A method for interference learning prediction suitable for a cooperative heterogeneous unmanned aerial vehicle and robot dog platform

Abstract

The present application relates to the field of machine learning, in particular to a kind of interference learning prediction method suitable for collaborative heterogeneous unmanned aerial vehicle and robot dog platform. By force analysis on unmanned aerial vehicle and robot dog, the dynamics model containing lift and external disturbance term is built for unmanned aerial vehicle, and the floating base model reflecting the coupling characteristics of base 6 degrees of freedom and leg joint is built for robot dog, and the dynamics state update of two types of platforms is completed under the action of known external force. Continuous time-varying disturbances such as cubic spline smoothing disturbance and Fourier superposition disturbance are added to the platform, and a disturbance prediction neural network is built. The historical operation data of the platform is used as the input feature, and the actual disturbance value calculated by the dynamics model is used as the label to train the network, so that the network has accurate disturbance prediction capability. The present application effectively improves the prediction accuracy of external time-varying disturbance of unmanned aerial vehicle and robot dog, and simplifies the implementation process of multi-platform interference learning.

Description

Technical Field

[0001] This invention relates to the field of machine learning, and more specifically to a method for interference learning and prediction applicable to collaborative heterogeneous drones and robot dog platforms. Background Technology

[0002] Today, drones and robotic dogs are rapidly developing, finding widespread and in-depth applications in numerous fields such as inspection and exploration, material transportation, and operations in complex environments. This application background and development trend places higher demands on the operational stability, interference adaptability, and control performance of multiple platforms in complex scenarios. However, in practical applications, drones are susceptible to interference such as airflow, and robotic dogs often face interference from uneven ground and friction variations. These interferences alter the platform's dynamic characteristics, causing its trajectory to deviate from expectations and even leading to safety issues. Traditional methods often focus on single platforms or specific interference types, lacking a universal interference learning and prediction scheme across platforms. This makes it difficult to efficiently adapt to the needs of different platforms in diverse interference environments, limiting the performance and application expansion of various types of actuators.

[0003] In the field of multi-platform anti-interference, how to improve the versatility and accuracy of interference prediction and control through data-driven approaches and cross-platform adaptation is a key issue that needs to be addressed. Existing methods either focus on a single anti-interference dimension or are limited to specific platforms, making it difficult to meet the needs of multiple platforms in complex interference environments. Patent publication number CN120567535A, entitled "An Anti-interference Method, System, Device, and Medium for Rotary-Wing Unmanned Aerial Vehicles," achieves anti-interference of the UAV communication link by determining the disturbance rhythm factor of the communication link, inserting camouflaged data packets, and encrypting them. However, this method does not employ data-driven learning methods such as neural networks, and therefore cannot actively learn the time-varying patterns and characteristics of interference. It lacks adaptability to continuously time-varying non-communication interference and is only applicable to a single UAV platform, lacking cross-platform scalability.

[0004] The patent publication number CN119512185A, entitled "Anti-interference learning control method for unmanned aerial vehicles with predetermined time for periodic attitude tracking tasks", achieves anti-interference control for periodic attitude tracking of unmanned aerial vehicles by designing a predetermined time disturbance observer and a fully saturated repetitive learning law. However, this method is only designed for unmanned aerial vehicles and lacks a cross-platform general disturbance learning framework.

[0005] The patent publication number CN116300882A, entitled "A Path Planning Method and System for Substation Equipment Inspection Based on Robot Dog", realizes robot dog inspection path planning through lidar point cloud processing and Euclidean clustering algorithm. However, this method focuses on trajectory planning and equipment identification in the inspection scenario, without involving the learning and prediction of external interference. The core technology direction is environmental perception and path optimization, rather than interference learning. Summary of the Invention

[0006] To address the shortcomings of existing methods, such as being limited to a single platform, lacking interference learning capabilities, and exhibiting poor cross-platform adaptability, this invention provides an interference learning and prediction method applicable to collaborative heterogeneous UAV and robot dog platforms. This method involves constructing a dynamic model incorporating external forces for the UAV and establishing a floating base model for the robot dog that reflects the coupling between the base and joints, enabling accurate updates to the dynamic states of both platforms. Based on these models, continuous time-varying interference is added, and an interference prediction neural network is built. The network is trained using historical platform state data as input and actual interference values ​​calculated from the dynamic model as labels. Ultimately, this enables the network to accurately predict external time-varying interference for both types of platforms.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] An interference learning and prediction method applicable to collaborative heterogeneous UAVs and robot dog platforms includes the following steps:

[0009] (1) Dynamic modeling of the mission platform: Force analysis is performed on the UAV and the robot dog, and a dynamic model of the UAV and a floating base model of the robot dog are established. The dynamic state update of the two types of platforms is completed under the known external forces.

