Heat dissipation drum brake dynamic control system and method based on reinforcement learning

Abstract

The invention discloses a heat dissipation drum brake dynamic control system and method based on reinforcement learning. The method comprises the following steps that S1, drum brake operation data are collected and preprocessed; s2, training by using a gating circulation unit neural network, and outputting temperature rise trend data; s3, training based on a near-end strategy optimization algorithm to obtain a control strategy network, and outputting a heat dissipation control strategy; s4, performing joint optimization on hyper-parameters of the thermal state prediction model and the control strategy network by adopting a Harris eagle optimization algorithm; s5, generating a heat dissipation control instruction according to the optimized model, and controlling a heat dissipation execution mechanism to execute corresponding operation; and S6, continuously collecting temperature feedback data, comparing the temperature feedback data with the temperature rise trend data, and carrying out dynamic adjustment. Through fusion of reinforcement learning, deep prediction and intelligent optimization technologies, accurate prediction and dynamic heat dissipation control of the temperature rise trend of the drum brake are realized, so that the safety, the energy efficiency and the stability of a brake system are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of reinforcement learning, and particularly to a dynamic control system and method for a heat dissipation drum brake based on reinforcement learning. Background Art

[0002] As an important part of a vehicle braking system, drum brakes are widely used in various passenger cars, commercial vehicles, and heavy-duty vehicles, and still dominate in rear-wheel braking and medium- and low-speed vehicle systems. Compared with disc brakes, drum brakes have the advantages of compact structure, low cost, good sealing, and convenient maintenance. However, under conditions of long-term braking, frequent braking, or high-load operation, the temperature of drum brakes is likely to rise rapidly. If heat dissipation is not timely or the control is unreasonable, it is extremely easy to cause the phenomenon of thermal fade, which will further lead to serious safety problems such as a decline in braking performance, thermal expansion and deformation, and abnormal wear of friction linings.

[0003] To address the problem of thermal fade of drum brakes, passive heat dissipation designs (such as adding heat sinks, enlarging ventilation openings, improving the thermal conductivity of materials, etc.) or simple active heat dissipation controls (such as turning on a fan or a liquid cooling device when the temperature exceeds a set threshold) are generally adopted in the prior art to assist in cooling. However, although passive designs can improve the heat dissipation efficiency at the structural level, it is difficult to cope with sudden situations where heat accumulates rapidly during dynamic driving. The active heat dissipation methods mostly adopt fixed-threshold trigger control, lacking the ability to predict the temperature rise trend, with a lag in response, and it is easy to cause untimely or excessive heat dissipation control, affecting the energy consumption of the whole vehicle and braking safety.

[0004] In recent years, with the development of sensors, communication, and computing platforms, vehicle thermal management systems have begun to evolve towards intelligence. Some studies have attempted to introduce advanced methods such as model predictive control, fuzzy control, and adaptive control to optimize the heat dissipation process. However, the above methods generally rely on accurate physical modeling, complex control law design, and are not very adaptable to changes in the driving environment and data dimension expansion. For example, model predictive control requires establishing an accurate mathematical model for vehicle load, speed, braking conditions, etc., and it is difficult to maintain long-term stability under actual complex working conditions; while fuzzy control is prone to problems of inconsistent control performance when facing high-dimensional dynamic systems due to the relatively subjective rule design.

[0005] In addition, the existing heat dissipation control methods generally have the following technical deficiencies: First, they fail to make full use of historical operation data to predict the temperature rise trend, unable to achieve early heat dissipation preparation, which reduces the foresight and initiative of control; Second, the control strategy lacks an intelligent evolution mechanism and fails to dynamically adjust with changes in working conditions and driving behaviors, resulting in fixed and rigid control effects and making it difficult to adapt to complex and changing driving environments; Third, there is a lack of a systematic optimization process for the parameters of the heat dissipation control model. Traditional methods often rely on empirical parameter setting and it is difficult to obtain the optimal performance combination.

[0006] In recent years, significant progress has been made in the application of reinforcement learning in fields such as automatic control, robot decision-making, and energy scheduling. Especially in systems with a large state space, multiple control objectives, and fast-changing environments, reinforcement learning demonstrates strong adaptability and self-learning capabilities. In the field of vehicle thermal management, reinforcement learning has the natural advantage of learning the optimal control strategy through reward and punishment feedback, and can adaptively optimize the control strategy through interaction with the environment under the condition of lacking an accurate physical model. However, there are still challenges in directly applying reinforcement learning to drum brake heat dissipation control, such as incomplete state space construction, slow policy convergence speed, and insufficient training samples.

[0007] At the same time, temperature change belongs to typical time series data. How to accurately extract the trend characteristics of the thermal state evolving over time is the key to achieving accurate prediction. Traditional methods mostly use simple methods such as moving average and first-order difference for trend analysis, ignoring the long-term dependence information in the time series data. The gated recurrent unit neural network (GRU) has become one of the mainstream technologies in the current time series prediction field due to its advantages such as simple structure, high computational efficiency, and strong long-term memory ability when dealing with sequence data. Introducing the GRU model into the thermal state prediction of drum brakes can effectively improve the accuracy of temperature rise trend prediction and provide a reliable basis for subsequent control strategies.

