Task-oriented self-adaptive hybrid control system for drive-by-wire chassis of unmanned vehicle

Abstract

The invention relates to the technical field of chassis control, and discloses an unmanned vehicle drive-by-wire chassis hybrid control system oriented to task self-adaption, which comprises an environment recognition module, a sensing regression module, a task analysis module, a hybrid self-adaption control module and a servo control module, wherein the task analysis module dynamically reconstructs a target function and a constraint condition, and optimizes and coordinates the priority and task weight of multiple tasks; the hybrid adaptive control module fuses the time sequence prediction capability of offline learning of the long and short-term memory network and the real-time optimization effect of online kinematics, dynamically optimizes the control weight according to the short-term prediction error of the two controllers, generates an optimal control instruction, and achieves the dynamic balance of offline learning and online control. And the task execution precision and the persistent operation capability of a complex scene are improved. The adaptability of the unmanned vehicle to a complex environment and a narrow indoor environment is enhanced, the unmanned vehicle is low in energy consumption, high in flexibility and compact in size, the performance guarantee is improved, and more efficient movement and operation are achieved.

Description

Technical Field

[0001] This invention relates to the field of chassis control technology, and more specifically, to a task-adaptive driverless vehicle drive-by-wire chassis hybrid control system. Background Technology

[0002] With the rapid development of mechanical automation and intelligent technologies, traditional mechanical chassis are gradually transforming into more powerful intelligent chassis. As an advanced stage of intelligent chassis development, distributed drive-by-wire chassis can achieve independent drive, steering, and braking of all four wheels, offering significant advantages such as precise transmission and a modular, scalable architecture. These advantages can meet the automation and intelligent upgrade needs of mechanical equipment in complex environments, such as logistics distribution, warehouse management, and agricultural harvesting.

[0003] In complex and confined indoor environments such as hospitals and warehouses, chassis are required to have low energy consumption, high flexibility, and compact size to adapt to complex terrain and space constraints, enabling more efficient movement and operation. However, current research on drive-by-wire chassis still faces many challenges in terms of endurance and mission execution capabilities, making it difficult to simultaneously meet the high-precision and sustained operational requirements of positioning, perception, prediction, and path planning in complex scenarios.

[0004] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0005] To address the problems in related technologies, this invention proposes a task-adaptive unmanned vehicle drive-by-wire chassis hybrid control system to overcome the aforementioned technical problems in existing related technologies.

[0006] Therefore, the specific technical solution adopted by the present invention is as follows:

[0007] A task-oriented adaptive driverless vehicle drive-by-wire chassis hybrid control system includes:

[0008] The environmental recognition module is used to acquire external environmental data through sensors on the chassis of the unmanned vehicle, and to extract features from the external environmental data to obtain environmental feature information;

[0009] The perception regression module is used to receive environmental feature information from the environment recognition module and map the environmental feature information into an environmental state vector to obtain the environmental coupling matrix of the unmanned vehicle's chassis position, attitude, speed, acceleration and relative distance to obstacles in the global or local map.

[0010] The task parsing module receives task instructions, decomposes them into several sub-tasks and assigns initial task weights. Combined with the environment coupling matrix of the perceptual regression module, it constructs a variable quadratic programming model composed of objective function and dual environmental and task adjustment constraints to dynamically reconstruct the objective function and constraints, and optimize and coordinate the priority and task weights of multiple tasks.

[0011] The hybrid adaptive control module receives the objective function and constraints dynamically reconstructed by the task parsing module, integrates the temporal prediction capability of offline learning from the long short-term memory network with the real-time optimization effect of online kinematics, and dynamically optimizes the control weights based on the short-term prediction errors of the two controllers. Based on the optimized control weights, the best control command is generated to achieve a dynamic balance between offline learning and online control, thereby improving the task execution accuracy and persistent operation capability in complex scenarios.

[0012] The servo control module receives the optimal control commands from the hybrid adaptive control module, drives the independent control of the four wheels of the unmanned vehicle, and feeds back the motion state information of the unmanned vehicle chassis to the hybrid adaptive control module to calculate the error between the actual control effect and the expected target, dynamically adjust the controller parameters, and optimize the control accuracy.

[0013] The environmental recognition module, perception regression module, task parsing module, hybrid adaptive control module, and servo control module form a closed loop to enhance the adaptability of the unmanned vehicle to complex and narrow indoor environments, and improve the task execution accuracy and long-term operation capability in complex scenarios.

[0014] Furthermore, the environmental recognition module includes LiDAR, millimeter-wave radar, front and rear depth cameras, and inertial sensors;

[0015] Among them, LiDAR is used to provide high-precision, high-resolution 3D environmental geometry contours and identify the location of obstacles in the environment;

[0016] Millimeter-wave radar is used to measure the range and radial velocity of targets.

