Parameter progressive updating method and system for electrically assisted supercharger boost pressure control

Abstract

The application belongs to the technical field of vehicle-mounted supercharging system control, and discloses a parameter gradual updating electric auxiliary supercharger supercharging pressure control method and system, an artificial intelligence agent framework serving an online updating safety link of system control parameters is constructed, the learning intelligent agent generates candidate parameters according to the current system state and the future reference state sequence, the checking intelligent agent respectively predicts the controlled output in the preset time domain by using the candidate parameters and the current application parameters, and whether the candidate parameters meet the winning condition in the current period is determined through a cost function; when the candidate parameters meet the winning condition in the continuous N determination periods and the abnormal threshold is not triggered, the current application parameters are updated to the candidate parameters; otherwise, the called application parameters are restored to the backup parameters or the last stable application parameters, and the robustness, reliability and long-term self-adaptive ability of the electric auxiliary supercharging control system are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control of power systems and control technology of on-board supercharger systems, and in particular to a method and system for controlling the boost pressure of an electric auxiliary supercharger with progressive parameter updates. Background Technology

[0002] In the control system of an electric assisted turbocharger (also known as "eTurbo"), the controller typically relies on a simplified model of the controlled object for design, state estimation, and parameter tuning. This control method has the following drawbacks: Firstly, in highly dynamic scenarios where the operating conditions of the electric auxiliary turbocharger frequently and rapidly change, the inability to adjust fixed parameters online in real time and the lack of predictive ability for the future state of the system will lead to severe mismatch of the control model in highly dynamic scenarios, resulting in a significant decrease in dynamic performance. Secondly, over long time scales such as equipment aging, environmental changes, operating condition migration, and degradation of actuator characteristics, the dynamic characteristics of the controlled object will continuously deviate from the initial modeling results, making it difficult for fixed model parameters to maintain control performance in the long term.

[0003] Secondly, although existing online parameter learning methods can generate new parameter estimates based on real-time input and output data, most solutions directly or quasi-directly use the new estimated parameters for controller updates, lacking a mechanism for usability verification and safe use of candidate parameters.

[0004] Especially under conditions of noise disturbance, insufficient excitation, local false convergence, sudden changes in operating conditions, or transient data pollution, even if the parameter learning algorithm provides numerical parameter estimation results, the results may not be practically usable for control.

[0005] If such parameters are written directly into the controller, it can easily lead to a decrease in control performance, frequent parameter switching, and serious closed-loop risks.

[0006] Third, existing solutions often adopt a single-module or loosely coupled approach. There is a lack of a unified state management mechanism between candidate parameter generation, parameter availability verification, controller parameter writing, exception rollback, and cloud-edge data interaction. This leads to difficulties in state synchronization between modules, incomplete exception handling chains, and difficulty in standardizing online update processes. It is also not conducive to engineering expansion across multiple devices and scenarios.

[0007] Therefore, there is an urgent need to propose a new online parameter update method that utilizes real-time operational data to learn control model parameters online. After multi-dimensional safety verification, the learned model parameters are progressively written into the controller. Furthermore, when the control effect deviates from expectations, an exception rollback management is adopted to ensure the safety and reliability of the control effect. At the same time, it is also necessary to introduce a collaborative mechanism that can support task orchestration, message interaction, state management, and event triggering between modules, thereby realizing the integrated engineering implementation of online parameter learning, verification, updating, and rollback. Summary of the Invention

[0008] The purpose of this invention is to address the technical deficiencies in the prior art by providing a method and system for controlling the boost pressure of an electric-assisted turbocharger with progressively updated parameters.