[0008] On the other hand, there are a large number of hyperparameters in the reinforcement learning model and neural network model (such as learning rate, number of hidden units, discount factor, reward weight, etc.). These hyperparameters directly affect the model performance and convergence speed. Traditional parameter adjustment methods mostly rely on manual experience or grid search, which is not only inefficient but also easily falls into local optima. Therefore, introducing swarm intelligence optimization algorithms (such as particle swarm, differential evolution, whale optimization, etc.) has become an important means to improve the parameter optimization efficiency in recent years. Among them, the Harris Hawks Optimization algorithm, as an emerging heuristic intelligent algorithm, simulates the behavior of a group of eagles hunting prey and has the characteristics of flexible search strategies, strong local development ability, and excellent global optimization ability, making it very suitable for hyperparameter optimization of high-dimensional nonlinear models.

[0009] However, the application of existing optimization algorithms in the cooling system is mostly limited to the parameter optimization of a single model (such as a neural network prediction model), and a mechanism for collaborative optimization of prediction and control models has not been formed. Moreover, the dynamic influence of external information such as driving mode and working condition characteristics on the optimization process has not been considered. In addition, the existing optimization strategies lack a surprise response mechanism for high-risk hot emergencies and cannot make large-step parameter adjustments in a timely manner during the rapid temperature rise stage, which limits the system's rapid adaptation ability to thermal emergencies.

[0010] Therefore, how to provide a dynamic control system and method for a cooling drum brake based on reinforcement learning is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0011] An object of the present invention is to propose a dynamic control system and method for a cooling drum brake based on reinforcement learning. The present invention fully integrates the time-series prediction ability of a gated recurrent unit neural network for thermal states, an intelligent control strategy learning mechanism based on the proximal policy optimization algorithm, and a joint optimization strategy of an improved Harris hawk optimization algorithm for key model hyperparameters, and details the whole process of the drum brake to realize temperature rise trend prediction, dynamic generation of heat dissipation strategies, and real-time control execution under complex working conditions. The invention has the advantages of high prediction accuracy, fast control response, strong adaptability, and high parameter optimization efficiency, and significantly improves the safety and stability of the drum brake in high-frequency use and extreme temperature rise environments.

[0012] A dynamic control method for a cooling drum brake based on reinforcement learning according to an embodiment of the present invention includes the following steps:

[0013] S1. Collect the operation data of the drum brake system, preprocess the operation data to form a thermal state data sequence;

[0014] S2. Use a gated recurrent unit neural network to train the thermal state data sequence to obtain a thermal state prediction model and output temperature rise trend data;

[0015] S3. Based on the thermal state data and temperature rise trend data, use the proximal policy optimization algorithm to train a control policy network. The control policy network constructs a reward function with the goals of temperature stability, minimum energy consumption, and shortest response time on the basis of the defined state space and action space, and outputs the heat dissipation control policy at each moment;

[0016] S4. Use the Harris hawk optimization algorithm to jointly optimize the key hyperparameters of the thermal state prediction model and the control policy network to obtain an optimal parameter combination, and form an optimized thermal state prediction model and control policy network;

[0017] S5. Based on the optimized thermal state prediction model and control strategy network, according to the output heat dissipation control strategy and the control parameters output by the control strategy network, control the heat dissipation actuator to perform corresponding operations;

[0018] S6. Continuously collect the temperature feedback data of the drum brake, compare it with the temperature rise trend data, and dynamically adjust the parameters of the thermal state prediction model and the control strategy network according to the prediction error, and update the training data samples.

[0019] Optionally, the operating data of the drum brake system specifically includes brake temperature, braking duration, brake pedal opening, vehicle speed, ambient temperature, and friction frequency, which are used to generate thermal state data and train the thermal state prediction model and the control strategy network.

[0020] Optionally, the preprocessing of the operating data specifically includes data cleaning, outlier removal, and normalization processing, which are used to eliminate noise and unify the data scale, and provide a standardized input for generating thermal state data.

[0021] Optionally, S2 specifically includes:

[0022] S21. Reorganize the operating data of the drum brake system in the order of collection time to construct a thermal state data sequence X = {x1, x2, …, x N}, where the i-th thermal state vector is x i , i ∈ [1, N], and N is the total length of the thermal state data sequence;

[0023] S22. Construct a gated recurrent unit neural network, which includes an input layer, at least one GRU hidden layer, and an output layer, and is used to extract the temporal features in the thermal state data sequence and perform temperature prediction modeling;

[0024] S23. In the GRU hidden layer, iteratively update the hidden state at each moment:

[0025] h t = (1 - z t ) ⊙ h t-1 + z t ⊙ tanh(Wx t + U(r t ⊙ h t-1 ) + b);

[0026] Among them, x t is the input vector at time t, h t is the hidden state vector at time t, z t is the update gate vector, r t is the reset gate vector, W is the input weight matrix, U is the recurrent weight matrix, b is the bias vector, ⊙ represents element-wise multiplication, ht-1 Represents the hidden state vector at time t-1;

[0027] S24. Train the gated recurrent unit neural network using the thermal state data sequence X, and optimize it using the mean square error loss function containing the thermal trend consistency term:

[0028]

[0029] Among them, y i Represents the true temperature value at the i-th moment, Represents the predicted temperature value at the i-th moment, λ is the thermal trend consistency penalty coefficient, L represents the total loss function value, N represents the total number of data points in the sample sequence, y i-1 Represents the true temperature value at the (i-1)-th moment, Represents the predicted temperature value at the (i-1)-th moment;

[0030] S25. After the training is completed, obtain the thermal state prediction model M for modeling the future temperature evolution law of the drum brake;

[0031] S26. Input the thermal state data sequence X into the thermal state prediction model to obtain the temperature rise trend data within a preset future time period.

