An unmanned excavator operation track planning method and device

Abstract

This invention discloses a method and apparatus for planning the operation trajectory of an unmanned excavator, comprising: inputting an excavation task; decomposing the excavation task to obtain excavation path points; using a fifth-order polynomial to plan the trajectory between every two excavation path points in a single operation to obtain the trajectory of a single operation; using a particle swarm optimization algorithm with time as the optimization parameter to optimize the trajectory of a single operation, and simultaneously collecting the optimized trajectory planning data as a dataset; using the dataset to train a physical neural network algorithm model; and using the trained physical neural network algorithm model to plan the operation trajectory of the unmanned excavator to generate a complete excavation trajectory. This invention uses a physical neural network algorithm for complete excavation task trajectory planning, which improves the real-time performance of operation trajectory planning and enables autonomous planning of the unmanned excavator's operation trajectory, thereby improving the intelligence level of the unmanned excavator.

Description

Technical Field

[0001] This invention relates to a method and apparatus for planning the working trajectory of an unmanned excavator, belonging to the field of excavator operation technology. Background Technology

[0002] Excavators, as efficient and multifunctional engineering machinery, are widely used in ports, mines, construction, and other fields. These scenarios involve harsh working environments, and the operation process is prone to extreme dangers such as landslides, posing a significant threat to the safety of operators. At the same time, the excavator industry generally suffers from labor shortages and long operator training cycles. Unmanned excavators (UAVs), on the other hand, can be remotely controlled, avoiding human work in hazardous environments and eliminating fatigue limitations, thus improving construction efficiency and solving the labor shortage problem. Furthermore, UAVs integrate multiple algorithms, resulting in high operational accuracy and simple operation. Therefore, the development of UAVs is of great significance, and trajectory planning is one of the key technologies for realizing UAVs. How to design autonomous trajectory planning is crucial for achieving UAVs. Current trajectory planning research mainly focuses on single-operation trajectory planning, failing to achieve complete trajectory planning for excavation tasks. Moreover, trajectory planning is time-consuming, has a low level of intelligence, and cannot meet the requirements of real-time operations. Summary of the Invention

[0003] The technical problem to be solved by this invention is that the existing technology cannot achieve complete trajectory planning for mining tasks, and the trajectory planning time is long and the level of intelligence is low.

[0004] To solve the above-mentioned technical problems, the present invention adopts a method and device for planning the operation trajectory of an unmanned excavator.

[0005] On one hand, the present invention provides a method for planning the working trajectory of an unmanned excavator, comprising:

[0006] Input the excavation task, decompose the excavation task, and obtain the excavation path points;

[0007] The trajectory of a single operation is obtained by planning the trajectory between every two excavation path points in a single operation using a fifth-degree polynomial.

[0008] The particle swarm optimization algorithm is used to optimize the trajectory of a single operation with time as the optimization parameter, and the optimized trajectory planning data is collected as a dataset.

[0009] The physical neural network algorithm model is trained using the dataset to obtain the trained physical neural network algorithm model;

[0010] A trained physical neural network algorithm model is used to plan the operation trajectory of the unmanned excavator and generate a complete excavation trajectory.

[0011] The method for training the physical neural network algorithm model includes:

[0012] The dataset is divided into a training set and a validation set, and the input and target data of the training and validation sets are standardized.

[0013] Configure neural network training parameters, including gradient clipping thresholds;

[0014] For each training batch, the loss is calculated by forward propagation of the standardized input data and target data, the gradient is obtained by back propagation, the gradient is clipped according to the gradient clipping threshold, and the training parameters of the neural network are updated.

[0015] After each round of training, the performance of the physical neural network algorithm model is evaluated using the validation set and the validation loss is recorded. The learning rate is adjusted based on the validation loss. If the validation loss does not decrease for several consecutive rounds, the early stopping mechanism is triggered to terminate the training.

[0016] After training is terminated, the model with the minimum validation loss is saved as the completed physical neural network algorithm model.

[0017] This invention uses a task decomposition strategy to break down a complete trench excavation task into excavation points. This method enables the autonomous generation of excavation task trajectory planning, avoiding repetitive or omitted trajectory planning actions and improving work efficiency.

