Articulated vehicle kinematic model modeling method

Abstract

An articulated vehicle kinematics model modeling method aims at the problems of model misalignment, control lag and the like under complex working conditions, articulated vehicle kinematics modeling based on a matrix memory long-short-term memory network (Pi-mLSTM) driven by physical information can accurately control an articulated vehicle under the complex working conditions, and the method specifically comprises six steps A-F, the method has the advantages that rapid online updating and accurate prediction of the kinematics model can be realized under complex working conditions, the capturing capability of the model on time-varying parameters is remarkably improved, meanwhile, a physical information driving strategy is utilized, it is ensured that the model always conforms to the basic rule of vehicle kinematics in the training and reasoning process, and the reliability of the model is improved. The precision of the kinematic model of the articulated vehicle can be obviously improved, and the real-time precise control of the vehicle state can be realized, so that the operation stability and the control performance of the articulated vehicle under the complex working condition are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and in particular to a method for modeling the kinematics of an articulated vehicle. Background Technology

[0002] Articulated vehicles are widely used in mining transportation due to their flexible steering characteristics. However, in actual operation, these vehicles are affected by multiple factors such as sudden changes in road conditions, dynamic load distribution, and nonlinear coupling of the articulated mechanism, resulting in kinematic models that often exhibit strong time-varying characteristics. Traditional kinematic modeling methods based on fixed parameters have two inherent drawbacks: First, offline static models are difficult to adapt to real-time operating parameter drifts such as changes in steering friction torque and tire slip ratio fluctuations, leading to an exponential decrease in model prediction accuracy over time. Second, while existing recurrent neural network modeling methods can capture dynamic characteristics, they are limited by the gating delay effect and parameter freezing mechanism of LSTM units, failing to meet the real-time requirements of onboard edge calculators. This superposition effect of model inaccuracy and control lag directly causes the cumulative amplification of articulation angle tracking errors, which in severe cases can lead to dynamic oscillations or even loss of control of the vehicle.

[0003] To address the aforementioned issues, several patents have been published, including CN 202410975931.7, "A Path Tracking Method for an Articulated Driving Vehicle in an Underground Mine." This method involves the following steps: In a forward-moving state, the front pose of the articulated vehicle is acquired as the overall pose for trajectory tracking. The position and attitude information of the front of the vehicle are obtained through onboard sensors and used as the overall position and attitude of the vehicle. A look-ahead point is obtained based on the front pose. The position of the look-ahead point is determined by pushing the vehicle's current position forward a certain distance along the current heading direction. This forward distance is determined based on the vehicle's speed and control response time. A distance is selected to keep the look-ahead point in front of the vehicle, and timely adjustments are made when the vehicle reaches the look-ahead point. The adjustments were made, but the structural characteristics of articulated vehicles were not utilized to improve the effect and performance of path tracking. Furthermore, the pose information of the front and rear segments of the vehicle was not acquired for trajectory tracking to separately obtain the pose of the front and rear segments, resulting in difficulty in capturing the overall motion state and attitude changes of the vehicle, and poor accuracy and stability in path tracking. CN202410774034.X, "A Lateral Control Method and System for an Unmanned Articulated Vehicle," employs unmanned articulated steering plus rear-wheel steering technology. Based on the desired articulation angle calculated by the lateral control algorithm, the inner and outer steering angles of the rear wheels are calculated under the condition of the same turning radius at the front and rear axle centers, achieving more flexible and stable lateral control for unmanned articulated vehicles. Stable, it can deduce the inner and outer steering angles of the rear wheels under the same turning radius at the front and rear axle centers based on the expected articulation angle calculated by the lateral control algorithm, achieving more flexible and stable lateral control for unmanned articulated vehicles; CN202210570780.8 "A parking trajectory planning and tracking control method and system for articulated vehicles", it designs an uncertain nonlinear system based on the kinematic model and tracking deviation model of the articulated vehicle; based on the uncertain nonlinear system, it constructs the objective function and constraint function to obtain the description of the optimization problem of the uncertain nonlinear system; solves the optimization problem of the uncertain nonlinear system to obtain the parking trajectory planning and tracking control scheme; the articulated vehicle according to the parking trajectory planning and tracking control scheme... The vehicle trajectory planning and tracking control scheme enables automatic parking. It considers uncertainties in the parking trajectory planning and tracking control process and uses a scenario tree to represent these uncertainties. The trajectory planning and tracking control process is described using constraint functions and multi-order objective functions. Because multi-order objective functions and constraint functions are more adaptable to the current system model and input parameters, the trajectory planning and tracking control method based on scenario trees is more adaptable than that based on constant objective functions and constant constraints. The planned trajectory and tracking control process better meets the needs of real-world scenarios and has high practicality.

