Control system and method for intelligent engine cooling water pump actuator

Abstract

The invention discloses a control system and method for an intelligent engine cooling water pump actuator, and the system integrates a multi-mode perception technology, dynamic thermal entropy feature extraction, a hierarchical reinforcement learning strategy, graph neural network collaborative optimization, adaptive model prediction control, and an edge calculation deployment and fault self-repairing mechanism. A whole-process intelligent cooling control framework is constructed; collecting multi-source sensor data, calculating dynamic thermal entropy through a hybrid mechanism, and performing weighted fusion and anomaly suppression on the data; a hierarchical control strategy is adopted to make a cooling water pump target rotating speed decision, a finite element analysis method is introduced to dynamically predict thermal stress, the thermal stress serves as a penalty term to be incorporated into a strategy optimization process, a graph neural network is constructed, and inter-subsystem coupling relation modeling and cooperative control instruction generation are achieved; a self-adaptive MPC optimization model is constructed, and a lightweight control model compression method based on knowledge distillation is adopted to adapt to vehicle-mounted edge computing resource limitation; and designing an online thermal entropy monitoring mechanism and an incremental learning self-repairing strategy.

Description

TECHNICAL FIELD

[0001] The application relates to a control system and method for an intelligent engine cooling water pump actuator, in particular to a control system and method for an automobile engine cooling water pump actuator based on multi-modal perception and deep reinforcement learning, and belongs to the technical field of automobile engine thermal management. BACKGROUND

[0002] With the rapid development of the automobile industry, the engine thermal management system has become a key component to ensure vehicle performance, reduce fuel consumption and meet increasingly stringent emission standards. The engine cooling system mainly circulates coolant through the operation of the water pump to dissipate the heat generated during engine operation. This process can prevent engine overheating and reduce thermal stress accumulation, which may otherwise accelerate component aging and lead to premature failure. However, traditional cooling pump control strategies still mainly rely on pre-defined fixed rules or basic switch logic. These methods lack real-time perception of the actual operating state of the engine, dynamic load conditions and surrounding environmental factors. Therefore, the system often cannot maintain optimal thermal balance under various driving scenarios.

[0003] In addition, as the powertrain system develops towards higher efficiency, precision and intelligence, the engine operating conditions also become more complex and variable. The limitations of traditional control methods are therefore more apparent. For example, under low load conditions, excessive coolant flow can prevent the engine from reaching optimal operating temperature, thereby reducing thermal efficiency and increasing fuel consumption. Conversely, under high load conditions, insufficient coolant flow can cause local overheating, uneven thermal stress, and accelerate the wear or failure of key components. Obviously, fixed or rule-based control strategies cannot meet the requirements of dynamic thermal regulation under complex actual operating conditions, ultimately hindering the performance and energy-saving potential of vehicles.

[0004] In recent years, research has explored intelligent control methods based on single-modal data such as coolant temperature or engine load. Although these methods have made preliminary progress, they often suffer from low data utilization efficiency, insufficient control accuracy and poor anti-interference ability. Although some artificial intelligence-based methods exhibit adaptive ability, their reliance on single-modal perception limits the understanding of the comprehensive thermal behavior of the engine, especially under multi-faceted operating scenarios. More importantly, they are difficult to capture the dynamic coupling between subsystems, which limits reliability and universality, thereby hindering large-scale industrial applications. SUMMARY

[0005] Purpose of the Invention: Addressing the problems and shortcomings of existing technologies, this invention provides a control system and method for an intelligent engine cooling water pump actuator. This invention integrates multimodal sensing data to enable real-time, efficient analysis and response to complex engine operating conditions. It can improve the accuracy, efficiency, and adaptability of water pump control, thereby improving overall engine thermal management, extending component lifespan, and significantly reducing vehicle fuel consumption and emissions.

