Power source and hydraulic mechanical transmission integrated matching and damping method based on bond graph

Abstract

The application relates to a power source and hydraulic mechanical transmission integrated matching and damping method based on a bond graph, and particularly relates to the field of power transmission. The scheme provides a reliable dynamic mirror by constructing a high-fidelity digital twin model synchronized with a physical entity. A reinforcement learning intelligent agent is deployed, which autonomously explores and optimizes in the interaction between the state space and the action space according to a predefined composite reward function. With the powerful parallel simulation capability of the cloud, an optimized action sequence can be efficiently evolved for the real-time vehicle state. The edge controller executes the action sequence and completes feedback adjustment to form a closed loop. The whole process enables the transmission system to have the ability of online perception, dynamic adjustment and continuous self-adaptation, thereby stably inhibiting the lock-up impact and torsional vibration under complex working conditions in the whole domain, guaranteeing smoothness and high efficiency, and fundamentally overcoming the inherent limitations of the static calibration strategy.

Description

Technical Field

[0001] This invention relates to the field of power transmission, and more specifically, to a method for integrated matching and vibration reduction of power source and hydraulic mechanical transmission based on bond graph. Background Technology

[0002] With the development of new energy technologies, construction machinery represented by hybrid excavators and electric loaders is being gradually promoted and applied. The operating environment of these machines is extremely complex, and their working cycles exhibit strong non-steady-state characteristics. For example, in typical excavation and loading operations, the equipment needs to frequently switch between multiple modes such as excavation, heavy-load lifting, slewing, and traveling in a short period of time, causing the load torque of the drive motor and the speed ratio of the hydraulic torque converter to fluctuate continuously and drastically. At the same time, the density and composition distribution of the work objects, such as soil and ore, are significantly uneven. In addition, the slope of the work site and the dynamic changes in ground adhesion conditions constitute complex and variable external excitations. The combined effect of these internal and external factors makes the transmission system bear a wide excitation spectrum and large amplitude variation, posing a severe challenge to the smoothness, reliability, and overall energy efficiency of the transmission chain.

[0003] To address the issues of torsional vibration and lock-up impact suppression under the aforementioned complex operating conditions, existing technologies typically employ a static parameter matching and control strategy based on "calibration-table lookup." Specifically, during the R&D phase, extensive bench tests are conducted to measure the dynamic characteristics of the system at several discrete operating points. This calibrates the optimal nonlinear stiffness and damping parameters of the integrated torsional damper during the engagement of the hydraulic torque converter's lock-up clutch, and determines the corresponding control threshold for motor torque compensation. This results in a multi-dimensional data table stored in the controller. During actual operation, the controller senses finite state parameters such as speed and load, and then looks up preset fixed parameters and strategies. However, this method has significant limitations: firstly, the construction of a high-precision data table is challenging. Firstly, the system relies heavily on massive and expensive physical testing, resulting in long development cycles and high costs. Secondly, the discrete data points cannot cover the infinite combinations of continuously changing operating conditions in actual operations. When the system operates in an uncalibrated operating condition region, the control effect based on static optimal parameters will significantly decrease or even fail, leading to insufficient suppression of locking shocks, abnormal loads on transmission components, and deviation of the system from its efficient operating range, thus damaging the energy-saving potential of the new energy system. Therefore, it is urgent to develop an adaptive method that can sense the system status and external excitation characteristics online and autonomously and dynamically adjust vibration reduction parameters and control strategies to continuously and stably ensure the smoothness and efficiency of the transmission system under complex operating conditions across the entire range. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing an integrated matching and vibration reduction method for power sources and hydraulic mechanical transmission based on bond graphs, thereby solving the problems mentioned in the background art.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically, it includes the following steps: Step S1: In response to vehicle startup, the integrated topology model of drive motor-torsional damper-locked hydraulic torque converter, which includes nonlinear stiffness characteristics and damping parameters, is established based on bond graph theory. It receives real-time data streams from real vehicle sensors and corrects the internal state of the integrated topology model through a state observer to construct a high-fidelity digital twin model synchronized with the physical entity. Step S2: Deploy a reinforcement learning agent in the high-fidelity digital twin model. Configure the state space of the reinforcement learning agent as a feature vector containing real-time load torque, speed ratio and vibration spectrum features. Configure the action space as fine-tuning instructions for the control parameters in the active intervention logic of motor speed. Configure its reward function as a predefined composite reward function to drive the agent to learn vibration reduction. Step S3: When the vehicle is running, the current vehicle state collected by the edge controller is uploaded to the cloud. The cloud uses multiple simulation instances corresponding to the high-fidelity digital twin model to simulate the future responses of the reinforcement learning agent to various action sequences generated by the agent in response to the current vehicle state in parallel. The results of the future responses are evaluated based on the predefined composite reward function. The agent's policy network is updated through the policy optimization algorithm. The optimized action sequence determined by the updated policy for the current vehicle state is then sent to the edge controller. Step S4: The edge controller interprets the received optimized action sequence into specific control parameters, dynamically adjusts the locally stored motor speed active intervention logic, and collects the actual response data of the vehicle after the adjustment of the motor speed active intervention logic. It calculates the deviation between the actual response data and the predicted response data generated by the high-fidelity digital twin model, and uses this deviation to adjust the model parameters in the high-fidelity digital twin model.

[0006] In a preferred embodiment, step S1 involves the state observer correcting the internal state of the integrated topology model, specifically including: The state observer continuously receives real-time data streams, which include at least the drive motor speed, drive motor output torque, and hydraulic torque converter turbine speed. Based on the received real-time data stream and the nonlinear stiffness characteristics and damping parameters built into the integrated topology model, the state observer performs an iterative filtering process that includes prediction and update steps to estimate the pump wheel speed of the hydraulic torque converter and the rotation angle of the torsional damper in the integrated topology model. Subsequently, the estimated pump impeller speed and angle values ​​are injected into the integrated topology model, replacing the corresponding state variables calculated by the integrated topology model through its own numerical integration. This completes the dynamic closed-loop correction of the internal state of the integrated topology model, forming a verified internal state set of the high-fidelity digital twin model.

[0007] In a preferred embodiment, an integrated topology model is established based on bond graph theory, specifically including: The pre-set model parameters are loaded from the non-volatile memory of the edge controller to initialize the integrated topology model. The pre-set model parameters are non-linear stiffness characteristic functions and damping coefficients used to calculate the restoring torque component based on non-linear stiffness characteristics and the damping torque component based on damping parameters. During operation, the integrated topology model uses the verified internal state set obtained after correction by the state observer, combined with the real-time changing load inertia during operation, to solve the system dynamics differential equations derived from bond graph theory. It then calculates the energy flow and signal flow interaction between the drive motor, torsional damper and lock-up hydraulic torque converter in real time, thereby dynamically generating predictive response data that is synchronized with the transmission behavior of the physical entities.

[0008] In a preferred embodiment, step S2 configures the state space of the reinforcement learning agent as a feature vector containing real-time load torque, speed ratio, and vibration spectrum characteristics, specifically including: From the output data obtained by real-time calculation of the high-fidelity digital twin model, the output torque of the drive motor after low-pass filtering is extracted as the real-time load torque feature; the ratio of the turbine speed of the hydraulic torque converter in the high-fidelity digital twin model to the estimated value of the pump wheel speed in the verified internal state set is calculated as the real-time speed ratio feature. Spectral analysis was performed on the rotation angle signal of the torsional vibration damper output in real time from the high-fidelity digital twin model. The proportion of vibration energy in a set bandwidth around the main resonant frequency of the integrated topology model was calculated as the vibration energy concentration characteristic. Finally, the feature vector composed of real-time load torque characteristics, real-time speed ratio characteristics, and vibration energy concentration characteristics is defined as the state space of the reinforcement learning agent.

