Unmanned clamping and holding vehicle goods stacking deviation learning method based on fusion of physical model and neural network

Abstract

The invention discloses an unmanned clamping and holding vehicle goods stacking deviation learning method based on fusion of a physical model and a neural network. The method comprises the following steps that 1, a vehicle clamping and holding vehicle dynamics micro-twinborn model is constructed to predict a clamping and holding vehicle track, and a track prediction result Ypy is obtained; step 2, constructing a trajectory prediction model based on a neural network to predict a clamping and holding vehicle trajectory to obtain a trajectory prediction result Yta; 3, an interactive multimode fusion module is built, Ypy and Yta are fused, a fused trajectory prediction result Yfso is output, and Yfso = wp * Ypy + wn * Yta; 4, the intersection point of the Yfso and the transverse extension line of the cotton bale goods shelf is a predicted goods stacking point, and the distance between the goods stacking point and the expected goods placing point is the goods stacking deviation of the clamping and holding truck.

Description

Technical Field

[0001] The present invention relates to the technical field of driverless clamping trucks, and particularly to a learning method for the deviation of goods stacking of a driverless clamping truck based on the fusion of a physical model and a neural network, aiming to improve the accuracy and efficiency of cotton bale stacking during the cotton warehousing process of a driverless clamping truck, reduce the operation risk, and promote the process of agricultural modernization. Background Technique

[0002] With the rapid development of automation and intelligent technologies, the application of driverless technology in the fields of logistics and warehousing is becoming increasingly widespread. As a key device in an automated warehousing system, a driverless clamping truck can significantly improve the efficiency and accuracy of goods handling and reduce the risk of manual operation. However, in practical applications, the driverless clamping truck still faces many challenges during the goods stacking process. Especially in complex environments and changing conditions, how to ensure the accuracy and stability of goods stacking has become an urgent problem to be solved.

[0003] Shao Junkai et al. proposed an adaptive PID path tracking control algorithm based on reinforcement learning. First, the kinematic model of an articulated vehicle was derived, and a model of the deviation between the actual driving path and the reference path was established. An adaptive PID path tracking controller based on reinforcement learning was designed. This controller takes the lateral position deviation, heading angle deviation, and curvature deviation as inputs and the steering angle control quantity as the output, and adaptively tunes the PID parameters online through the reinforcement learning algorithm. (Shao Junkai, Zhao Xuan, Yang Jue, Zhang Wenming, Kang Yiting, Zhao Xinxin. Reinforcement Learning Path Tracking Control Algorithm for Driverless Articulated Vehicles [J]. Transactions of the Chinese Society for Agricultural Machinery, 2017, 48(3): 376-382)

[0004] Fujii et al. proposed a trajectory prediction model based on a two-block RNN to handle the problem of incomplete trajectory prediction caused by missed detections. This model draws on the Bayesian filtering framework and processes the cases of successful and failed detections through two RNNs respectively to optimize the estimation of the hidden state. Experimental results show that this method has a significant performance improvement compared with the baseline method on the public datasets ETH and UCY. (R. Fujii, J. Vongkulbhisal, R. Hachiuma and H. Saito, "A Two-Block RNN-Based Trajectory Prediction From Incomplete Trajectory," in IEEE Access, vol. 9, pp. 56140-56151, 2021, doi: 10.1109 / ACCESS.2021.3072135.)