[0004] In summary, while some existing technologies have attempted to improve model adaptability through online parameter identification (such as Extended Kalman Filter, EKF) or fuzzy logic compensation, these methods still have certain limitations. For example, EKF-based methods rely on the Gaussian noise assumption, which is often not true in actual operating conditions; fuzzy logic-based methods suffer from insufficient rule base completeness and real-time performance. Furthermore, while traditional recurrent neural networks (such as LSTM, Long Short-Term Memory) excel at capturing dynamic characteristics, their computational complexity in gating mechanisms and limitations in parameter freezing mechanisms make them difficult to meet the real-time requirements of onboard controllers for online model updates. Such models suffer from a superposition effect of inaccuracies and control lag, further exacerbating the accumulation of articulation angle tracking errors and severely impacting vehicle control performance.

[0005] Therefore, it is particularly urgent and significant to seek an articulated vehicle kinematics modeling method that can solve the dual coupling problem of rapid online parameter self-calibration and real-time optimization of model dynamic matching and control under complex working conditions. Summary of the Invention

[0006] The objective of this invention is to overcome the shortcomings of existing technologies and provide a kinematic modeling method for articulated vehicles. This method enables rapid online updates and accurate predictions of kinematic models under complex operating conditions, significantly improving the model's ability to capture time-varying parameters. Simultaneously, by utilizing a physical information-driven strategy, it ensures that the model always conforms to the basic laws of vehicle kinematics during training and inference. This not only significantly improves the accuracy of articulated vehicle kinematic models but also enables real-time and precise control of vehicle states, thereby effectively enhancing the operational stability and control performance of articulated vehicles under complex operating conditions.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] A kinematic modeling method for articulated vehicles addresses issues such as model inaccuracy and control lag under complex operating conditions. Based on a Physics Informed-matrix memory (Pi-mLSTM) network, this method enables precise control of articulated vehicles under complex conditions. The specific method includes the following steps:

[0009] A. Vehicle motion information collected by vehicle-mounted lidar and vehicle-mounted controller; the raw data is preprocessed after data collection.

[0010] B. Standardize the X and Y data according to formula (1) so that the mean of the data is 0 and the standard deviation is 1.

[0011]

[0012] C. Use sliding windows with window lengths H1 and H2 to continuously sample X and Y, and obtain the sequence of X sampled data and Y sampled data at time t according to formula (2) and formula (3) respectively;

[0013]

[0014] In the formula, Let m be the kinematic parameter or controlled variable of the articulated vehicle at time t. Let be the k-th kinematic parameter of the articulated vehicle at time t;

[0015] D. Calculate the predicted values ​​of the vehicle kinematic parameters according to formula (4). loss function L mLSTM :

[0016]

[0017] E. Based on the state-space equations of the fundamental kinematics of the articulated vehicle, the sum of the squares of the physical residuals at time t+1 is used as the physical loss to construct the physical information loss function L. phy

[0018]

[0019] Where Δ k The residual is obtained by subtracting the physical state variable from the derivative of the discrete approximation of the k-th state variable, where M is the number of state variables with kinematic geometry constraints.

[0020] F. According to formula (6), use the data fitting loss L mLSTM With physical information loss L phy Construct a total loss function and use this function to train the kinematic model.

[0021] L = L mLSTM +λL phy (6)

[0022] Where λ is the weight hyperparameter balancing physical constraints and data fitting.