[0032] Optionally, the specific content of S3 includes:

[0033] S31. Combine the thermal state data x t with the temperature rise trend data T′ t and the driving mode data M t to form an extended state vector:

[0034]

[0035] Among them, x t Represents the thermal state vector at time t, T′ t Represents the temperature rise trend data at time t, M t Represents the driving mode state vector at time t;

[0036] S32. Construct the heat dissipation control action space A = {a1, a2,..., a M}, where each action a j represents a heat dissipation control instruction, and M is the total number of actions;

[0037] S33. Train the control policy network π through the proximal policy optimization algorithm to establish the mapping relationship between the extended state vector s t and the action a t , π: s t → a t ;

[0038] S34. Define the reward function R t for evaluating the performance of the control strategy:

[0039] R t = -α|T t - T set |- βE t - γτ t - δφ(M t );

[0040] where T t represents the actual temperature at time t, T set represents the predetermined temperature, E t represents the energy consumption at time t, τ t represents the response time at time t, φ(M t ) represents the driving mode deviation value, and α, β, γ, and δ are the weight coefficients of the temperature deviation, energy consumption, response time, and driving mode deviation respectively;

[0041] S35. During the process of training the control strategy network π, define the training objective function J(θ) through the proximal policy optimization algorithm:

[0042]

[0043] where r t (θ) represents the policy ratio, represents the advantage function at time t, ∈ represents the clipping parameter, μ represents the weight coefficient of the driving mode consistency penalty term, Δ(M t ) represents the driving mode deviation amount at time t, θ represents the parameter of the control strategy network π, represents the calculation of the expected value for all time steps t, clip() represents the clipping function, and min represents the minimum operation;

[0044] S36. After training is completed, the control strategy network π outputs the corresponding heat dissipation control strategy a t for each extended state vector s t .

[0045] Optionally, the S4 specifically includes:

[0046] S41. Determine the key hyperparameter set Θ GRU of the thermal state prediction model;

[0047] S42. Determine the key hyperparameter set Θ PPO of the control strategy network;

[0048] S43. Construct the joint hyperparameter set Θ = Θ GRU ∪ Θ PPO, which is used to uniformly represent all the parameters to be optimized of the thermal state prediction model and the control strategy network;

[0049] S44. Construct a temperature control physical model through the principles of heat conduction and energy conservation:

[0050] Collect physical parameters related to brake heat conduction, heat dissipation efficiency, and frictional heat generation to form a parameter set θ p = {k, h, A, m, c}, where k is the heat conduction coefficient, h is the convective heat transfer coefficient, A is the heat dissipation area, m is the brake mass, and c is the specific heat capacity;

[0051] Based on the principles of energy conservation and heat conduction, construct a temperature control physical model, and obtain the temperature evolution function T exp (Θ):

[0052] T exp (Θ) = f(T0, θ p );

[0053] Among them, T0 is the initial temperature, and the function f() is obtained from the heat conduction and convective heat transfer formulas;

[0054] Construct a working condition index C according to vehicle operation data, where the working condition index is composed of braking duration, vehicle speed, and ambient temperature;

[0055] S45. By fusing the training error and physical deviation term of the temperature control physical model and the control strategy network, define the fitness function F(Θ) as the objective evaluation function of the Harris hawk optimization algorithm:

[0056] F(Θ) = L GRU (Θ GRU ) + L PPO (Θ PPO ) + α o |T exp (Θ) - T obs | + κ|C - C opt |;

[0057] Among them, L GRU (Θ GRU ) represents the mean square prediction error in the training process of the control strategy network, L PPO (Θ PPO ) represents the policy error in the training process of the control strategy network in the proximal policy optimization algorithm, T obs represents the observed actual temperature, C opt represents the preset ideal working condition index, and α o and κ are the weight factors of the temperature control physical constraint term and the working condition deviation term respectively;

[0058] S46. Combine the temperature rise trend prediction data and driving mode information, and construct an adaptive dynamic step factor λ through the thermal risk estimation mechanism k :

[0059]

[0060] where λ0 represents the initial step size, ΔT is the temperature rise trend, T thr is the temperature warning threshold, T set represents the predetermined temperature, φ(M t ) represents the driving mode deviation value, φ ref is the reference driving mode complexity, ζ is the scaling coefficient affected by the driving mode, and exp is the exponential function;

[0061] S47. When the temperature rise trend ΔT exceeds the warning threshold T thr , perform an enhanced raid update on the candidate hyperparameter solution through the dynamic step factor λ k :

[0062]

[0063] where X (k) represents the candidate hyperparameter combination at the k-th iteration, U (k) represents the reference solution obtained based on the current local optimal information, is the indicator function, which takes the value of 1 when ΔT exceeds T thr , and 0 otherwise, and X (k+1) represents the candidate hyperparameter combination at the (k + 1)-th iteration;

[0064] S48. Through the improved Harris hawk optimization algorithm, under the guidance of the fitness function F(Θ), perform iterative search and update on the candidate hyperparameter set Θ until the preset termination condition is met, so as to obtain the optimal hyperparameter combination Θ * ;

[0065] S49. Update the parameters of the thermal state prediction model and the control strategy network respectively according to the optimal hyperparameter combination Θ * ;

[0066] S410. Use the updated hyperparameters to construct an optimized thermal state prediction model and a control strategy network, and form an optimized model jointly guaranteed by the temperature control physical constraints, adaptive driving conditions, and improved raid mechanism.