[0009] In a preferred embodiment, the predefined composite reward function is specifically configured as follows: The value of the composite reward function consists of the weighted negative sum of the vibration penalty term, the peak impact penalty term, the motion smoothness penalty term, and the model uncertainty penalty term; Among them, the vibration penalty term is calculated based on the square integral of the difference between the drive motor angular acceleration and the pump wheel angular acceleration in the high-fidelity digital twin model within an adjustable time window; the impact peak penalty term is taken as the maximum absolute value of the drive shaft impact torque calculated by the high-fidelity digital twin model within the same time window. The action smoothness penalty term is defined by the Euclidean norm of the vector consisting of the fine-tuning instructions for the control parameters in the active intervention logic of the motor speed currently output by the reinforcement learning agent. The model uncertainty penalty term directly adopts the confidence index output by the state observer during the construction of the high-fidelity digital twin model; Each penalty is weighted and summed using pre-set positive weighting coefficients, and the negative value of the final result is the reward value at the current moment.

[0010] In a preferred embodiment, in step S3, the cloud utilizes multiple simulation instances corresponding to the high-fidelity digital twin model to simulate in parallel the future responses of the reinforcement learning agent to various action sequences generated by the agent in response to the current vehicle state, specifically including: The cloud server receives the current vehicle status and the complete current runtime status of the high-fidelity digital twin model uploaded from the edge controller; based on the received complete current runtime status, the cloud server quickly clones hundreds to thousands of simulation instances that are completely identical to the high-fidelity digital twin model synchronized with the physical entity. The cloud-deployed reinforcement learning agent generates multiple candidate action sequences with subtle differences in number, matching the number of simulation instances, based on the policy network parameters of the reinforcement learning agent. Each candidate action sequence defines the adjustment instructions for each control parameter in the active intervention logic of motor speed in each control cycle within a preset short time domain in the future. The cloud scheduler dynamically assigns each candidate action sequence to an independent simulation instance. Driven by their respective assigned action sequences, all simulation instances independently and in parallel simulate the preset short time domain, thereby synchronously obtaining the future state response trajectory corresponding to each candidate action sequence.

[0011] In a preferred embodiment, the result of evaluating future responses based on a predefined composite reward function is used to update the agent's policy network through a policy optimization algorithm. The updated policy-determined, optimized action sequence for the current vehicle state is then distributed to the edge controller. Specifically, this includes: The cloud server aggregates the complete trajectory data, including state, action, and instantaneous reward value calculated by a predefined composite reward function, output by all simulation instances after the simulation is completed, forming a large-scale experience pool for policy network updates; The policy optimization algorithm employs a near-end policy optimization algorithm, which guides the parameter updates of the policy network by constructing an alternative objective function that includes a probability ratio pruning mechanism. After the policy network parameters are updated, the cloud server filters all candidate action sequences in the simulation according to the multi-objective optimization principle: the cumulative value of the composite reward function, the peak value of the vibration penalty term, and the peak value of the impact torque corresponding to the impact peak penalty term in the simulation results of each candidate action sequence are jointly constructed into a three-dimensional evaluation vector. The fast non-dominated sorting algorithm is used to process the evaluation vectors of all candidate action sequences to find the Pareto optimal solution set. Finally, based on the preset engineering priority rules, a candidate action sequence with the best overall performance is selected from the Pareto optimal solution set, marked as an optimized action sequence, and sent to the edge controller through the vehicle-cloud communication link.

[0012] In a preferred embodiment, in step S4, the edge controller interprets the received optimized action sequence into specific control parameters and dynamically adjusts the locally stored motor speed active intervention logic, specifically including: The optimized sequence of actions received by the edge controller contains multiple consecutive control commands, each of which corresponds to the adjustment increment of the control parameters in the active intervention logic for motor speed within the next control cycle. During the interpretation process, a preset mapping table is first used to convert each adjustment increment in the action sequence into a specific target change in the corresponding control parameter. To ensure control stability and safety, incremental limiting processing is performed on each interpreted parameter target change, that is, limiting it to a preset upper limit of absolute value. Subsequently, a first-order inertial filtering algorithm is used to gradually transition the target parameter value after the limiting processing, so that the new control parameter smoothly approaches the target parameter value after the limiting processing. Finally, the new control parameter value after the smooth transition is written into the corresponding storage unit of the motor speed active intervention logic stored locally in the edge controller, and takes effect at the beginning of the next control cycle.

[0013] In a preferred embodiment, the actual response data of the vehicle's active intervention logic after adjusting the motor speed is collected, and the deviation between the actual response data and the predicted response data generated by the high-fidelity digital twin model is calculated, specifically including: After the new active intervention logic for motor speed takes effect, the edge controller initiates a synchronous monitoring window with the same preset short time domain length as the cloud simulation. During this window, actual response data is synchronously collected from the vehicle's sensors. The actual response data includes at least the real-time speed of the drive motor, the output torque of the drive motor, and the output torque of the drive shaft. At the same time, the edge controller drives a high-fidelity digital twin model locally, using the currently verified internal state set as the initial condition and employing the exact same control parameters as the adjusted active intervention logic for motor speed, to perform a fully synchronous forward simulation, thereby generating predicted response data within the preset short time domain. The deviation is calculated using a multimodal response deviation measurement method, which includes: extracting frequency domain features representing vibration energy, time domain features representing impact peak value, and time series curve features representing torque trend from actual response data and predicted response data, respectively, and constructing multiple sets of corresponding feature vector pairs; For each pair of feature vectors, the specified feature distance function value is calculated. For the frequency domain feature pairs of vibration energy, the relative entropy is used to measure the difference in their probability distribution; for the time-series curve feature pairs of torque trends, the dynamic time warping distance is used to measure their morphological difference; and for the impact peak feature pairs, the absolute value difference of their impact peak sequences is calculated. Finally, the calculated feature distance function values ​​are multiplied by a pre-set positive weighting coefficient that sums to one, and all weighted distance values ​​are summed. The sum is the comprehensive multimodal response deviation metric.

[0014] In a preferred embodiment, the model parameters in the high-fidelity digital twin model are adjusted using a multimodal response deviation metric, specifically including: Using the calculated multimodal response deviation metric and the model-entity consistency residual from the state observer as inputs, a robustness metric is constructed. The robustness metric is defined as a negative exponential function of the product of the multimodal response deviation metric and the model-entity consistency residual, with a value between zero and one. Subsequently, the damping coefficient, which characterizes the energy dissipation of the torsional damper in the high-fidelity digital twin model, is taken as the target parameter to be adjusted. A Bayesian inference method based on adaptive learning rate is used to update it: First, it is assumed that the current value of the damping coefficient is prior knowledge, and its uncertainty is described by a prior variance. Then, the robustness metric is regarded as a likelihood measure of the reasonableness of the damping coefficient under the observed current data conditions; the maximum a posteriori probability estimate is approximated by an adaptive learning rate, which is obtained by multiplying the base learning rate by one, subtracting the robustness metric, and then dividing by the prior standard deviation. Finally, a correction amount for the damping coefficient is calculated, which is proportional to the product of the adaptive learning rate and the gradients of the log-likelihood and the log-prior. This correction amount is added to the current damping coefficient to obtain the updated new damping coefficient value. At the same time, the updated parameter uncertainty estimate is calculated. The new damping coefficient and the new uncertainty estimate are written back to the parameter storage area of ​​the high-fidelity digital twin model to complete the online feedback adjustment.

[0015] The beneficial effects of this invention are as follows: by constructing a high-fidelity digital twin model synchronized with the physical entity, a reliable dynamic mirror is provided; and a reinforcement learning agent is deployed, which autonomously explores and optimizes in the interaction between the state space and the action space based on a predefined composite reward function; with the help of the powerful parallel simulation capabilities of the cloud, an optimized action sequence can be efficiently evolved for the real-time vehicle state; the edge controller executes the action sequence and completes feedback adjustment to form a closed loop; the whole process enables the transmission system to have the ability of online perception, dynamic adjustment and continuous self-adaptation, thereby stably suppressing locking shock and torsional vibration under complex working conditions in the whole domain, ensuring smoothness and efficiency, and fundamentally overcoming the inherent limitations of static calibration strategies. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention.