[0017] Front and rear depth cameras are used to simultaneously output RGB images for recognition, so as to assign semantic labels to key areas of road boundaries, pedestrians and vehicles;

[0018] Inertial sensors are used to output real-time corrected and compensated triaxial angular velocity, acceleration, and triaxial magnetometer data to determine the orientation and attitude of the equipment.

[0019] Furthermore, the task parsing module receives task instructions from the upper-level system through a touch-screen display. Upon receiving the task instructions, it decomposes them into several sub-tasks and assigns initial task weights to construct a task mapping matrix. Based on the environment coupling matrix of the perceptual regression module, it constructs a variable quadratic programming model composed of objective function and dual adjustment constraints of environment and task. This variable quadratic programming model updates task weights and environmental feature vectors according to the feedback of task performance indicators and environmental status, driving the dynamic adjustment of objective function and constraint conditions.

[0020] Furthermore, the expression for the objective function in the quadratic programming model is:

[0021]

[0022]

[0023]

[0024]

[0025]

[0026]

[0027]

[0028] In the formula, Let T be the decision variable, and T be the transpose of a matrix or vector. This is a quadratic weighted matrix of the control variables. Let the vector be the coefficients of the first-order term of the control quantity. This is the sensitivity matrix of the error to the control. The environment coupling matrix, Here is the error weight matrix. This is the task mapping matrix. For the control quantity weight matrix, For bias compensation term, Represents a diagonal matrix function. For the nth mapping function, Let n be the environment vector. Let i be the weight of the i-th error term. Let the weight of the j-th control variable be . For a set of tasks, To control the task under period t The weight, This is the task basis matrix.

[0029] Furthermore, the constraints in the variable quadratic programming model include control boundaries and linear constraints, where the expressions for the control boundaries and linear constraints are:

[0030]

[0031]

[0032] In the formula, This means that the following conditions must be met. , These are the lower and upper limits of the control quantity, respectively. Based on the basic constraint matrix, Based on the basic constraint vector, , These are the environment coefficient vector and the task coefficient vector, respectively.

[0033] Furthermore, the hybrid adaptive control module includes a deep temporal offline modeling adaptive controller, a real-time kinematic constraint dynamic optimization controller, and a dual-path error feedback adaptive weight fusion unit;

[0034] Among them, the deep temporal offline modeling adaptive controller collects multi-dimensional historical state data of the vehicle, learns the dynamic characteristics of the chassis system and establishes a dynamic model after preprocessing, and outputs the predictive control quantities of the four-wheel steering angle and driving torque at the current moment; after the task is executed, the meta-control unit built into the controller adjusts the output parameters based on the historical task performance indicators to improve the controller's generalization ability across task scenarios.

[0035] The real-time kinematic constraint dynamic optimization controller uses optimization algorithms to solve for the optimal control action in each control cycle based on real-time sensor data. Its initial value is provided by the predictive control quantity. The controller updates the vehicle kinematic matrix parameters in real time and incorporates chassis kinematic constraints and the current state during the solution process. With the goal of minimizing the tracking error between the expected trajectory and the actual trajectory, it achieves dynamic optimization control of vehicle attitude and path.

[0036] The dual-path error feedback adaptive weight fusion unit dynamically generates adaptive control weights based on the short-term prediction errors of the two controllers in each control cycle. It outputs the optimal control quantity through linear weighted fusion and sends this optimal control quantity as the final optimal control command to the servo control module to drive the vehicle chassis to perform corresponding actions.

[0037] Furthermore, the expression for the deep temporal offline modeling adaptive controller is:

[0038]

[0039]

[0040]

[0041]

[0042]

[0043]

[0044] The expression for the meta-control unit in the deep temporal offline modeling adaptive controller is:

[0045]

[0046]

[0047] In the formula, These represent the forget gate, input gate, cell state, and output gate of the deep temporal offline modeling adaptive controller in the t-th control cycle, respectively. The inputs are the chassis state and environmental state for the t-th control cycle. The weights are calculated for the forget gate, input gate, cell state, and output gate, respectively. These are the bias parameters for the forget gate, input gate, cell state, and output gate, respectively. The output is the hidden state during the t-th control cycle. Here, is the sigmoid function, and tanh is the hyperbolic tangent function. To model the output values ​​of the adaptive controller offline for deep temporal data, Let this be the end position of the chassis in the t-th control cycle. Let be the control input value in the t-th control cycle. The output value of the depth-time offline modeling adaptive controller is the result of correction by the meta-controller. Indicates the current task type. Indicates task The scaling factor is controlled during the k-th execution. Indicates task The bias compensation term during the k-th execution. , These represent the incrementing and decrementing adjustment step sizes, respectively, when control performance is insufficient. Indicates task The task performance metrics at the k-th execution time. For the task The reference threshold, For the task Performance fluctuation tolerance.