[0067] Optionally, the key hyperparameter set of the thermal state prediction model specifically includes the number of units in the GRU hidden layer, the time window length, and the learning rate of the GRU network, which are used to adjust the structural complexity and the range of time series data processing of the thermal state prediction model.

[0068] Optionally, the set of key hyperparameters of the control policy network specifically includes a discount factor, a learning rate of the control policy network, and weight coefficients of temperature deviation, energy consumption, response time, and driving mode deviation in the reward function, to balance and adjust the learning rate of the control policy network and the reward weights.

[0069] Optionally, S5 specifically includes:

[0070] S51. Obtain real-time thermal state data and input it into the optimized thermal state prediction model, and use the thermal state prediction model to predict the temperature rise trend of the drum brake in the short term in the future, for judging the temperature change trend and control requirements;

[0071] S52. Input the thermal state data at the current moment, the predicted temperature rise trend result, and the current driving mode state into the optimized control policy network together;

[0072] S53. The control policy network outputs a heat dissipation control policy and control parameters according to the input information. The control policy includes the selection of a heat dissipation mode, and the control parameters include the start signal of the heat dissipation device, the wind speed level, and the flow rate of the cooling medium;

[0073] S54. Generate a control instruction according to the control parameters output by the control policy network, and send the control instruction to the corresponding heat dissipation execution mechanism through the communication interface. The heat dissipation execution mechanism includes a fan, a heat dissipation air duct, or a liquid cooling device;

[0074] S55. After receiving the control instruction, the heat dissipation execution mechanism performs heat dissipation operations according to the specified parameters, including starting the heat dissipation device, adjusting the wind speed and air volume, and regulating the flow rate of the heat dissipation liquid;

[0075] S56. After the heat dissipation execution mechanism completes the control task, transmit the execution status information and real-time temperature feedback data to the system.

[0076] A dynamic control system for a heat-dissipating drum brake based on reinforcement learning according to an embodiment of the present invention includes the following modules:

[0077] A data acquisition and preprocessing module, configured to collect the operation data of the drum brake in real time and preprocess the operation data;

[0078] A temperature rise prediction module, configured to process the input thermal state data and predict the temperature change trend of the drum brake in a future time period;

[0079] A control policy generation module, configured to output a heat dissipation control policy and control parameters according to the thermal state data, the temperature rise trend, and the driving mode state;

[0080] A heat dissipation execution module, configured to receive the control instruction sent by the control policy generation module and control the operation state of the heat dissipation execution mechanism.

[0081] A feedback and update module, which is used to receive the execution status and temperature feedback information returned by the heat dissipation actuator, and dynamically update the parameters of the thermal state prediction model and the control strategy network in combination with the prediction error;

[0082] A parameter optimization module, which is used to jointly optimize the key hyperparameters in the temperature rise prediction module and the control strategy generation module.

[0083] The beneficial effects of the present invention are:

[0084] A dynamic control system and method for a heat dissipation drum brake based on reinforcement learning proposed by the present invention break through the limitations of the prior art in aspects such as heat dissipation response lag, rigid control strategy, inaccurate temperature rise prediction, and low parameter tuning efficiency, and achieve various beneficial technical effects. By constructing a gated recurrent unit neural network, time series modeling and learning are carried out on multi-dimensional thermal state data during the operation of the drum brake, realizing the forward-looking prediction of the temperature rise trend, and overcoming the problem that it is difficult for traditional moving average or fixed threshold triggering methods to dynamically adapt to thermal changes. Combining the control strategy network trained by the proximal policy optimization algorithm enables the system to generate optimal heat dissipation control strategies and specific parameters in real time according to different thermal states, driving conditions, and temperature rise trends, thereby realizing more intelligent and efficient heat dissipation control decisions.

[0085] In terms of optimization, the present invention introduces a temperature control physical model constructed based on the theory of heat conduction and convective heat transfer, and uses it as part of the fitness evaluation, and cooperates with an improved Harris hawk optimization algorithm to realize the joint optimization of the key hyperparameters of the thermal state prediction model and the control strategy network. Especially when the temperature rise trend prediction result exceeds the warning threshold, an enhanced raid update is guided for the candidate solution through a dynamic step size adjustment mechanism, thereby significantly improving the response speed and global search ability of the algorithm in case of sudden high temperature, ensuring that the system can output the optimal heat dissipation strategy in time and avoiding the risk of heat fade.

