·).

[0059] General Constraints: The above describes the portion of the driver's control of vehicle dynamics specific to a particular vehicle model. The following section introduces general rules of control, also known as rules of thumb. This paper analyzes and extracts these general rules from the perspective of three influencing factors, yielding the following patterns:

[0060]

[0061] In summary, the driver's control of vehicle drive involves three steps under the constraints of empirical rules: sensing vehicle information, estimating and calculating pedal opening (including determining the desired acceleration and pedal opening), and feedback correction.

[0062] like Figure 2 and 3 As shown, in a preferred embodiment of the present invention, in step 2, the human-like vehicle longitudinal control neural network possesses rule-based active interpretability because of its structure (e.g., Figure 3 The diagram shown is designed based on the formula proposed in step 1.

[0063] The control response delay estimation layer is designed based on a feature-layered input structure. By reproducing the structure in equation (1), rule-based interpretability and mimicking human tolerance control patterns are achieved. The nonlinear parameter estimation layer approximates the multi-valued nonlinear function in equation (2) through a radial basis function neural network, thereby estimating the weights k1, k2, and k3. The pedal opening output layer is designed based on equation (2). The structure of this layer mimics the process by which the driver determines the accelerator pedal position, and its operating mechanism is consistent with the rules in equation (2), thus possessing proactive rule-based interpretability. Notably, this layer uses a fixed weight design and selects weights according to the calculation rules in equation (2) to ultimately output the pedal position. Furthermore, the network introduces general laws as constraints in the nonlinear parameter estimation layer. The above structural design is based on the aforementioned driver behavior mechanism, making this network particularly suitable for the field of human-like vehicle control and providing it with proactive rule-based interpretability.

[0064] The network error function is shown in the following equation:

[0065]

[0066] In the formula, p and p r These represent the accelerator pedal position calculated by the model and the pedal opening position given by the driver (both expressed as percentages). The three sub-networks—control response delay estimation layer, nonlinear parameter estimation layer, and pedal opening output layer—will be described below.

[0067] Control response delay estimation layer:

[0068] By designing a hierarchical feature input structure, the model can autonomously estimate control response delays and achieve tolerance control, much like a human. The hierarchical feature input structure means that the features required by the network are input at different levels. The advantage of this design is that a portion of the data can be preprocessed and the knowledge from this data can be learned for subsequent learning and prediction by the network.

[0069] The feature-layered input structure is reflected in the design of the response delay estimation layer, such as... Figure 3 As shown, it is deployed before the main network (nonlinear parameter estimation layer and pedal opening output layer). The features input into this network are divided into two parts: one part (current acceleration and current vehicle speed) is directly input into the main network for training, and the other part is input into the control response delay estimation layer. After pre-learning and preprocessing in the control response delay estimation layer, it is then fed into the main network.

[0070] The input to the control response delay estimation layer is the desired acceleration sequence, and its structure is a fully connected feedforward network to reproduce the process by which the driver determines the desired acceleration. In this method, the process of solving for the mathematical expectation of acceleration according to the acceptance distribution function p(t) is simplified to a weighted calculation of different acceleration values, and the corresponding weights are learned through a fully connected layer. The acceptance is denoted as m. i The output of this layer is then obtained as follows:

[0071]

[0072] In the formula, a i Let m be the acceleration value at time i. i The acceptable level corresponds to this acceleration value, where n represents the state the vehicle needs to reach within a time period of n seconds. In this method, n is set to 1.25 seconds.

[0073] The control response delay estimation layer estimates the vehicle's control response delay and performs tolerance control by adjusting the weights of each input, much like a human would. Instead of directly tracking the point with the highest acceptance as the desired acceleration, the network outputs a weighted desired acceleration value. This incorporates the driver's tolerance control mode into the model, as the model no longer tracks the precise acceleration value at a given moment, but rather tracks an acceleration sequence through weighted calculation.