[0010] (2) External interference modeling, interference network construction and interference feedback application: Based on the dynamic model constructed in step 1, continuous time-varying interference is added to the task platform, and an interference prediction neural network is built. The network is trained with the platform's historical operating data as input features and the actual interference value as a label, so that it has the ability to predict interference. The platform's historical operating data includes speed, acceleration, motor speed and torque.

[0011] Furthermore, step (1) includes establishing dynamic equations for the heterogeneous actuators of the drone and the robot dog, clarifying the mapping relationship between the output force / torque of the actuator and the motion state of the device. The heterogeneous actuators of the drone and the robot dog are the drone propeller and the robot dog joint drive unit. The motion state of the device is position, speed and attitude.

[0012] Furthermore, step (1) is as follows:

[0013] Step 1.1: Based on the dynamic characteristics of the quadcopter UAV, establish the following dynamic model: ;

[0014] ;

[0015] in, The location of the drone. The derivative of the drone's position is its velocity. For the quality of drones, For the speed of the drone, The derivative of the drone's velocity is the drone's acceleration. It is the acceleration due to gravity. Let R be the unit vector along the z-axis in the world coordinate system, with the superscript T indicating the transpose of the vector or matrix, and R be the rotation matrix of the UAV. The lift generated by the drone's propellers The continuous time-varying external interference force experienced by the drone body; Let be the derivative of the rotation matrix of the UAV. The angular velocity of the drone, For the angular acceleration of the drone, Let be the rotational inertia matrix of the UAV, and τ be the external torque acting on the UAV. Angular velocity of the drone The corresponding antisymmetric matrix;

[0016] Step 1.2: Establishment of the floating base dynamic model for the robot dog: The core feature of the floating base model is that the base has no fixed constraints, possesses 6 degrees of freedom, and couples with the leg joint motion to form a multibody dynamic system. The floating base system consists of one floating base and four legs. The forces acting on the base and legs need to be analyzed separately, clarifying the force transmission path from the feet to the leg joints and then to the base. The dynamic equations of the floating base system are established using the Newton-Euler recursive method, in two steps: "forward recursion" and "backward recursion," ultimately forming a complete dynamic model coupling the 6 degrees of freedom of the base and the 12 degrees of freedom of the leg joints.

[0017] ;

[0018] ;

[0019] in, For the overall configuration space of the mechanical dog, Its degrees of freedom are 12 + 6 = 18. The six degrees of freedom representing the body's floating base. This represents the 12 degrees of freedom of a joint. For generalized position, velocity, and acceleration vectors, For the generalized mass matrix, For Cocteau force terms, For gravity, The joint torque provided by the motors configured for the robot dog Here is the kinematic pair matrix, which describes the direction of the joint degrees of freedom. For foot power, For the number of contacting legs, For the contact constraint Jacobian matrix, The external disturbance forces and torques experienced by the robot dog's base. The spatial transformation matrix transforms the disturbance forces on the base from the world coordinate system to the rigid body local coordinate system.

[0020] Furthermore, the 6 degrees of freedom include 3 translational degrees of freedom and 3 rotational degrees of freedom, and each of the 4 legs contains 3 rotational joints. The forward recursion is used to calculate the inertial force / torque of each link, and the reverse recursion is used to calculate the joint torque and the force on the base.

[0021] Furthermore, step (2) includes:

[0022] Step 2.1: The cubic spline smoothing disturbance force is used as a mathematical model to simulate the continuous time-varying disturbances experienced by the mission platform during operation. Its core feature is that it generates a smooth and continuous disturbance force curve through piecewise quadratic polynomial interpolation, which can effectively simulate the gradual disturbances in the actual environment.

[0023] Step 2.2: The core function of the interference prediction network is to learn from the historical operation data of the task platform and predict external time-varying interference in real time, so as to provide interference prior information for subsequent trajectory planning.