[0086] In addition, the system structure designed by the present invention covers a complete closed-loop control process including data acquisition, preprocessing, prediction, control, optimization, and feedback, and has high integration and engineering adaptability. The system can continuously and stably operate under a variety of complex working conditions, and has good self-learning ability and self-adaptive ability. Overall, the invention realizes the accurate modeling and active control of the temperature rise process of the drum brake, improves the intelligent level of the thermal management of the vehicle braking system, and has significant engineering application value in improving braking safety, extending the service life of components, and reducing energy consumption losses. Description of the Drawings

[0087] The accompanying drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, and do not constitute a limitation to the present invention. In the accompanying drawings:

[0088] Figure 1 is a flowchart of a dynamic control method for a heat-dissipating drum brake based on reinforcement learning proposed by the present invention;

[0089] Figure 2 is a schematic structural diagram of a dynamic control system for a heat-dissipating drum brake based on reinforcement learning proposed by the present invention. Detailed Embodiments

[0090] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0091] Referring to Figure 1 , a dynamic control method for a heat-dissipating drum brake based on reinforcement learning includes the following steps:

[0092] S1. Collect the operation data of the drum brake system, preprocess the operation data to form a thermal state data sequence;

[0093] S2. Use a gated recurrent unit neural network to train the thermal state data sequence to obtain a thermal state prediction model, and output the temperature rise trend data;

[0094] S3. Based on the thermal state data and the temperature rise trend data, use the proximal policy optimization algorithm to train and obtain a control policy network. Based on the defined state space and action space, the control policy network constructs a reward function with the goals of temperature stability, minimum energy consumption, and shortest response time, and outputs the heat dissipation control strategy at each moment;

[0095] S4. Use the Harris hawk optimization algorithm to jointly optimize the key hyperparameters of the thermal state prediction model and the control policy network to obtain the optimal parameter combination, and form the optimized thermal state prediction model and control policy network;

[0096] S5. Based on the optimized thermal state prediction model and control policy network, according to the output heat dissipation control strategy and the control parameters output by the control policy network, control the heat dissipation actuator to perform corresponding operations;

[0097] S6. Continuously collect the temperature feedback data of the drum brake, compare it with the temperature rise trend data, and dynamically adjust the parameters of the thermal state prediction model and the control policy network according to the prediction error, and update the training data samples.

[0098] The present invention realizes the accurate prediction of the temperature rise trend and the intelligent generation of real-time heat dissipation strategies by dynamically controlling the heat dissipation process of drum brakes using reinforcement learning technology. The high-quality thermal state data sequence is obtained by the data acquisition and preprocessing module, and it is trained by the gated recurrent unit neural network. The system can accurately predict the future temperature change trend, avoiding the response lag and the risk of sudden temperature rise in the traditional fixed threshold control method. Combining the control strategy network trained by the proximal policy optimization algorithm, the optimal heat dissipation control strategy is automatically generated in the multi-dimensional state space to ensure that the heat dissipation device starts at the best time, thereby reducing energy consumption and improving the safety of the brake. Further, the key hyperparameters of the temperature rise prediction model and the control strategy network are jointly optimized by the improved Harris hawk optimization algorithm, enabling the entire system to quickly adjust the control strategy in the face of complex working conditions and sudden high temperatures, realizing adaptive closed-loop control. The system continuously collects temperature feedback data and compares it with the predicted temperature rise trend, dynamically adjusting the model parameters according to the prediction error, thereby continuously improving the heat dissipation control accuracy. In summary, the present invention has the advantages of accurate temperature rise prediction, fast response speed, low energy consumption, flexible control strategy, and strong adaptability, significantly improving the heat dissipation performance and safety of drum brakes and being applicable to practical applications under various complex working conditions.

[0099] In this embodiment, the operation data of the drum brake system specifically includes brake temperature, braking duration, brake pedal opening, vehicle speed, ambient temperature, and friction frequency, which are used to generate thermal state data and train the thermal state prediction model and the control strategy network.

[0100] In this embodiment, the operation data preprocessing specifically includes data cleaning, outlier removal, and normalization processing, which are used to eliminate noise and unify the data scale, providing a standardized input for generating thermal state data.

[0101] In this embodiment, S2 specifically includes:

[0102] S21. Reorganize the operation data of the drum brake system in the order of collection time to construct a thermal state data sequence X = {x1, x2, …, x N}, where the i-th thermal state vector is x i , i ∈ [1, N], and N is the total length of the thermal state data sequence;

[0103] S22. Construct a gated recurrent unit neural network, which includes an input layer, at least one GRU hidden layer, and an output layer, and is used to extract the temporal features in the thermal state data sequence and perform temperature prediction modeling;

[0104] S23. In the GRU hidden layer, iteratively update the hidden state at each moment:

[0105] h t = (1 - z t ) ⊙ h t-1 + z t ⊙ tanh(Wx t + U(r t ⊙ h t-1 ) + b);

[0106] where x t is the input vector at time t, h t is the hidden state vector at time t, z t is the update gate vector, r t is the reset gate vector, W is the input weight matrix, U is the recurrent weight matrix, b is the bias vector, ⊙ represents element-wise multiplication, and h t-1 represents the hidden state vector at time t - 1;

[0107] S24. Train the gated recurrent unit neural network using the thermal state data sequence X, and optimize it using the mean squared error loss function containing the thermal trend consistency term:

[0108]

[0109] where y i represents the true temperature value at the i-th moment, represents the predicted temperature value at the i-th moment, λ is the thermal trend consistency penalty coefficient, L represents the total loss function value, N represents the total number of data points in the sample sequence, y i-1 represents the true temperature value at the (i - 1)-th moment, represents the predicted temperature value at the (i - 1)-th moment;

[0110] S25. After the training is completed, obtain the thermal state prediction model M for modeling the future temperature evolution law of the drum brake;

[0111] S26. Input the thermal state data sequence X into the thermal state prediction model to obtain the temperature rise trend data within a future preset time period.