[0023] This invention overcomes the limitations of traditional methods that separate physical models from data-driven approaches by constructing a hybrid modeling framework with deep embedding of physical information. By transforming the essential laws of vehicle kinematics (such as nonholonomic constraints and the kinematic coupling relationships of articulated mechanisms) into the structural constraints and loss functions of a neural network, the Pi-mLSTM network maintains its data adaptability while strictly adhering to physical conservation laws. This fusion mechanism effectively avoids the "black box" characteristic of purely data-driven models, ensuring reasonable physical consistency even under drastic changes in operating conditions and significantly improving generalization performance in unknown scenarios. Furthermore, considering the strong coupling and dynamic characteristics of multiple parameters in articulated vehicles, the matrix-memory long short-term memory network model can establish the pressure dynamics of the hydraulic system and the tire-ground phase... The interaction and other multi-physics coupling relationships, as well as the dynamic feature correlation across time scales, simultaneously capture the rapidly changing dynamics of the hydraulic system and the slowly changing characteristics of the mechanical structure. Therefore, this multi-level state representation capability enables the model to more accurately reflect the nonlinear mapping relationship between the articulation angle change and the overall vehicle motion, laying the foundation for precise control under complex working conditions. In addition, by reconstructing the computational paradigm of the neural network, adopting a parallel matrix operation strategy, and using an incremental learning mechanism to support the online and rapid updating of model parameters, Pi-mLSTM achieves a balance between modeling accuracy and computational efficiency. While retaining high-dimensional memory capabilities, it significantly reduces the computational burden, ensuring that the control system can respond promptly to sudden working conditions such as load changes and changes in road adhesion conditions, meeting the real-time requirements of the vehicle-mounted embedded platform.

[0024] (1) The innovation of this invention lies in constructing a hybrid modeling framework with physical information constraints, breaking through the limitation of the separation between physical models and data-driven approaches in traditional methods. By encoding the essential laws of vehicle kinematics (such as nonholonomic constraints, motion coupling relationships of articulated mechanisms, etc.) into the model's loss function, the Pi-mLSTM network, while maintaining its data adaptability, forces the model to approximate physical conservation laws through the constraints of the loss function. This mechanism significantly suppresses the physically unreasonable output of purely data-driven models and improves the generalization performance in unknown scenarios.

[0025] (2) To address the strong coupling and dynamic characteristics of articulated vehicles with multiple parameters, a matrix-memory long short-term memory network model can be used to establish multi-physics coupling relationships such as the pressure dynamics of the hydraulic system and the tire-ground interaction, as well as the dynamic feature correlation across time scales, simultaneously capturing the fast-changing dynamics of the hydraulic system and the slow-changing characteristics of the mechanical structure. This multi-level state representation capability enables the model to more accurately reflect the nonlinear mapping relationship between the articulation angle change and the overall motion of the vehicle, laying the foundation for precise control under complex working conditions.

[0026] (3) Pi-mLSTM adopts a parallel matrix operation strategy, which significantly reduces the computational burden while retaining high-dimensional memory capabilities. At the same time, it uses an incremental learning mechanism to support online and rapid updates of model parameters, ensuring that the control system can respond promptly to sudden conditions such as load changes and changes in road surface adhesion conditions, thus meeting the real-time requirements of the vehicle-mounted embedded platform.

[0027] The innovation of this invention lies in achieving accurate modeling of the kinematic characteristics of articulated vehicles through a Pi-mLSTM network, and combining it with a physical information-driven strategy to ensure the robustness and real-time performance of the model in practical applications. This method effectively solves the problems of model inaccuracy and control lag in traditional modeling methods under complex operating conditions, providing reliable technical support for the intelligent control of articulated vehicles. This invention aims to reduce the technical problem that changes in various operating conditions during the operation of articulated vehicles cause the established kinematic model to fail to accurately reflect the actual operating mode of the vehicle, thus hindering precise control. Therefore, the purpose of this invention is to propose a kinematic modeling method for articulated vehicles based on a physical information-driven matrix memory long short-term memory network (Pi-mLSTM), which can establish an accurate kinematic model of articulated vehicles and achieve precise control of articulated vehicles.