[0009] The technical solution adopted to achieve the purpose of this invention is: A method for controlling the boost pressure of an electrically assisted supercharger with progressively updated parameters, comprising the following steps: Step 1: Collect system status data during the operation of the electric auxiliary turbocharger, including boost pressure, rotor angular velocity, compressor outlet mass flow rate, and engine intake manifold gas temperature. Step 2: Construct a physical mechanism model of the electric booster to provide physical boundaries for parameter learning of the electric booster control model; Step 3: Construct an electric auxiliary turbocharger control model containing the parameters to be updated, and use the model parameters already written to the controller as the current application parameters. The controlled input is the motor torque or target speed, and the controlled output is the boost pressure. Step 4: Construct an AI agent framework that includes a learning agent and a verification agent to serve the secure link for online updating of system control parameters; Step 5: The learning agent generates candidate parameters using a physical perception reinforcement learning method based on the data collected in Step 1, the current system state, and the sequence of future reference states. Step 6: The verification agent uses the candidate parameters and the current application parameters to predict the controlled output in the preset time domain, and determines whether the candidate parameters meet the winning conditions in the current period through the cost function. Step 7, when the candidate parameters are in continuous If the winning condition is met within each prediction step and no abnormal threshold is triggered, the current application parameters are gradually updated to candidate parameters according to the progressive update law, and the updated parameters are written to the controller; otherwise, an abnormal rollback is triggered, and the application parameters called by the controller are restored to the backup parameters or the previous stable application parameters. Step 8: The controller generates electric auxiliary turbocharger control commands based on the parameters updated in step 7, thus completing the control closed loop.

[0010] In the above technical solution, in step 2, the physical mechanism model of the electric auxiliary turbocharger includes a rotor dynamics model, a boost pressure model, and an engine intake manifold gas dynamics model, wherein: The rotor dynamics model is as follows: ; In the formula, The equivalent rotational inertia of the rotor system, For rotor angular acceleration, For the motor output torque, The aerodynamic torque provided to the turbine side This is the compressor load torque. This is the equivalent resistance torque corresponding to bearing friction and mechanical damping.

[0011] The boost pressure model is as follows: ; In the formula, This represents the change in the mass flow rate at the compressor outlet. This refers to the change in the angular velocity of the turbocharger rotor. This represents the change in boost pressure. This indicates the sensitivity of flow rate to rotational speed. This indicates the sensitivity of flow rate to pressure.

[0012] The gas dynamics model of the engine intake manifold is as follows: ; In the formula, The first derivative of the boost pressure. The gas constant is This refers to the air temperature in the engine intake manifold. For the intake manifold equivalent volume, This refers to the engine's air intake volume. The equivalent loss flow caused by leakage or bypass.

[0013] In the above technical solution, the electric auxiliary turbocharger control model in step 3 is as follows: ; In the formula, To control the output, , for Constant pressure increase; and These are the first and second derivatives of the boost pressure at the corresponding time points; for The motor input torque or target speed at all times. This is the first derivative of the motor input torque or target speed at the corresponding moment; , , , For the control model parameters, let represent the parameters of the control model damping term, pressure recovery term, input gain term, and differential feedforward term, respectively; This is the lumped disturbance term.

[0014] In the above technical solution, in step 4, the learning agent includes a physical perception network module and a policy network module, wherein: At each prediction step, firstly, the physical sensing network module, given the current system state and system control parameters, recursively generates the predicted system state value for the next step and sends it to the policy network module. The policy network module then generates the corresponding system control parameters for the next step based on the predicted system state value and the corresponding reference state sequence. Subsequently, the system control parameters are sent back to the physical sensing network module to obtain the predicted system state value for the next step. This predicted system state value is then used as the input to the policy network module for the next prediction step to generate new system control parameters. Through this alternating recursive process, the final state prediction sequence and system control parameter sequence in the prediction time domain can be obtained. The specific recursive process can be represented as follows: ; ; ; ; in, For a moment The system status, , , They are respectively Generated in time time, time, The predicted system state at time t. , This represents the total number of prediction steps. For at any time The generated effect is on the first System control parameters for each prediction step From From the start of time to the number The reference state sequence for each prediction step. From From the start of time to the number The system control parameter sequence generated in each prediction step This represents the candidate parameters obtained through learning. This represents the operation of taking the first value of the sequence. Indicates from The mapping relationship between the predicted system state values ​​and the corresponding reference state sequences to the system control parameter sequences. For the nominal model, For parameters The residual network.

[0015] In the above technical solution, the cost function in step 6 is: ; This is the evaluation value for the improvement effect of the candidate parameter in the current period. , , , These are the weighting coefficients for each evaluation indicator; This represents the improvement in the prediction error of the candidate parameters relative to the currently applied parameters. This is the volatility criterion for candidate parameters; This is the criterion for determining whether candidate parameters satisfy the physical mechanism constraints; This is the metric for determining whether the current data meets the conditions for sufficient incentive. when When the candidate parameter is determined to meet the winning condition in the current period, This is a preset threshold.