[0018] The excavation task is decomposed into the following steps: setting the operation parameters for the trench excavation task; dividing the overall excavation task into several discrete single-bucket excavation operations based on the capacity parameters of the excavator bucket; determining the coordinates of the key positions in Cartesian space for each single-bucket excavation operation; converting the coordinates of the key positions into the angles of each joint of the excavator through inverse kinematics; and then solving for the parameters of the corresponding excavation path points.

[0019] The method of using a fifth-order polynomial to plan the trajectory between every two excavation path points in a single operation includes:

[0020] Fifth-order polynomials are used to interpolate the joint angles of the links to generate smooth trajectories. The polynomial interpolation function... as follows:

[0021] ;

[0022] in, It is a polynomial interpolation function. The boom joint angle of the link member. 2 represents the stick joint angle of the rod. The bucket joint angle of the rod, The rotation angle of the member. For rods Joint angle interpolation function coefficient of the secondary term, The time required to complete the trajectory execution, For the required time Power of 1.

[0023] The step of optimizing the trajectory of a single operation using the particle swarm optimization algorithm with time as the optimization parameter includes: taking the shortest excavation time t as the objective function, setting the constraint ranges for the angle, velocity, acceleration, and torque of each joint, establishing a trajectory planning optimization model for the unmanned excavator working device, and optimizing the trajectory of a single operation, as shown in the following formula:

[0024] ;

[0025] in, For the first The angle of each joint For the first angular velocity of each joint For the first angular acceleration of each joint, For the first The driving torque of each joint For the first The maximum angular amplitude of each joint For the first Maximum angular velocity amplitude of each joint For the first The maximum angular acceleration amplitude of each joint For the first The maximum driving torque amplitude of each joint For the first The maximum permissible angle of each joint For the first The maximum permissible angular velocity of each joint For the first The maximum permissible angular acceleration of each joint For the first The maximum permissible driving torque of each joint;

[0026] The objective function for optimization is defined as follows:

[0027] ;

[0028] in, As an indicator to measure the efficiency of robotic arm movement, For the first The time required for each joint to complete its trajectory movement, where 7 represents the number of joints.

[0029] After simultaneously collecting the optimized trajectory planning data as a dataset, the method further includes: establishing a simulation model of the unmanned excavator working device to verify whether the optimized trajectory planning data is feasible. If it is feasible, the dataset is used to train the physical neural network algorithm model. If it is not reliable, the trajectory is re-optimized.

[0030] The trajectory planning and optimization model for the unmanned excavator working device includes the kinematic model and the dynamic model of the working device;

[0031] The kinematic model is a 4-DOF kinematic model of the unmanned excavator working device established using the DH method, including a forward kinematic model and an inverse kinematic model.

[0032] The dynamic model is a multi-rigid-body dynamic model of the unmanned excavator working device established using the Lagrange equation.

[0033] In the forward propagation loss calculation step for each training batch of standardized input and target data, a loss function is used to calculate the loss. This loss function includes: data loss, physical constraint loss, performance loss, and torque constraint loss, expressed as follows:

[0034] ;

[0035] in, For loss function, For data loss, These are the weighting coefficients for the physical constraint loss. For physical constraint loss, The weighting coefficients for performance loss For performance loss, The weighting coefficient for torque constraint loss. This is the torque constraint loss.

[0036] The data loss The formula is as follows: ; in, For Huber's losses, This represents the mean square error loss.

[0037] The physical constraint loss The formula is as follows: ; in, For angle limit constraints, For speed constraints, For acceleration constraints, To add jerk constraints.

[0038] The performance loss The formula is as follows: ; in, Number the joints. For the first Time deviation of each joint.

[0039] The torque constraint loss The formula is as follows: ; in, The total number of torque samples. For the first Each joint at any time The actual torque, For the first The maximum permissible rated torque of each joint.

[0040] The formula is as follows: ; in, Let be the minimum allowable angle for the j-th joint. Let be the actual joint angle of the j-th joint at time t. Let be the maximum allowed angle of the j-th joint, and Softplus() be a non-linear activation function.

[0041] ; Where n is the total number of time steps for the trajectory. Let be the angular velocity of the joint at time t. This represents the maximum permissible angular velocity of the joint.

[0042] ; in, Let be the angular acceleration of the joint at time t. This represents the maximum permissible angular acceleration of the joint.