[0005] Lan You et al. proposed an extended sequence-to-sequence model based on AIS data. The gated recurrent unit (GRU) network encodes the historical spatio-temporal sequence into a context vector, which not only preserves the sequential relationship between trajectory positions but also alleviates the gradient descent problem. The GRU network serves as the decoder and outputs the target trajectory position sequence. Real AIS data from the Chongqing section and Wuhan section of the Yangtze River are selected as the typical experimental areas for evaluation. The proposed ST-Seq2Seq model is compared and tested with the LSTM-RNN and GRU-RNN baseline models in the short-term trajectory prediction experiment. A 10-minute historical trajectory sequence is used to predict the next 5-minute trajectory sequence. The overall research results show that when using the recursive method to predict the sequence of continuous trajectory points, the accuracy of the LSTM and GRU networks decreases as the number of predicted trajectory points increases. In contrast, the extended sequence-to-sequence model shows satisfactory stability in different ship channels. (L. You et al., "ST-Seq2Seq: A Spatio-Temporal Feature-Optimized Seq2Seq Model for Short-Term Vessel Trajectory Prediction," in IEEE Access, vol. 8, pp. 218565-218574, 2020, doi: 10.1109 / ACCESS.2020.3041762.)

[0006] In summary, the traditional control system of unmanned clamping vehicles mainly relies on preset physical models and control algorithms. Although these methods can ensure basic operation requirements to a certain extent, they often show great limitations when facing internal dynamic changes of the vehicle, such as vehicle speed steering response and precision uncertainty. These uncertain factors will lead to deviations during the cargo stacking process, thereby affecting the overall operation efficiency and safety. The data-driven approach provides a new idea for solving the control problem of unmanned clamping vehicles in complex environments. However, most of the existing data-driven methods rely on a large amount of training data and have problems such as insufficient model generalization ability and high computational complexity in practical applications. In addition, relying solely on data-driven methods often lacks sufficient robustness and interpretability when facing new environments and emergencies.

[0007] To overcome the above problems, this patent proposes a method for learning cargo stacking deviation of unmanned clamping vehicles based on the fusion of physical models and neural networks. Summary of the Invention

[0008] The purpose of the present invention is to provide a method for learning cargo stacking deviation of unmanned clamping vehicles based on the fusion of physical models and neural networks to address the technical defects existing in the prior art.

[0009] The technical solution adopted to achieve the purpose of the present invention is as follows:

[0010] A learning method for the deviation of goods stacking by an unmanned clamping vehicle based on the fusion of a physical model and a neural network, comprising the following steps:

[0011] Step 1, constructing a dynamic micro-twin model of the vehicle clamping vehicle to predict the trajectory of the clamping vehicle to obtain a trajectory prediction result Y py :

[0012] The dynamic micro-twin model of the vehicle clamping vehicle includes a neural network model Neural network model And a kinematic model of the clamping vehicle. The neural network model Predicts the front wheel steering angle to obtain δ a (t). The neural network model Predicts the vehicle speed to obtain v a (t). The kinematic model of the clamping vehicle combines dynamic factors and uses δ a (t) and the tire cornering stiffness to calculate the front and rear wheel cornering angles and the front and rear wheel lateral forces, and then combines δ a (t), v a (t) to predict the trajectory of the clamping vehicle and obtain a trajectory prediction result Y py ;

[0013] Step 2, constructing a trajectory prediction model based on a neural network and training the trajectory prediction model based on a neural network using the driving data of the unmanned clamping vehicle under different working conditions. The trajectory prediction model based on a neural network predicts the trajectory of the clamping vehicle to obtain a trajectory prediction result Y ta :

[0014] The trajectory prediction model based on a neural network includes an input layer, an LSTM encoder, an attention layer, and an output layer. The historical feature sequence of the clamping vehicle is used as the input, and the trajectory prediction result Y ta ;

[0015] Step 3, building an interactive multi-modal fusion module to fuse the trajectory prediction result Y py obtained in Step 1 ta and the trajectory prediction result Y fso obtained in Step 2 fso , and output the fused trajectory prediction result Y p , Y py = w n ·Y ta , w p , w n are the weights of the dynamic micro-twin model of the clamping vehicle and the weights of the trajectory prediction model based on a neural network that are dynamically adjusted. w p + wn = 1;

[0016] Step 4, the trajectory prediction result Y obtained in Step 3 fso The intersection point with the horizontal extension line of the bale rack is the predicted cargo stacking point, and the distance between the predicted cargo stacking point and the desired stacking point is the deviation of the forklift's cargo stacking.