[0028] This invention achieves accurate modeling of the kinematic characteristics of articulated vehicles through a Pi-mLSTM network and combines it with a physical information-driven strategy to ensure the robustness and real-time performance of the model in practical applications. This method effectively solves the problems of model inaccuracy and control lag in traditional modeling methods under complex operating conditions, providing reliable technical support for the intelligent control of articulated vehicles. This invention aims to reduce the technical problem that the kinematic model cannot accurately reflect the actual operating mode of the vehicle due to changes in various operating conditions during the operation of articulated vehicles, thus hindering precise control. Therefore, the purpose of this invention is to propose a kinematic modeling method for articulated vehicles based on a physical information-driven matrix memory long short-term memory network (Pi-mLSTM), which can establish an accurate kinematic model of articulated vehicles and achieve precise control of articulated vehicles. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the MLSTM unit structure, which is a hinged vehicle kinematics modeling method proposed in this invention.

[0030] Figure 2 This is a kinematic modeling method for articulated vehicles. A schematic diagram of the kinematic structure of an articulated vehicle is shown.

[0031] Figure 3This is a schematic diagram of the process for modeling articulated vehicle kinematics based on the Pi-mLSTM model.

[0032] Figure 4 This is a schematic diagram illustrating the application of a centrally articulated steering mining truck based on the articulated vehicle kinematics modeling method proposed in this invention.

[0033] Figure 5 This is a schematic diagram of the application system architecture of a centrally articulated steering mining truck based on the articulated vehicle kinematics modeling method proposed in this invention.

[0034] Figure 6 This is a schematic diagram illustrating the application of a centrally articulated steering loader based on the articulated vehicle kinematics modeling method proposed in this invention.

[0035] The present invention will now be described in further detail with reference to the accompanying drawings. Detailed Implementation

[0036] like Figures 1-6 As shown, this invention provides a method for modeling the kinematics of an articulated vehicle, which specifically includes the following steps:

[0037] 1. Collect vehicle motion information, including trajectory information and operation information, from the vehicle-mounted LiDAR and vehicle controller to construct a training dataset. (Select...) The input variable X∈R N×i , The dataset Y∈R is constructed as the output variable. N×j Among them, I TC For vehicle steering control current, I ACC For the accelerator pedal current, I bra Gs represents the braking current, and Gs represents the vehicle gear position. These are the vehicle's kinematic parameters.

[0038] 2. After data collection, the raw data is preprocessed, including the removal and alignment of outlier data;

[0039] 3. Standardize the X and Y data according to the following formula:

[0040]

[0041] 4. Use sliding windows with window lengths H1 and H2 to continuously sample X and Y, and obtain the sequence of X sampled data and Y sampled data at time t according to formula (2) and formula (3) respectively;

[0042] (2); (3);

[0043] In the formula, Let m be the kinematic parameters and operating variables of the articulated vehicle at time t. Let be the k-th kinematic parameter of the articulated vehicle at time t;

[0044] 5. Use the mLSTM method to establish the relationship between the vehicle's kinematic parameters before and after the control data intervention process:

[0045] Figure 1 For the structure of the mLSTM network, the input vector is projected using equation (4) to obtain the input x. t Mapped to query vector q through linear transformation t Key vector k t Value vector v t ,

[0046]

[0047] Where d represents the scaling factor.

[0048] For the current input x t and the hidden state h from the previous moment t-1 The input gate is obtained by performing a linear transformation. Forgotten Gate and output gate The activated i is obtained by using exponential activation. t f t and o t :

[0049]

[0050] Update matrix memory unit C t

[0051] C t =f t C t-1 +i t (v t k t (6)

[0052] Where i t and f t These are the element-wise outputs of the forget gate and the input gate, respectively. C t-1 This is the memory matrix from the previous time step.

[0053] The formula for calculating the normalized state can be obtained from the following formula:

[0054] n t =f t n t-1 +i t kt (7)

[0055] This normalized state can be used to perform denominator normalization when calculating hidden states, preventing numerical explosion or excessive growth.