[0048] Furthermore, the expression for the real-time kinematically constrained dynamic optimization controller is:

[0049]

[0050]

[0051]

[0052]

[0053] In the formula, Given a 2x2 real matrix to be optimized, To represent the 2-norm, It is a real number matrix containing historical location information. This is a control quantity matrix that includes historical control information. A real matrix Position error estimation, For position tracking error, For the target location, The control quantity output by the controller is dynamically optimized based on real-time kinematic constraints.

[0054] Furthermore, the expression for the dual-path error feedback adaptive weight fusion unit is:

[0055]

[0056]

[0057]

[0058] In the formula, To linearly mix the output values ​​of the two controllers according to their weights, , These are the fusion weight parameters of the deep temporal offline modeling adaptive controller and the real-time kinematic constraint dynamic optimization controller during the control period t. , These represent the short-term prediction errors of the corrected deep temporal offline modeling adaptive controller and the real-time kinematic constraint dynamic optimization controller, respectively. To avoid numerically robust constants with a denominator of zero.

[0059] Furthermore, after receiving the optimal control command from the hybrid adaptive control module, the servo control module transforms the target position from the sensor coordinate system to the chassis coordinate system. It then drives the steering motor, rotation motor, and coupling through the vehicle controller to achieve independent and precise control of the four wheels. Simultaneously, it feeds back the chassis motion status information to the hybrid adaptive control module, calculates the error between the actual control effect and the expected target, and dynamically adjusts the parameters of the deep temporal offline modeling adaptive controller and the real-time kinematic constraint dynamic optimization controller to optimize control accuracy, thereby forming a closed-loop control.

[0060] The beneficial effects of this invention are as follows:

[0061] This invention integrates real-time environmental data from multiple sensors, combining the temporal prediction capabilities of LSTM offline learning with the real-time optimization effect of online kinematics. Under the action of dynamically reconstructed objective functions and constraints, the dual-path error feedback adaptive weight fusion unit adaptively adjusts the weights based on the short-term prediction errors of the two controllers, achieving a dynamic balance between offline learning and online control. This effectively improves the vehicle's execution accuracy for tasks and trajectories, significantly enhances the vehicle's motion performance and interactive capabilities, and improves the task execution accuracy and sustained operation capability in complex scenarios. Compared to traditional drive-by-wire chassis technology, this invention enhances the adaptability of unmanned vehicles to complex and confined indoor environments, possesses low energy consumption, high flexibility, and compact size, improves performance assurance, and achieves more efficient movement and operation. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a schematic diagram of a task-oriented adaptive unmanned vehicle drive-by-wire chassis hybrid control system according to an embodiment of the present invention;

[0064] Figure 2 This is a structural block diagram of the environment recognition module in a task-oriented adaptive unmanned vehicle drive-by-wire chassis hybrid control system according to an embodiment of the present invention;

[0065] Figure 3 This is a control schematic diagram of a task-oriented adaptive unmanned vehicle drive-by-wire chassis hybrid control system according to an embodiment of the present invention. Detailed Implementation

[0066] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0067] According to an embodiment of the present invention, a task-adaptive unmanned vehicle drive-by-wire chassis hybrid control system is provided.

[0068] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figures 1-3 As shown, the task-adaptive unmanned vehicle drive-by-wire chassis hybrid control system according to an embodiment of the present invention includes:

[0069] The environmental recognition module is used to acquire external environmental data through multiple sensors installed on the chassis of the unmanned vehicle, and to extract features from the external environmental data to obtain environmental feature information.

[0070] Specifically, the environmental recognition module includes a lidar, a millimeter-wave radar, front and rear depth cameras, and an inertial sensor (i.e., an IMU sensor). Each sensor is mounted on the chassis body. The millimeter-wave radar is connected to the industrial control computer via a CAN-to-USB converter. The front depth camera and lidar are connected to the router via a network port and are also connected to the battery. The outputs of the other sensors are connected to the industrial control computer on the chassis via USB. The industrial control computer is connected to the router and the battery respectively.

[0071] The environmental recognition module provides high-precision, high-resolution 3D environmental geometric contours through LiDAR to identify the location of obstacles in the environment; millimeter-wave radar accurately measures the distance and radial velocity of targets; front and rear depth cameras simultaneously output RGB images for recognition, assigning semantic labels to key areas such as road boundaries, pedestrians, and vehicles, while also linking the dynamic targets locked by millimeter-wave radar with the image detection results; the IMU integrates a three-axis gyroscope, accelerometer, and magnetometer, with the gyroscope and accelerometer outputting real-time corrected and compensated three-axis angular velocity and acceleration, and the magnetometer outputting three-axis magnetometer data to determine the device's orientation and attitude.