[0017] In the above technical solution, the neural network model is:

[0018]

[0019] where: δ a (t) is the front wheel steering angle predicted by the neural network, δ d (t) is the target front wheel steering angle, is the change rate of the target steering angle, δ a (t - 1) is the actual steering angle at the previous moment.

[0020] In the above technical solution, the neural network model is:

[0021]

[0022] where: v a (t) is the vehicle speed predicted by the neural network, v d (t) is the target vehicle speed, is the change rate of the target vehicle speed, v a (t - 1) is the actual vehicle speed at the previous moment.

[0023] In the above technical solution, the front wheel side slip angle the rear wheel side slip angle where: v x , v y are the longitudinal and lateral speeds of the forklift in the vehicle body coordinate system, l f , l r are the distances from the center of mass of the forklift to the front axle and the rear axle, ω z is the yaw angular velocity of the forklift.

[0024] In the above technical solution, the front wheel lateral force F yf = C f ·α f , the rear wheel lateral force F yr = C r ·α r , where, C f , C r are the side slip stiffnesses of the front wheel and the rear wheel.

[0025] In the above technical solution, in the kinematic model of the clamping vehicle, the change rates of the position (x, y) in the global coordinate system are updated through v a (t) and the heading angle θ:

[0026]

[0027] where: v y is the lateral velocity and θ is the heading angle;

[0028] The change rate of the heading angle θ is updated through the vehicle speed v a (t) and the front wheel steering angle δ a (t):

[0029]

[0030] where: L is the wheelbase of the clamping vehicle, L = l f + l r ;

[0031] The change rate of the longitudinal velocity in the vehicle body coordinate system and the change rate of the lateral velocity are calculated by the following formulas:

[0032]

[0033] where: a x is the longitudinal acceleration of the clamping vehicle, m is the mass of the clamping vehicle, v x , v y are the longitudinal and lateral velocities in the vehicle body coordinate system, ω z is the yaw angular velocity;

[0034] The dynamic change of the yaw angular velocity ω z is calculated by torque balance: where: I

[0035]

[0036] where: I z is the moment of inertia of the clamping vehicle about the center of mass;

[0037] The state of the clamping vehicle is gradually calculated by the Euler method to achieve trajectory prediction:

[0038]

[0039] where: q = [x, y, θ, v x , v y , ω z is the state variable of the clamping vehicle, is the change rate of the clamping vehicle state;

[0040] Trajectory prediction results of the clamping vehicle dynamics micro-twin model: Y py = [y 1 , y 2 ,..., y T , where y t represents the position predicted by the dynamics micro-twin model at time t, t = 1, 2,..., T, T is the number of predicted time steps, and each position y t is a two-dimensional coordinate vector y t = [x t , y t , x t , y t is the position in the global coordinates at time t.

[0041] In the above technical solution, the historical feature sequence includes the clamping vehicle pose, speed, accelerator pedal travel, and brake pedal travel.

[0042] In the above technical solution, the input of the input layer is the historical feature sequence X = [x 1 , x 2 ,..., x T of the clamping vehicle;

[0043] The LSTM encoder captures the dynamic information in the time series: the input sequence is X = [x 1 , x 2 ,..., x T , and the hidden state is calculated through the LSTM encoder:

[0044] h t , c t = LSTM(x t , h t-1 , c t-1 )

[0045] where h t is the hidden state at time step t; c t is the cell state at time step t; x t is the input feature at time step t;

[0046] The attention layer calculates the importance weights of each time step t in the input sequence, dynamically weights the relevant information, and for each hidden state h t , calculates the attention weight α t and the context vector c t :

[0047] e t = v T tanh(W h h t + bh )

[0048]

[0049] where W h and b h are the trainable parameters of the attention network, v is a trainable vector, and e t is the attention score at time t, used to evaluate the contribution of the hidden state h t to the final output. e i is the attention score at time step i; α t is the attention weight at time step t, satisfying c t is the context vector, representing the weighted sum of all hidden states;