[0056] Multiply the updated memory matrix Ct by the query vector qt, and then combine this with the output gate to obtain the hidden state:

[0057]

[0058] Where is the linear transformation result of the output gate. This indicates a comparison with the dot product of the normalized state and 1 to ensure that the denominator is not 0.

[0059] The predicted values ​​of vehicle kinematic parameters can be obtained from formula (9).

[0060]

[0061] The predicted values ​​of the vehicle kinematic parameters are calculated according to formula (10). loss function L mLSTM :

[0062]

[0063] 6. Construct physical information constraints:

[0064] Based on the kinematics of articulated vehicles, the following relationship can be obtained between the steering angles of the front and rear frames:

[0065] θ f =θ r +γ (11)

[0066]

[0067] The predicted values ​​are compared with the kinematic physics formulas to obtain the residuals:

[0068]

[0069] or

[0070] The summation range covers all training sample time steps.

[0071] Based on the state-space equations of the articulated vehicle kinematics, the sum of the squares of the physical residuals from time t+1 to t+H2 is used as the physical loss, and the physical loss function is constructed as follows:

[0072]

[0073] During training, the total loss is typically the data fitting loss L.mLSTM With physical loss L phy Weighted sum: L = L mLSTM +λL phy

[0074] Where λ is the weight hyperparameter balancing physical constraints and data fitting.

[0075] 7. Adaptive online incremental learning method for changing operating conditions:

[0076] (1) The model shows the vehicle state at time t. The predicted value, Y(t), is the true value obtained from the sensor or actual measurement. The interval average error is calculated using formula (15):

[0077]

[0078] (2) Set a threshold ò, when an error e is detected at a certain time t t When the threshold is greater than ò, it indicates that the current model's predictive ability is insufficient and an incremental learning mechanism needs to be triggered. The value of the threshold ò can be adjusted according to the vehicle control accuracy requirements, sensor noise levels, etc.

[0079] (3) In order to update the model online, it is necessary to retain the original observation data D(t) for the most recent period of time. When incremental learning is triggered, the recent data is extracted from the sliding window with a fixed size H to form the dataset D. new D new ={X(τ1),Y(τ2)∣τ1∈[t-H+1,...,t],τ2=t+1};

[0080] (4) The training process is divided into mainstream and auxiliary streams. The mainstream stream is only fine-tuned using incremental data from the new process, and the loss function is:

[0081] L final =αL new +βL replay (8)

[0082] Among them, L new It is the loss of new data; L replay It is the loss of playback data;

[0083] α and β are dynamic weights, calculated using formula (9):

[0084]

[0085] (5) If the loss of new data and historical data is close (stable change), α≈0.5, balanced training. When the error of new data is much greater than that of historical data (sudden change in working conditions), α becomes larger, and the model is more inclined to adapt to the new working conditions.

[0086] (6) Simultaneously, combining self-supervised enhancement, the model utilizes reconstructed input data to strengthen the extraction of new data features, accelerating the model's adaptation to new dynamic features of operating conditions. Input reconstruction loss:

[0087]

[0088] Final loss:

[0089] L final =αL new +βL replay +δL reconstruct

[0090] (7) The hyperparameter weight δ can be adjusted on-site to control the contribution of the self-supervised task to the optimization. After training, the new model is used for on-site prediction and the error detection process is repeated.

[0091] Example 1

[0092] See Figure 4 and Figure 5 This embodiment provides a kinematic modeling method for a centrally articulated steering mining truck, and the actual deployment and application of this model in the control of articulated mining trucks. The application of this method involves lidar, controllers, angle encoders, steer-by-wire solenoid valves, edge computers, base stations, etc. The multimodal sensor system includes multiple millimeter-wave radar detectors and multiple lidar detectors. A lidar is installed on the top of the articulated mining truck's cab, high-precision encoders are deployed at the front / rear frame hinge points, and millimeter-wave radar is deployed in the vehicle's blind spots. Each sensor aligns its data timestamps via a time synchronization module, and the onboard industrial control computer collects the raw data stream in real time via the CAN bus, forming a vehicle dynamic state observation matrix. (n represents the vehicle's kinematic parameters, and d represents the feature dimension). The weights of the pre-trained Pi-mLSTM model are embedded into the mining truck's embedded platform. At the start of the control cycle, normalized sensor data is input into the Pi-mLSTM model, and the output layer generates a prediction of the vehicle's state for the next N steps. The QP solver is used to solve the problem online and output the optimal control quantity to the electro-hydraulic proportional valve.