[0072] The perception regression module is used to receive environmental feature information from the environment recognition module and map the environmental feature information into an environmental state vector to obtain the environmental coupling matrix of the unmanned vehicle's chassis position, attitude, speed, acceleration, and relative distance to obstacles in the global or local map.

[0073] Specifically, the perception regression module receives environmental feature data from the sensor and maps the environmental identification results into an environmental state vector. It can calculate the environmental coupling matrix of the vehicle in real time in the global or local map, including the chassis position, attitude, speed, acceleration and relative distance to obstacles. .

[0074] The task parsing module receives task instructions, decomposes them into several sub-tasks, and assigns initial task weights to represent their priorities. Combined with the environment coupling matrix of the perceptual regression module, it constructs a variable quadratic programming model (QP) composed of objective function and dual environmental and task adjustment constraints to dynamically reconstruct the objective function and constraints, and optimize and coordinate the priorities and task weights of multiple tasks.

[0075] Specifically, the task parsing module connects to the industrial control computer via an HDMI interface and is powered by a battery. It receives task instructions from the upper-level system through a touch-screen display, including tasks such as emergency obstacle avoidance, high-precision docking, and constant-speed cruising. Upon receiving a task instruction, it decomposes it into several sub-tasks and assigns initial task weights to construct a task mapping matrix. Environmental coupling matrix based on the perceptual regression module A variable quadratic programming model is constructed, consisting of an objective function and dual adjustment constraints related to the environment and task. During task execution, the variable quadratic programming model adjusts the task weight coefficients based on the feedback of task performance indicators and environmental status. Mapping function with environmental features Updates are performed to drive dynamic adjustments to the objective function and constraints.

[0076] Specifically, the expression for the objective function in the quadratic programming model is:

[0077]

[0078]

[0079]

[0080]

[0081]

[0082]

[0083]

[0084] In the formula, Let T be the decision variable, representing the control quantity that needs to be optimized, and let T be the transpose of a matrix or vector. This is a quadratic weighted matrix of control quantities, which comprehensively reflects the joint adjustment of the error term by the environment coupling matrix and the task mapping matrix. The vector of coefficients for the first-order terms of the control variables describes the linear influence of the environment and task on the error bias, while the quadratic term of the objective function represents the second-order terms. Used to smooth control actions and ensure system stability, primary term This involves introducing a linear gradient for performance metrics such as tracking error to reduce deviation from the target state. This is the sensitivity matrix of the error to the control. The environment coupling matrix is ​​formed by the environment vector. Each component is mapped by the function Function composition, Let be an n-dimensional real vector space, representing the set of column vectors consisting of all n real components. This is the error weight matrix, used to adjust the weights of the error terms. This is the task mapping matrix. This is the control weight matrix, used to adjust the weights of the control variables. For error compensation, Represents a diagonal matrix function. For the nth mapping function, Let n be the environment vector. Let i be the weight of the i-th error term. Let the weight of the j-th control variable be . For a set of tasks, To control the task under period t The weight, This is the task basis matrix.

[0085] Specifically, the constraints in the variable quadratic programming model include control boundaries and linear constraints, where the expressions for the control boundaries and linear constraints are:

[0086]

[0087]

[0088] In the formula, That is, subject to, means that the following conditions must be met. , These are the lower and upper limits of the control quantity, used to restrict the range of the control quantity. The basic constraint matrix covers linear constraint relationships such as velocity, acceleration, and steering angle. Define the static constraint values ​​of the system using the basic constraint vector. , These are the environment coefficient vector and the task coefficient vector, respectively. The constraint vector is adjusted based on the environment coupling matrix and the task mapping matrix.

[0089] The hybrid adaptive control module receives the objective function and constraints dynamically reconstructed by the task parsing module. It integrates the temporal prediction capability of offline learning from Long Short-Term Memory (LSTM) networks with the real-time optimization effect of online kinematics. In each control cycle, it dynamically optimizes the control weights based on the short-term prediction errors of the two controllers. Based on the optimized control weights, it generates the best control command through linear weighting to achieve a dynamic balance between offline learning and online control, thereby improving the task execution accuracy and persistent operation capability in complex scenarios.

[0090] Specifically, the hybrid adaptive control module includes a deep temporal offline modeling adaptive controller (i.e., an LSTM offline learning controller), a real-time kinematic constraint dynamic optimization controller (i.e., an online kinematic optimization controller), and a dual-path error feedback adaptive weight fusion unit (i.e., an error ratio fusion unit).