[0024] Step 2.3: After obtaining the network-predicted interference value in the UAV mission platform, the interference estimate needs to be used as a real-time feedback signal in the baseline controller.

[0025] Furthermore, in step 2.1, the process of constructing the interference force mainly includes three stages:

[0026] ① Determine the time interval of the interference force, and uniformly select several time control points within this interval;

[0027] ② Assign a corresponding interference force amplitude to each time control point. These amplitudes are randomly generated based on the intensity range of the actual interference scenario.

[0028] ③ Between two adjacent time control points, a cubic polynomial function is used to perform interpolation; this ensures that the interpolation curve satisfies the continuity of the second derivative throughout the entire time interval, and finally obtains a continuous time-varying disturbance force that conforms to physical characteristics.

[0029] Furthermore, in step 2.1, the specific form of the interference force is described as follows:

[0030] ;

[0031] ;

[0032] ;

[0033] ;

[0034] in To add interference forces to the mission platform, interference forces in three directions. They are independent of each other. This represents the time-varying disturbance values ​​in its three directions. This represents the cubic spline disturbance force, where the parameter... For time; parameters The time control point vector is uniformly distributed over the time interval. Internal; Parameter For amplitude control point vectors, This indicates that the amplitude control point is within the interference amplitude range. The internal order follows a uniform distribution.

[0035] Furthermore, in step 2.2, the design of the interference prediction network revolves around "temporal feature capture - multi-source data fusion - dual-platform adaptation";

[0036] For UAV platforms: The input features are selected from key operational data that can directly reflect the impact of interference, specifically a three-dimensional velocity sequence and a three-dimensional expected control force sequence with nine consecutive time steps. The two types of sequences are spliced ​​together to form a 54-dimensional time-series input feature vector.

[0037] ;

[0038] ;

[0039] ;

[0040] in, The net thrust of the drone is the thrust of the propellers minus the weight of the drone itself. It is a velocity sequence. These are velocity vectors at nine different time points. For thrust sequence, These are the net thrust vectors at nine different time points. To integrate the network input sequence of velocity and net thrust, Represents a neural network for predicting interference. Its weights and biases, etc. These represent the interference values ​​in the three directions predicted by the final network.

[0041] For the robot dog platform: the three-dimensional velocity of the base and the rotational speed sequence of each joint over nine consecutive time steps are spliced ​​together with the three-dimensional desired control torque sequence of each joint as the input to the network;

[0042] ;

[0043] ;

[0044] ;

[0045] in, It is a velocity sequence. These are the base velocity and joint rotation speed vectors at nine different time points. It is a torque sequence. These are the torque vectors at nine different times. To integrate the network input sequences of velocity and torque, Represents a neural network for predicting interference. Its weights and biases, etc. These represent the interference values ​​in the three directions predicted by the final network.

[0046] The network structure adopts a lightweight architecture of "embedding layer - 1D convolutional layer - global pooling layer - regression layer". The embedding layer maps the input to a low-dimensional feature space to reduce computation. The 1D convolutional layer extracts temporal features, the global pooling layer compresses redundant features while retaining key information, and the regression layer finally outputs a 3D interference prediction value. Finally, during the training phase, the true interference value calculated by the platform dynamics model is used as the label. The loss function, through iterative optimization of network weights, enables real-time and accurate prediction of interference;

[0047] The loss function is defined as follows:

[0048] .

[0049] Furthermore, step 2.3 is detailed as follows:

[0050] In the generation of the desired control force of the controller, an interference compensation mechanism is introduced. Based on the ideal control force corresponding to the desired motion of the UAV, the inverse compensation amount of the interference term is directly superimposed. That is, the interference estimate obtained by the interference learning prediction method is incorporated into the calculation process of the desired control force in an equal and inverse manner, so that the desired control force output by the controller can offset the influence of the interference on the actual force of the UAV, thereby ensuring the consistency between the UAV's motion state and the desired state.

[0051] ;

[0052] in, The desired location for the drone. For the desired speed of the drone, For the quality of drones, Given a first diagonal matrix where all elements are greater than 0, Given a second diagonal matrix in which all elements are greater than 0.