[0112] The present invention uses a gated recurrent unit neural network to perform a time series modeling on the operation data of a drum brake. By reorganizing the data in the order of acquisition time to form a hot state data sequence, it realizes the comprehensive utilization of multi-dimensional data such as the temperature of the brake, braking duration, pedal opening, vehicle speed, ambient temperature, and friction frequency. On this basis, a mean square error loss function containing a thermal trend consistency term is used to train the neural network, effectively reducing the prediction error caused by data noise and environmental changes, and ensuring the accuracy and robustness of temperature prediction. The obtained hot state prediction model can accurately capture the dynamic law of the future temperature evolution of the drum brake, and output the temperature rise trend data in advance, providing a reliable forward-looking input for the subsequent heat dissipation control strategy based on reinforcement learning. This method not only achieves the goal of real-time predicting temperature changes, but also has the advantages of fast response speed, high prediction accuracy, and strong adaptability, thus effectively preventing the risk of thermal fade under high-frequency braking and complex working conditions, and improving the overall safety and durability of the system.

[0113] In this embodiment, step S3 specifically includes:

[0114] S31. Combine the hot state data x t with the temperature rise trend data T' t and the driving mode data M t to form an extended state vector:

[0115]

[0116] where x t represents the hot state vector at time t, T' t represents the temperature rise trend data at time t, and M t represents the driving mode state vector at time t;

[0117] S32. Construct a heat dissipation control action space A = {a1, a2,..., a M}, where each action a j represents a heat dissipation control instruction, and M is the total number of actions;

[0118] S33. Train a control policy network π through the proximal policy optimization algorithm to establish a mapping relationship between the extended state vector s t and the action a t , π: s t → a t ;

[0119] S34. Define a reward function R t for evaluating the performance of the control policy:

[0120] R t = -α|T t - Tset |-βE t -γτ t -δφ(M t );

[0121] Among them, T t represents the actual temperature at time t, T set represents the preset temperature, E t represents the energy consumption at time t, τ t represents the response time at time t, φ(M t ) represents the driving mode deviation value, and α, β, γ, and δ are the weight coefficients of temperature deviation, energy consumption, response time, and driving mode deviation respectively;

[0122] S35. During the process of training the control policy network π, the training objective function J(θ) is defined by the proximal policy optimization algorithm:

[0123]

[0124] Among them, r t (θ) represents the policy ratio, represents the advantage function at time t, ε represents the clipping parameter, μ represents the weight coefficient of the driving mode consistency penalty term, Δ(M t ) represents the driving mode deviation amount at time t, θ represents the parameter of the control policy network π, represents the calculation of the expected value for all time steps t, clip() represents the clipping function, and min represents the minimum operation;

[0125] S36. After the training is completed, the control policy network π outputs the corresponding heat dissipation control policy a t for each extended state vector s t .

[0126] The present invention realizes multi-dimensional perception and expression of the current operating state of drum brakes by constructing an extended state vector that includes thermal state, temperature rise trend, and driving mode information, enhancing the context relevance of control strategy generation. On this basis, by constructing a heat dissipation control action space and training the control strategy network using the proximal policy optimization algorithm, an efficient mapping relationship from the extended state vector to the optimal heat dissipation control action is established, significantly improving the stability and convergence speed of strategy generation. By designing a multi-objective reward function and using temperature stability, energy consumption level, response time, and driving mode deviation as comprehensive evaluation criteria, it is ensured that the heat dissipation control strategy of the system is not only accurate and effective under complex working conditions but also takes into account energy efficiency and comfort. Further, a clipping function and a driving mode consistency constraint term are introduced during the strategy training process to effectively avoid unstable behaviors caused by excessive strategy updates and improve the matching degree of the control strategy with actual driving behaviors. After training is completed, the control strategy network can output refined and dynamic heat dissipation control strategies in real time according to the extended state vector at any moment, realizing feedforward control and real-time adjustment of temperature rise. Overall, this method has the advantages of strong strategy stability, high control accuracy, and strong adaptability, and can be widely applied to the intelligent vehicle thermal management system.

[0127] In this embodiment, step S4 specifically includes:

[0128] S41. Determine the key hyperparameter set Θ of the thermal state prediction model GRU ;

[0129] S42. Determine the key hyperparameter set Θ of the control strategy network PPO ;

[0130] S43. Construct the joint hyperparameter set Θ = Θ GRU ∪ Θ PPO , which is used to uniformly represent all the parameters to be optimized of the thermal state prediction model and the control strategy network;

[0131] S44. Construct a temperature control physical model through the principles of heat conduction and energy conservation:

[0132] Collect physical parameters related to brake heat conduction, heat dissipation efficiency, and frictional heat generation to form a parameter set θ p = {k, h, A, m, c}, where k is the thermal conductivity, h is the convective heat transfer coefficient, A is the heat dissipation area, m is the brake mass, and c is the specific heat capacity;