[0091] The deep temporal offline modeling adaptive controller collects multi-dimensional historical state data of the vehicle, preprocesses it to learn the dynamic characteristics of the chassis system and establish a dynamic model, and outputs predictive control quantities of the four-wheel steering angle and driving torque at the current moment. After task execution, the controller's built-in meta-control unit adjusts the output parameters based on historical task performance indicators to improve the controller's generalization ability across task scenarios; specifically including:

[0092] Historical state sequence data of the vehicle chassis, including position, speed, and steering angle, is collected. After preprocessing such as normalization and denoising, a Long Short-Term Memory (LSTM) network is used to learn the dynamic characteristics of the chassis offline and construct a data-driven dynamic model. The trained network parameters implicitly encode the nonlinear mapping relationship between the state and the control quantity, enabling it to output predictive control quantities of the four-wheel steering angle and driving torque at the current moment based on current and historical state information. After the task is executed, the controller's built-in meta-control unit adjusts the output parameters based on historical task performance metrics. Specifically, when the task performance metrics fall below a set threshold, the scaling factor is updated. This allows for adjusting the control gain, enabling adaptive parameter updates across different tasks and enhancing the controller's generalization ability across cross-task scenarios.

[0093] The expression for the adaptive controller in deep temporal offline modeling is:

[0094]

[0095]

[0096]

[0097]

[0098]

[0099]

[0100] In the formula, These represent the forget gate, input gate, cell state, and output gate of the deep temporal offline modeling adaptive controller in the t-th control cycle, respectively. The inputs are the chassis state and environmental state for the t-th control cycle. The weights are calculated for the forget gate, input gate, cell state, and output gate, respectively. These are the bias parameters for the forget gate, input gate, cell state, and output gate, respectively. The hidden state output for the t-th control cycle contains information from the past to the present. The sigmoid function is an activation function that maps a variable to the range 0-1, while the hyperbolic tangent function is an activation function that maps values ​​to the range -1-1. To model the output values ​​of the adaptive controller offline for deep temporal data, Let this be the end position of the chassis in the t-th control cycle. Let be the control input value in the t-th control cycle; in order to use the LSTM network as the controller, the network needs to be trained offline using the collected dataset, with the previous activation values. and end position As input, the current activation value As the output value of the LSTM offline learning controller.

[0101] The expression for the meta-control unit in the deep temporal offline modeling adaptive controller is:

[0102]

[0103] Parameter update mechanism:

[0104]

[0105] In the formula, The output value of the depth-time offline modeling adaptive controller is the result of correction by the meta-controller. Indicates the current task type, belonging to the task set. , Indicates task The control scaling factor at the k-th execution is used to adjust the control quantity. Indicates task The bias compensation term during the k-th execution. and These are adjustable parameters for specific tasks, used to balance the responsiveness and stability of the LSTM control output. The initial value is set to... , =0 means that the bias term will be reset to 0 at the end of each task cycle. , These represent the incrementing and decrementing adjustment step sizes, respectively, when control performance is insufficient. Indicates task The task performance metrics at the k-th execution time. For the task The reference threshold, For the task Performance fluctuation tolerance.

[0106] A real-time kinematic constraint dynamic optimization controller, based on real-time sensor data, uses an optimization algorithm to solve for the optimal control action in each control cycle. Its initial values ​​are provided by predictive control variables. This controller updates the vehicle's kinematic matrix parameters in real time and incorporates chassis kinematic constraints and the current state during the solution process. Its objective is to minimize the tracking error between the desired trajectory and the actual trajectory, thereby achieving dynamic optimization control of vehicle attitude and path. Specifically, it includes:

[0107] In each control cycle t, the controller first collects real-time sensor data to obtain the current position. Control quantity compared to the previous moment And update the historical data matrix. and Then, the first optimization problem is solved to update the kinematic matrix. Based on the updated K and the target position at the next time step. Solve the second optimization problem to obtain the current optimal control quantity. The output is sent to the weight fusion unit and enters the next control cycle. The initial values ​​for the numerical iterations in this solution process are the predictive control variables provided by the deep temporal offline modeling adaptive controller. To accelerate convergence;

[0108] The expression for the real-time kinematically constrained dynamic optimization controller is:

[0109]

[0110]

[0111]

[0112]

[0113] In the formula, This represents a 2x2 real matrix to be optimized, used to establish a linear mapping relationship between the two-dimensional control input and the two-dimensional position output. Describing the 2-norm, It is a real matrix containing 2D position vectors of the most recent 5 historical moments (i.e., a real matrix containing historical position information). It is a real matrix containing the 2D control vectors of the most recent 5 historical moments (i.e., a control quantity matrix containing historical control information). A real matrix Position error estimation, For position tracking error, It is a 2-dimensional position vector (i.e., the target position). This is the 2D control vector output by the controller dynamically optimized by real-time kinematic constraints.