[0050] The output layer: combines the context vector c t with the last hidden state h T of the LSTM and predicts the future trajectory points through a fully connected layer:

[0051] y′t = FC([c t ; h T )

[0052] where [c t ; h T represents concatenating the context vector c t and the last hidden state h T ; y′ t is the output of the trajectory prediction model based on the neural network, representing the predicted future trajectory points, t = 1, 2,..., T, and the trajectory prediction result Y ta of the trajectory prediction model based on the neural network is 1 = [y′ 2 , y′ T ,...,[y′

[0053] In the above technical solution, the objective of training the trajectory prediction model based on the neural network is to minimize the mean square error MSE between the prediction deviation and the actual deviation:

[0054]

[0055] where n is the number of samples, yi is the actual trajectory sequence value of the forklift, is the predicted forklift trajectory sequence value of the trajectory prediction model based on the neural network, and the parameters of the trajectory prediction model based on the neural network are adjusted through the gradient descent optimization algorithm.

[0056] In the above technical solution, w p , w nAdjusted by the following method:

[0057]

[0058] Where: E phys = |Y tu - Y py | is the error of the dynamic micro-twin model of the forklift, and Y tu is the actual trajectory of the forklift; γ is the weight adjustment rate.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] Utilize the high-precision description ability of forklift dynamics and the non-linear learning ability of neural networks to achieve accurate prediction and dynamic adjustment of the deviation of goods stacking by unmanned forklifts. By fusing the physical model of the forklift and the prediction results of the neural network, this method can dynamically optimize the model weights and improve the robustness and adaptability of the system, specifically as follows:

[0061] 1. The present invention combines the advantages of the physical dynamics model and the neural network model, which can not only utilize the high-precision characteristics of the physical model but also make up for the limitations of the physical model in complex environments through the non-linear learning ability of the neural network.

[0062] 2. Through multi-sensor data acquisition and fusion, the dynamic response characteristics of the forklift can be accurately described to adapt to different loads and road conditions.

[0063] 3. The dynamic weight adjustment mechanism enables the present invention to automatically optimize the contribution ratio of the dynamic micro-twin model of the forklift and the trajectory prediction model based on the neural network when the prediction error is large, thereby improving the prediction accuracy and system robustness.

[0064] 4. Real-time feedback closed-loop control is achieved, and after the forklift executes the movement, it can correct the model parameters according to the actual trajectory to achieve continuous optimization.

[0065] 5. Improve the efficiency and accuracy of goods stacking by unmanned forklifts under complex working conditions, reduce the goods stacking deviation rate, and enhance the overall performance of logistics automation. Brief Description of the Drawings

[0066] Figure 1 The structural framework diagram of the present invention is shown. Detailed Embodiments

[0067] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0068] A learning method for the deviation of goods stacking by an unmanned clamping vehicle based on the fusion of physical models and neural networks, comprising the following steps:

[0069] Step 1: Construct a dynamic micro twin model of the clamping vehicle to predict the trajectory of the clamping vehicle and describe the motion law of the clamping vehicle during the goods stacking process from a physical perspective. Based on the dynamic equation and initial state of the clamping vehicle, combined with the operation instructions of the clamping vehicle (such as steering wheel angle, acceleration, and braking instructions), calculate the position and attitude of the clamping vehicle at future moments.

[0070] The dynamic micro twin model of the clamping vehicle includes a neural network model and a neural network model as well as the kinematic model of the clamping vehicle. The neural network model is used to learn the steering response, and the neural network model is used to learn the vehicle speed response. The kinematic model of the clamping vehicle combines dynamic factors, calculates the side slip angle and lateral force using the tire cornering stiffness, and predicts the trajectory.