[0016] In the above technical solutions, The calculation formula is: ; in, and These represent the current application parameters and candidate parameters in the future, respectively. The prediction error index of the step; In the above technical solution, the asymptotic update law in step 7 is: ; in, for Application parameters at time, for Application parameters at time, Candidate parameters, For smoothing gain.

[0017] In the above technical solution, in step 7, The threshold is 2-10. The triggering thresholds include prediction error deterioration exceeding the threshold, parameters exceeding the allowable range of physical constraints, or pressure control performance deteriorating beyond the threshold.

[0018] Another aspect of the present invention includes a system for implementing the electric-assisted turbocharger boost pressure control method for progressively updating the parameters, comprising: An AI agent framework serving the online update security link of system control parameters includes a message interaction module, a state management module, and an event triggering module. The state management module maintains the current application parameters, candidate parameters, backup parameters, consecutive win counters, and abnormal state flags. The event triggering module triggers parameter updates, parameter rollbacks, or stimulus enhancements based on the verification results. The data acquisition module is used to collect real-time input and output data from the electric auxiliary turbocharger and transmit it to the learning agent; The learning agent is used to generate candidate parameters based on the real-time input and output data and the reinforcement learning algorithm, and then send them to the verification agent. The verification agent is used to perform multi-dimensional quality evaluation of candidate parameters, continuous superiority judgment, incremental update control and abnormal rollback decision, and transmit the decision results to the controller. The controller invokes application parameters from the verification agent to execute closed-loop control.

[0019] In another aspect of the invention, there is an electrically assisted turbocharger comprising the aforementioned system.

[0020] Compared with the prior art, the beneficial effects of the present invention are: 1. This method constructs an artificial intelligence agent framework that serves the online update security link of system control parameters, mobilizes the learning agent and the verification agent to work together, and realizes real-time online learning, availability determination, progressive update application and anomaly protection of control model parameters. It is applicable to internal combustion engine turbocharging system, motor system and other industrial control scenarios with parameter drift, operating condition change and model mismatch problems. 2. This invention combines online parameter learning with a parameter safety verification mechanism to transform candidate parameters from being estimable to being safely applicable; it avoids performance degradation caused by directly writing candidate parameters into the controller through continuous winner determination and incremental update mechanisms; it improves the system's operational stability under complex operating conditions through an anomaly rollback mechanism; and it enhances the robustness, reliability, and long-term adaptive capability of the electric auxiliary boost control system by realizing the modularity and scalability of the parameter management process through a dual-agent collaborative structure. Attached Figure Description

[0021] Figure 1 The diagram shown is a flowchart of the present invention.

[0022] Figure 2 The diagram shown is an overall block diagram of the present invention.

[0023] Figure 3 The figure shows the curve of the online learning and updating process of the system control parameters of the present invention.

[0024] Figure 4 The figure shows a comparison curve of the boost pressure control effect of the present invention and the system under fixed control parameters. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to specific embodiments.

[0026] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0027] Example 1 like Figure 1 and Figure 2 As shown, a method for controlling the boost pressure of an electrically assisted turbocharger with progressively updated parameters includes the following steps: Step 1: Collect system status data during the operation of the electric auxiliary turbocharger, including boost pressure, rotor angular velocity, compressor outlet mass flow rate, and engine intake manifold gas temperature. Step 2: Construct a physical mechanism model of the electric-assisted supercharger, including at least a rotor dynamics model, a boost pressure model, and an engine intake manifold gas dynamics model, to provide physical boundaries for learning system control parameters.

[0028] The rotor dynamics model is as follows: ; In the formula, The equivalent rotational inertia of the rotor system, For rotor angular acceleration, For the motor output torque, The aerodynamic torque provided to the turbine side This is the compressor load torque. This is the equivalent resistance torque corresponding to bearing friction and mechanical damping.

[0029] The boost pressure model is as follows: ; In the formula, This represents the change in the mass flow rate at the compressor outlet. This refers to the change in the angular velocity of the turbocharger rotor. This represents the change in boost pressure. This indicates the sensitivity of flow rate to rotational speed. This indicates the sensitivity of flow rate to pressure.