[0043] ; in, Let be the angular acceleration of the joint at time t. The maximum permissible angular acceleration of the joint.

[0044] ; ; in: The total number of samples, For the first The true value of each sample For the first The model prediction value for each sample. This is the threshold parameter for Huber loss.

[0045] The input and target data of the training and validation sets are standardized using the following formula:

[0046] ;

[0047] ;

[0048] in, The numerical values ​​for each feature dimension, The standardized values ​​for each feature dimension, The mean value for each feature dimension. The corrected standard deviation for each feature dimension. The original standard deviation of each feature dimension The standard deviation is set as a minimum threshold for the numerical stability constant. .

[0049] The configuration of the neural network training parameters includes: using the AdamW optimizer, setting the initial learning rate to 0.0005, and setting the weight decay to... Set the maximum number of training rounds to 120 and the gradient clipping threshold to 1.0.

[0050] On the other hand, the present invention provides an unmanned excavator operation trajectory planning device, comprising:

[0051] The task input and decomposition module is used to decompose the input task into mining path points.

[0052] The job trajectory planning module is used to: plan the trajectory between every two excavation path points in a single job using a fifth-order polynomial, and output the trajectory of a single job.

[0053] The job trajectory optimization module is used to optimize the trajectory of a single job using the particle swarm optimization algorithm with time as the optimization parameter, and at the same time collect the optimized trajectory planning data as a dataset.

[0054] The physical neural network algorithm module is used to: plan the operation trajectory of the unmanned excavator using a physical neural network algorithm model and generate a complete excavation trajectory;

[0055] The physical neural network algorithm model training module is used for: dividing the dataset into a training set and a validation set; standardizing the input and target data of the training and validation sets and inputting them into the physical neural network algorithm model; configuring neural network training parameters, including gradient clipping thresholds; calculating the loss through forward propagation and updating the neural network training parameters through backpropagation for each training batch, and applying gradient clipping; evaluating the performance of the physical neural network algorithm model and recording the validation loss using the validation set after each round of training; adjusting the learning rate based on the validation loss; triggering an early stopping mechanism to terminate training if the validation loss does not decrease for several consecutive rounds; and saving the model with the smallest validation loss as the completed physical neural network algorithm model after training is terminated.

[0056] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: The present invention uses a physical neural network algorithm for complete task trajectory planning. Using this algorithm improves the computational efficiency of operation trajectory planning and realizes autonomous planning of the operation trajectory of unmanned excavators, thereby improving the intelligence level of unmanned excavators. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the unmanned excavator operation trajectory planning process shown in Embodiment 1 of the present invention;

[0058] Figure 2 This is a schematic diagram of the unmanned excavator operation trajectory planning method shown in Embodiment 3 of the present invention. Detailed Implementation

[0059] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0060] Example 1

[0061] This embodiment introduces a method for planning the working trajectory of an unmanned excavator, including:

[0062] Input the excavation task, break it down, and obtain the excavation path points;

[0063] The trajectory of a single operation is obtained by planning the trajectory between every two excavation path points in a single operation using a fifth-degree polynomial.

[0064] The particle swarm optimization algorithm is used to optimize the trajectory of a single operation with time as the optimization parameter, and the optimized trajectory planning data is collected as a dataset.

[0065] The physical neural network algorithm model is trained using the dataset to obtain the trained physical neural network algorithm model;

[0066] A trained physical neural network algorithm model is used to plan the operation trajectory of the unmanned excavator and generate a complete excavation trajectory.

[0067] The methods for training physical neural network algorithm models include:

[0068] The dataset is divided into a training set and a validation set, and the input and target data of the training and validation sets are standardized.

[0069] Configure neural network training parameters, including gradient clipping thresholds;

[0070] For each training batch, the loss is calculated by forward propagation of the standardized input data and target data, the gradient is obtained by backpropagation, the gradient is clipped according to the gradient clipping threshold, and the neural network training parameters are updated.

[0071] After each round of training, the performance of the physical neural network algorithm model is evaluated using the validation set and the validation loss is recorded. The learning rate is adjusted based on the validation loss. If the validation loss does not decrease for several consecutive rounds, the early stopping mechanism is triggered to terminate the training.

[0072] After training is terminated, the model with the minimum validation loss is saved as the completed physical neural network algorithm model.