[0017] Figure 2 This is a model diagram of the bond graph of the entire system under hydraulic conditions.

[0018] Figure 3 This is a model diagram of the bond graph of the entire system under locked-down conditions.

[0019] Figure 4 A high-fidelity digital twin model synchronized with the physical entity. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0022] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0023] Example 1 This embodiment provides, for example Figure 1 The method for integrated matching and vibration reduction of power source and hydraulic mechanical transmission based on bond graph, as shown, specifically includes the following steps: Step S1: In response to vehicle start-up or self-test commands, the integrated topology model of drive motor-torsional damper-locked hydraulic torque converter, which includes nonlinear stiffness characteristics and damping parameters, is established based on bond graph theory. It receives real-time data streams from real vehicle sensors and corrects the internal state of the integrated topology model through a state observer to construct a high-fidelity digital twin model synchronized with the physical entity. Step S2: Deploy a reinforcement learning agent in the high-fidelity digital twin model. Configure the state space of the reinforcement learning agent as a feature vector containing real-time load torque, speed ratio and vibration spectrum features. Configure the action space as fine-tuning instructions for the control parameters in the active intervention logic of motor speed. Configure its reward function as a predefined composite reward function to drive the agent to learn vibration reduction. Step S3: When the vehicle is running, the current vehicle state collected by the edge controller is uploaded to the cloud. The cloud uses multiple simulation instances corresponding to the high-fidelity digital twin model to simulate the future responses of the reinforcement learning agent to various action sequences generated by the agent in response to the current vehicle state in parallel. The results of the future responses are evaluated based on the predefined composite reward function. The agent's policy network is updated through the policy optimization algorithm. The optimized action sequence determined by the updated policy for the current vehicle state is then sent to the edge controller. Step S4: The edge controller interprets the received optimized action sequence into specific control parameters, dynamically adjusts the locally stored motor speed active intervention logic, and collects the actual response data of the vehicle after the adjustment of the motor speed active intervention logic. It calculates the deviation between the actual response data and the predicted response data generated by the high-fidelity digital twin model, and uses this deviation to adjust the model parameters in the high-fidelity digital twin model.

[0024] In this embodiment, it is particularly important to explain that in step S1, the state observer corrects the internal state of the integrated topology model, which specifically includes: The state observer continuously receives real-time data streams, which include at least the drive motor speed, drive motor output torque, and hydraulic torque converter turbine speed. Based on the received real-time data stream and the nonlinear stiffness characteristics and damping parameters built into the integrated topology model, the state observer performs an iterative filtering process that includes prediction and update steps to estimate the pump wheel speed of the hydraulic torque converter and the rotation angle of the torsional damper in the integrated topology model. The iterative filtering process specifically adopts the extended Kalman filter algorithm; the prediction step is as follows: based on the optimal state estimate of the previous control cycle and the state transition equation describing the change of the system state over time, the prior estimate of the system state in the current cycle is calculated; at the same time, the prior value of the state estimation error covariance matrix is ​​updated according to the preset process noise covariance matrix. The update steps are as follows: First, using the observation equation describing the relationship between the system state and sensor measurements, calculate the theoretical prediction of the sensor measurements for the current period. Next, calculate the difference between the actual sensor measurements and the theoretical prediction; this difference is called the innovation. Then, calculate the Kalman gain matrix based on the prior values ​​of the observation matrix, the state estimation error covariance matrix, and the preset observation noise covariance matrix. After that, weight the innovation using the Kalman gain matrix and correct the state prior estimate with the weighted innovation to obtain the optimal state estimate for the current period, i.e., the optimal estimate of the pump wheel speed and the optimal estimate of the damper angle. Finally, update the posterior value of the state estimation error covariance matrix to prepare for the filtering calculation in the next period. The initial values ​​and update rules of the process noise covariance matrix and the observation noise covariance matrix need to be pre-calibrated according to the sensor accuracy and system disturbance level. For example, the observation noise variance of the drive motor speed sensor can be set to the order of 0.1 radians per second, and the observation noise variance of the torque sensor can be set to the order of 5 Newton-meters. Subsequently, the estimated pump impeller speed and angle values ​​are injected into the integrated topology model, replacing the corresponding state variables calculated by the integrated topology model through its own numerical integration. This completes the dynamic closed-loop correction of the internal state of the integrated topology model, forming a verified internal state set of the high-fidelity digital twin model. The "injection" operation is performed in each control cycle. Its technical effect is to force the alignment of the key internal states of the digital model with the actual state of the physical entity after optimal estimation, thereby eliminating the cumulative state deviation caused by model initial condition errors, simplification assumptions, or parameter drift. The resulting "verified internal state set" not only contains the optimal estimates of the pump impeller speed and damper angle, but also the posterior value of its state estimation error covariance matrix serves as a quantitative indicator characterizing the confidence of the state set. This dynamic correction mechanism ensures that the high-fidelity digital twin model can serve as a highly reliable dynamic mirror of the physical transmission system at any time, providing a virtual training and testing foundation with realistic state and reliable environment for the reinforcement learning agent in subsequent steps. This is a prerequisite for achieving online adaptive optimization. An integrated topology model based on bond graph theory is established, specifically including: The pre-set model parameters are loaded from the non-volatile memory of the edge controller to initialize the integrated topology model. The pre-set model parameters are non-linear stiffness characteristic functions and damping coefficients used to calculate the restoring torque component based on non-linear stiffness characteristics and the damping torque component based on damping parameters. The nonlinear stiffness characteristic function can be specifically expressed as a polynomial function or a piecewise spline interpolation function with respect to the damper rotation angle. Its specific coefficients are obtained through torsional damper bench characteristic tests. For example, a nonlinear stiffness characteristic function in cubic polynomial form is expressed as follows: the stiffness value is equal to the constant term coefficient plus the first term coefficient of the rotation angle multiplied by the rotation angle, plus the second term coefficient of the rotation angle multiplied by the square of the rotation angle, plus the third term coefficient of the rotation angle multiplied by the cube of the rotation angle; the damping coefficient is a positive constant. During model operation, the calculation process of the recovery torque component is as follows: for the current damper rotation angle value, the nonlinear stiffness characteristic function is called to calculate the corresponding instantaneous stiffness value, and then the stiffness value is integrated with respect to the rotation angle to obtain the recovery torque; The calculation process for the damping torque component is as follows: multiply the rate of change of the damper angle at the current moment by the damping coefficient to obtain the damping torque; the total torque output by the torsional damper is the sum of the restoring torque component and the damping torque component. During operation, the integrated topology model, based on the verified internal state set obtained after correction by the state observer and combined with the real-time changing load inertia during operation, calculates the energy flow and signal flow interaction between the drive motor, torsional damper and lock-up hydraulic torque converter in real time by solving the system dynamics differential equations derived from bond graph theory, thereby dynamically generating predictive response data synchronized with the transmission behavior of the physical entity. The specific process of solving the system dynamics differential equations derived from bond graph theory is as follows: First, based on the system power flow model established by bond graph theory, several first-order ordinary differential equations are listed with the rotor angular velocity of the drive motor, the rotation angle of the damper, and the pump wheel angular velocity of the hydraulic torque converter as state variables. These equations include the electromagnetic torque equation of the drive motor, the total torque of the torsional damper calculated above, the pump wheel torque and turbine torque characteristic functions of the hydraulic torque converter (usually expressed in the form of torque coefficients), and the real-time load inertia. The solution process uses numerical integration methods, such as the fourth-order Runge-Kutta method. The detailed calculation steps within each simulation step are as follows: First, the verified internal state set at the current moment (such as the estimated pump wheel speed and the estimated damper rotation angle) is used as the initial state vector of the differential equation system. The second step is to calculate the electromagnetic torque output by the drive motor, the total torque generated by the torsional damper, and the pump wheel torque and turbine torque generated by the hydraulic torque converter based on the current speed ratio, according to the current state vector. The third step involves calling the computation flow of the fourth-order Runge-Kutta solver: First, calculate the derivative vector based on the current state vector; then, use this derivative vector to calculate an intermediate state vector, and calculate the second derivative vector based on this intermediate state vector; next, combine the first two derivative vectors to calculate the new intermediate state vector and the third derivative vector; then, use the third derivative vector to calculate the final intermediate state vector and the fourth derivative vector; finally, add the current state vector to the weighted sum of these four derivative vectors to obtain the new state vector after the next simulation step. The fourth step is to use the updated state vector as the input for the next calculation loop and output the drive motor speed, torque converter turbine speed, and drive shaft torque as "predicted response data". This real-time solution process enables the integrated topology model to simulate the dynamic response of the physical transmission system under various control strategies with high fidelity. The generated predicted response data is the benchmark for evaluating the quality of the control strategy and for subsequent comparison with the actual response data.