[0030] The gas dynamics model of the engine intake manifold is as follows: ; In the formula, The first derivative of the boost pressure. The gas constant is This refers to the air temperature in the engine intake manifold. For the intake manifold equivalent volume, This refers to the engine's air intake volume. The equivalent loss flow caused by leakage or bypass.

[0031] Step 3: Construct a control model of the electric auxiliary turbocharger containing the parameters to be updated, and use the model parameters that have been written into the controller as the current application parameters. The control input is either the motor torque or the target speed command. A second-order dynamic model of the boost pressure control loop of the electric-assisted turbocharger is constructed. Based on the physical nature of the electric-assisted turbocharger (rotor mechanical inertia + intake manifold volume effect), a second-order linearized model is established to determine the dominant dynamic from the target speed command to the manifold pressure output.

[0032] The control model for the electric-assisted turbocharger is as follows: ; in: To control the output (boost pressure); For control input ( (motor input torque or target speed at any time). for The first derivative, that is, the first derivative of the motor input torque or target speed at the corresponding moment; , , , For the control model parameters, let represent the parameters of the control model damping term, pressure recovery term, input gain term, and differential feedforward term, respectively; The total disturbance is lumped together, including unmodeled dynamics, parameter uncertainties, and external disturbances.

[0033] Step 4: Construct an AI agent framework that serves the online update security link of system control parameters. This framework includes a learning agent and a verification agent, which are used to maintain candidate parameters, current application parameters, backup parameters, consecutive win counters, and abnormal status flags in the vehicle controller or edge computing unit, and trigger parameter writing or abnormal rollback based on the verification results.

[0034] The learning agent and the verification agent use an artificial intelligence agent framework to orchestrate tasks, exchange messages, synchronize states, and manage abnormal events.

[0035] in: 1. The learning agent is responsible for subscribing to the real-time input and output data streams of the controlled object, learning parameters through physical perception reinforcement learning methods, generating candidate parameters, and sending them to the verification agent; 2. The verification agent is responsible for receiving candidate parameters, current application parameters and operating status data, performing safety verification and continuous success decision. When the candidate parameters meet the continuous success condition, they are progressively updated to the controller. When the prediction error deteriorates beyond the threshold, the parameters exceed the allowable range of physical constraints, or the pressure control performance of the boost pressure decreases beyond the threshold, an abnormal rollback is triggered to restore the controller parameters to the backup parameters or the previous stable application parameters. 3. The AI ​​agent framework is responsible for the intelligent coordination of functions such as parameter maintenance, continuous win counting, parameter updates, parameter rollback, or incentive enhancement.

[0036] First, the control model is discretized: ; in, The controlled output (i.e., boost pressure). For control input (i.e., motor input torque or target speed). for System status at all times The parameters of the control model to be learned include , , and , This is the lumped disturbance term, which includes unmodeled dynamics, parameter perturbations, and external disturbances.

[0037] The AI ​​agent framework includes a message interaction module, a state management module, and an event triggering module. The message interaction module is responsible for data transmission and interaction between the sensor, the learning agent, the verification agent, and the controller. The state management module maintains the current application parameters, candidate parameters, backup parameters, continuous win counters, and abnormal state flags. The event triggering module triggers parameter updates, parameter rollbacks, or stimulus enhancements based on the verification results.

[0038] The abnormal rollback includes: when the prediction error deteriorates beyond a threshold, the candidate parameter or the applied parameter exceeds the physical allowable range, or the boost pressure overshoot or response error exceeds a threshold, the applied parameter is rolled back to the backup parameter or the previous stable applied parameter.

[0039] Step 5: The learning agent generates candidate parameters based on the real-time input / output data, the current system state, and the future reference state sequence; The learning agent consists of a Perception module and an Actor module, which are responsible for predicting the state of the electric booster system and learning the system control parameters online, respectively.

[0040] Actor module

[0041] The Actor module is used to generate the sequence of system control parameters for the electric booster in the prediction time domain.

[0042] Considering that the control of the electric auxiliary turbocharger not only needs to be adjusted in real time according to the current system state, but also needs to take into account the speed reference target in the future time domain, this invention uses the predicted system state value and the reference state sequence in the future prediction time domain as input to the Actor network.