[0053] The advantages of this invention compared to the prior art are:

[0054] This invention discloses an interference learning and prediction method applicable to collaborative heterogeneous UAVs and robot dog platforms. It models the differentiated dynamic characteristics of UAVs and robot dogs separately: a dynamic model incorporating external forces is constructed for the UAV, and a floating base model reflecting the coupling between the base and joints is established for the robot dog. This enables accurate updates to the dynamic states of both platforms, avoiding errors caused by adapting a single model to multiple platforms. Based on the above models, continuous time-varying interference is added, and an interference prediction neural network is built. The network is trained by directly learning the actual interference value as a label using historical platform operation data as input. The interference prediction error can converge to a smaller range, ensuring the stability and accuracy of interference prediction.

[0055] This method directly achieves interference learning through a combination of "dynamic model + data-driven approach," which simplifies the data processing flow and avoids computational latency. Furthermore, during network training, only the input features (such as the motor speed of a drone and the joint torque of a robot dog) need to be adjusted for different platforms, without changing the network topology, enabling rapid cross-platform adaptation. This effectively improves the operational stability of multiple platforms under interference environments. Attached Figure Description

[0056] Figure 1 The flowchart shows the design of the interference learning prediction method applicable to collaborative heterogeneous UAVs and robot dog platforms of the present invention.

[0057] Figure 2 The image shows the UAV hardware used for method verification in a specific embodiment.

[0058] Figure 3 The diagram shows the hardware of the robot dog used for method verification in a specific embodiment.

[0059] Figure 4 The image shows the interference prediction diagram of the quadcopter drone used for method verification in a specific embodiment.

[0060] Figure 5 The image shows a 3D model of the quadcopter drone trajectory tracking used for method verification in a specific embodiment.

[0061] Figure 6 This is a comparison chart of the trajectory tracking position error of the quadcopter drone used for method verification in a specific embodiment.

[0062] Figure 7 The image shows the interference prediction graph of the robot dog used for method verification in a specific embodiment. Detailed Implementation

[0063] An interference learning and prediction method applicable to collaborative heterogeneous UAVs and robot dog platforms includes the following steps:

[0064] (1) Dynamic modeling of the mission platform: Force analysis is performed on the UAV and the robot dog, and a dynamic model of the UAV and a floating base model of the robot dog are established. The dynamic state update of the two types of platforms is completed under the known external forces.

[0065] (2) External interference modeling, interference network construction and interference feedback application: Based on the dynamic model constructed in step 1, continuous time-varying interference is added to the task platform, and an interference prediction neural network is built. The network is trained with the platform's historical operating data as input features and the actual interference value as a label, so that it has the ability to predict interference. The platform's historical operating data includes speed, acceleration, motor speed and torque.

[0066] Furthermore, step (1) includes establishing dynamic equations for the heterogeneous actuators of the drone and the robot dog, clarifying the mapping relationship between the output force / torque of the actuator and the motion state of the device. The heterogeneous actuators of the drone and the robot dog are the drone propeller and the robot dog joint drive unit. The motion state of the device is position, speed and attitude.

[0067] Furthermore, step (1) is as follows:

[0068] Step 1.1: Based on the dynamic characteristics of the quadcopter UAV, establish the following dynamic model: ;

[0069] ;

[0070] in, The location of the drone. The derivative of the drone's position is its velocity. For the quality of drones, For the speed of the drone, The derivative of the drone's velocity is the drone's acceleration. It is the acceleration due to gravity. Let R be the unit vector along the z-axis in the world coordinate system, with the superscript T indicating the transpose of the vector or matrix, and R be the rotation matrix of the UAV. The lift generated by the drone's propellers The continuous time-varying external interference force experienced by the drone body; Let be the derivative of the rotation matrix of the UAV. The angular velocity of the drone, For the angular acceleration of the drone, Let be the rotational inertia matrix of the UAV, and τ be the external torque acting on the UAV. Angular velocity of the drone The corresponding antisymmetric matrix;

[0071] Step 1.2: Establishment of the floating base dynamic model for the robot dog: The core feature of the floating base model is that the base has no fixed constraints, possesses 6 degrees of freedom, and couples with the leg joint motion to form a multibody dynamic system. The floating base system consists of one floating base and four legs. The forces acting on the base and legs need to be analyzed separately, clarifying the force transmission path from the feet to the leg joints and then to the base. The dynamic equations of the floating base system are established using the Newton-Euler recursive method, in two steps: "forward recursion" and "backward recursion," ultimately forming a complete dynamic model coupling the 6 degrees of freedom of the base and the 12 degrees of freedom of the leg joints.