[0025] In this embodiment, it is specifically necessary to explain that in step S2, the state space of the reinforcement learning agent is configured as a feature vector containing real-time load torque, speed ratio, and vibration spectrum features, specifically including: From the output data obtained by real-time calculation of the high-fidelity digital twin model, the output torque of the drive motor after low-pass filtering is extracted as the real-time load torque feature. The low-pass filtering process adopts a first-order inertial filtering algorithm with a time constant set to 0.1 to 0.5 seconds to filter out high-frequency noise in the output torque signal of the drive motor and retain the low-frequency trend component that reflects the real load change. The ratio of the turbine speed of the hydraulic torque converter in the high-fidelity digital twin model to the estimated pump wheel speed in the verified internal state set is calculated as a real-time speed ratio feature. Spectral analysis was performed on the rotation angle signal of the torsional vibration damper output in real time from the high-fidelity digital twin model. The proportion of vibration energy in a set bandwidth around the main resonant frequency of the integrated topology model was calculated as the vibration energy concentration characteristic. The specific process of spectrum analysis is as follows: First, a short-time Fourier transform is performed on the real-time acquired torsional damper angle signal. The operation of this transform is to divide the angle signal into a series of overlapping time windows, apply the Hanning window function to the signal in each time window to reduce spectral leakage, and then perform a fast Fourier transform on the windowed signal to obtain the frequency components and energy distribution of the signal in that time period, i.e., the power spectral density. The defined bandwidth refers to the frequency range extending from the main resonance frequency point identified by theoretical calculation or experiment through the integrated topology model to both sides by 5 Hz to 15 Hz. The formula for calculating the vibration energy concentration characteristic is: the sum of the power spectral density values ​​of all frequency points within the defined bandwidth is divided by the sum of the power spectral density values ​​of all frequency points from zero Hz to the Nyquist frequency range. The ratio obtained is the vibration energy concentration characteristic value. The closer this characteristic value is to one, the more concentrated the vibration energy is near the system's sensitive main resonance frequency band.

[0026] Finally, the feature vector composed of real-time load torque characteristics, real-time speed ratio characteristics, and vibration energy concentration characteristics is defined as the state space of the reinforcement learning agent. The predefined composite reward function is configured as follows: The value of the composite reward function consists of the weighted negative sum of the vibration penalty term, the peak impact penalty term, the motion smoothness penalty term, and the model uncertainty penalty term; The vibration penalty term is calculated based on the square integral of the difference between the angular acceleration of the drive motor and the angular acceleration of the pump wheel in the high-fidelity digital twin model within an adjustable time window. The specific process of "square integral calculation" is as follows: within an adjustable time window of 0.1 to 0.3 seconds, the real-time difference between the angular acceleration of the drive motor and the angular acceleration of the pump wheel, as simulated by the high-fidelity digital twin model, is sampled; the angular difference at each sampling point is squared; then the square values ​​corresponding to all sampling points within the time window are accumulated, or the trapezoidal integral method is used to numerically integrate the continuous square value curve. The result is the value of the vibration penalty term. This value directly quantifies the total energy of the torsional vibration of the transmission system within a specified time period. The peak impact penalty term is taken as the maximum absolute value of the drive shaft impact torque calculated by the high-fidelity digital twin model within the same time window; The action smoothness penalty term is defined by the Euclidean norm of the vector consisting of the fine-tuning instructions for the control parameters in the active intervention logic of motor speed output by the reinforcement learning agent. The calculation process of the "Euclidean norm definition" is as follows: the set of fine-tuning instructions output by the reinforcement learning agent at the current decision moment, such as the adjustment amount of three parameters including torque compensation coefficient, speed ratio intervention threshold and time constant, is regarded as a multi-dimensional vector; the square of each component of the vector is calculated; the square values ​​of all components are added together; and then the square root of the sum is taken. The resulting value is the value of the action smoothness penalty term. This penalty is used to encourage the agent to output smooth and small-amplitude parameter adjustment instructions to avoid drastic changes in control instructions. The model uncertainty penalty term is directly adopted during the construction of a high-fidelity digital twin model, with the structure as follows: Figure 4 As shown, the confidence index is derived from the posterior value of the state estimation error covariance matrix generated by the state observer. One specific way to derive the "confidence index" is to calculate the trace of the posterior value matrix of the state estimation error covariance matrix, which is the sum of the elements on the main diagonal of the matrix. The magnitude of this trace value reflects the overall uncertainty of the state observer in estimating the current state. After processing this trace value with a preset scaling factor, it can be used as the input value of the model uncertainty penalty term. When the state estimation error is large and the synchronization between the model and the entity is poor, this penalty value increases, prompting the agent to adopt a conservative strategy in areas with low model confidence. Each penalty term is weighted and summed using pre-set positive weight coefficients. The negative value of the final result is the reward value at the current moment. This collectively drives the agent to learn to effectively suppress vibrations and shocks while maintaining stable control and avoiding unreliable regions of the model. The "pre-set positive weight coefficients" need to be tuned according to the priority of the control objectives. For example, when vibration suppression is the most critical, the weight coefficient of the vibration penalty term can be set in the range of 0.5 to 0.7. When suppressing transient shocks is secondary, the weight coefficient of the shock peak penalty term can be set in the range of 0.3 to 0.5. To maintain stable control, the weight coefficient of the action smoothness penalty term can be set in the range of 0.1 to 0.2. To avoid risks when the model is uncertain, the weight coefficient of the model uncertainty penalty term can be set in the range of 0.05 to 0.15, and can be dynamically correlated with the model uncertainty index, automatically increasing the weight when the uncertainty is high.