[0043] Set time The system status is The total number of prediction steps is Then the future reference state sequence can be represented as: ; in, For future reference state sequence, , , , They represent the future number of... , , , The reference state corresponding to the prediction step.

[0044] The Actor network is used to generate the sequence of system control parameters for the electric auxiliary turbocharger in the prediction time domain: ; in, From Time to the Future The system prediction parameter sequence of the step, , , At time respectively The generated effect is on the first , , System control parameters for each prediction step.

[0045] The system control parameters include , , , .

[0046] The optimization objective of the Actor network is defined as: ; in, For Actor network parameters, For optimal parameters, This is the loss function for the Actor network.

[0047] It should be noted that, although Actor networks in The output at any given time is the sequence of system control parameters for the entire prediction time domain, but in actual online control, only the first parameter vector corresponding to the current time is output. It is applied to the electric boost control system.

[0048] At the next sampling time, the input is reconstructed and a new parameter sequence is generated based on the new actual state of the system.

[0049] Perception module

[0050] The Perception module is used to model the state dynamics of the electric booster. Its goal is to recursively predict the state of the electric booster system in the prediction time domain, given the current system state and system control parameters.

[0051] During training, the Actor network interacts with the Perception module to obtain the future state response sequence under the predicted system control parameter sequence, thereby reducing interaction with the original electric-assisted boost simulation environment and improving the efficiency of policy update.

[0052] To balance model interpretability and prediction accuracy, the Perception module employs a composite perception structure of "nominal model + residual network".

[0053] Therefore, the one-step prediction of the Perception module can be expressed as: ; in, For the nominal model, For parameters residual network, This is the predicted value of the system state at the next moment.

[0054] This invention employs a deduction mechanism that alternates between generating system control parameters and predicting states in the prediction time domain.

[0055] Specifically, at each prediction step, firstly, the physical sensing network module, given the current system state and system control parameters, recursively generates the predicted system state value for the next step and sends it to the policy network module. The policy network module then generates the corresponding system control parameters for the next step based on the predicted system state value and the corresponding reference state sequence. Subsequently, the system control parameters are sent back to the physical sensing network module to obtain the predicted system state value for the next step. This predicted system state value is then used as the input to the policy network module for the next prediction step to generate new system control parameters. Through this alternating recursive process, the final state prediction sequence and system control parameter sequence in the prediction time domain can be obtained. The specific recursive process can be represented as follows: ; ; ; ; in, For a moment The system status, , , They are respectively Generated in time time, time, The predicted system state at time t. , For at any time The generated effect is on the first System control parameters for each prediction step From From the start of time to the number The reference state sequence for each prediction step. From From the start of time to the number The system control parameter sequence generated in each prediction step This represents the candidate parameters obtained through learning. This represents the operation of taking the first value of the sequence. Indicates from The mapping relationship between the predicted system state values ​​and the corresponding reference state sequences and the system control parameter sequences.

[0056] The learning law of the Perception module can be expressed as: ; in, These are parameters for the Perception module. Update the gradient for the residual network parameters. For optimal parameters, The learning rate is updated for the residual network. This is the loss function for the state prediction module.

[0057] Step 6: Based on the verification agent, predict the controlled output within the preset time domain using the candidate parameters and the current application parameters respectively. Calculate the prediction error index corresponding to the candidate parameters and the prediction error index corresponding to the current application parameters, and calculate the parameter fluctuation index, physical constraint satisfaction index, and data incentive sufficiency index of the candidate parameters. When the prediction error index corresponding to the candidate parameters is better than the prediction error index corresponding to the current application parameters, and the parameter fluctuation index, physical constraint satisfaction index, and data incentive sufficiency index of the candidate parameters all meet preset conditions, it is determined that the candidate parameters meet the winning conditions in the current judgment period. The verification agent includes an evaluation (MPC-critic) module for constructing the model predictive control cost function, which is used for value assessment.

[0058] The MPC-critic module is used to evaluate the future control effect under the current system control parameters and is a key link connecting the Perception module and the Actor network parameter updates.