[0073] Specifically, such as Figure 1 As shown, a simulation model of the unmanned excavator's working device is established, and dynamic and kinematic models are also established. The excavation task is input and decomposed. A fifth-order polynomial is used to plan the trajectory of a single operation. The particle swarm optimization algorithm is used to optimize the trajectory of a single operation with time as the optimization parameter, and the optimized trajectory planning data is collected as a dataset. The Physical Neural Network (PINN) algorithm is used to train the dataset. The trained PINN algorithm model is used to perform trajectory planning for the unmanned excavator, generating the complete excavation trajectory of the target task.

[0074] Specifically, the particle swarm optimization algorithm simulates the collaborative cooperation and information exchange among individuals in a biological community. In the process of solving a problem, it continuously and dynamically updates the position and velocity parameters of the particles, driving the entire particle swarm to evolve towards the optimal solution. The algorithm's iterative formula is as follows:

[0075] ;

[0076] ;

[0077] in, For particles No. Flight speed at the next iteration For particles No. Flight speed at the next iteration For particles No. Position at the next iteration For particles No. Position at the next iteration For inertial weights, and As a learning factor, , ~U(0,1), and The optimal position experienced by a particle during its flight is called the individual extreme value, and the optimal position that has appeared for all particles in the entire population is called the global extreme value.

[0078] The excavation task is decomposed into the following steps: setting the operation parameters for the trench excavation task; dividing the overall excavation task into several discrete single-bucket excavation operations based on the capacity parameters of the excavator bucket; determining the coordinates of the key positions in Cartesian space for each single-bucket excavation operation; and converting the coordinates of the key positions into the angles of each joint of the excavator through inverse kinematics solution, thereby solving for the parameters of the corresponding excavation path points.

[0079] The trajectory between every two excavation path points in a single operation is planned using a fifth-order polynomial, including:

[0080] Fifth-order polynomials are used to interpolate the joint angles of the links to generate smooth trajectories. The polynomial interpolation function... as follows:

[0081] ;

[0082] in, It is a polynomial interpolation function. The boom joint angle of the link member. 2 represents the stick joint angle of the rod. The bucket joint angle of the rod, The rotation angle of the member. For rods Joint angle interpolation function coefficient of the secondary term, The time required to complete the trajectory execution, For the required time Power of 1.

[0083] Specifically, joint space trajectory planning is performed on the path points of the single operation trajectory planning. A fifth-order polynomial is used as the interpolation function to fit the joint angle change curve of the path points of the single operation trajectory planning. The single operation trajectory includes seven path points: starting point, entry point (i.e., the excavation point of the complete task division), excavation point 1, excavation point 2, exit point, unloading point, and unloading operation point.

[0084] The steps for optimizing the trajectory of a single operation using the particle swarm optimization algorithm with time as the optimization parameter include: taking the shortest excavation time t as the objective function, setting the constraint ranges for the angles, velocities, accelerations, and torques of each joint, establishing a trajectory planning optimization model for the unmanned excavator's working device, and optimizing the trajectory of a single operation. The formula is as follows:

[0085] ;

[0086] in, For the first The angle of each joint For the first angular velocity of each joint For the first angular acceleration of each joint, For the first The driving torque of each joint For the first The maximum angular amplitude of each joint For the first Maximum angular velocity amplitude of each joint For the first The maximum angular acceleration amplitude of each joint For the first The maximum driving torque amplitude of each joint For the first The maximum permissible angle of each joint For the first The maximum permissible angular velocity of each joint For the first The maximum permissible angular acceleration of each joint For the first The maximum permissible driving torque of each joint;

[0087] The objective function for optimization is defined as follows:

[0088] ;

[0089] in, As an indicator to measure the efficiency of robotic arm movement, For the first The time required for each joint to complete its trajectory movement.

[0090] Simultaneously, after collecting the optimized trajectory planning data as a dataset, the process also includes: establishing a simulation model of the unmanned excavator's working device to verify whether the optimized trajectory planning data is feasible. If it is feasible, the dataset is used to train the physical neural network algorithm model; if it is not reliable, the trajectory is re-optimized.

[0091] The trajectory planning and optimization model for the unmanned excavator's working device includes the kinematic and dynamic models of the working device.