[0027] In this embodiment, it is particularly important to explain step S3, where the cloud utilizes multiple simulation instances corresponding to the high-fidelity digital twin model to simulate in parallel the future responses of the reinforcement learning agent to various action sequences generated by the agent in response to the current vehicle state. Specifically, this includes: The cloud server receives the current vehicle state and the complete current runtime state of the high-fidelity digital twin model uploaded from the edge controller. The complete current runtime state includes the verified internal state set and the internal intermediate state of the numerical integrator used to solve the dynamic differential equations of the integrated topology model. The cloud server uses the received complete current runtime state as a blueprint to quickly clone hundreds to thousands of simulation instances that are completely identical to the high-fidelity digital twin model synchronized with the physical entity. The internal numerical state of each simulation instance is precisely initialized to be completely identical to the model state synchronized with the physical entity state. Subsequently, the cloud-deployed reinforcement learning agent, based on the policy network parameters of the reinforcement learning agent, generates multiple candidate action sequences with subtle differences, matching the number of simulation instances, by superimposing random noise conforming to a truncated normal probability distribution onto the deterministic output. The specific process of superimposing random noise conforming to a truncated normal probability distribution onto the deterministic output is as follows: First, the policy network of the cloud-based reinforcement learning agent calculates a deterministic basic action output based on the input current vehicle state. This basic action output is a multi-dimensional vector, with each dimension corresponding to an adjustment amount of a control parameter. Next, random noise needs to be generated for exploration. "Truncation normal probability distribution" refers to a standard normal distribution. However, its random sampling values ​​are restricted to the range of -3 to +3, that is, any sampling value less than -3 is set to -3, and any sampling value greater than +3 is set to +3. For the generation of each candidate action sequence, a set of random numbers conforming to this truncated normal distribution is independently drawn, and the dimension of the random numbers is the same as the dimension of the action vector. Then, this set of random numbers is multiplied by a preset noise coefficient, which is a small positive number, such as 0.05, to control the intensity of exploration. Finally, the scaled random noise vector is added element-wise to the deterministic basic action vector output by the policy network to obtain the final candidate actions with exploration noise. This process is executed in parallel, generating hundreds to thousands of slightly different candidate actions. Each candidate action sequence defines the adjustment instructions for each control parameter in the active intervention logic of motor speed within a preset short time domain and each control cycle in the future; the preset short time domain is usually set to 0.3 to 0.8 seconds, which corresponds to the time length that can cover a typical hydraulic torque converter lock-up impact process, such as... Figure 2 As shown; Subsequently, the cloud scheduler dynamically assigns each candidate action sequence to an independent simulation instance. All simulation instances, driven by their respective assigned action sequences, independently and in parallel simulate a preset short time domain, and perform state evolution based on the integrated topological dynamics equations of the drive motor-torsional damper-locked hydraulic torque converter, built into the high-fidelity digital twin model and based on bond graph theory. Figure 3As shown, this allows for the synchronous acquisition of the future state response trajectory corresponding to each candidate action sequence. The cloud scheduler is a software module that manages and allocates computing tasks within a distributed computing framework. Its specific workflow is as follows: First, the scheduler maintains a resource pool containing all idle simulation instances. When a batch of candidate action sequences is generated, the scheduler allocates an idle simulation instance from the resource pool for each candidate action sequence and establishes a one-to-one mapping relationship. Next, the scheduler packages the action sequence data and initial state data into tasks and distributes them to the simulation instances running on each computing node via message queues or remote procedure calls. After receiving the tasks, all simulation instances start simulation calculations simultaneously, and after the simulation ends, they return the result trajectory to the scheduler for aggregation. This parallel scheduling mechanism enables large-scale simulations to be completed within seconds, meeting the real-time requirements of online optimization. The future response is evaluated based on a predefined composite reward function. The agent's policy network is updated using a policy optimization algorithm. The optimized action sequence determined by the updated policy for the current vehicle state is then sent to the edge controller. Specifically, this includes: The cloud server aggregates the complete trajectory data, including state, action, and instantaneous reward value calculated by a predefined composite reward function, output by all simulation instances after the simulation is completed, forming a large-scale experience pool for policy network updates; The policy optimization algorithm employs a near-end policy optimization algorithm, which guides the parameter updates of the policy network by constructing an alternative objective function that includes a probability ratio pruning mechanism. The detailed computational steps for constructing an alternative objective function that includes a probability ratio pruning mechanism are as follows: First, randomly sample a small batch of empirical data from a large-scale experience pool, including state, action, reward, and next state; The second step is to use an independent value network to estimate the value of each state based on the current policy network parameters. The third step is to calculate the advantage function estimate. The calculation process is as follows: First, calculate the time-series difference error, which is equal to the current reward plus the discount factor multiplied by the value estimate of the next state, and then subtract the value estimate of the current state. Then, using the generalized advantage estimation algorithm, the time-series difference errors of the current time and multiple future time points are summed exponentially to obtain the advantage function estimate of the current state-action pair. This value reflects the performance of the current action relative to the average performance. The fourth step is to calculate the probability ratio, which is the probability of the updated policy network selecting a sampling action in the current state, divided by the probability of the old policy network selecting the same action in the same state. Fifth step, apply the cropping operation: limit the probability ratio to the range of 0.8 to 1.2, that is, if the probability ratio is less than 0.8, calculate it as 0.8; if the probability ratio is greater than 1.2, calculate it as 1.2. Step 6: Construct two quantities: the first quantity is the unpruned probability ratio multiplied by the dominance function estimate; the second quantity is the pruned probability ratio multiplied by the dominance function estimate. The seventh step is to take the smaller of these two values ​​as the contribution of the current empirical sample to the alternative objective function; finally, the average of the contributions of all sampled empirical samples is calculated to obtain the value of the alternative objective function that needs to be maximized. The alternative objective function first calculates the probability ratio of the new and old policy networks choosing the same action under the same conditions, and then multiplies this probability ratio by an estimated advantage function value, which is used to quantify the superiority or inferiority of the current action relative to the average performance. To prevent excessive policy changes in a single update, the near-end policy optimization algorithm uses a pruning operation to limit the probability ratio to a preset range centered at 0.01 and fluctuating between 0.01 and 0.02, thereby ensuring training stability. The cloud server uses stochastic gradient ascent to maximize this alternative objective function and iteratively updates the parameters of the policy network. After the policy network parameters are updated, the cloud server filters all candidate action sequences in the simulation according to the multi-objective optimization principle: the cumulative value of the composite reward function, the peak value of the vibration penalty term, and the peak value of the impact torque corresponding to the impact peak penalty term in the simulation results of each candidate action sequence are jointly constructed into a three-dimensional evaluation vector. The fast non-dominated sorting algorithm is used to process the evaluation vectors of all candidate action sequences to find the Pareto optimal solution set. The specific process of the fast non-dominated sorting algorithm is as follows: First, the sorting level and dominance counter of all candidate action sequences are initialized; for each candidate action sequence, the set of other sequences it dominates and the number of sequences it dominates are calculated; if a candidate sequence is not dominated by any other sequence (i.e., the number of dominant sequences is zero), it is placed in the first Pareto front; then, the sequences in the first layer are removed, and the number of dominant sequences of the remaining sequences is updated; at this time, the sequences whose number of dominant sequences is reduced to zero constitute the second Pareto front; this process is repeated until all sequences are assigned to a certain Pareto front; the Pareto optimal solution set refers to the set of candidate action sequences in the first Pareto front obtained by sorting. Finally, based on preset engineering priority rules, such as prioritizing candidate action sequences with peak impact torque below 150 Nm, a candidate action sequence with the best overall performance is selected from the Pareto optimal solution set and marked as the optimized action sequence, which is then sent to the edge controller via the vehicle-cloud communication link. The specific rules for selecting a candidate action sequence with the best overall performance can be as follows: First, from the first-level Pareto optimal solution set, all sequences with peak impact torque exceeding the 150 Nm threshold are removed to ensure safety; then, among the remaining sequences, their cumulative composite reward function values ​​are compared, and the sequence with the highest cumulative value is selected as the final optimized action sequence; if the peak impact of all sequences exceeds the threshold, the sequence with the smallest peak impact is selected to ensure damage minimization.