[0006] Technical Solution: A smart engine cooling water pump actuator control method based on multimodal perception and deep reinforcement learning aims to solve the problems of slow response, high energy consumption, and low thermal management efficiency of traditional water pump control strategies when dealing with complex operating conditions. This method integrates multimodal perception technology, dynamic thermal entropy feature extraction, hierarchical reinforcement learning strategy, graph neural network collaborative optimization, adaptive model predictive control, and edge computing deployment and fault self-healing mechanism to construct a full-process intelligent cooling control framework. First, multi-source sensor data, including coolant temperature, engine oil temperature, engine load, ambient temperature, water pump speed, and vehicle speed, are collected. Dynamic thermal entropy is calculated through an innovative "thermodynamic entropy-information entropy" hybrid mechanism. The data is weighted and fused, and anomaly suppression is performed to obtain a stable and reliable state feature vector. Subsequently, a hierarchical control strategy based on meta-reinforcement learning is used to determine the target speed of the cooling water pump. Simultaneously, finite element analysis is introduced to dynamically predict thermal stress, which is incorporated as a penalty term into the strategy optimization process to improve control robustness. Next, a graph neural network is constructed with engine cooling system actuators such as water pumps, thermostats, fans, and radiators as nodes to model the coupling relationships between various subsystem components and generate collaborative control commands. The listed components are merely examples; other relevant components can be added as needed in specific applications. Based on this, an adaptive model predictive control optimization with dynamic thermal entropy constraints is constructed to achieve fine-grained control of cooling behavior. A lightweight control model compression method based on knowledge distillation is proposed to adapt to the limitations of onboard edge computing resources. Finally, an online thermal entropy monitoring mechanism and an incremental learning self-healing strategy are designed, possessing the capabilities of anomaly detection, fault isolation, and rapid strategy recovery. This invention overcomes the limitations of traditional rule-based control methods, achieving intelligent, adaptive, and highly efficient control of the engine thermal management system, significantly improving engine reliability and thermal efficiency, and possessing promising engineering application prospects and industrial promotion value.

[0007] A control method for an intelligent engine cooling water pump actuator includes the following steps:

[0008] S1, Multimodal perception and dynamic thermal entropy fusion feature extraction. Engine coolant temperature T is collected. coolant Ambient temperature T ambient Engine load, oil temperature oil Pump speed Tpump and vehicle speed V vehicle Using multi-source sensor data, construct the state vector x at time t. t =[T coolant (t),T ambient (t),Load(t),T oil (t),T pump (t),V vehicle (t)]. Combining thermodynamic entropy and information entropy, a dynamic thermal entropy is proposed to quantify the thermal efficiency and energy dissipation of a cooling system, defined as:

[0009]

[0010] Thermodynamic entropy is defined as:

[0011]

[0012] in, Q is the thermodynamic entropy, used to describe the heat flow distribution characteristics of an engine cooling system. i (t) represents the heat flow rate of the i-th heat source at time t, such as cylinder heat, exhaust heat, etc.; T ambient (t) represents the ambient temperature; T i (t) represents the temperature of the i-th heat source at time t; n represents the total number of heat sources considered in the system. The information entropy is:

[0013]

[0014] Where j represents the index of different sensing data modes, P j (t) represents the normalized probability distribution of the j-th mode at time t, and the formula uses the natural logarithm (base e); the weighting factor α is adaptively determined based on the entropy fluctuation.

[0015]

[0016] Where, σ info (t) represents the degree of information entropy fluctuation, σ thermo (t) represents the degree of thermodynamic entropy fluctuation. Finally, by using dynamic thermodynamic entropy weighting, the state characteristics are obtained by fusing multimodal data:

[0017] X fused (t)=w(t)⊙x(t),

[0018] Among them, X fused (t) represents the fusion feature, W i (t) represents the fusion weights. Let be the comprehensive entropy of the i-th mode. This method achieves efficient fusion of multimodal data and suppression of abnormal interference, effectively improving the stability and intelligence of the control strategy.

[0019] S2: Meta-reinforcement learning-driven hierarchical deep reinforcement control strategy. Based on the fusion state in step S1, this invention proposes a hierarchical deep reinforcement control framework based on meta-reinforcement learning, specifically including:

[0020] S2.1 Constructs a two-layer control architecture, defined as an upper-layer meta-learner and a lower-layer policy network. The upper-layer meta-learner is trained on data from different vehicle models and operating conditions to quickly generalize and adapt to the control requirements of the new engine. Its optimization objective is:

[0021]

[0022] In the formula, τ represents the samples of different vehicle models and tasks, p(τ) represents the probability distribution of different engine models and operating conditions τ, and θmeta is the meta-learning parameter. Indicates the meta-learning parameter θ meta The control strategy generated for a specific task τ (i.e., the initial strategy for different vehicle models or operating conditions) is given below, and L(.) is the task loss function used to evaluate the performance of the strategy.