[0074] Nonlinear parameter estimation layer:

[0075] Its inputs are the desired acceleration, the current acceleration, and the current vehicle speed, where the desired acceleration is given by the control response delay estimation layer, and the output is the estimated weights of the three influencing factors, namely k1, k2, and k3 in formula (3). The hidden layer uses Gaussian radial basis functions, which are defined as follows:

[0076]

[0077] In the formula, x p Given the input feature vector, c i The coordinates of the center points of the radial basis functions are given, and the distance between them is measured using the 2-norm. The output layer uses the Rectified Linear Function (ReLU) as the activation function. The overall structure and principle of this layer are the same as those of the standard RBF network, but its output should satisfy the range constraints described by formula (3) to conform to the general laws of driver drive control.

[0078] Pedal opening output layer:

[0079] The pedal opening output layer ultimately calculates and determines the accelerator pedal position (percentage), mimicking the process by which a driver obtains the pedal opening through weighted calculation. Its input consists of the weights k1, k2, and k3 of three influencing factors, which are also the output of the nonlinear parameter estimation layer. To reproduce the calculation process of formula (2), and to make the network structure more consistent with the vehicle's delay response characteristics and achieve clearer rule-based interpretability, it employs a special fixed-weight design, where the weights are the three influencing factors a... exp a now and v now This means that its weights are only related to the current input and do not change with training. In other words, the structural design of the pedal opening output layer takes into account the influence of the current vehicle state on future motion, reflecting an adaptation to the delay response characteristics.

[0080] By linking the control response delay estimation layer, the nonlinear parameter estimation layer, and the pedal opening output layer, the driver's drive control mechanism is reproduced, and the network structure design reflects tolerance control characteristics and delay response characteristics. Specifically:

[0081] The driver's determination of the desired acceleration and the tolerance control characteristics therein are reproduced by the control response delay estimation layer;

[0082] The driver's estimation of the weights of the three influencing factors and the general patterns therein are reproduced by the nonlinear parameter estimation layer;

[0083] The driver's determination of pedal opening and his understanding of its delay response characteristics are reproduced by the pedal opening output layer;

[0084] The process by which the driver adjusts their control based on feedback to quickly adapt to a new vehicle is achieved through feedback training of a neural network.

[0085] In summary, the proposed network model fully considers the specific application scenario of vehicle drive control. The design of each sub-module corresponds to the sub-processes in the driving mechanism, which gives it rule-based active interpretability in the overall network structure.

[0086] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A human-like vehicle longitudinal control method, characterized in that, Includes the following steps: Step 1: Based on the QN cognitive system and hierarchical task structure, construct the driver's control process of vehicle dynamics characteristics; the driver's control of vehicle drive includes three steps under the constraints of empirical rules: sensing vehicle information, pedal opening estimation and calculation, and feedback correction. Among them, pedal opening estimation and calculation includes determining the desired acceleration, determining the importance of influencing factors, and determining the pedal opening; Step 2: Construct a human-like vehicle longitudinal control neural network. This neural network is used to calculate and correct the pedal opening so that the longitudinal control law can quickly adapt to the new vehicle model and become consistent with the driver's control law. Its input is the expected acceleration sequence, current acceleration, and current vehicle speed over a period of time. This neural network imitates and implements the three parts of pedal opening estimation and calculation, feedback correction, and general law constraints in Step 1. The neural network comprises three sub-networks: a control response delay estimation layer, a nonlinear parameter estimation layer, and a pedal opening output layer; the three sub-networks correspond to the three sub-processes of the driver determining the desired acceleration, the driver judging the importance of influencing factors, and the driver calculating the accelerator pedal position.