[0028] In this embodiment, it is particularly important to explain that in step S4, the edge controller interprets the received optimized action sequence into specific control parameters and dynamically adjusts the locally stored motor speed active intervention logic, specifically including: The optimized sequence of actions received by the edge controller contains multiple consecutive control commands, each of which corresponds to the adjustment increment of the control parameters in the active intervention logic for motor speed within the next control cycle. During interpretation, a pre-defined mapping table is first used to convert each adjustment increment in the action sequence into a specific target change in the corresponding control parameter. Control parameters include, for example, the speed ratio intervention threshold for torque converter lockout, the amplitude coefficient of torque compensation, and the filtering time constant of the compensation command. The mapping table is a data structure stored in the edge controller that defines the linear or nonlinear scaling relationship between the action space (adjustment increment) and the physical changes in the control parameters. For example, an adjustment increment in the action sequence might be a normalized value between -1 and +1. The mapping table multiplies this value by a scaling factor corresponding to the physical range of the parameter to obtain the target change in physical units. For the speed ratio intervention threshold, the scaling factor might be 0.01, meaning that an action increment of "1" corresponds to a speed ratio change of 0.01. For the torque compensation coefficient, the scaling factor might be 0.05, meaning that an action increment of "1" corresponds to a compensation coefficient change of 5%. To ensure control stability and safety, incremental limiting is applied to the target change of each interpreted parameter, restricting it to a preset upper limit of absolute value. For example, the single adjustment of the speed ratio threshold cannot exceed 0.05%, and the single adjustment of the torque compensation coefficient cannot exceed 10% of its current value. The specific calculation process for incremental limiting is as follows: First, the maximum absolute change allowed for a single adjustment for each control parameter is read. Then, the interpreted target change is compared with this maximum change. If the absolute value of the target change is less than or equal to the maximum change, it is allowed to proceed. If the absolute value of the target change is greater than the maximum change, its sign is retained, but its value is replaced with the maximum change. For example, if the interpreted target change of the torque compensation coefficient is an increase of 15%, and the single upper limit is 10%, then after limiting, its target change is corrected to an increase of 10%. This processing effectively prevents aggressive strategies issued from the cloud from causing excessive impact on the physical system in a single step. Subsequently, a first-order inertial filtering algorithm is used for a gradual transition of the target parameter value after amplitude limiting, rather than directly replacing the original parameter. The time constant of the first-order inertial filtering algorithm is set between 0.1 and 0.5 seconds, so that the new control parameter smoothly approximates the target parameter value after amplitude limiting. The detailed calculation process of the first-order inertial filtering algorithm adopts a recursive form; assuming the current control parameter value is the old value, the target parameter value after amplitude limiting is the new target value, and the filtering time constant is a preset value; in each control cycle, a filtering coefficient is calculated, and this coefficient is equal to... The control cycle length is divided by the sum of the control cycle length and the filter time constant. Then, the new parameter value is calculated, which is equal to the old parameter value plus the filter coefficient multiplied by the difference between the new target value and the old parameter value. This calculation is repeated in each control cycle, so that the parameter value smoothly approaches the target value in an exponential decay manner. A filter time constant of 0.1 seconds means that the parameter value reaches 98% of the target value in about 0.4 seconds, and 0.5 seconds corresponds to about 2 seconds. This gradual process completely eliminates the step jump of the control parameters when switching, ensuring the continuity of the drive system control. Finally, the new control parameter values ​​after the smooth transition are written into the corresponding storage unit of the motor speed active intervention logic stored locally on the edge controller, and take effect and execute at the beginning of the next control cycle; The actual response data of the vehicle's active intervention logic for adjusted motor speed is collected, and the deviation between the actual response data and the predicted response data generated by the high-fidelity digital twin model is calculated, specifically including: After the new active intervention logic for motor speed takes effect, the edge controller starts a synchronous monitoring window with the same preset short time domain length as the cloud simulation. During this window period, actual response data is synchronously collected from the vehicle's sensors. The actual response data includes at least the real-time speed of the drive motor, the output torque of the drive motor, and the output torque of the drive shaft. Meanwhile, the edge controller drives a high-fidelity digital twin model locally, using the currently verified internal state set as the initial condition and adopting the same control parameters as the adjusted motor speed active intervention logic, to perform a fully synchronized forward simulation, thereby generating predicted response data in the preset short time domain. The deviation is calculated using a multimodal response deviation measurement method, which includes: extracting frequency domain features representing vibration energy, time domain features representing impact peak value, and time series curve features representing torque trend from actual response data and predicted response data, respectively, and constructing multiple sets of corresponding feature vector pairs; The specific process for extracting the frequency domain features characterizing vibration energy is as follows: Fast Fourier Transform is performed on the actual acquired and simulated transmission shaft output torque signals to calculate their respective power spectral densities. Then, on the power spectral density curve, a frequency band from 10 Hz to 150 Hz (covering the main torsional vibration frequencies) is identified. The power values ​​at all frequency points within this frequency band are normalized to ensure that the total energy within the frequency band is one, thus obtaining two probability vectors describing the distribution of vibration energy within the frequency band, corresponding to the actual and predicted values, respectively. The specific process for extracting the time-domain features characterizing the impact peak is as follows: within the synchronous monitoring window, the local maximum points of the actual and predicted drive shaft torque signals are detected respectively; a minimum peak threshold is set, and only the peak points exceeding the threshold are retained, and the amplitudes of these peak points are recorded in chronological order to form the actual impact peak sequence and the predicted impact peak sequence. The specific process for extracting the time-series curve features that characterize the torque trend is as follows: directly take the complete time series of the drive motor output torque within the entire synchronous monitoring window as the actual torque trend curve and the predicted torque trend curve; For each pair of feature vectors, the specified feature distance function value is calculated. The frequency domain feature pairs of vibration energy are measured using relative entropy to determine their probability distribution differences. The calculation process for relative entropy to measure the probability distribution difference is as follows: Let the actual vibration energy distribution probability vector be P, and the predicted distribution probability vector be Q, both having the same length. The formula for calculating relative entropy is: for each corresponding element in the vector, calculate the value of that element in P multiplied by the logarithm of the value of that element in P divided by the value of the corresponding element in Q (base 2), and then sum the results for all elements. To prevent calculation errors caused by zero elements in Q, the Q vector is smoothed before calculation to ensure that each element has a very small positive value. The larger the relative entropy value, the greater the difference between the two probability distributions, meaning the greater the difference between the model-predicted vibration energy distribution and the actual distribution. The time-series curve characteristics of torque trends are analyzed using dynamic time warping distance (RTD) to measure their morphological differences. The calculation process for this RTD is as follows: Let the actual torque trend curve be sequence A with length N, and the predicted torque trend curve be sequence B with length M. First, construct an N x M cumulative cost matrix. Then, initialize the first row and first column of the matrix. Next, starting from the second row and second column, fill the matrix row by row and column by column. For the element in the i-th row and j-th column of the matrix, its value is equal to the Euclidean distance between the i-th point of sequence A and the j-th point of sequence B, plus the minimum value among its three adjacent elements to its left, above, and upper left. After completing the filling of the entire matrix, the matrix... The value of the element in the lower right corner is the dynamic time warping distance; this algorithm allows two curves to be non-linearly scaled and aligned on the time axis to find the best matching path. Its distance value can more accurately reflect the difference in the overall shape of the two curves, rather than just the point-to-point difference; the impact peak feature pairs calculate the absolute difference of their impact peak sequences; the method for calculating the absolute difference of their impact peak sequences is as follows: first, ensure that the actual and predicted impact peak sequence lengths are consistent. If the lengths are different, truncate the longer sequence or pad it by interpolation; then, calculate the absolute value of the difference between the peaks at corresponding positions in the two sequences; finally, add the absolute values ​​of the differences at all corresponding positions to obtain the total difference. Finally, the calculated feature distance function values ​​are multiplied by a pre-set positive weighting coefficient that sums to one, and all weighted distance values ​​are summed. The sum is the comprehensive multimodal response deviation metric. The larger the multimodal response deviation metric, the higher the inconsistency between the model prediction and the actual response. The pre-set positive weighting coefficient that sums to one can be set based on engineering experience. For example, if vibration suppression is the most important factor in the current working condition, the weighting coefficient for the relative entropy of the frequency domain characteristics of vibration energy can be 0.5; if high torque tracking accuracy is required, the weighting coefficient for the torque trend dynamic time warping distance can be 0.3; if the focus on the difference in impact peak values ​​is relatively low, the weighting coefficient for the absolute difference in impact peak values ​​can be 0.2. The sum of these weighting coefficients is one. The model parameters in the high-fidelity digital twin model are adjusted using a multimodal response deviation metric, specifically including: Using the calculated multimodal response bias metric and the model-entity consistency residual from the state observer as inputs, a robustness metric is constructed. The robustness metric is defined as a negative exponential function of the product of the multimodal response bias metric and the model-entity consistency residual, with a value between zero and one. A smaller robustness metric value indicates lower reliability of the current model's predictions. The specific expression for the negative exponential function is: the robustness metric equals the base of the natural logarithm, with its exponent being a negative sensitivity coefficient multiplied by the multimodal response bias metric and then by the model-entity consistency residual. The sensitivity coefficient is a preset positive number, such as 0.5, used to adjust the metric's sensitivity to bias and residuals. When the bias or residual is zero, the robustness metric is one, indicating that the model is completely reliable; as the bias or residual increases, the robustness metric rapidly decays from one to zero. Subsequently, the damping coefficient, which characterizes the energy dissipation of the torsional damper in the high-fidelity digital twin model, is taken as the target parameter to be adjusted. A Bayesian inference method based on adaptive learning rate is used for updating: First, it is assumed that the current value of the damping coefficient is prior knowledge, and its uncertainty is described by a prior variance. The initial value of the prior variance can be set according to the nominal value range of the damping coefficient. For example, if the nominal value of the damping coefficient is ten Newton-meter-seconds per radian, and its possible range is five to fifteen, then the initial prior variance can be set to 2.5 squared. In subsequent updates, the prior variance is passed from the posterior variance obtained from the previous update. Then, the robustness metric is regarded as a likelihood measure of the reasonableness of the damping coefficient under the observed data conditions. This step means that, in the Bayesian inference framework, "observed data" is interpreted as the reliability of the model's prediction under given parameters. The higher the robustness metric, the better the model's prediction matches reality under the current damping coefficient, that is, the more likely the parameter value is to be correct, and therefore the higher its likelihood. It can be approximated that the log-likelihood function is proportional to the logarithm of the robustness metric. Next, an adaptive learning rate is used to approximate the maximum a posteriori probability estimate. The adaptive learning rate is obtained by multiplying the base learning rate by one, subtracting the robustness metric, and then dividing by the prior standard deviation. This allows for faster updates when the model is unreliable and slower updates when the model itself has high uncertainty. The base learning rate is a small positive constant, such as 0.01, which serves as the baseline for the update step size. The prior standard deviation is the arithmetic square root of the prior variance. The calculation of the adaptive learning rate ensures that: when the robustness metric is very low (close to zero, indicating an unreliable model), the numerator minus the robustness metric is close to one, resulting in a larger learning rate that prompts rapid parameter correction; when the prior standard deviation is large (indicating high parameter uncertainty), the denominator increases, and the learning rate decreases, making updates more cautious when parameter confidence is low and avoiding blind, large adjustments. Finally, a correction amount for the damping coefficient is calculated. This correction amount is proportional to the product of the adaptive learning rate and the sum of the gradients of the log-likelihood and the log-prior. The current damping coefficient is added to this correction amount to obtain the updated damping coefficient value. Simultaneously, the updated parameter uncertainty estimate is calculated. The new damping coefficient and the new uncertainty estimate are written back to the parameter storage area of ​​the high-fidelity digital twin model to complete the online feedback adjustment. The approximate calculation process of the sum of the gradients of the log-likelihood and the log-prior is as follows: Assuming the prior distribution of the damping coefficient is normal, its logarithmic gradient is proportional to the difference between the current damping coefficient and the prior mean, with the proportionality coefficient being the negative reciprocal of the prior variance; the gradient of the log-likelihood is approximately... It seems that the sensitivity coefficient is multiplied by the model-entity consistency residual, and then multiplied by the partial derivative of the multimodal response bias metric with respect to the damping coefficient. The partial derivative of the multimodal response bias metric with respect to the damping coefficient can be approximated by recalculating the model response and bias metric after making a small perturbation near the current value of the damping coefficient, and then using the finite difference method. The gradients of these two parts are added together and then multiplied by the negative adaptive learning rate to obtain the correction amount of the damping coefficient. This process is essentially performing a gradient descent step with an adaptive step size to maximize the posterior probability. The update of parameter uncertainty is carried out according to the formula, reflecting the re-evaluation of the confidence of the parameter values ​​after obtaining new "observations" (robustness metrics).