[0092] The kinematic model is a 4-DOF kinematic model of the unmanned excavator working device established using the DH method, including the forward kinematic model and the inverse kinematic model;

[0093] The dynamic model is a multi-rigid-body dynamic model of the unmanned excavator working device established using the Lagrange equation.

[0094] Specifically, the forward kinematics solution solves for the transformation between the joint space and pose space of the working device. The forward kinematic model of the excavator's working device, obtained through calculation, is as follows:

[0095]

[0096]

[0097]

[0098]

[0099] in, For the rotation joint angle, The boom joint angle, For the pole joint angle, For bucket joint angle, For boom length, For the length of the pole, For the length of the bucket boom, coordinate system With coordinate system Distance in the X direction, coordinate system With coordinate system Distance in the Y direction, x This represents the position of the endpoint in the X-axis direction of the base coordinate system. y This represents the position of the endpoint in the Y-axis direction of the base coordinate system. z This represents the position of the endpoint in the Z-axis direction of the base coordinate system. This refers to the angle of the bucket teeth.

[0100] Specifically, the inverse kinematics solution involves precisely calculating the corresponding boom joint angle based on the desired position and orientation of the bucket teeth of the excavator's working device. , pole joint angle Bucket joint angle and the angle of the bucket teeth Let the tip of the bucket tooth be in the coordinate system. The coordinates in [ In the coordinate system The coordinates in are , For the bucket tooth tip in the end coordinate system of Coordinates on the axis For the bucket tooth tip in the end coordinate system of Coordinates on the axis For the bucket tooth tip in the end coordinate system of The coordinates on the axis. From the above geometric relationships, we can deduce:

[0101]

[0102]

[0103]

[0104]

[0105]

[0106]

[0107]

[0108]

[0109] in, The coordinate system of the rotary device is set as the base coordinate system. Let the boom coordinate system be... For the boom coordinate system, Let's use the bucket coordinate system. Let be the coordinate system for the bucket tooth tip.

[0110] Specifically, in establishing the dynamic model, it is assumed that the position of the center of mass of each member is known and the friction of mechanical transmission is not considered. The digging resistance during the operation is taken into account, which can be transformed into a matrix form:

[0111] ;

[0112] in, The inertia matrix, The term representing the square of the velocity is called the centripetal force influence matrix. The gravity effect matrix, The resistance torque of the material on the bucket during excavation operations. For the angles of each joint , For the torque of each joint , Angular velocity of each joint , Represented as the angular acceleration of each joint . This is the driving torque of the boom joint. This refers to the driving torque of the boom joint. The driving torque of the bucket joint. for angular velocity, for angular velocity, for angular velocity, for angular acceleration, for angular acceleration, for angular acceleration.

[0113] Specifically, the digging load generated by the bucket tooth tip can be divided into tangential digging load. and normal excavation load

[0114] ;

[0115] ;

[0116] In the formula, For the specific drag coefficient of excavation, The width of the bucket teeth, This is the corrected excavation depth. To extract the resistance coefficient, This is the corrected tangential excavation load.

[0117] In the forward propagation loss calculation step for each training batch of standardized input and target data, a loss function is used to calculate the loss. This loss function includes: data loss, physical constraint loss, performance loss, and torque constraint loss, expressed as:

[0118] ;

[0119] in, For loss function, For data loss, These are the weighting coefficients for the physical constraint loss. For physical constraint loss, The weighting coefficients for performance loss For performance loss, The weighting coefficient for torque constraint loss. For torque constraint loss;

[0120] Physical constraint loss Including angular limits, velocity, acceleration, and jerk constraints, the formulas are as follows:

[0121] ;

[0122] in, For angle limit constraints, , For the first The minimum permissible angle of each joint. For the first Each joint at any time The actual joint angle, For the first The maximum permissible angle of each joint For speed constraints, , The total number of time steps for the trajectory. For joints at all times angular velocity, The maximum permissible angular velocity of the joint. For acceleration constraints, , For joints at all times angular acceleration, This represents the maximum permissible angular acceleration of the joint. For angular acceleration constraints, , For joints at all times angular acceleration, The maximum permissible angular acceleration of the joint;

[0123] The performance loss formula is: ,in, Let be the performance loss function. Number the joints. For the first Time deviation of each joint;

[0124] The formula for torque constraint loss is: ,in, The torque constraint loss function, The total number of torque samples. For the first Each joint at any time The actual torque, For the first The maximum permissible rated torque of each joint.