2. The human-like vehicle longitudinal control method according to claim 1, characterized in that, Step 1 includes the following specific steps: Step 1.1: Sensing vehicle information, i.e., the process by which the driver acquires vehicle speed and acceleration data and performs noise reduction on it; Step 1.2, Estimation and Calculation of Accelerator Pedal Opening: Specifically, it includes three steps: determining the desired acceleration, determining the importance of influencing factors, and determining the pedal position; among which, the desired acceleration a exp It is the specific value that the driver expects the vehicle to reach at a certain time t0. The process by which the driver determines the expected acceleration is expressed by equation (1): a exp =E m～p(t) [a(m)] (1) In the formula, different accelerations a correspond to different driver acceptance m, and the driver acceptance m follows a distribution function p(t) with respect to time t; After determining the desired acceleration, the driver calculates the accelerator pedal position. The driver's calculation of the accelerator pedal position is only related to three influencing factors: desired acceleration, current vehicle speed, and current acceleration. The driver first judges the importance of these three factors based on the vehicle information currently perceived, then comprehensively considers the importance of the three factors, and finally calculates the value of the accelerator pedal position, which is the determination of the pedal position. In the above two steps, the estimation of the importance of the influencing factors is obtained through the operation of a human neural network. The human neural network is represented as a multivalued nonlinear function φ(·), whose input is the desired acceleration, the current acceleration, and the current vehicle speed. The driver's calculation process for the accelerator pedal position is described as shown in Equation (2): In the formula, p represents the pedal position as a percentage; a exp a now and v now These represent the desired acceleration, current acceleration, and current vehicle speed, respectively; φ(a exp ,a now ,v now ) [i] The i-th output of the multivalued function represents the driver's level of attention to the i-th influencing factor, denoted as ki; Step 1.3, Feedback Correction: After estimating the pedal position, the driver conducts multiple trials and adjusts the estimation of the vehicle's longitudinal drive characteristics based on the trial results to improve the control effect. Mathematically, this is reflected in the adjustment and final determination of the internal weights of the multivalued function φ(·). Step 1.4, General Constraints: Steps 1.1 to 1.3 cover the portion of the driver's control of vehicle dynamics specific to a particular vehicle model; general rules, also known as empirical rules, are analyzed and extracted from the perspective of three influencing factors, yielding the following rules:

3. The human-like vehicle longitudinal control method according to claim 2, characterized in that, In step 2, the error function of the network is as follows: In the formula, p is the accelerator pedal position calculated by the model, p r The pedal opening position given by the driver, and p and p r All figures are expressed as percentages.

4. The human-like vehicle longitudinal control method according to claim 3, characterized in that, In step 2, a feature-layered input structure is designed and deployed before the nonlinear parameter estimation layer and the pedal opening output layer. The features input into the main network are divided into two parts. The current acceleration and current vehicle speed are directly input into the nonlinear parameter estimation layer and the pedal opening output layer for training. The other part is input into the control response delay estimation layer. After pre-learning and preprocessing in the control response delay estimation layer, it is then sent to the nonlinear parameter estimation layer and the pedal opening output layer. The input to the control response delay estimation layer is the desired acceleration sequence, and its structure is a fully connected feedforward network to reproduce the process of the driver determining the desired acceleration in step 1.

2. The process of solving the mathematical expectation of acceleration according to the acceptance distribution function p(t) is simplified to a weighted calculation of different acceleration values, and the corresponding weights are learned through a fully connected layer. The acceptance is denoted as m. i The output of this layer is as follows: In the formula, a i Let m be the acceleration value at time i. i The value of acceleration corresponds to the level of acceptance, where n represents the state that the vehicle needs to reach within a time period of n seconds, where n is 1.25 seconds.

5. The human-like vehicle longitudinal control method according to claim 4, characterized in that, In step 2, the inputs to the nonlinear parameter estimation layer are the desired acceleration, the current acceleration, and the current vehicle speed. The desired acceleration is given by the control response delay estimation layer, and the output is the estimated weights of the three influencing factors, namely k1, k2, and k3 in formula (3). The hidden layer uses Gaussian radial basis functions, which are defined as follows: In the formula, x p Given the input feature vector, c i The coordinates of the center point of the radial basis function are given, and the distance between them is measured by the 2-norm. The output layer uses a linear rectified function as the activation function, and it should satisfy the range constraint described by formula (3) when outputting.

6. The human-like vehicle longitudinal control method according to claim 5, characterized in that, In step 2, the pedal opening output layer simulates the process by which a driver obtains the pedal opening through weighted calculation, and ultimately calculates and determines the accelerator pedal position. Its input is the weights k1, k2, and k3 of three influencing factors, which is also the output of the nonlinear parameter estimation layer. It adopts a fixed weight design, and its weights are the three influencing factors a. exp a now and v now .

Images