[0023] S2.2 The lower-layer policy network receives the merged state characteristics X in real time. fused (t), target speed of output cooling water pump The objective function is innovatively optimized by introducing a thermal stress penalty term:

[0024]

[0025] Where, π θ The control strategy output by the lower-level policy network for specific real-time states; For state features X fused(t) Take the expected value; γ is the discount factor (0 < γ < 1) to weigh the future costs against the present costs; c energy (t) represents the energy consumption cost, σ thermal (t) represents the predicted value of thermal stress calculated in real time, λ σ For penalty weights.

[0026] S2.3 Real-time calculation of predicted thermal stress σ of key engine components based on finite element analysis thermal (t) is provided to the lower-level policy network as a constraint and feedback, which significantly improves the robustness and reliability of the lower-level control policy network.

[0027] S3: Graph Neural Network Cooperative Control Optimization. In step S2, the lower-level policy network optimizes the fused state features X. fused(t) The target speed of the cooling water pump was output. Building upon this, an innovative collaborative control optimization method based on graph neural networks is further proposed to achieve intelligent linkage between multiple subsystems of the engine cooling system:

[0028] S3.1 Establish a system diagram network structure with cooling water pump, thermostat, cooling fan and radiator as nodes. The node status represents the real-time operating parameters of each component, such as opening degree, speed and temperature.

[0029] S3.2 innovatively proposes a dynamic edge weight calculation method to describe the dynamic coupling relationship between subsystems:

[0030] W ij (t)=exp(-β|T i (t)-T j (t)|)

[0031] Among them, W ij (t) represents the edge weight between node i and node j, reflecting the real-time strength of thermal coupling; T i (t), T j (t) represents the temperature characteristics of nodes i and j, respectively, and β is the adjustment coefficient.

[0032] S3.3 The graph neural network iteratively updates the real-time states (such as temperature, opening degree, and speed) of each component node in the cooling system. Through inter-node state information propagation and the calculation of the nonlinear aggregation function f(·), it obtains coordinated control commands for each node (such as water pump speed, fan speed, and thermostat opening degree) to ensure the coordinated and optimized operation of the entire cooling system. The iterative update formula for the hidden state of a node is defined as follows:

[0033]

[0034] in, Let N(i) be the hidden state of node i in the k-th iteration, N(i) be the set of neighboring nodes of node i, and f(.) be the nonlinear aggregation function of the graph neural network.

[0035] S4: Entropy-Constrained Adaptive Model Predictive Control Optimization. In step S3, the graph neural network outputs the coordinated control commands for various components of the engine cooling system (such as water pump speed, fan speed, thermostat opening, etc.). Building upon this, this step innovatively proposes an entropy-constrained adaptive model predictive control method, using the coordinated control commands output by the graph neural network as the control objective to achieve real-time optimization and fine control of the cooling water pump speed. The specific implementation steps are as follows:

[0036] S4.1 Constructs the model predictive control optimization objective function, and innovatively introduces a dynamic thermal entropy constraint term based on traditional energy consumption and performance indicators:

[0037]

[0038] Among them, c energy (k|t) and c tracking (k|t) represent the energy consumption and tracking performance cost within the prediction interval, respectively, N. p To predict the time domain length, H m (t) represents the dynamic thermal entropy, and μ(t) is the adaptive entropy constraint coefficient, which dynamically adjusts the penalty intensity on the entropy.

[0039] S4.2 innovatively proposes an adaptive adjustment mechanism for entropy constraint coefficients:

[0040]

[0041] Where μ0 is the initial entropy constraint coefficient. and σ H These are the sliding window mean and standard deviation of the entropy feature, respectively, which reflect the fluctuation level of the system's thermal state in real time and adaptively adjust the intensity of the entropy penalty.