[0093]

[0094] stx t+k|t =f Pi-mLSTM (·), k=1,...,N p

[0095] u min ≤u t+k ≤u max k = 0, ..., N c -1n

[0096] Δu min ≤Δu t+k ≤Δu max k = 0, ..., N c -1

[0097] Δu t+k =u t+k -u t+k-1

[0098] N p For prediction in the time domain, N c To control the time domain (N) c ≤N p Q is the state tracking weight matrix, R is the control weight matrix, and x ref For reference trajectory, For the prediction of the state at time t+k, u t+k This serves as the future control input to be solved. The controller receives the status information of the steering solenoid valve current signal and the throttle and brake current signals obtained from the mining truck, enabling the truck to perform mining operations according to the driving route. When extreme road conditions or changes in the working characteristics of the mining truck's mechanical components due to wear are detected, the incremental learning module is triggered: it collects abnormal working condition data from the buffer, transmits it to the surface server, completes the online update of key weights through online incremental learning, and then transmits and deploys the updated model to the onboard edge calculator, ensuring that the model dynamically adapts to load changes without violating the vehicle's kinematics.

[0099] Example 2

[0100] See Figure 6This embodiment describes a kinematic modeling and control method for a centrally articulated steering loader. The method involves core components such as a lidar unit, controller, angle encoder, inertial navigation unit, steer-by-wire solenoid valve, edge computer, and base station. The system architecture is similar to Embodiment 1, but improvements are made to suit the characteristics of the loader. Because the loader frequently turns back and forth in confined spaces, and the bucket load significantly affects the vehicle's attitude, this solution specifically introduces an inertial navigation unit. This unit effectively suppresses sensor jitter noise interference, reduces motion distortion, and enhances the system's robustness under dynamic conditions. To adapt to these complex conditions, the modeling process employs higher sampling accuracy requirements, with shorter sampling and control cycles compared to mining truck systems to meet high-frequency response needs. The vehicle is also equipped with a multi-modal sensor system, including multiple millimeter-wave radar and lidar detectors. After time synchronization, the sensor data is collected and processed in real-time by the onboard computing unit to form dynamic state observation information of the vehicle. After the kinematic prediction model is trained, it is deployed on the onboard edge computer. Within a shorter control cycle, normalized sensor data is input into the prediction model, outputting a future state sequence. The model predictive controller constructs a rolling optimization problem based on the predicted state and solves for the optimal control value in real time within a shortened cycle. The controller ultimately outputs steering, throttle, and braking commands to the corresponding actuators. The system has adaptive update capabilities. When extreme road conditions or component wear cause characteristic changes, an online incremental learning mechanism is triggered. Abnormal operating condition data is transmitted to the server for updating key model weights. After the update, the model is fed back to the onboard edge device to ensure the system continuously adapts to load changes and conforms to vehicle kinematic constraints.

[0101] As described above, the present invention can be well implemented. The above embodiments are only the best implementations of the present invention, but the implementation of the present invention is not limited to the above embodiments. Other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention should be considered equivalent substitutions and are all included within the protection scope of the present invention.