[0071] (1) The neural network model is:

[0072]

[0073] where: δ a (t) is the front wheel steering angle predicted by the neural network (the steering angle obtained after the chassis response delay and dynamic characteristics), and δ d (t) is the target front wheel steering angle. is the change rate of the target steering angle, used to capture the steering dynamic characteristics. δ a (t - 1) is the actual steering angle at the previous moment, used to describe the historical state of the steering response.

[0074] (2) The neural network model is:

[0075]

[0076] where: v a (t) is the vehicle speed predicted by the neural network (the vehicle speed obtained after the power system delay and dynamic characteristics). v d (t) is the target vehicle speed (output of the PID controller). is the change rate of the target vehicle speed, used to capture the dynamic characteristics of the speed response. v a (t - 1) is the actual vehicle speed at the previous moment, used to describe the historical state of the vehicle speed response.

[0077] (3) Calculation of the side slip angle and lateral force:

[0078] Front wheel side slip angle:

[0079] Rear wheel sideslip angle:

[0080] Where: v x , v y are the longitudinal and lateral speeds of the clamping vehicle in the vehicle body coordinate system. l f , l r are the distances from the center of mass of the clamping vehicle to the front axle and the rear axle. ω z is the yaw angular velocity of the vehicle which is a clamping vehicle.

[0081] Front wheel lateral force: F yf = C f ·α f ;

[0082] Rear wheel lateral force: F yr = C r ·α r ;

[0083] Where: C f , C r are the cornering stiffnesses of the front and rear wheels, describing the linear relationship between the lateral force of the tire and the sideslip angle.

[0084] (4) Combine the output δ of the neural network model a (t), the output v of the neural network model a (t) and C f , C r to perform trajectory prediction:

[0085] The rate of change of the position (x, y) in the global coordinate system is updated by the predicted vehicle speed v of the neural network model a (t) and the heading angle θ:

[0086]

[0087] Where: v y is the lateral speed (to be calculated from the dynamic equation), and θ is the heading angle.

[0088] The rate of change of the heading angle θ is updated by the predicted vehicle speed v of the neural network model a (t) and the predicted front wheel steering angle δ of the neural network model a (t):

[0089]

[0090] Where: L is the wheelbase of the forklift, and L = l f + l r .

[0091] The rate of change of the longitudinal speed in the vehicle body coordinate system and the rate of change of the lateral speed are calculated as follows:

[0092]

[0093] Where: a x is the longitudinal acceleration of the forklift, m is the mass of the forklift, v x , v y are the longitudinal and lateral speeds in the vehicle body coordinate system, and ω z is the yaw angular velocity.

[0094] The dynamic change of the yaw angular velocity ω z is calculated from the moment balance:

[0095]

[0096] Where: I z is the moment of inertia of the forklift about the center of mass;

[0097] Numerical integration to calculate the trajectory: The state (position, heading angle, speed) of the forklift is calculated step by step through the Euler method to achieve trajectory prediction:

[0098]

[0099] Where: q = [x, y, θ, v x , v y , ω z is the state variable of the forklift. is the rate of change of the forklift state.

[0100] The trajectory prediction result of the forklift dynamics micro-twin model:

[0101] Y py = [y 1 , y 2 ,..., y T

[0102] Where y t represents the position predicted by the dynamics micro-twin model at time t, where t = 1, 2,..., T and T is the number of predicted time steps. Each position y t is a two-dimensional coordinate vector:

[0103] y t = [x t , y​t .

[0104] Step 2, build a trajectory prediction model based on a neural network. The trajectory prediction model based on a neural network includes an input layer, an LSTM encoder, an attention layer, and an output layer, where:

[0105] The input of the input layer is the historical feature sequence X = [x 1 , x 2 ,..., x T , where x 1 , x 2 ,..., x T are features such as the pose, speed, accelerator pedal travel, and brake pedal travel of the clamping truck.