[0042] S4.3 employs optimization algorithms (such as SQP or interior point method) to solve the above entropy-constrained adaptive model predictive control optimization problem, in order to obtain the optimal control input for the pump speed, effectively improving the real-time performance and anti-disturbance capability of the control strategy.

[0043] S5: Lightweight Model Knowledge Distillation and Edge Computing Deployment. Considering the limited computing resources in real-world automotive environments, this invention innovatively proposes a staged knowledge distillation method for lightweight network model training and edge computing deployment:

[0044] S5.1 employs a teacher-student network architecture for knowledge distillation, using the graph neural network (step S3) and the hierarchical deep reinforcement control strategy driven by meta-reinforcement learning (step S2) from the previous steps as the teacher model to train a lightweight student model, making the student model's performance as close as possible to that of the teacher model.

[0045]

[0046] Among them, D KL Let p be the KL divergence. teacher With p student These represent the probability distributions of the control strategies output by the network for teachers and students, respectively.

[0047] S5.2 performs pruning and quantization on the lightweight student model obtained by knowledge distillation to reduce the computational resource requirements;

[0048] S5.3 utilizes edge computing nodes to assist in completing some computational tasks during the lightweight student model inference process, thereby achieving reasonable allocation of computing resources and ensuring the real-time operation of the control strategy.

[0049] S6: Dynamic Fault Detection and Online Self-Repair Mechanism. To ensure the reliability and safety of the control system, this invention proposes an online fault detection and self-repair mechanism based on dynamic thermal entropy characteristics:

[0050] S6.1 defines the dynamic thermal entropy threshold. Used to identify abnormal operating conditions and fault states in the engine cooling system in real time;

[0051] S6.2 When the entropy value exceeds the threshold, i.e. The system automatically triggers a dynamic fault diagnosis program to quickly locate abnormal sensors or actuators. Based on the abnormal characteristics, it activates a predefined safety backup strategy to ensure the engine cooling function remains effective.

[0052] S6.3 proposes a self-healing method based on online incremental learning, which updates the control policy network parameters in real time:

[0053]

[0054] Where η is the incremental learning rate; L(.) is the control error loss function, which quickly adapts to system changes or sensor drift, enabling online self-repair and adaptive updating of the policy network; ω p (t) represents the target speed of the cooling water pump output by the lower-level policy network at time t; X fus ed(t) is the fused state feature vector; θupdate(t) is the parameter to be updated in the policy network at time t.

[0055] A smart engine cooling water pump actuator control system includes the following modules:

[0056] A multimodal perception and dynamic thermal entropy fusion feature extraction module is used to collect data from multiple sensor sources, including engine coolant temperature T. coolant Ambient temperature T ambient Engine load, oil temperature oil Pump speed T pump and vehicle speed V vehicle Construct the state vector x at time t t =[T coolant (t),T ambient (t),Load(t),T oil (t),T pump (t),V vehicle (t)]; Combine the thermodynamic entropy and information entropy hybrid mechanism to calculate dynamic thermal entropy, perform weighted fusion and anomaly suppression on multi-source sensor data, and obtain state feature vector;

[0057] The hierarchical deep reinforcement control strategy module driven by meta-reinforcement learning adopts a hierarchical control strategy based on meta-reinforcement learning to make the target speed decision of the cooling water pump. At the same time, the finite element analysis method is introduced to dynamically predict thermal stress and incorporate it as a penalty term into the strategy optimization process.

[0058] Graph Neural Network Cooperative Control Optimization Module: A graph neural network with components as nodes is established, where the node state represents the real-time working parameters of each component. A dynamic edge weight calculation method is proposed to describe the dynamic coupling relationship between subsystems. The graph neural network iteratively updates the system state and outputs cooperative control commands.

[0059] An entropy-constrained adaptive model predictive control optimization module is proposed. For the cooperative control objective, an entropy-constrained adaptive model predictive control method is proposed to achieve real-time optimization and control of the pump speed.

[0060] The module includes a lightweight model knowledge distillation and edge computing deployment module. It employs a teacher-student network architecture for knowledge distillation, and performs lightweight student model pruning and quantization based on the knowledge distillation. Edge computing nodes are used to assist in completing some computational and reasoning tasks in the lightweight student model inference process.