[0029] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0030] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0031] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0032] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0033] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0034] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0035] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for integrated matching and vibration reduction of power source and hydraulic mechanical transmission based on bond graph, characterized in that, Specifically, the steps include the following: Step S1: In response to vehicle startup, the integrated topology model of drive motor-torsional damper-locked hydraulic torque converter, which includes nonlinear stiffness characteristics and damping parameters, is established based on bond graph theory. It receives real-time data streams from real vehicle sensors and corrects the internal state of the integrated topology model through a state observer to construct a high-fidelity digital twin model synchronized with the physical entity. Step S2: Deploy a reinforcement learning agent in the high-fidelity digital twin model. Configure the state space of the reinforcement learning agent as a feature vector containing real-time load torque, speed ratio and vibration spectrum features. Configure the action space as fine-tuning instructions for the control parameters in the active intervention logic of motor speed. Configure its reward function as a predefined composite reward function to drive the agent to learn vibration reduction. Step S3: When the vehicle is running, the current vehicle state collected by the edge controller is uploaded to the cloud. The cloud uses multiple simulation instances corresponding to the high-fidelity digital twin model to simulate the future responses of the reinforcement learning agent to various action sequences generated by the agent in response to the current vehicle state in parallel. The results of the future responses are evaluated based on the predefined composite reward function. The agent's policy network is updated through the policy optimization algorithm, and the optimized action sequence determined by the updated policy for the current vehicle state is sent to the edge controller. Step S4: The edge controller interprets the received optimized action sequence into specific control parameters and dynamically adjusts the locally stored motor speed active intervention logic. The system collects actual response data of the vehicle's active intervention logic for motor speed adjustment, calculates the deviation between the actual response data and the predicted response data generated by the high-fidelity digital twin model, and uses this deviation to adjust the model parameters in the high-fidelity digital twin model.

2. The method for integrated matching and vibration reduction of power source and hydraulic mechanical transmission based on bond graph as described in claim 1, characterized in that: In step S1, the state observer corrects the internal state of the integrated topology model, specifically including: The state observer continuously receives real-time data streams, which include at least the drive motor speed, drive motor output torque, and hydraulic torque converter turbine speed. Based on the received real-time data stream and the nonlinear stiffness characteristics and damping parameters built into the integrated topology model, the state observer performs an iterative filtering process that includes prediction and update steps to estimate the pump wheel speed of the hydraulic torque converter and the rotation angle of the torsional damper in the integrated topology model. Subsequently, the estimated pump impeller speed and angle values ​​are injected into the integrated topology model, replacing the corresponding state variables calculated by the integrated topology model through its own numerical integration. This completes the dynamic closed-loop correction of the internal state of the integrated topology model, forming a verified internal state set of the high-fidelity digital twin model.

3. The method for integrated matching and vibration reduction of power source and hydraulic mechanical transmission based on bond graph as described in claim 2, characterized in that: An integrated topology model based on bond graph theory is established, specifically including: The pre-set model parameters are loaded from the non-volatile memory of the edge controller to initialize the integrated topology model. The pre-set model parameters are non-linear stiffness characteristic functions and damping coefficients used to calculate the restoring torque component based on non-linear stiffness characteristics and the damping torque component based on damping parameters. During operation, the integrated topology model uses the verified internal state set obtained after correction by the state observer, combined with the real-time changing load inertia during operation, to solve the system dynamics differential equations derived from bond graph theory. It then calculates the energy flow and signal flow interaction between the drive motor, torsional damper and lock-up hydraulic torque converter in real time, thereby dynamically generating predictive response data that is synchronized with the transmission behavior of the physical entities.

4. The method for integrated matching and vibration reduction of power source and hydraulic mechanical transmission based on bond graph as described in claim 3, characterized in that: In step S2, the state space of the reinforcement learning agent is configured as a feature vector containing real-time load torque, speed ratio, and vibration spectrum characteristics, specifically including: From the output data obtained by real-time calculation of the high-fidelity digital twin model, the output torque of the drive motor after low-pass filtering is extracted as the real-time load torque feature; the ratio of the turbine speed of the hydraulic torque converter in the high-fidelity digital twin model to the estimated value of the pump wheel speed in the verified internal state set is calculated as the real-time speed ratio feature. Spectral analysis was performed on the rotation angle signal of the torsional vibration damper output in real time from the high-fidelity digital twin model. The proportion of vibration energy in a set bandwidth around the main resonant frequency of the integrated topology model was calculated as the vibration energy concentration characteristic. Finally, the feature vector composed of real-time load torque characteristics, real-time speed ratio characteristics, and vibration energy concentration characteristics is defined as the state space of the reinforcement learning agent.