[0114] The dual-path error feedback adaptive weighted fusion unit dynamically generates adaptive control weights based on the short-term prediction errors of the two controllers within each control cycle. It outputs the optimal control quantity through linear weighted fusion and sends this optimal control quantity as the final optimal control command to the servo control module to drive the vehicle chassis to perform corresponding actions. Specifically, it includes:

[0115] In each control cycle t, the fusion unit acquires the corrected LSTM prediction value. With real-time kinematic optimization value Calculate the short-term prediction errors of the two. and The error is derived from the statistical value of the deviation between the recent predicted output and the execution result of the two controllers. The weights are calculated using the error weighting formula and then linearly weighted and fused to generate the hybrid control quantity. The optimal control command for the current moment is sent to the servo control module to drive the vehicle chassis to perform corresponding actions. When the short-term error of the deep temporal offline modeling adaptive controller is small and the prediction is stable, the system adaptively increases its weight to improve control efficiency. When the output of the deep temporal offline modeling adaptive controller is abnormal or the error exceeds the safety threshold, the weight of the real-time kinematic constraint dynamic optimization controller is automatically increased to ensure the safety of vehicle driving and the stability of the system.

[0116] The expression for the dual-path error feedback adaptive weight fusion unit is:

[0117]

[0118]

[0119]

[0120] In the formula, The output values ​​of the two controllers are linearly mixed according to weights and then output to the servo control module as control commands. , These are the fusion weight parameters for the deep temporal offline modeling adaptive controller and the real-time kinematic constraint dynamic optimization controller during the control period t, used to adjust the contribution values ​​of the LSTM controller and the kinematic controller. , These represent the short-term prediction errors of the corrected deep temporal offline modeling adaptive controller and the real-time kinematic constraint dynamic optimization controller, respectively, with non-negative scalar values. To avoid numerically robust constants with a denominator of zero.

[0121] The servo control module receives the optimal control commands from the hybrid adaptive control module, drives the independent control of the four wheels of the unmanned vehicle, and feeds back the motion state information of the unmanned vehicle chassis to the hybrid adaptive control module to calculate the error between the actual control effect and the expected target, dynamically adjust the controller parameters, and optimize the control accuracy.

[0122] Specifically, such as Figure 3 As shown, the servo control module includes the vehicle control unit (VCU), steering motor, rotation motor, and coupling, and adopts PCAN2. + A professional-grade CAN-USB adapter converts CAN bus signals into USB interface signals, enabling the industrial control computer to connect to the chassis CAN bus for data exchange and chassis movement control; the VCU connects via PCAN2. + The VCU acquires the control CAN signal transmitted by the hybrid adaptive control module, precisely controls the steering motor, rotation motor and coupling of the four-wheel drive four-turn drive-by-wire chassis, and provides real-time feedback on the current motion status information; the VCU is equipped with a remote control transceiver terminal, and is also connected to the headlights, front and rear turn signals and brake lights respectively;

[0123] Specifically, after receiving the optimal control command from the hybrid adaptive control module, the servo control module transforms the target position from the sensor coordinate system to the chassis coordinate system. It then drives the steering motor, rotation motor, and coupling via the vehicle controller to achieve independent control of all four wheels. Simultaneously, it feeds back the chassis motion status information to the hybrid adaptive control module, calculates the error between the actual control effect and the expected target, and dynamically adjusts the parameters of the deep temporal offline modeling adaptive controller and the real-time kinematic constraint dynamic optimization controller to optimize control accuracy and form a closed-loop control. This includes:

[0124] The servo control module receives the optimal control commands from the hybrid adaptive control module. Then, the coordinate system is transformed from the sensor-based coordinate system to the vehicle coordinate system with the chassis center as the origin. This transformation is achieved through a pre-calibrated rotation and translation matrix. The transformed command is sent to the vehicle control unit (VCU). The VCU analyzes the target steering angle and driving torque of each wheel according to the command, and then sends drive signals to the corresponding steering motor, rotation motor and coupling through the CAN bus to achieve independent and precise control of the four wheels.

[0125] Meanwhile, the servo module feeds back the chassis's real-time position, attitude, and other motion states to the hybrid adaptive control module, which calculates the tracking error between the actual position and the target position. This drives the adaptive updating of the controller parameters; the error signal is used, on the one hand, to update the matrix of the kinematic controller online. On the other hand, the driving element control unit updates its scaling factor and bias, forming a complete closed-loop control, thereby continuously optimizing the overall control accuracy of the system.