[0061] The module for dynamic fault detection and online self-repair mechanism proposes an online fault detection and self-repair mechanism based on dynamic thermal entropy characteristics.

[0062] Define a dynamic thermal entropy threshold for real-time identification of abnormal operating conditions and fault states in the engine cooling system;

[0063] When the entropy value exceeds the threshold, the system automatically triggers a dynamic fault diagnosis program to quickly locate abnormal sensors or actuators; based on the abnormal characteristics, it activates a predefined safety backup strategy; and proposes an online incremental learning self-healing method to update the control strategy network parameters in real time.

[0064] The implementation process of the system and the method is the same, so it will not be described again.

[0065] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described intelligent engine cooling water pump actuator control method.

[0066] A computer-readable storage medium storing a computer program that performs the above-described intelligent engine cooling water pump actuator control method.

[0067] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0068] (1) A dynamic thermal entropy weighting method based on the fusion of thermodynamic entropy and information entropy is proposed to achieve efficient fusion and anomaly suppression of multimodal sensing data, and significantly improve control accuracy and stability.

[0069] (2) Innovatively construct a hierarchical intelligent control architecture that combines meta-reinforcement learning and graph neural networks, which has the advantages of cross-vehicle adaptation and real-time collaborative control.

[0070] (3) Introducing an entropy-constrained adaptive MPC framework effectively enhances the robustness of the control strategy to complex working conditions and disturbances.

[0071] (4) Employ multi-stage knowledge distillation and model pruning quantization techniques to achieve lightweight vehicle deployment of high-performance intelligent control models and meet the needs of automotive edge computing.

[0072] (5) Design an online dynamic fault detection and incremental learning self-repair mechanism to ensure the long-term reliable operation of the control system. Attached Figure Description

[0073] Figure 1 This is a schematic diagram of the overall technical solution flow of an embodiment of the present invention;

[0074] Figure 2 This is a schematic diagram illustrating the implementation principle of the dynamic thermal entropy feature fusion module;

[0075] Figure 3 This is a schematic diagram of meta-reinforcement learning and thermal stress prediction;

[0076] Figure 4 This is a diagram of a graph neural network collaborative control structure.

[0077] Figure 5 This is a flowchart of online fault detection and self-repair. Detailed Implementation

[0078] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0079] A smart engine cooling water pump actuator control system includes a multimodal perception and dynamic thermal entropy fusion feature extraction module, a meta-reinforcement learning-driven hierarchical deep reinforcement control strategy module, a graph neural network collaborative control optimization module, an entropy-constrained adaptive model predictive control optimization module, a lightweight model knowledge distillation and edge computing deployment module, and a dynamic fault detection and online self-repair mechanism module. All modules are connected via CAN bus or Ethernet to achieve information synchronization and control closed loop.

[0080] likeFigure 1 As shown, a method for controlling an intelligent engine cooling water pump actuator includes the following steps:

[0081] Step 1: Fusion of multimodal sensing and dynamic thermal entropy. For example... Figure 2 As shown, during vehicle operation, the system collects multi-source data in real time, including coolant temperature, engine oil temperature, ambient temperature, engine load, water pump speed, and vehicle speed, through sensors, and constructs a state vector x. t =[T coolant (t),T ambient (t),Load(t),T oil (t),T pump (t),V vehicle (t)]. Feature fusion is performed using the following dynamic thermal entropy model:

[0082]

[0083] Among them, thermodynamic entropy is used to reflect the relationship between heat flow and temperature difference, information entropy is used to assess the uncertainty and noise level of the sensed data, and the entropy weighting coefficient α is automatically adjusted according to entropy fluctuations. After inverse entropy weighting, the fused feature X is generated. fused It is then sent to the downstream control module.

[0084] Step 2: Meta-reinforcement learning-driven hierarchical deep reinforcement control strategy. For example... Figure 3 As shown, the control strategy adopts a two-layer structure: the upper layer is a meta-learner, which pre-trains parameters θ on different vehicle models or work tasks. meta Improve strategy transfer capabilities;

[0085] The lower layer is the policy network, which receives the fusion status in real time and outputs the target speed of the water pump.