[0072] ;

[0073] ;

[0074] in, For the overall configuration space of the mechanical dog, Its degrees of freedom are 12 + 6 = 18. The six degrees of freedom representing the body's floating base. This represents the 12 degrees of freedom of a joint. For generalized position, velocity, and acceleration vectors, For the generalized mass matrix, For Cocteau force terms, For gravity, The joint torque provided by the motors configured for the robot dog Here is the kinematic pair matrix, which describes the direction of the joint degrees of freedom. For foot power, For the number of contacting legs, For the contact constraint Jacobian matrix, The external disturbance forces and torques experienced by the robot dog's base. The spatial transformation matrix transforms the disturbance forces on the base from the world coordinate system to the rigid body local coordinate system.

[0075] Furthermore, the 6 degrees of freedom include 3 translational degrees of freedom and 3 rotational degrees of freedom, and each of the 4 legs contains 3 rotational joints. The forward recursion is used to calculate the inertial force / torque of each link, and the reverse recursion is used to calculate the joint torque and the force on the base.

[0076] Furthermore, step (2) includes:

[0077] Step 2.1: The cubic spline smoothing disturbance force is used as a mathematical model to simulate the continuous time-varying disturbances experienced by the mission platform during operation. Its core feature is that it generates a smooth and continuous disturbance force curve through piecewise quadratic polynomial interpolation, which can effectively simulate the gradual disturbances in the actual environment.

[0078] Step 2.2: The core function of the interference prediction network is to learn from the historical operation data of the task platform and predict external time-varying interference in real time, so as to provide interference prior information for subsequent trajectory planning.

[0079] Step 2.3: After obtaining the network-predicted interference value in the UAV mission platform, the interference estimate needs to be used as a real-time feedback signal in the baseline controller.

[0080] Furthermore, in step 2.1, the process of constructing the interference force mainly includes three stages:

[0081] ① Determine the time interval of the interference force, and uniformly select several time control points within this interval;

[0082] ② Assign a corresponding interference force amplitude to each time control point. These amplitudes are randomly generated based on the intensity range of the actual interference scenario.

[0083] ③ Between two adjacent time control points, a cubic polynomial function is used to perform interpolation; this ensures that the interpolation curve satisfies the continuity of the second derivative throughout the entire time interval, and finally obtains a continuous time-varying disturbance force that conforms to physical characteristics.

[0084] Furthermore, in step 2.1, the specific form of the interference force is described as follows:

[0085] ;

[0086] ;

[0087] ;

[0088] ;

[0089] in To add interference forces to the mission platform, interference forces in three directions. They are independent of each other. This represents the time-varying disturbance values ​​in its three directions. This represents the cubic spline disturbance force, where the parameter... For time; parameters The time control point vector is uniformly distributed over the time interval. Internal; Parameter For amplitude control point vectors, This indicates that the amplitude control point is within the interference amplitude range. The internal order follows a uniform distribution.

[0090] Furthermore, in step 2.2, the design of the interference prediction network revolves around "temporal feature capture - multi-source data fusion - dual-platform adaptation";

[0091] For UAV platforms: The input features are selected from key operational data that can directly reflect the impact of interference, specifically a three-dimensional velocity sequence and a three-dimensional expected control force sequence with nine consecutive time steps. The two types of sequences are spliced ​​together to form a 54-dimensional time-series input feature vector.

[0092] ;

[0093] ;

[0094] ;

[0095] in, The net thrust of the drone is the thrust of the propellers minus the weight of the drone itself. It is a velocity sequence. These are velocity vectors at nine different time points. For thrust sequence, These are the net thrust vectors at nine different time points. To integrate the network input sequence of velocity and net thrust, Represents a neural network for predicting interference. Its weights and biases, etc. These represent the interference values ​​in the three directions predicted by the final network.

[0096] For the robot dog platform: the three-dimensional velocity of the base and the rotational speed sequence of each joint over nine consecutive time steps are spliced ​​together with the three-dimensional desired control torque sequence of each joint as the input to the network;

[0097] ;

[0098] ;

[0099] ;

[0100] in, It is a velocity sequence. These are the base velocity and joint rotation speed vectors at nine different time points. It is a torque sequence. These are the torque vectors at nine different times. To integrate the network input sequences of velocity and torque, Represents a neural network for predicting interference. Its weights and biases, etc. These represent the interference values ​​in the three directions predicted by the final network.