[0106] The LSTM encoder captures the dynamic information in the time series: the input sequence is X = [x 1 , x 2 ,..., x T , and the hidden state is calculated through the LSTM encoder:

[0107] h t , c t = LSTM(x t , h t-1 , c t-1 )

[0108] where h t is the hidden state at time step t; c t is the cell state at time step t; x t is the input feature at time step t.

[0109] The attention layer calculates the importance weights of each time step t in the input sequence, dynamically weights the relevant information, and for each hidden state h t , calculates the attention weight α t and the context vector c t :

[0110] e t = v T tanh(W h h t + b h )

[0111]

[0112] where W h and b h are trainable parameters of the attention network; v is a trainable vector; e t is the attention score at time t, used to evaluate the hidden state h tContribution to the final output, e i is the attention score at time step i; α t is the attention weight at time step t, satisfying c t is the context vector, representing the weighted sum of all hidden states.

[0113] The output layer: combines the context vector c t with the last hidden state h of the LSTM T to predict future trajectory points through a fully connected layer:

[0114] y′ t = FC([c t ; h T )

[0115] where [c t ; h T represents concatenating the context vector c t and the last hidden state h T ; y′ t is the output of the trajectory prediction model based on the neural network, representing the predicted future trajectory points, t = 1, 2,..., T, and the trajectory prediction result Y of the trajectory prediction model based on the neural network ta = [y′ 1 , y′ 2 ,..., y′ T .

[0116] Training and optimization of the trajectory prediction model based on the neural network:

[0117] Use the driving data of the unmanned clamping vehicle under different working conditions to train the trajectory prediction model based on the neural network. The training objective is to minimize the mean square error (MSE) between the prediction deviation and the actual deviation:

[0118]

[0119] where n is the number of samples. y i is the actual trajectory sequence value of the clamping vehicle. is the trajectory sequence value of the clamping vehicle predicted by the trajectory prediction model based on the neural network. Adjust the parameters of the trajectory prediction model based on the neural network through the gradient descent optimization algorithm.

[0120] Step 3, build an interactive multi-modal fusion module to fuse the trajectory prediction result Y of the clamping vehicle dynamics micro-twin model in Step 1 py and the trajectory prediction result Y of the trajectory prediction model based on the neural network in Step 2 ta , dynamically adjust the weights, and achieve high-precision prediction.

[0121] (1) Feature fusion:

[0122] Y fso = w p ·Y py + w n ·Y ta

[0123] Where Y fso is the fused trajectory prediction result, w p , w n are the weights of the dynamically adjusted kinematic micro-twin model of the clamping vehicle and the weights of the trajectory prediction model based on neural network, satisfying w p + W n = 1.

[0124] (2) Dynamically adjust the weights through prediction error feedback:

[0125]

[0126] Where: E phys = |Y tu - Y py | is the error of the kinematic micro-twin model of the clamping vehicle, i.e., error compensation. Use this error compensation to update w p , w n , Y tu is the actual trajectory of the clamping vehicle; γ is the weight adjustment rate.

[0127] (3) Real-time feedback and correction:

[0128] After the clamping vehicle executes, the sensor collects the actual trajectory Y tu of the clamping vehicle in real time, compares it with the predicted value Y py of the kinematic micro-twin model of the clamping vehicle, and updates the weights w p of the kinematic micro-twin model of the clamping vehicle and the weights W n of the trajectory prediction model based on neural network.

[0129] Step 4, the intersection point of the trajectory prediction result Y fso obtained in Step 3 and the lateral extension line of the bale rack is the predicted cargo stacking point, and the distance between the cargo stacking point and the desired stacking point is the cargo stacking deviation of the clamping vehicle.

[0130] Step 5, System integration and testing

[0131] (1) System integration: Integrate the kinematic micro-twin model of the clamping vehicle, the trajectory prediction model based on neural network, and the interactive multi-modal fusion module into the control system of the unmanned clamping vehicle, and use ROS for system integration.

[0132] (2) Simulation test: Conduct tests in a simulation environment, using MATLAB and Simulink for simulation tests to verify the performance of the algorithm under different load weights and road surface conditions.