[0086] The objective function introduces a thermal stress penalty term:

[0087]

[0088] Among them, thermal stress σ thetmal The calculations are performed in real time by the finite element model of the engine components and fed back to the strategy network.

[0089] Step 3: Graph Neural Network Cooperative Control Optimization. For example... Figure 4 As shown, a cooling system topology diagram is constructed, with nodes including water pumps, thermostats, fans, radiators, etc. Node coupling is modeled in the following way:

[0090] W ij (t)=exp(-β|T i -T j |)

[0091] The graph neural network updates the node state and generates collaborative control commands (such as thermostat opening and fan speed) to regulate the linkage of subsystems in a globally optimal manner.

[0092] Step 4: Entropy-constrained adaptive MPC optimization. A dynamic thermal entropy penalty is introduced into the predictive control framework:

[0093]

[0094] Among them, the entropy weight μ(t) is adaptively adjusted, that is... The controller is optimized through rolling prediction and feedback correction, and its robustness is improved by combining physical constraints.

[0095] Step 5: Lightweight Knowledge Distillation and Edge Deployment. The original large-scale policy network is used as the teacher model, and a simplified student model, D, is trained by minimizing the KL divergence. KL (p teacher ||p student Further pruning, quantization, and structural optimization methods are used to compress the model, which is then deployed on the vehicle-mounted edge computing unit to achieve low-latency, high-response control.

[0096] Step 6: Dynamic fault detection and online self-repair. For example... Figure 5 As shown, the system calculates the fusion entropy value H in real time. m (t), when satisfying This triggers anomaly diagnosis, automatically switches to the security control strategy, and initiates an incremental learning mechanism to update the strategy parameters. The normal control strategy will be resumed once the monitoring status recovers.

[0097] This invention provides a highly modular, intelligent, and adaptive cooling water pump control method, particularly suitable for thermal management tasks in complex environments with multiple vehicle models and operating conditions. It possesses broad engineering applicability and excellent scalability. All control modules can be embedded into the vehicle control system, supporting collaborative communication and integration with the engine ECU and TCU.

[0098] Obviously, those skilled in the art should understand that the steps of the control method for the intelligent engine cooling water pump actuator or the modules of the control system for the intelligent engine cooling water pump actuator in the above embodiments of the present invention can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by the computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any particular combination of hardware and software.

Claims

1. A control method for an intelligent engine cooling water pump actuator, characterized in that, First, multi-source sensor data is collected, including coolant temperature, engine oil temperature, engine load, ambient temperature, water pump speed, and vehicle speed. Dynamic thermal entropy is calculated through a thermodynamic entropy-information entropy hybrid mechanism. The multi-source sensor data is then weighted, fused, and anomaly suppressed to obtain a state feature vector. Subsequently, a hierarchical control strategy based on meta-reinforcement learning was adopted to determine the target speed of the cooling water pump. At the same time, the finite element analysis method was introduced to dynamically predict thermal stress and incorporated as a penalty term into the strategy optimization process. Next, by constructing a graph neural network with components as nodes, the coupling relationship between subsystems is modeled and collaborative control commands are generated. On this basis, an adaptive model predictive control optimization with dynamic thermal entropy constraints is constructed to achieve fine control of cooling behavior. Furthermore, a lightweight control model compression method based on knowledge distillation is proposed to adapt to the limitations of on-board edge computing resources. Finally, an online thermal entropy monitoring mechanism and an incremental learning self-healing strategy were designed.

2. The intelligent engine cooling water pump actuator control method according to claim 1, characterized in that, Collect data from multiple sensor sources, including engine coolant temperature T. coolant Ambient temperature T ambient Engine load, oil temperature oil Pump speed T pump and vehicle speed V vehicle Construct the state vector at time t x t =[T coolant (t),T ambient (t),Load(t),T oil (t),T pump (t),V vehicle [t] is used to comprehensively reflect the real-time state of engine operation, serving as the input basis for subsequent control strategies; combining thermodynamic entropy and information entropy, a dynamic thermal entropy is proposed to quantify the thermal efficiency and energy dissipation of the cooling system, defined as: Thermodynamic entropy is defined as: Among them, Q i (t) represents the heat flow rate of the i-th heat source, T i (t) represents the corresponding temperature; the information entropy is: Among them, P j (t) represents the normalized probability distribution of each modality; the weighting factor α is adaptively determined based on entropy fluctuations. Finally, state features are obtained by fusing multimodal data using dynamic thermal entropy weights:

3. The intelligent engine cooling water pump actuator control method according to claim 1, characterized in that, In the hierarchical control strategy based on meta-reinforcement learning, a hierarchical deep reinforcement control framework based on meta-reinforcement learning is proposed on the basis of fusion state, specifically including: S2.1 Constructs a two-layer control architecture, defined as an upper-layer meta-learner and a lower-layer policy network; The upper-level meta-learner is trained on data from different vehicle models and operating conditions to quickly generalize and adapt to the control requirements of new engines. Its optimization objective is: In the formula, θ meta Let τ be the meta-learning parameter, τ represent the samples for different vehicle models, and L(.) be the task loss function; S2.2 The lower-layer policy network receives the merged state characteristics X in real time. fused (t), Target speed of output cooling water pump The objective function is innovatively optimized by introducing a thermal stress penalty term: Among them, c energy (t) represents the energy consumption cost, σ thermal (t) represents the predicted value of thermal stress calculated in real time, λ σ For penalty weighting; S2.3 Real-time calculation of predicted thermal stress σ of key engine components based on finite element analysis thermal (t) is provided to the lower-level policy network as constraints and feedback.

4. The intelligent engine cooling water pump actuator control method according to claim 1, characterized in that, Intelligent linkage between multiple subsystems of the engine cooling system is achieved through a collaborative control optimization method based on graph neural networks. S3.1 Establish a system diagram network structure with cooling water pump, thermostat, cooling fan and radiator as nodes, and the node status represents the real-time working parameters of each component; S3.2 proposes a dynamic edge weight calculation method to describe the dynamic coupling relationship between subsystems: W ij (t)=exp(-β|T i (t)-T j (t)|) Among them, W ij (t) represents the edge weight between node i and node j, reflecting the real-time strength of thermal coupling; T i (t), T j (t) represents the temperature characteristics of nodes i and j, respectively, and β is the adjustment coefficient; S3.3 After iteratively updating the real-time states of each component node in the cooling system using a graph neural network, the collaborative control commands for each node are obtained through inter-node state information propagation and the calculation of the nonlinear aggregation function f(·), ensuring the coordinated and optimized operation of the entire cooling system. The iterative update formula for the hidden state of a node is defined as follows: in, Let N(i) be the hidden state of node i in the k-th iteration, N(i) be the set of neighboring nodes of node i, and f(.) be the nonlinear aggregation function of the graph neural network.

5. The intelligent engine cooling water pump actuator control method according to claim 1, characterized in that, An entropy-constrained adaptive model predictive control method is proposed to achieve real-time optimization and control of pump speed; the specific implementation steps are as follows: S4.1 Construct the model predictive control optimization objective function, introducing a dynamic thermal entropy constraint term based on energy consumption and performance indicators: Among them, c energy (k|t) and c tracking (k|t) represent the energy consumption and tracking performance cost within the prediction interval, respectively, N. p To predict the time domain length, H m (t) represents the dynamic thermal entropy, and μ(t) is the adaptive entropy constraint coefficient, which dynamically adjusts the penalty intensity on the entropy. S4.2 proposes an adaptive adjustment mechanism for entropy constraint coefficients: Where μ0 is the initial entropy constraint coefficient. and σ H These are the sliding window mean and standard deviation of the entropy feature, respectively, which reflect the fluctuation level of the system's thermal state in real time and adaptively adjust the intensity of the entropy penalty. S4.3 employs an optimization algorithm to solve the entropy-constrained adaptive model predictive control optimization problem, thereby obtaining the optimal control input for the pump speed and effectively improving the real-time performance and anti-disturbance capability of the control strategy.