[0101] The network structure adopts a lightweight architecture of "embedding layer - 1D convolutional layer - global pooling layer - regression layer". The embedding layer maps the input to a low-dimensional feature space to reduce computation. The 1D convolutional layer extracts temporal features, the global pooling layer compresses redundant features while retaining key information, and the regression layer finally outputs a 3D interference prediction value. Finally, during the training phase, the true interference value calculated by the platform dynamics model is used as the label. The loss function, through iterative optimization of network weights, enables real-time and accurate prediction of interference;

[0102] The loss function is defined as follows:

[0103] .

[0104] Furthermore, step 2.3 is detailed as follows:

[0105] In the generation of the desired control force of the controller, an interference compensation mechanism is introduced. Based on the ideal control force corresponding to the desired motion of the UAV, the inverse compensation amount of the interference term is directly superimposed. That is, the interference estimate obtained by the interference learning prediction method is incorporated into the calculation process of the desired control force in an equal and inverse manner, so that the desired control force output by the controller can offset the influence of the interference on the actual force of the UAV, thereby ensuring the consistency between the UAV's motion state and the desired state.

[0106] ;

[0107] in, The desired location for the drone. For the desired speed of the drone, For the quality of drones, Given a first diagonal matrix where all elements are greater than 0, Given a second diagonal matrix in which all elements are greater than 0.

[0108] The following specific embodiments illustrate the implementation principle of the present invention:

[0109] The experiment is based on the accompanying diagram in the instruction manual. Figure 2 quadcopter drones and Figure 3 The interference estimation and trajectory tracking performance of the robot dog were verified.

[0110] The specific experimental configuration and parameters of the quadcopter UAV used in the simulation are as follows: the quadcopter UAV has a mass of 1.2 kg, and the moments of inertia of the x, y, and z axes in the mechanical system are 0.0035 kg·m², 0.0045 kg·m², and 0.004 kg·m², respectively. In the experiment, the controller adopts a PID control architecture, with a position loop proportional gain of 2.0 and a velocity loop proportional gain of 0.5. The iteration step size of the neural network predictor is 0.02 s, and the number of neurons in the input layer and hidden layer are 64 and 128, respectively. The ReLU function is used as the activation function.

[0111] Experimental results related to quadcopter drones are as follows: Figures 4 to 6 As shown.

[0112] Figure 4 The interference estimation results for the quadcopter UAV show that the estimated disturbance closely matches the actual disturbance, demonstrating good interference estimation accuracy. Figures 5 to 6 To improve trajectory tracking performance, the disturbance estimated by the proposed neural network prediction method was incorporated into the controller, resulting in a significant improvement in the trajectory tracking performance of the quadcopter UAV. Comparative data on position tracking errors show that the average error after disturbance compensation is only 0.0627 m, far lower than the 0.3929 m without disturbance compensation, demonstrating the significant effect of disturbance compensation on improving trajectory tracking accuracy.

[0113] The robot dog platform weighs 3.3 kg, with a length of 0.38 m, a width of 0.098 m, and a height of 0.1 m. The reduction ratios for hip abduction, hip joint, and knee joint are 6, 6, and 9.33, respectively. The maximum output torque of the motor is 3 N·m, the battery voltage is 24 volts, the joint damping coefficient is 0.01, and the joint dry friction coefficient is 0.2. The lengths of the hip abduction link, hip joint link, and knee joint link are 0.195 m, with a maximum leg extension length of 0.409 m. The controller's joint angle loop proportional gain is set to 20, and the speed loop proportional gain is set to 0.8.

[0114] The machine dog interference estimation results are as follows Figure 7 As shown, the proposed interference estimation method can accurately capture the dynamic changes of actual interference (including fluctuation amplitude, frequency and other characteristics). In the 200-second continuous test, the estimation results are always consistent with the actual interference, without drift or amplification of deviation, which demonstrates the reliability of the method under long-term and complex working conditions.

[0115] In summary, the application verification of the proposed interference estimation technology on quadcopters and robot dogs shows that it has cross-platform adaptability and can accurately and stably capture multi-dimensional dynamic interference, providing a reliable guarantee for the stable control of robots in complex scenarios.

[0116] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made according to the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention. Contents not described in detail in this specification are prior art known to those skilled in the art.