Claims

1. A method for modeling the kinematics of an articulated vehicle, addressing problems such as model inaccuracy and control lag under complex operating conditions, characterized in that... Kinematic modeling of articulated vehicles based on a physically-informed matrix memory long short-term memory network (Pi-mLSTM) enables precise control of articulated vehicles under complex operating conditions. The specific method includes the following steps: A. Vehicle motion information collected by vehicle-mounted lidar and vehicle-mounted controller; the raw data is preprocessed after data collection. B. Standardize the X and Y data according to formula (1) so that the mean of the data is 0 and the standard deviation is 1. C. Use sliding windows with window lengths H1 and H2 to continuously sample X and Y, and obtain the sequence of X sampled data and Y sampled data at time t according to formula (2) and formula (3) respectively; In the formula, Let m be the kinematic parameter or controlled variable of the articulated vehicle at time t. Let be the k-th kinematic parameter of the articulated vehicle at time t; D. Calculate the predicted values ​​of the vehicle kinematic parameters according to formula (4). loss function L mLSTM : E. Based on the state-space equations of the fundamental kinematics of the articulated vehicle, the sum of the squares of the physical residuals at time t+1 is used as the physical loss to construct the physical information loss function L. phy Where Δ k The residual is obtained by subtracting the physical state variable from the derivative of the discrete approximation of the k-th state variable, where M is the number of state variables with kinematic geometry constraints. F. According to formula (6), use the data fitting loss L mLSTM With physical information loss L phy Construct a total loss function and use this function to train the kinematic model. L=L mLSTM +λL phy (6) Where λ is the weight hyperparameter balancing physical constraints and data fitting.

2. The method according to claim 1, characterized in that: Step A involves collecting vehicle motion information from the vehicle-mounted lidar and vehicle controller.

3. The method according to claim 1, characterized in that: Step F encodes the essential laws of vehicle kinematics into the model's loss function, and forces the model to approximate the physical conservation laws through the constraints of the loss function, thereby suppressing the physically unreasonable output of the purely data-driven model.

4. The method according to claim 1, characterized in that: The training method for the kinematic model in step C is as follows: a. Construct the mLSTM network architecture, and randomly initialize the mLSTM (matrix memory Long Short-Term Memory) network parameters, including the connection weights and biases of each network layer; b. Input the historical state sequence X, which consists of the vehicle pose and vehicle control parameters. t The preliminary predicted state is output after calculation by the mLSTM network; c. Calculate the error L between the predicted output and the actual observed value. mLSTM (such as mean squared error) and physical consistency penalty term L phy (such as losses due to violations of kinematic constraints); d. Combine physical constraints and prediction errors to calculate gradients, and use the Adam optimization algorithm to update network weights to ensure that the training process takes into account both data fitting and physical compliance; e. Repeat the above steps to gradually optimize the model and reduce prediction errors.

5. The method according to claims 1 and 4, characterized in that after the kinematic model is deployed, the network is fine-tuned by combining real-time sensor data to adapt to adaptive updates under dynamic environments and changing operating conditions. The online incremental learning method is as follows: (1) Definition of prediction error: y represents the model's predicted values ​​of the vehicle's kinematic parameters at time t. t This is the true value obtained from the sensor or actual measurement. The instantaneous error is calculated using formula (7): (2) Set threshold When an error is detected at a certain time t When this happens, it indicates that the current model's predictive ability is insufficient, and an incremental learning mechanism needs to be triggered. (Threshold) The size can be adjusted according to vehicle control precision requirements, sensor noise levels, etc. (3) In order to update the model online, it is necessary to retain the original observation data D(t) for the most recent period of time. When incremental learning is triggered, the recent data is extracted from the sliding window with a fixed size H to form the dataset D. new D new ={X(τ1),Y(τ2)∣τ1∈[t-H+1,...,t],τ2=t+1}; (4) The training process is divided into mainstream and auxiliary streams. The mainstream stream is only fine-tuned using incremental data from the new process, and the loss function is: L final =αL new +βL replay (8) in, L new It is the loss of new data; L replay It is the loss of playback data; α and β are dynamic weights, calculated using formula (9): (5) If the loss of new data and historical data is close (stable change), α≈0.5, balanced training. When the error of new data is much greater than that of historical data (sudden change in working conditions), α becomes larger, and the model is more inclined to adapt to the new working conditions. (6) Simultaneously, combining self-supervised enhancement, the model utilizes reconstructed input data to strengthen the extraction of new data features, accelerating the model's adaptation to new dynamic features of operating conditions. Input reconstruction loss: Final loss: L final =αL new +βL replay +δL reconstruct (7) The hyperparameter weight δ can be adjusted on-site to control the contribution of the self-supervised task to the optimization. After training, the new model is used for on-site prediction and the error detection process is repeated.

Images