6. The intelligent engine cooling water pump actuator control method according to claim 1, characterized in that, Lightweight model knowledge distillation and edge computing deployment: Considering the limited computing resources in the actual automotive environment, a lightweight network model training and edge computing deployment method with staged knowledge distillation is proposed: S5.1 employs a teacher-student network architecture for knowledge distillation, using a graph neural network and a hierarchical control strategy based on meta-reinforcement learning as the teacher model to train a lightweight student model, thereby approximating the performance of the teacher model. Among them, D KL Let p be the KL divergence. teacher With p student The probability distributions of the control strategies output by the network for teachers and students are respectively. S5.2 Pruning and quantization are performed on the lightweight student model obtained by knowledge distillation; S5.

3. Utilize edge computing nodes to assist in completing some computational tasks during the lightweight student model inference process, thereby achieving reasonable allocation of computing resources and ensuring the real-time operation of the control strategy.

7. The intelligent engine cooling water pump actuator control method according to claim 1, characterized in that, An online fault detection and self-repair mechanism based on dynamic thermal entropy characteristics is proposed: S6.1 defines the dynamic thermal entropy threshold. Used to identify abnormal operating conditions and fault states in the engine cooling system in real time; S6.2 When the entropy value exceeds the threshold, i.e. Automatically triggers dynamic fault diagnosis procedures to quickly locate abnormal sensors or actuators; based on abnormal characteristics, activates predefined safety backup strategies to ensure the engine cooling function remains effective. S6.3 proposes a self-healing method based on online incremental learning, which updates the control policy network parameters in real time: Where η is the incremental learning rate; L(.) is the control error loss function, which quickly adapts to system changes or sensor drift, enabling online self-repair and adaptive updating of the policy network; ω p (t) represents the target speed of the cooling water pump output by the lower-level policy network at time t; X fus ed(t) is the fused state feature vector; θupdate(t) is the parameter to be updated in the policy network at time t.

8. A smart engine cooling water pump actuator control system, characterized in that, Includes the following modules: A multimodal perception and dynamic thermal entropy fusion feature extraction module is used to collect multi-source sensor data, including engine coolant temperature (Tcoolant), ambient temperature (Tambient), engine load (Load), oil temperature (Toil), water pump speed (Tpump), and vehicle speed (Vvehicle), to construct the state vector at time t. xt = [Tcoolant(t),Tambient(t),Load(t),Toil(t),Tpump(t),Vvehicle(t)]; Dynamic thermal entropy is calculated by combining thermodynamic entropy and information entropy, and weighted fusion and anomaly suppression are performed on multi-source sensor data to obtain state feature vector; The hierarchical deep reinforcement control strategy module driven by meta-reinforcement learning adopts a hierarchical control strategy based on meta-reinforcement learning to make the target speed decision of the cooling water pump. At the same time, the finite element analysis method is introduced to dynamically predict thermal stress and incorporate it as a penalty term into the strategy optimization process. Graph Neural Network Cooperative Control Optimization Module: A graph neural network with components as nodes is established, where the node state represents the real-time working parameters of each component. A dynamic edge weight calculation method is proposed to describe the dynamic coupling relationship between subsystems. The graph neural network iteratively updates the system state and outputs cooperative control commands. Entropy-constrained adaptive model predictive control optimization module; To achieve the goal of coordinated control, an entropy-constrained adaptive model predictive control method is proposed to realize the real-time optimization and control of pump speed. The module includes a lightweight model knowledge distillation and edge computing deployment module; it employs a teacher-student network architecture for knowledge distillation, and performs model pruning and quantization based on the knowledge distillation; it utilizes edge computing nodes to assist in completing some model inference tasks. The module for dynamic fault detection and online self-repair mechanism proposes an online fault detection and self-repair mechanism based on dynamic thermal entropy characteristics. Define a dynamic thermal entropy threshold for real-time identification of abnormal operating conditions and fault states in the engine cooling system; When the entropy value exceeds the threshold, the system automatically triggers a dynamic fault diagnosis program to quickly locate abnormal sensors or actuators; based on the abnormal characteristics, it activates a predefined safety backup strategy; and proposes an online incremental learning self-healing method to update the control strategy network parameters in real time.

9. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the intelligent engine cooling water pump actuator control method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program that performs the intelligent engine cooling water pump actuator control method as described in any one of claims 1-7.

Images