[0133] Based on the principles of energy conservation and heat conduction, construct a temperature control physical model, and obtain the temperature evolution function T exp (Θ):

[0134] T exp (Θ) = f(T0, θp );

[0135] Among them, T0 is the initial temperature, and the function f() is obtained from the heat conduction and convective heat transfer formulas;

[0136] Construct the operating condition index C according to the vehicle operation data, where the operating condition index is comprehensively composed of the braking duration, vehicle speed, and ambient temperature;

[0137] S45. By fusing the training error and physical deviation term of the temperature control physical model and the control strategy network, define the fitness function F(Θ) as the objective evaluation function of the Harris hawk optimization algorithm:

[0138] F(Θ)=L GRU (Θ GRU )+L PPO (Θ PPO )+α o |T exp (Θ)-T obs |+κ|C-C opt |;

[0139] Among them, L GRU (Θ GRU ) represents the mean square prediction error in the training process of the control strategy network, L PPO (Θ PPO ) represents the policy error of the control strategy network in the training process of the proximal policy optimization algorithm, T obs represents the observed actual temperature, C opt represents the preset ideal operating condition index, α o and κ are the weight factors of the temperature control physical constraint term and the operating condition deviation term respectively;

[0140] S46. Combine the temperature rise trend prediction data and driving mode information, and construct an adaptive dynamic step factor λ through the thermal risk prediction mechanism k :

[0141]

[0142] Among them, λ0 represents the initial step size, ΔT is the temperature rise trend, T thr is the temperature warning threshold, T set represents the predetermined temperature, φ(M t ) represents the driving mode deviation value, φ ref is the reference driving mode complexity, ζ is the scaling coefficient affected by the driving mode, and exp is the exponential function;

[0143] S47. When the temperature rise trend ΔT exceeds the warning threshold T thr , perform an enhanced raid update on the candidate hyperparameter solution through the dynamic step factor λ k :

[0144]

[0145] Among them, X (k) represents the candidate hyperparameter combination at the k-th iteration, and U (k) represents the reference solution obtained based on the current local optimal information. is an indicator function that takes the value of 1 when ΔT exceeds T thr and 0 otherwise, and X (k+1) represents the candidate hyperparameter combination at the (k + 1)-th iteration;

[0146] S48. Through the improved Harris hawk optimization algorithm, under the guidance of the fitness function F(Θ), iteratively search for and update the candidate hyperparameter set Θ until the preset termination condition is met, so as to obtain the optimal hyperparameter combination Θ * ;

[0147] S49. Update the parameters of the thermal state prediction model and the control strategy network respectively according to the optimal hyperparameter combination Θ * ;

[0148] S410. Use the updated hyperparameters to construct an optimized thermal state prediction model and a control strategy network to form an optimized model jointly guaranteed by the temperature control physical constraints, adaptive driving conditions and improved surprise mechanism.

[0149] By introducing the improved Harris hawk optimization algorithm, the present invention realizes the collaborative optimization of the key hyperparameters of the thermal state prediction model and the control strategy network, and solves the problems of low efficiency, poor global search ability and difficulty in adapting to dynamic working conditions existing in the traditional manual parameter setting or single-model optimization method. First, a temperature control physical model is constructed based on the principles of heat conduction and energy conservation, key physical parameters such as the heat conduction coefficient, heat transfer area, and mass are introduced, and indicators reflecting the actual working conditions are constructed in combination with the vehicle operation data, so that the model performance can be evaluated on the premise of meeting the physical constraints and working condition adaptability. Secondly, a comprehensive fitness function including prediction error, strategy error, physical deviation and working condition deviation is constructed as the optimization goal, so that the algorithm can optimize the prediction accuracy, strategy reliability and system adaptability at the same time. In terms of the optimization strategy, an adaptive dynamic step factor is constructed by combining the prediction result of the temperature rise trend and the driving mode deviation. When the predicted high temperature risk appears, an enhanced surprise update is triggered to accelerate the algorithm convergence and improve the response ability to sudden risks. Finally, by jointly updating the parameters of the thermal state prediction model and the control strategy network, an intelligent thermal control system integrating physical mechanism and data-driven is constructed. This method effectively improves the scientificity and efficiency of parameter tuning, and has the remarkable advantages of stable control effect, reliable optimization result and strong working condition adaptability.

[0150] In this embodiment, the key hyperparameter set of the thermal state prediction model specifically includes the number of units in the GRU hidden layer, the time window length, and the learning rate of the GRU network, which are used to adjust the structural complexity of the thermal state prediction model and the range of time series data processing.

[0151] In this embodiment, the key hyperparameter set of the control strategy network specifically includes the discount factor, the learning rate of the control strategy network, and the weight coefficients of temperature deviation, energy consumption, response time, and driving mode deviation in the reward function, which balance and adjust the learning rate of the control strategy network and the reward weights.