[0133] (3) Real vehicle test: Conduct tests in an actual operation environment, using an unmanned clamping vehicle for real vehicle tests to verify the effectiveness of the algorithm in actual applications, and optimize and adjust according to the test results.

[0134] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A method for learning cargo stacking deviation of an unmanned clamping vehicle based on the fusion of physical model and neural network, characterized in that: The following steps are involved: Step 1: Construct a vehicle clamping vehicle dynamics micro twin model to predict the clamping vehicle trajectory and obtain the trajectory prediction result Y py : The vehicle clamping vehicle dynamics micro twin model includes a neural network model Neural Network Model And the kinematic model of the clamping vehicle, the neural network model Predict the front wheel steering angle to get δ a (t), the neural network model Predict the vehicle speed to get v a (t), the kinematic model of the clamping vehicle combines the dynamic factors and uses δ a (t) and tire cornering stiffness to calculate the front and rear wheel side slip angles and the front and rear wheel lateral forces, and then combine δ a (t), v a (t) Predict the trajectory of the clamping vehicle and obtain the trajectory prediction result Y py ; Step 2: construct a trajectory prediction model based on a neural network, and use the driving data of the unmanned clamping vehicle under different working conditions to train the trajectory prediction model based on the neural network. The trajectory prediction model based on the neural network predicts the trajectory of the clamping vehicle to obtain a trajectory prediction result Y ta : The trajectory prediction model based on neural network includes an input layer, an LSTM encoder, an attention layer and an output layer. It takes the historical feature sequence of the clamping vehicle as input and outputs the trajectory prediction result Y ta ; Step 3: Build an interactive multi-modal fusion module to integrate the trajectory prediction result Y obtained in step 1 py And the trajectory prediction result Y obtained in step 2 ta , output the fused trajectory prediction result Y fso , Y fso =w p ·Y py +w n ·Y ta , w p ,w n is the weight of the dynamically adjusted micro-twin model of the clamping vehicle dynamics and the weight of the trajectory prediction model based on the neural network, w p +w n =1; Step 4: The trajectory prediction result Y obtained in step 3 fso The intersection of the horizontal extension line of the cotton bale rack and the predicted cargo stacking point is the predicted cargo stacking point, and the distance between the cargo stacking point and the expected cargo release point is the cargo stacking deviation of the clamping vehicle.

2. The method for learning cargo stacking deviation of an unmanned clamping vehicle according to claim 1, characterized in that: The neural network model for: Where: a (t) is the front wheel steering angle predicted by the neural network, δ d (t) is the target front wheel steering angle, is the rate of change of the target steering angle, δ a (t-1) is the actual steering angle at the previous moment.

3. The method for learning cargo stacking deviation of an unmanned clamping vehicle according to claim 1, characterized in that: Neural Network Model for: Where: v a (t) is the vehicle speed predicted by the neural network, v d (t) is the target vehicle speed, is the rate of change of target vehicle speed, v a (t-1) is the actual vehicle speed at the previous moment.

4. The method for learning cargo stacking deviation of an unmanned clamping vehicle according to claim 1, characterized in that: Front wheel slip angle Rear wheel slip angle Where: v x ,v y are the longitudinal and lateral velocities of the clamping vehicle in the vehicle body coordinate system, l f ,l r is the distance from the center of mass of the clamped vehicle to the front and rear axles, ω z is the yaw angular velocity of the clamping vehicle.

5. The method for learning cargo stacking deviation of an unmanned clamping vehicle according to claim 1, characterized in that: Front wheel lateral force F yf =C f α f , rear wheel lateral force F yr =C r α r , where C f ,C r is the cornering stiffness of the front and rear wheels.