5. The method for integrated matching and vibration reduction of power source and hydraulic mechanical transmission based on bond graph as described in claim 4, characterized in that: The predefined composite reward function is configured as follows: The value of the composite reward function consists of the weighted negative sum of the vibration penalty term, the peak impact penalty term, the motion smoothness penalty term, and the model uncertainty penalty term; Among them, the vibration penalty term is calculated based on the square integral of the difference between the drive motor angular acceleration and the pump wheel angular acceleration in the high-fidelity digital twin model within an adjustable time window; the impact peak penalty term is taken as the maximum absolute value of the drive shaft impact torque calculated by the high-fidelity digital twin model within the same time window. The action smoothness penalty term is defined by the Euclidean norm of the vector consisting of the fine-tuning instructions for the control parameters in the active intervention logic of the motor speed currently output by the reinforcement learning agent. The model uncertainty penalty term directly adopts the confidence index output by the state observer during the construction of the high-fidelity digital twin model; Each penalty is weighted and summed using pre-set positive weighting coefficients, and the negative value of the final result is the reward value at the current moment.

6. The method for integrated matching and vibration reduction of power source and hydraulic mechanical transmission based on bond graph as described in claim 5, characterized in that: In step S3, the cloud utilizes multiple simulation instances corresponding to the high-fidelity digital twin model to simulate in parallel the future responses of the reinforcement learning agent to various action sequences generated by the agent in response to the current vehicle state. Specifically, this includes: The cloud server receives the current vehicle status and the complete current runtime status of the high-fidelity digital twin model uploaded from the edge controller; based on the received complete current runtime status, the cloud server quickly clones hundreds to thousands of simulation instances that are completely identical to the high-fidelity digital twin model synchronized with the physical entity. The cloud-deployed reinforcement learning agent generates multiple candidate action sequences with subtle differences in number, matching the number of simulation instances, based on the policy network parameters of the reinforcement learning agent. Each candidate action sequence defines the adjustment instructions for each control parameter in the active intervention logic of motor speed in each control cycle within a preset short time domain in the future. The cloud scheduler dynamically assigns each candidate action sequence to an independent simulation instance. Driven by their respective assigned action sequences, all simulation instances independently and in parallel simulate the preset short time domain, thereby synchronously obtaining the future state response trajectory corresponding to each candidate action sequence.

7. The method for integrated matching and vibration reduction of power source and hydraulic mechanical transmission based on bond graph as described in claim 6, characterized in that: The future response is evaluated based on a predefined composite reward function. The agent's policy network is updated using a policy optimization algorithm. The optimized action sequence determined by the updated policy for the current vehicle state is then sent to the edge controller. Specifically, this includes: The cloud server aggregates the complete trajectory data, including state, action, and instantaneous reward value calculated by a predefined composite reward function, output by all simulation instances after the simulation is completed, forming a large-scale experience pool for policy network updates; The policy optimization algorithm employs a near-end policy optimization algorithm, which guides the parameter updates of the policy network by constructing an alternative objective function that includes a probability ratio pruning mechanism. After the policy network parameters are updated, the cloud server filters all candidate action sequences in the simulation according to the multi-objective optimization principle: the cumulative value of the composite reward function, the peak value of the vibration penalty term, and the peak value of the impact torque corresponding to the impact peak penalty term in the simulation results of each candidate action sequence are jointly constructed into a three-dimensional evaluation vector. The fast non-dominated sorting algorithm is used to process the evaluation vectors of all candidate action sequences to find the Pareto optimal solution set. Finally, based on the preset engineering priority rules, a candidate action sequence with the best overall performance is selected from the Pareto optimal solution set, marked as an optimized action sequence, and sent to the edge controller through the vehicle-cloud communication link.

8. The method for integrated matching and vibration reduction of power source and hydraulic mechanical transmission based on bond graph as described in claim 7, characterized in that: In step S4, the edge controller interprets the received optimized action sequence into specific control parameters and dynamically adjusts the locally stored motor speed active intervention logic, specifically including: The optimized sequence of actions received by the edge controller contains multiple consecutive control commands, each of which corresponds to the adjustment increment of the control parameters in the active intervention logic for motor speed within the next control cycle. During the interpretation process, a preset mapping table is first used to convert each adjustment increment in the action sequence into a specific target change in the corresponding control parameter. To ensure control stability and safety, incremental limiting processing is performed on each interpreted parameter target change, that is, limiting it to a preset upper limit of absolute value. Subsequently, a first-order inertial filtering algorithm is used to gradually transition the target parameter value after the limiting processing, so that the new control parameter smoothly approaches the target parameter value after the limiting processing. Finally, the new control parameter value after the smooth transition is written into the corresponding storage unit of the motor speed active intervention logic stored locally in the edge controller, and takes effect at the beginning of the next control cycle.

9. The method for integrated matching and vibration reduction of power source and hydraulic mechanical transmission based on bond graph as described in claim 8, characterized in that: The actual response data of the vehicle's active intervention logic for adjusted motor speed is collected, and the deviation between the actual response data and the predicted response data generated by the high-fidelity digital twin model is calculated, specifically including: After the new active intervention logic for motor speed takes effect, the edge controller initiates a synchronous monitoring window with the same preset short time domain length as the cloud simulation. During this window, actual response data is synchronously collected from the vehicle's sensors. The actual response data includes at least the real-time speed of the drive motor, the output torque of the drive motor, and the output torque of the drive shaft. At the same time, the edge controller drives a high-fidelity digital twin model locally, using the currently verified internal state set as the initial condition and employing the exact same control parameters as the adjusted active intervention logic for motor speed, to perform a fully synchronous forward simulation, thereby generating predicted response data within the preset short time domain. The deviation is calculated using a multimodal response deviation measurement method, which includes: extracting frequency domain features representing vibration energy, time domain features representing impact peak value, and time series curve features representing torque trend from actual response data and predicted response data, respectively, and constructing multiple sets of corresponding feature vector pairs; For each pair of feature vectors, the specified feature distance function value is calculated. For the frequency domain feature pairs of vibration energy, the relative entropy is used to measure the difference in their probability distribution; for the time-series curve feature pairs of torque trends, the dynamic time warping distance is used to measure their morphological difference; and for the impact peak feature pairs, the absolute value difference of their impact peak sequences is calculated. Finally, the calculated feature distance function values ​​are multiplied by a pre-set positive weighting coefficient that sums to one, and all weighted distance values ​​are summed. The sum is the comprehensive multimodal response deviation metric.

10. The method for integrated matching and vibration reduction of power source and hydraulic mechanical transmission based on bond graph as described in claim 9, characterized in that: The model parameters in the high-fidelity digital twin model are adjusted using a multimodal response deviation metric, specifically including: Using the calculated multimodal response deviation metric and the model-entity consistency residual from the state observer as inputs, a robustness metric is constructed. The robustness metric is defined as a negative exponential function of the product of the multimodal response deviation metric and the model-entity consistency residual, with a value between zero and one. Subsequently, the damping coefficient, which characterizes the energy dissipation of the torsional damper in the high-fidelity digital twin model, is taken as the target parameter to be adjusted. A Bayesian inference method based on adaptive learning rate is used to update it: First, it is assumed that the current value of the damping coefficient is prior knowledge, and its uncertainty is described by a prior variance. Then, the robustness metric is regarded as a likelihood measure of the reasonableness of the damping coefficient under the observed current data conditions; the maximum a posteriori probability estimate is approximated by an adaptive learning rate, which is obtained by multiplying the base learning rate by one, subtracting the robustness metric, and then dividing by the prior standard deviation. Finally, a correction amount for the damping coefficient is calculated, which is proportional to the product of the adaptive learning rate and the gradients of the log-likelihood and the log-prior. This correction amount is added to the current damping coefficient to obtain the updated new damping coefficient value. At the same time, the updated parameter uncertainty estimate is calculated. The new damping coefficient and the new uncertainty estimate are written back to the parameter storage area of ​​the high-fidelity digital twin model to complete the online feedback adjustment.

Images