[0125] Data loss Using a combination of Huber loss and mean squared error, the formula is as follows:

[0126] ;

[0127] ;

[0128] in, For data loss, For Huber's losses, For mean square error loss, , The total number of samples, For the first The true value of each sample For the first The model prediction value for each sample. This is the threshold parameter for Huber loss.

[0129] The input and target data for the training and validation sets are standardized using the following formula:

[0130] ;

[0131] ;

[0132] in, The numerical values ​​for each feature dimension, The standardized values ​​for each feature dimension, The mean value for each feature dimension. The corrected standard deviation for each feature dimension. The original standard deviation of each feature dimension The standard deviation is set as a minimum threshold for the numerical stability constant. .

[0133] Configuring neural network training parameters includes: using the AdamW optimizer, setting the initial learning rate to 0.0005, and setting the weight decay to... Set the maximum number of training rounds to 120 and the gradient clipping threshold to 1.0.

[0134] Specifically, the AdamW optimizer is used with an initial learning rate of 0.0005 and weight decay of . The learning rate scheduling employs the ReduceLROnPlateau strategy. When the validation loss fails to improve for 10 consecutive training epochs, the learning rate decays by a factor of 0.5. The maximum number of training epochs is 120, and the early stopping patience value is 15. The training process automatically terminates when the validation loss fails to show significant improvement for 15 consecutive epochs.

[0135] Specifically, the trained neural network receives eight path points as input and outputs a 175-dimensional vector, including seven time intervals and 164 polynomial coefficients. Based on these parameters, a trajectory function for each joint can be generated.

[0136] Example 2

[0137] This embodiment describes a device for planning the working trajectory of an unmanned excavator, including:

[0138] The task input and decomposition module is used to decompose the input task into mining path points.

[0139] The job trajectory planning module is used to: plan the trajectory between every two excavation path points in a single job using a fifth-order polynomial, and output the trajectory of a single job.

[0140] The job trajectory optimization module is used to optimize the trajectory of a single job using the particle swarm optimization algorithm with time as the optimization parameter, and at the same time collect the optimized trajectory planning data as a dataset.

[0141] The physical neural network algorithm module is used to: plan the operation trajectory of the unmanned excavator using a physical neural network algorithm model and generate a complete excavation trajectory;

[0142] The physical neural network algorithm model training module is used for: dividing the dataset into training and validation sets; standardizing the input and target data in both sets; configuring neural network training parameters, including gradient clipping thresholds; calculating the loss through forward propagation and updating the neural network training parameters through backpropagation for each training batch, and applying gradient clipping; evaluating the performance of the physical neural network algorithm model using the validation set and recording the validation loss, adjusting the learning rate based on the validation loss, and triggering an early stopping mechanism to terminate training if the validation loss does not decrease for several consecutive rounds; and saving the model with the minimum validation loss as the completed physical neural network algorithm model after training termination.

[0143] Example 3

[0144] This embodiment introduces a method for planning the working trajectory of an unmanned excavator, such as... Figure 2First, the task is input, then it is decomposed into several sub-drilling points. The standardized path points are input into the trained Physical Neural Network (PINN) trajectory planning model, and forward propagation is performed. The PINN trajectory planning model generates standardized trajectory parameter vectors. Then, it is determined whether trajectory planning has been completed for all sub-drilling points: if not, the task decomposition and path point solving steps are returned, and the path point solving and PINN trajectory planning are repeated for the next drilling point; if all are completed, the compliant trajectories of all drilling points are integrated to form a complete operation sequence, the drilling operation is executed, and the task ends.

[0145] As shown in Table 1, the results comparison table verifies the significant effect of this invention through measured data from six typical mining tasks: Regarding trajectory planning time (t), the planning time of the PSO algorithm fluctuates between 130.685 seconds and 580.465 seconds, while the planning time of the PINN algorithm remains stable in the lower range of 54.10 seconds to 135.24 seconds. This proves that PINN has an absolute advantage in computational efficiency and can complete trajectory solving faster. In terms of program running time (t... In terms of execution time: the PSO algorithm runs between 11.410 seconds and 24.656 seconds, while the PINN algorithm runs for less than 0.16 seconds. This order-of-magnitude difference indicates that the PINN algorithm can send the planning results to the controller with almost no delay, greatly improving the system's real-time control capability.