[0059] The MPC-critic designed in this invention draws on the idea of ​​model predictive control, and directly substitutes the candidate parameters generated based on the state trajectory and system control parameter sequence in the predicted time domain into the explicitly constructed cost function for evaluation.

[0060] After the Actor network and Perception module alternately recursively obtain the system control parameter sequence and the predicted state sequence, the MPC-critic module substitutes the predicted state sequence and the system control parameter sequence in the prediction time domain into the finite-time cumulative cost function to construct the current time step. The cost function.

[0061] The cost function consists of the following aspects: the prediction error corresponding to the candidate parameter is less than the prediction error corresponding to the current application parameter, and the reduction in error is greater than a preset threshold; the fluctuation index of the candidate parameter is less than a preset fluctuation threshold; the candidate parameter satisfies the physical constraints; and the current data satisfies the incentive sufficiency condition.

[0062] The cost function expression within a single period is: ; in, This is the evaluation value for the improvement effect of the candidate parameter in the current period. , , , These are the weighting coefficients for each evaluation indicator; This represents the improvement in the prediction error of the candidate parameters relative to the currently applied parameters; This is the volatility criterion for candidate parameters; This is the criterion for determining whether candidate parameters satisfy the physical mechanism constraints; This is the criterion for determining whether the current data meets the conditions for sufficient incentive.

[0063] The improvement in prediction error can be expressed as: ; in, and These represent the current application parameters and candidate parameters in the future, respectively. The prediction error index of the step; A candidate parameter is considered superior to the currently applied parameter only if it meets the following conditions: ; in, This is a preset threshold.

[0064] Step 7, when the candidate parameters are in continuous If the winning condition is met within a certain number of prediction steps and no abnormal threshold is triggered, the current application parameters are progressively updated to candidate parameters according to the asymptotic update law, and the updated parameters are written to the controller. In this embodiment, It is 2-10; Specifically, if and only if the above conditions are continuous If all prediction steps are successful, and no physical constraint anomaly, volatility anomaly, or insufficient excitation anomaly is triggered within the continuous period, the candidate parameters are determined to meet the continuous winning condition, and the next parameter update work is carried out.

[0065] When a candidate parameter satisfies the condition of continuous superiority, the currently applied parameter is not directly replaced. Instead, the following asymptotic update law is adopted: ; in, for Application parameters at time, for Application parameters at time, Candidate parameters, To smooth the gain, in this embodiment, we take... .

[0066] During parameter application, when it is detected that the prediction error deteriorates beyond the threshold, the parameter exceeds the allowable range of physical constraints, or the pressure control performance of the booster exceeds the threshold, an abnormal rollback is triggered, restoring the application parameters called by the controller to the backup parameters or the previous stable application parameters. Step 8: The controller generates electric auxiliary turbocharger control commands based on the parameters updated in step 7, thus completing the control closed loop.

[0067] During parameter application, the verification agent continuously monitors prediction errors, parameter physical constraints, and control performance indicators.

[0068] When a prediction error deteriorates beyond a threshold, a parameter goes out of bounds, or control performance significantly degrades, a parameter rollback operation is immediately triggered to restore the controller's currently applied parameters to backup parameters. Or the previous stable application parameters.

[0069] In the implementation, to ensure the safety management of the parameter update process, the parameter management process is controlled by a finite state machine, which includes a learning monitoring state, a parameter application state, a backup state, and an incentive enhancement state.

[0070] In the learning monitoring state, the learning agent runs continuously to generate candidate parameters, and the controller uses the currently stable application parameters or preset backup parameters. In the parameter application state, when the candidate parameters meet the continuous superiority judgment conditions, the candidate parameters are written to the controller through a progressive update method. In the backup state, when an anomaly is detected during parameter application, the controller reverts to the backup parameters. In the stimulus enhancement state, when the system stimulus is insufficient, the parameter identifiability is improved by superimposing test disturbance signals on the control input. In this state, the candidate parameters do not participate in the controller update.

[0071] In this implementation, the state transition relationship of the state machine can be represented by the following function: ; in, Indicates the first The state machine states for each cycle. Controlling the effects and costs.

[0072] like Figure 3 This paper demonstrates a process in which the system control parameters converge to the target value under a step condition through online learning and incremental updating based on the present invention.