[0126] The environmental recognition module, perception regression module, task parsing module, hybrid adaptive control module, and servo control module form a closed loop to enhance the adaptability of the unmanned vehicle to complex and narrow indoor environments, and improve the task execution accuracy and long-term operation capability in complex scenarios.

[0127] 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, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A task-adaptive unmanned vehicle drive-by-wire chassis hybrid control system, characterized in that, include: The environmental recognition module is used to acquire external environmental data through sensors on the chassis of the unmanned vehicle, and to extract features from the external environmental data to obtain environmental feature information; The perception regression module is used to receive environmental feature information from the environment recognition module and map the environmental feature information into an environmental state vector to obtain the environmental coupling matrix of the unmanned vehicle's chassis position, attitude, speed, acceleration and relative distance to obstacles in the global or local map. The task parsing module receives task instructions, decomposes them into several sub-tasks and assigns initial task weights. Combined with the environment coupling matrix of the perceptual regression module, it constructs a variable quadratic programming model composed of objective function and dual environmental and task adjustment constraints to dynamically reconstruct the objective function and constraints, and optimize and coordinate the priority and task weights of multiple tasks. The hybrid adaptive control module receives the objective function and constraints dynamically reconstructed by the task parsing module, integrates the temporal prediction capability of offline learning from the long short-term memory network with the real-time optimization effect of online kinematics, and dynamically optimizes the control weights based on the short-term prediction errors of the two controllers. Based on the optimized control weights, the best control command is generated to achieve a dynamic balance between offline learning and online control, thereby improving the task execution accuracy and persistent operation capability in complex scenarios. The servo control module receives the optimal control commands from the hybrid adaptive control module, drives the independent control of the four wheels of the unmanned vehicle, and feeds back the motion state information of the unmanned vehicle chassis to the hybrid adaptive control module to calculate the error between the actual control effect and the expected target, dynamically adjust the controller parameters, and optimize the control accuracy. The environmental recognition module, perception regression module, task parsing module, hybrid adaptive control module, and servo control module form a closed loop to enhance the adaptability of the unmanned vehicle to complex and narrow indoor environments, and improve the task execution accuracy and long-term operation capability in complex scenarios.

2. The task-oriented adaptive unmanned vehicle drive-by-wire chassis hybrid control system according to claim 1, characterized in that, The environmental recognition module includes a lidar, a millimeter-wave radar, front and rear depth cameras, and an inertial sensor. The lidar is used to provide high-precision, high-resolution 3D environmental geometry contours and identify the location of obstacles in the environment. The millimeter-wave radar is used to measure the target's distance and radial velocity; The front and rear depth cameras are used to synchronously output RGB images for recognition, so as to assign semantic labels to key areas of road boundaries, pedestrians and vehicles. The inertial sensor is used to output real-time corrected and compensated triaxial angular velocity, acceleration, and triaxial magnetometer data to determine the orientation and attitude of the device.

3. The task-oriented adaptive unmanned vehicle drive-by-wire chassis hybrid control system according to claim 1, characterized in that, The task parsing module receives task instructions from the upper-level system through a touch screen. Upon receiving the task instructions, it decomposes them into several sub-tasks and assigns initial task weights to construct a task mapping matrix. Based on the environment coupling matrix of the perceptual regression module, it constructs a variable quadratic programming model composed of objective function and dual adjustment constraints of environment and task. This variable quadratic programming model updates task weights and environmental feature vectors according to the feedback of task performance indicators and environmental status, driving the dynamic adjustment of objective function and constraint conditions.

4. The task-oriented adaptive unmanned vehicle drive-by-wire chassis hybrid control system according to claim 1, characterized in that, The objective function in the variable quadratic programming model is expressed as: ； ； ； ； ； ； ； In the formula, Let T be the decision variable, and T be the transpose of a matrix or vector. This is a quadratic weighted matrix of the control variables. Let the vector be the coefficients of the first-order term of the control quantity. This is the sensitivity matrix of the error to the control. The environment coupling matrix, Here is the error weight matrix. This is the task mapping matrix. For the control quantity weight matrix, For bias compensation term, Represents a diagonal matrix function. For the nth mapping function, Let n be the environment vector. Let i be the weight of the i-th error term. Let the weight of the j-th control variable be . For a set of tasks, To control the task under period t The weight, This is the task basis matrix.

5. A task-oriented adaptive unmanned vehicle drive-by-wire chassis hybrid control system according to claim 4, characterized in that, The constraints in the variable quadratic programming model include control boundaries and linear constraints. The expressions for the control boundaries and linear constraints are as follows: ； ； In the formula, This means that the following conditions must be met. , These are the lower and upper limits of the control quantity, respectively. Based on the basic constraint matrix, Based on the basic constraint vector, , These are the environment coefficient vector and the task coefficient vector, respectively.