[0146] Table 1 Simulation results of autonomous operation trajectory planning for PINN and PSO

[0147]

[0148] The system was analyzed from multiple perspectives, and the results are shown in Table 2. In the real-time evaluation, the model demonstrated a significant advantage in computational efficiency. Test data shows that the average time for a single trajectory planning operation is only 0.471 ms.

[0149] Table 2 Training results of PINN neural network

[0150]

[0151] In summary, the physical neural network trajectory planning model of the present invention demonstrates significant performance in unmanned excavator operation trajectory planning, reducing the complete trajectory planning time of the operation task to the millisecond level, and greatly improving the real-time performance of unmanned excavators.

Claims

1. A method for planning the working trajectory of an unmanned excavator, characterized in that, include: Input the excavation task, decompose the excavation task, and obtain the excavation path points; The trajectory of a single operation is obtained by planning the trajectory between every two excavation path points in a single operation using a fifth-degree polynomial. The particle swarm optimization algorithm is used to optimize the trajectory of a single operation with time as the optimization parameter, and the optimized trajectory planning data is collected as a dataset. The physical neural network algorithm model is trained using the dataset to obtain the trained physical neural network algorithm model; A trained physical neural network algorithm model is used to plan the operation trajectory of the unmanned excavator and generate a complete excavation trajectory. The method for training the physical neural network algorithm model includes: The dataset is divided into a training set and a validation set, and the input and target data of the training and validation sets are standardized. Configure neural network training parameters, including gradient clipping thresholds; For each training batch, the loss is calculated by forward propagation of the standardized input data and target data, the gradient is obtained by back propagation, the gradient is clipped according to the gradient clipping threshold, and the training parameters of the neural network are updated. After each round of training, the performance of the physical neural network algorithm model is evaluated using the validation set and the validation loss is recorded. The learning rate is adjusted based on the validation loss. If the validation loss does not decrease for several consecutive rounds, the early stopping mechanism is triggered to terminate the training. After training is terminated, the model with the minimum validation loss is saved as the completed physical neural network algorithm model.

2. The unmanned excavator operation trajectory planning method according to claim 1, characterized in that, The excavation task is decomposed into the following steps: setting the operation parameters for the trench excavation task; dividing the overall excavation task into several discrete single-bucket excavation operations based on the capacity parameters of the excavator bucket; determining the coordinates of the key positions in Cartesian space for each single-bucket excavation operation; converting the coordinates of the key positions into the angles of each joint of the excavator through inverse kinematics; and then solving for the parameters of the corresponding excavation path points.

3. The unmanned excavator operation trajectory planning method according to claim 1, characterized in that, The method of using a fifth-order polynomial to plan the trajectory between every two excavation path points in a single operation includes: The joint angles of the members are interpolated using a fifth-order polynomial interpolation function to obtain the trajectory of a single operation. The polynomial interpolation function... as follows: ； in, It is a polynomial interpolation function. The boom joint angle of the link member. 2 represents the stick joint angle of the rod. The bucket joint angle of the rod, The rotation angle of the member. For rods Joint angle interpolation function coefficient of the secondary term, For the required time Power of 1.

4. The unmanned excavator operation trajectory planning method according to claim 1, characterized in that, The step of optimizing the trajectory of a single operation using the particle swarm optimization algorithm with time as the optimization parameter includes: taking the shortest excavation time t as the objective function, setting the constraint ranges for the angle, velocity, acceleration, and torque of each joint, establishing a trajectory planning optimization model for the unmanned excavator working device, and optimizing the trajectory of a single operation, as shown in the following formula: ； in, For the first The angle of each joint For the first angular velocity of each joint For the first angular acceleration of each joint, For the first The driving torque of each joint For the first The maximum permissible angle of each joint For the first The maximum permissible angular velocity of each joint For the first The maximum permissible angular acceleration of each joint For the first The maximum permissible driving torque of each joint; The objective function for optimization is defined as follows: ； in, As an indicator to measure the efficiency of robotic arm movement, For the first The time required for each joint to complete its trajectory movement, where 7 represents the number of joints.