Claims

1. A method for interference learning and prediction applicable to collaborative heterogeneous UAVs and robot dog platforms, characterized in that, Includes the following steps: (1) Dynamic modeling of the mission platform: Force analysis is performed on the UAV and the robot dog, and a dynamic model of the UAV and a floating base model of the robot dog are established. The dynamic state update of the two types of platforms is completed under the known external forces. (2) External interference modeling, interference network construction and interference feedback application: Based on the dynamic model constructed in step 1, continuous time-varying interference is added to the task platform, and an interference prediction neural network is built. The network is trained with the platform's historical operating data as input features and the actual interference value as a label, so that it has the ability to predict interference. The platform's historical operating data includes speed, acceleration, motor speed and torque. In step (2), the design of the interference prediction network revolves around "temporal feature capture - multi-source data fusion - dual-platform adaptation"; For UAV platforms: The input features are selected from key operational data that can directly reflect the impact of interference, specifically a three-dimensional velocity sequence and a three-dimensional expected control force sequence with nine consecutive time steps. The two types of sequences are spliced ​​together to form a 54-dimensional time-series input feature vector. ； ； ； in, The net thrust of the drone is the thrust of the propellers minus the weight of the drone itself. It is a velocity sequence. These are velocity vectors at nine different time points. For thrust sequence, These are the net thrust vectors at nine different time points. To integrate the network input sequence of velocity and net thrust, Represents a neural network for predicting interference. Its weights and biases, etc. These represent the interference values ​​in the three directions predicted by the final network. For the robot dog platform: the three-dimensional velocity of the base and the rotational speed sequence of each joint over nine consecutive time steps are spliced ​​together with the three-dimensional desired control torque sequence of each joint as the input to the network; ； ； ； in, It is a velocity sequence. These are the base velocity and joint rotation speed vectors at nine different time points. It is a torque sequence. These are the torque vectors at nine different times. To integrate the network input sequences of velocity and torque, Represents a neural network for predicting interference. Its weights and biases, etc. These represent the interference values ​​in the three directions predicted by the final network. The network structure adopts a lightweight architecture of "embedding layer - 1D convolutional layer - global pooling layer - regression layer". The embedding layer maps the input to a low-dimensional feature space to reduce computation. The 1D convolutional layer extracts temporal features, the global pooling layer compresses redundant features while retaining key information, and the regression layer finally outputs a 3D interference prediction value. Finally, during the training phase, the true interference value calculated by the platform dynamics model is used as the label. The loss function, through iterative optimization of network weights, enables real-time and accurate prediction of interference; The loss function is defined as follows: 。 2. The interference learning and prediction method applicable to collaborative heterogeneous UAVs and robot dog platforms according to claim 1, characterized in that, Step (1) includes establishing dynamic equations for the heterogeneous actuators of the drone and the robot dog, clarifying the mapping relationship between the output force / torque of the actuator and the motion state of the device. The heterogeneous actuators of the drone and the robot dog are the drone propeller and the robot dog joint drive unit. The motion state of the device is position, speed and attitude.

3. The interference learning and prediction method applicable to collaborative heterogeneous UAVs and robot dog platforms according to claim 2, characterized in that, The specific steps (1) are as follows: Step 1.1: Based on the dynamic characteristics of the quadcopter UAV, establish the following dynamic model: ; ； in, The location of the drone. The derivative of the drone's position is its velocity. For the quality of drones, For the speed of the drone, The derivative of the drone's velocity is the drone's acceleration. It is the acceleration due to gravity. Let R be the unit vector along the z-axis in the world coordinate system, with the superscript T indicating the transpose of the vector or matrix, and R be the rotation matrix of the UAV. The lift generated by the drone's propellers The continuous time-varying external interference force experienced by the drone body; Let be the derivative of the rotation matrix of the UAV. The angular velocity of the drone, For the angular acceleration of the drone, Let be the rotational inertia matrix of the UAV, and τ be the external torque acting on the UAV. Angular velocity of the drone The corresponding antisymmetric matrix; Step 1.2: Establishment of the floating base dynamic model for the robot dog: The core feature of the floating base model is that the base has no fixed constraints, possesses 6 degrees of freedom, and couples with the leg joint motion to form a multibody dynamic system. The floating base system consists of one floating base and four legs. The forces acting on the base and legs need to be analyzed separately, clarifying the force transmission path from the feet to the leg joints and then to the base. The dynamic equations of the floating base system are established using the Newton-Euler recursive method, in two steps: "forward recursion" and "backward recursion," ultimately forming a complete dynamic model coupling the 6 degrees of freedom of the base and the 12 degrees of freedom of the leg joints. ； ； in, For the overall configuration space of the mechanical dog, Its degrees of freedom are 12 + 6 = 18. The six degrees of freedom represent the floating base of the body. Represents the 12 degrees of freedom of a joint. For generalized position, velocity, and acceleration vectors, For the generalized mass matrix, For the Cocteau force term, For gravity, The joint torque provided by the motors configured for the robot dog Here is the kinematic pair matrix, which describes the direction of the joint degrees of freedom. For foot power, For the number of contacting legs, For the contact constraint Jacobian matrix, The external disturbance forces and torques experienced by the robot dog's base. The spatial transformation matrix transforms the disturbance forces on the base from the world coordinate system to the rigid body local coordinate system.