[0073] like Figure 4 As shown, under a working condition where the target boost pressure pulsates periodically, the system control parameters obtained by the parameter learning, safety verification, and incremental update of this invention can ensure that the boost pressure tracks the target value quickly and accurately in each pulsation cycle; however, when the system control parameters are fixed, the boost pressure exhibits significant overshoot and oscillation over time, eventually leading to an uncontrolled state.

[0074] Example 2 A system for implementing the electric-assisted turbocharger boost pressure control method of Embodiment 1 includes: An AI agent framework serving the online update security link of system control parameters includes a message interaction module, a state management module, and an event triggering module. The state management module maintains the current application parameters, candidate parameters, backup parameters, consecutive win counters, and abnormal state flags. The event triggering module triggers parameter updates, parameter rollbacks, or stimulus enhancements based on the verification results.

[0075] The data acquisition module is used to collect real-time input and output data from the electric auxiliary turbocharger and transmit it to the learning agent; The learning agent is used to generate candidate parameters based on the real-time input and output data and the reinforcement learning algorithm, and then send them to the verification agent. The verification agent is used to perform multi-dimensional quality evaluation of candidate parameters, continuous superiority judgment, incremental update control and abnormal rollback decision, and transmit the decision results to the controller. The controller invokes application parameters from the verification agent to execute closed-loop control.

[0076] Example 3 An electric booster includes the electric booster pressure control system with progressively updated parameters as described in Example 2.

[0077] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for controlling the boost pressure of an electrically assisted supercharger with progressively updated parameters, characterized in that, Includes the following steps: Step 1: Collect system status data during the operation of the electric auxiliary turbocharger, including boost pressure, rotor angular velocity, compressor outlet mass flow rate, and engine intake manifold gas temperature. Step 2: Construct a physical mechanism model of the electric booster to provide physical boundaries for parameter learning of the electric booster control model; Step 3: Construct an electric auxiliary turbocharger control model containing the parameters to be updated, and use the model parameters already written to the controller as the current application parameters. The controlled input is the motor torque or target speed, and the controlled output is the boost pressure. Step 4: Construct an AI agent framework that includes a learning agent and a verification agent to serve the secure link for online updating of system control parameters; Step 5: The learning agent generates candidate parameters using a physical perception reinforcement learning method based on the data collected in Step 1, the current system state, and the sequence of future reference states. Step 6: The verification agent uses the candidate parameters and the current application parameters to predict the controlled output in the preset time domain, and determines whether the candidate parameters meet the winning conditions in the current period through the cost function. Step 7, when the candidate parameters are in continuous If the winning condition is met within each prediction step and no abnormal threshold is triggered, the current application parameters are gradually updated to candidate parameters according to the progressive update law, and the updated parameters are written to the controller. Otherwise, an abnormal rollback is triggered, restoring the application parameters called by the controller to the backup parameters or the previous stable application parameters; Step 8: The controller generates electric auxiliary turbocharger control commands based on the parameters updated in step 7, thus completing the control closed loop.

2. The electric auxiliary turbocharger boost pressure control method with progressive parameter updates as described in claim 1, characterized in that, In step 2, the physical mechanism model of the electrically assisted supercharger includes a rotor dynamics model, a boost pressure model, and an engine intake manifold gas dynamics model, wherein: The rotor dynamics model is as follows: ； In the formula, The equivalent rotational inertia of the rotor system, For rotor angular acceleration, For the motor output torque, The aerodynamic torque provided to the turbine side This is the compressor load torque. This is the equivalent resistance torque corresponding to bearing friction and mechanical damping; The boost pressure model is as follows: ； In the formula, This represents the change in the mass flow rate at the compressor outlet. This represents the change in the angular velocity of the turbocharger rotor. This represents the change in boost pressure. This indicates the sensitivity of flow rate to rotational speed. Indicates the sensitivity of flow rate to pressure; The gas dynamics model of the engine intake manifold is as follows: ； In the formula, The first derivative of the boost pressure. The gas constant is This refers to the air temperature in the engine intake manifold. For the intake manifold equivalent volume, This refers to the engine's air intake volume. The equivalent loss flow caused by leakage or bypass.