5. The unmanned excavator operation trajectory planning method according to claim 1, characterized in that, After simultaneously collecting the optimized trajectory planning data as a dataset, the method further includes: establishing a simulation model of the unmanned excavator working device to verify whether the optimized trajectory planning data is feasible. If it is feasible, the dataset is used to train the physical neural network algorithm model. If it is not reliable, the trajectory is re-optimized. The simulation model of the unmanned excavator working device includes the kinematic model and the dynamic model of the working device; The kinematic model is a 4-DOF kinematic model of the unmanned excavator working device established using the DH method, including a forward kinematic model and an inverse kinematic model. The dynamic model is a multi-rigid-body dynamic model of the unmanned excavator working device established using the Lagrange equation.

6. The unmanned excavator operation trajectory planning method according to claim 1, characterized in that, In the forward propagation loss calculation step for each training batch of standardized input and target data, a loss function is used to calculate the loss. This loss function includes: data loss, physical constraint loss, performance loss, and torque constraint loss, expressed as follows: ； in, For loss function, For data loss, These are the weighting coefficients for the physical constraint loss. For physical constraint loss, The weighting coefficients for performance loss For performance loss, The weighting coefficient for torque constraint loss. For torque constraint loss; The data loss The formula is as follows: ； in, For Huber's losses, This is the mean square error loss; The physical constraint loss The formula is as follows: ； in, For angle limit constraints, For speed constraints, For acceleration constraints, To add jerk constraints; The performance loss The formula is as follows: ； in, Number the joints. For the first Time deviation of each joint; The torque constraint loss The formula is as follows: ； in, The total number of torque samples. For the first Each joint at any time The actual torque, For the first The maximum permissible rated torque of each joint; ； in, Let be the minimum allowable angle for the j-th joint. Let be the actual joint angle of the j-th joint at time t. Let be the maximum allowed angle of the j-th joint, and Softplus() be a non-linear activation function; ； Where n is the total number of time steps for the trajectory. Let be the angular velocity of the joint at time t. This represents the maximum permissible angular velocity of the joint. ； in, Let be the angular acceleration of the joint at time t. This represents the maximum permissible angular acceleration of the joint. ； in, Let be the angular acceleration of the joint at time t. The maximum permissible angular acceleration of the joint; ； ； in: The total number of samples, For the first The true value of each sample For the first The model prediction value for each sample. This is the threshold parameter for Huber loss.

7. The unmanned excavator operation trajectory planning method according to claim 1, characterized in that, The input and target data of the training and validation sets are standardized using the following formula: ； ； in, The numerical values ​​for each feature dimension, The standardized values ​​for each feature dimension, The mean value for each feature dimension. The corrected standard deviation for each feature dimension. The original standard deviation of each feature dimension Here is the numerical stability constant. Set a minimum threshold for the standard deviation.

8. The unmanned excavator operation trajectory planning method according to claim 1, characterized in that, The configuration of the neural network training parameters includes: using the AdamW optimizer, setting the initial learning rate to 0.0005, and setting the weight decay to... Set the maximum number of training rounds to 120 and the gradient clipping threshold to 1.

0.

9. A device for planning the working trajectory of an unmanned excavator, characterized in that, include: The task input and decomposition module is used to decompose the input task into mining path points. The job trajectory planning module is used to: plan the trajectory between every two excavation path points in a single job using a fifth-order polynomial, and output the trajectory of a single job. The job trajectory optimization module is used to optimize the trajectory of a single job using the particle swarm optimization algorithm with time as the optimization parameter, and at the same time collect the optimized trajectory planning data as a dataset. The physical neural network algorithm module is used to: plan the operation trajectory of the unmanned excavator using a physical neural network algorithm model and generate a complete excavation trajectory; The physical neural network algorithm model training module is used to: divide the dataset into a training set and a validation set, and to standardize the input data and target data of the training set and the validation set. Configure neural network training parameters, including gradient clipping thresholds; for each training batch, perform forward propagation to calculate the loss on the standardized input data and target data, backpropagate to update the neural network training parameters, and apply gradient clipping; after each round of training, use the validation set to evaluate the performance of the physical neural network algorithm model and record the validation loss, adjust the learning rate according to the validation loss, and if the validation loss does not decrease for several consecutive rounds, trigger the early stopping mechanism to terminate the training. After training is terminated, the model with the minimum validation loss is saved as the completed physical neural network algorithm model.

Images