4. The interference learning and prediction method applicable to collaborative heterogeneous UAVs and robot dog platforms according to claim 3, characterized in that, The 6 degrees of freedom include 3 translational degrees of freedom and 3 rotational degrees of freedom. Each of the 4 legs contains 3 rotational joints. The forward recursion is used to calculate the inertial force / torque of each link, and the reverse recursion is used to calculate the joint torque and the force on the base.

5. The interference learning and prediction method applicable to collaborative heterogeneous UAVs and robot dog platforms according to claim 4, characterized in that, Step (2) includes: Step 2.1: The cubic spline smoothing disturbance force is used as a mathematical model to simulate the continuous time-varying disturbances experienced by the mission platform during operation. Its core feature is that it generates a smooth and continuous disturbance force curve through piecewise quadratic polynomial interpolation, which can effectively simulate the gradual disturbances in the actual environment. Step 2.2: The core function of the interference prediction network is to learn from the historical operation data of the task platform and predict external time-varying interference in real time, so as to provide interference prior information for subsequent trajectory planning. Step 2.3: After obtaining the network-predicted interference value in the UAV mission platform, the interference value needs to be used as a real-time feedback signal in the baseline controller.

6. The interference learning and prediction method applicable to collaborative heterogeneous UAVs and robot dog platforms according to claim 5, characterized in that, In step 2.1, the process of constructing the interference force mainly includes three stages: ① Determine the time interval of the interference force, and uniformly select several time control points within this interval; ② Assign a corresponding interference force amplitude to each time control point. These amplitudes are randomly generated based on the intensity range of the actual interference scenario. ③ Between two adjacent time control points, interpolation is performed using a cubic polynomial function; To ensure that the interpolation curve satisfies the continuity of the second derivative throughout the entire time interval, a continuous time-varying disturbance force that conforms to physical properties is finally obtained.

7. The interference learning and prediction method applicable to collaborative heterogeneous UAVs and robot dog platforms according to claim 6, characterized in that, In step 2.1, the specific form of the interference force is described as follows: ； ； ； ； in To add interference forces to the mission platform, interference forces in three directions. They are independent of each other. This represents the time-varying disturbance values ​​in its three directions. This represents the cubic spline disturbance force, where the parameter... For time; parameters The time control point vector is uniformly distributed over the time interval. Internal; Parameter For amplitude control point vectors, This indicates that the amplitude control point is within the interference amplitude range. The internal order follows a uniform distribution.

8. The interference learning and prediction method applicable to collaborative heterogeneous UAVs and robot dog platforms according to claim 7, characterized in that, Step 2.3 is as follows: In the generation of the desired control force of the controller, an interference compensation mechanism is introduced. Based on the ideal control force corresponding to the desired motion of the UAV, the reverse compensation amount of the interference term is directly superimposed. That is, the interference estimate obtained by the interference learning prediction method is incorporated into the calculation process of the desired control force in an equal and reverse manner, so that the desired control force output by the controller can offset the influence of the interference on the actual force of the UAV, thereby ensuring the consistency between the UAV's motion state and the desired state. ； in, The desired location for the drone. For the desired speed of the drone, For the quality of drones, Given a first diagonal matrix where all elements are greater than 0, Given a second diagonal matrix in which all elements are greater than 0.

Images