6. The method for learning cargo stacking deviation of an unmanned clamping vehicle according to claim 1, characterized in that: In the kinematic model of the clamping vehicle, the rate of change of the position (x, y) in the global coordinates By v a (t) and heading angle θ are updated: Where: v y is the lateral velocity, θ is the heading angle; The rate of change of heading angle θ By vehicle speed v a (t) and the front wheel steering angle δ a (t) Update: Where: L is the wheelbase of the clamping vehicle, L = l f +l r ; The rate of change of longitudinal velocity in the vehicle coordinate system and the rate of change of lateral velocity The calculation formula is as follows: Among them: a x is the longitudinal acceleration of the clamping vehicle, m is the mass of the clamping vehicle, v x ,v y The longitudinal and lateral velocities in the vehicle coordinate system, ω z Yaw angular velocity; Yaw angular velocity ω z Dynamic changes Calculated from moment balance: Where: I z is the moment of inertia of the clamping vehicle around its center of mass; The state of the clamping vehicle is calculated step by step through the Euler method to achieve trajectory prediction: Where: q = [x, y, θ, v x ,v y ,ω z ] is the state variable of the clamping car, is the rate of change of the clamping vehicle state; Trajectory prediction results of the clamp vehicle dynamics micro-twin model:Y py =[y1,y2,…,y T ], where y t represents the position predicted by the dynamic micro-twin model at time t, t = 1, 2, ..., T, T is the number of predicted time steps, and each position y t is a two-dimensional coordinate vector y t =[x t ,y t ], x t ,y t is the position in global coordinates at time t.

7. The method for learning cargo stacking deviation of an unmanned clamping vehicle according to claim 1, characterized in that: The historical feature sequence includes the clamping vehicle's position, speed, accelerator pedal travel, and brake pedal travel.

8. The method for learning cargo stacking deviation of an unmanned clamping vehicle according to claim 1, characterized in that: The input layer is the historical feature sequence X = [x1, x2, ..., x T ]; The LSTM encoder captures dynamic information in the time series: the input sequence is X = [x1, x2, ..., x T ], and calculate the hidden state through the LSTM encoder: h t ,c t =LSTM(x t ,h t-1 ,c t-1 ) where h t is the hidden state at time step t; c t is the cell state at time step t; x t is the input feature at time step t; The attention layer calculates the importance weight of each time step t in the input sequence, dynamically weighting the relevant information, for each hidden state h t , calculate the attention weight α t and the context vector c t : e t =v T tanh(W h h t +b h ) Where W h and b h is a trainable parameter of the attention network, v is a trainable vector, e t is the attention score used to evaluate the hidden state h t Contribution to the final output; α t is the attention weight at time step t, satisfying c t is the context vector, representing the weighted sum of all hidden states; The output layer: the context vector c t With the last hidden state h of LSTM T Combined, the future trajectory points predicted by the fully connected layer: y′ t =FC([c t ;h T ]) Where [c t ;h T ] means to convert the context vector c t and the final hidden state h T splicing; y′ t is the output of the trajectory prediction model based on the neural network, representing the predicted future trajectory point, t = 1, 2, ..., T, the trajectory prediction result Y of the trajectory prediction model based on the neural network ta =[y′1,y′2,…,y′ T ].

9. The method for learning cargo stacking deviation of an unmanned clamping vehicle according to claim 1, characterized in that: The goal of training the trajectory prediction model based on the neural network is to minimize the mean square error (MSE) between the predicted deviation and the actual deviation: Where n is the number of samples, y i is the actual trajectory sequence value of the clamping vehicle, It is the trajectory sequence value of the clamping vehicle predicted by the trajectory prediction model based on the neural network, and the parameters of the trajectory prediction model based on the neural network are adjusted by the gradient descent optimization algorithm.

10. The method for learning cargo stacking deviation of an unmanned clamping vehicle according to claim 1, characterized in that: w p ,w n Adjust by: In n =1-in p Where: E phys =|Y tu -Y py | is the error of the micro-twin model of the clamping vehicle dynamics, Y tu is the actual trajectory of the clamping vehicle; γ is the weight adjustment rate.

Images