6. A task-oriented adaptive unmanned vehicle drive-by-wire chassis hybrid control system according to claim 1, characterized in that, The hybrid adaptive control module includes a deep temporal offline modeling adaptive controller, a real-time kinematic constraint dynamic optimization controller, and a dual-path error feedback adaptive weight fusion unit. The deep temporal offline modeling adaptive controller collects multi-dimensional historical state data of the vehicle, learns the dynamic characteristics of the chassis system and establishes a dynamic model after preprocessing, and outputs the predicted control quantities of the four-wheel steering angle and driving torque at the current moment. After the task is executed, the meta-control unit built into the controller adjusts the output parameters based on the historical task performance indicators to improve the controller's generalization ability across task scenarios. The real-time kinematic constraint dynamic optimization controller, based on real-time sensor data, uses an optimization algorithm to solve for the optimal control action in each control cycle. Its initial value is provided by the predictive control quantity. The controller updates the vehicle kinematic matrix parameters in real time and incorporates chassis kinematic constraints and the current state during the solution process. With the goal of minimizing the tracking error between the expected trajectory and the actual trajectory, it achieves dynamic optimization control of vehicle attitude and path. The dual-path error feedback adaptive weight fusion unit dynamically generates adaptive control weights based on the short-term prediction errors of the two controllers in each control cycle, outputs the optimal control quantity through linear weighted fusion, and sends the optimal control quantity as the final optimal control command to the servo control module to drive the vehicle chassis to perform corresponding actions.

7. A task-oriented adaptive unmanned vehicle drive-by-wire chassis hybrid control system according to claim 6, characterized in that, The expression for the adaptive controller in deep temporal offline modeling is: ； ； ； ； ； ； The expression for the meta-control unit in the deep temporal offline modeling adaptive controller is: ； ； In the formula, These represent the forget gate, input gate, cell state, and output gate of the deep temporal offline modeling adaptive controller in the t-th control cycle, respectively. The inputs are the chassis state and environmental state for the t-th control cycle. The weights are calculated for the forget gate, input gate, cell state, and output gate, respectively. These are the bias parameters for the forget gate, input gate, cell state, and output gate, respectively. The output is the hidden state during the t-th control cycle. Here, is the sigmoid function, and tanh is the hyperbolic tangent function. To model the output values ​​of the adaptive controller offline for deep temporal data, Let this be the end position of the chassis in the t-th control cycle. Let be the control input value in the t-th control cycle. The output value of the depth-time offline modeling adaptive controller is the result of correction by the meta-controller. Indicates the current task type. Indicates task The scaling factor is controlled during the k-th execution. Indicates task The bias compensation term during the k-th execution. , These represent the incrementing and decrementing adjustment step sizes, respectively, when control performance is insufficient. Indicates task The task performance metrics at the k-th execution time. For the task The reference threshold, For the task Performance fluctuation tolerance.

8. A task-oriented adaptive unmanned vehicle drive-by-wire chassis hybrid control system according to claim 7, characterized in that, The expression for the real-time kinematically constrained dynamic optimization controller is: ； ； ； ； In the formula, Given a 2x2 real matrix to be optimized, To represent the 2-norm, It is a real number matrix containing historical location information. This is a control quantity matrix that includes historical control information. A real matrix Position error estimation, For position tracking error, For the target location, The control quantity output by the controller is dynamically optimized based on real-time kinematic constraints.

9. A task-oriented adaptive unmanned vehicle drive-by-wire chassis hybrid control system according to claim 8, characterized in that, The expression for the dual-path error feedback adaptive weight fusion unit is: ； ； ； In the formula, To linearly mix the output values ​​of the two controllers according to their weights, , These are the fusion weight parameters of the deep temporal offline modeling adaptive controller and the real-time kinematic constraint dynamic optimization controller during the control period t. , These represent the short-term prediction errors of the corrected deep temporal offline modeling adaptive controller and the real-time kinematic constraint dynamic optimization controller, respectively. To avoid numerically robust constants with a denominator of zero.

10. A task-oriented adaptive unmanned vehicle drive-by-wire chassis hybrid control system according to claim 1, characterized in that, After receiving the optimal control command from the hybrid adaptive control module, the servo control module transforms the target position from the sensor coordinate system to the chassis coordinate system. It then drives the steering motor, rotation motor, and coupling through the vehicle controller to achieve independent control of the four wheels. Simultaneously, it feeds back the chassis motion status information to the hybrid adaptive control module, calculates the error between the actual control effect and the expected target, and dynamically adjusts the parameters of the deep temporal offline modeling adaptive controller and the real-time kinematic constraint dynamic optimization controller to optimize control accuracy and form a closed-loop control.

Images