3. The electric auxiliary turbocharger boost pressure control method with progressive parameter updates as described in claim 1, characterized in that, In step 3, the electric auxiliary turbocharger control model is as follows: ； In the formula, To control the output, , for Constant pressure increase; and These are the first and second derivatives of the boost pressure at the corresponding time points; for The constant motor input torque or target speed. This is the first derivative of the motor input torque or target speed at the corresponding moment; , , , For the control model parameters, let represent the parameters of the control model damping term, pressure recovery term, input gain term, and differential feedforward term, respectively; This is the lumped disturbance term.

4. The electric auxiliary turbocharger boost pressure control method with progressive parameter updates as described in claim 1, characterized in that, In step 4, the learning agent includes a physical perception network module and a policy network module, wherein: At each prediction step, firstly, the physical sensing network module, given the current system state and system control parameters, recursively generates the predicted system state value for the next step and sends it to the policy network module. The policy network module then generates the corresponding system control parameters for the next step based on the predicted system state value and the corresponding reference state sequence. Subsequently, the system control parameters are sent back to the physical sensing network module to obtain the predicted system state value for the next step. This predicted system state value is then used as the input to the policy network module for the next prediction step to generate new system control parameters. Through this alternating recursive process, the final state prediction sequence and system control parameter sequence in the prediction time domain can be obtained. The specific recursive process can be represented as follows: ； ； ； ； in, For a moment The system status, , , They are respectively Generated in time time, time, The predicted system state at time t. , This represents the total number of prediction steps. For at any time The generated effect is on the first System control parameters for each prediction step From From the start of time to the number The reference state sequence for each prediction step. From From the start of time to the number The system control parameter sequence generated in each prediction step This represents the candidate parameters obtained through learning. This represents the operation of taking the first value of the sequence. Indicates from The mapping relationship between the predicted system state values ​​and the corresponding reference state sequences to the system control parameter sequences. For the nominal model, For parameters The residual network.

5. The electric-assisted turbocharger boost pressure control method with progressive parameter updates as described in claim 1, characterized in that, In step 6, the cost function is: ； This is the evaluation value for the improvement effect of the candidate parameter in the current period. , , , These are the weighting coefficients for each evaluation indicator; This represents the improvement in the prediction error of the candidate parameters relative to the currently applied parameters. This is the volatility criterion for candidate parameters; This is the criterion for determining whether candidate parameters satisfy the physical mechanism constraints; This is the metric for determining whether the current data meets the conditions for sufficient incentive. when When the candidate parameter is determined to meet the winning condition in the current period, This is a preset threshold.

6. The electric-assisted turbocharger boost pressure control method with progressive parameter updates as described in claim 5, characterized in that, The calculation formula is: ； in, and These represent the current application parameters and candidate parameters in the future, respectively. The prediction error index of the step, This represents the total number of prediction steps.

7. The electric-assisted turbocharger boost pressure control method with progressive parameter updates as described in claim 1, characterized in that, In step 7, the asymptotic update law is: ； in, for Application parameters at time, for Application parameters at time, Candidate parameters, To smooth the gain.

8. The electric-assisted turbocharger boost pressure control method with progressive parameter updates as described in claim 1, characterized in that, In step 7 The threshold is 2-10. The triggering thresholds include prediction error deterioration exceeding the threshold, parameters exceeding the allowable range of physical constraints, or pressure control performance deteriorating beyond the threshold.

9. A system for implementing the electric-assisted turbocharger boost pressure control method with progressive parameter updates as described in any one of claims 1-8, characterized in that, include: An AI agent framework serving the online update security link of system control parameters, including a message interaction module, a state management module, and an event triggering module; The status management module maintains current application parameters, candidate parameters, backup parameters, consecutive win counters, and abnormal status flags; The event triggering module triggers parameter updates, parameter rollbacks, or stimulus enhancements based on the verification results; The data acquisition module is used to collect real-time input and output data from the electric auxiliary turbocharger and transmit it to the learning agent; The learning agent is used to generate candidate parameters based on the real-time input and output data and the reinforcement learning algorithm, and then send them to the verification agent. The verification agent is used to perform multi-dimensional quality evaluation of candidate parameters, continuous superiority judgment, incremental update control and abnormal rollback decision, and transmit the decision results to the controller. The controller invokes application parameters from the verification agent to execute closed-loop control.

Images