[0152] In this embodiment, S5 specifically includes:

[0153] S51. Obtain real-time thermal state data and input it into the optimized thermal state prediction model, and use the thermal state prediction model to predict the temperature rise trend of the drum brake in the short term in the future, which is used to judge the temperature change trend and control requirements;

[0154] S52. Input the thermal state data at the current moment, the predicted temperature rise trend result, and the current driving mode state into the optimized control strategy network;

[0155] S53. The control strategy network outputs a heat dissipation control strategy and control parameters according to the input information. The control strategy includes the selection of the heat dissipation mode, and the control parameters include the start signal of the heat dissipation device, the wind speed level, and the flow rate of the cooling medium;

[0156] S54. Generate a control instruction according to the control parameters output by the control strategy network, and send the control instruction to the corresponding heat dissipation actuator through the communication interface. The heat dissipation actuator includes a fan, a heat dissipation air duct, or a liquid cooling device;

[0157] S55. After receiving the control instruction, the heat dissipation actuator performs heat dissipation operations according to the specified parameters, including starting the heat dissipation device, adjusting the wind speed and air volume, and regulating the flow rate of the cooling liquid;

[0158] S56. After the heat dissipation actuator completes the control task, it transmits the execution status information and real-time temperature feedback data to the system.

[0159] By constructing a closed-loop real-time control mechanism, the present invention realizes the intelligent execution and dynamic regulation of the heat dissipation process of drum brakes, overcoming the problems of traditional systems such as response lag, control rigidity, and feedback absence. After obtaining the thermal state data, the system immediately calls the optimized thermal state prediction model to predict the temperature rise trend in the short term in the future, thereby perceiving the temperature change risk in advance and providing feedforward support for subsequent control. By jointly inputting the current thermal state, temperature rise trend, and driving mode into the optimized control strategy network, the system can output a heat dissipation control strategy and parameters highly matching the current working conditions, realizing the fine regulation of multi-dimensional control variables such as fan startup, wind speed level, and cooling medium flow rate. The control instructions are quickly sent to the heat dissipation actuators such as fans, air ducts, or liquid cooling through the communication module to ensure timely instruction response and accurate execution actions. After the heat dissipation execution is completed, the system can also obtain the temperature feedback and execution status information in real time, providing a basis for closed-loop verification and dynamic correction of the prediction model and strategy network. This method has the advantages of fast response speed, accurate control strategy, complete feedback mechanism, and strong adaptability, and can significantly improve the temperature control safety and system operation stability of drum brakes under high-frequency braking and complex working conditions.

[0160] A dynamic control system for a heat dissipation drum brake based on reinforcement learning according to an embodiment of the present invention includes the following modules:

[0161] A data acquisition and preprocessing module for real-time collecting the operation data of the drum brake and preprocessing the operation data;

[0162] A temperature rise prediction module for processing the input thermal state data and predicting the temperature change trend of the drum brake in the future time period;

[0163] A control strategy generation module for outputting a heat dissipation control strategy and control parameters according to the thermal state data, temperature rise trend, and driving mode state;

[0164] A heat dissipation execution module for receiving the control instructions sent by the control strategy generation module and controlling the operation state of the heat dissipation actuator;

[0165] A feedback and update module for receiving the execution status and temperature feedback information returned by the heat dissipation actuator and dynamically updating the parameters of the thermal state prediction model and the control strategy network in combination with the prediction error;

[0166] A parameter optimization module for jointly optimizing the key hyperparameters in the temperature rise prediction module and the control strategy generation module.

[0167] Embodiment 1:

[0168] To verify the feasibility of this invention, a certain brand of light truck was tested on a long, mountainous downhill section. This section, approximately 21 kilometers long, included numerous long slopes, sharp bends, tunnels, and frequent braking, posing a significant risk of thermal degradation. The test vehicle was a rear-wheel drive light truck equipped with drum brakes, with a fully loaded gross mass of 4.5 tons. The braking system consisted of a rear drum brake with liquid cooling and auxiliary heating. The vehicle was equipped with a standard temperature acquisition system, an ECU control unit, an intelligent fan actuator module, and a brake system data analysis platform.

[0169] Traditional systems for these vehicles typically use a fixed threshold to trigger the fan for cooling control. This involves automatically activating the cooling device when the brake drum temperature exceeds a set value (e.g., 220°C). However, this approach has significant drawbacks: delayed control response, a single strategy, high energy consumption, and the risk of repeated high temperatures and even thermal decay on complex slopes.

[0170] In this example, the system was integrated into a test vehicle and connected to the vehicle's CAN bus via a front-mounted deployment. The system also enabled intelligent prediction of drum brake temperature rise trends, dynamic control strategy generation, and parameter optimization modules. During operation, the system continuously collected vehicle operating status information, including multi-dimensional signals such as speed, braking intensity, braking duration, ambient temperature, and friction frequency. This information was then generated into a complete thermal state vector, which was then fed into an optimized gated recurrent unit neural network. The network then outputted temperature rise trend data for the next 60 seconds, which the control strategy network used to generate the optimal cooling control instructions.

[0171] The control strategy network, trained using a proximal strategy optimization algorithm, dynamically generates control parameters such as fan activation level and liquid cooling medium flow rate adjustment value, based on a multi-objective function that considers temperature stability, response speed, energy efficiency, and driving mode matching. The system also incorporates an improved Harris Eagle optimization module. When the system detects a significant temperature fluctuation or high-temperature risk, it immediately invokes an adaptive preemptive strategy to fine-tune key hyperparameters online, ensuring optimal control strategy response in emergencies.

[0172] In the actual road test for 6 consecutive days, the control system of the present invention (experimental group) and the traditional threshold control system (comparison group) were respectively used to run on the same route to test key indicators such as temperature peak value, fan response time, and energy consumption during heat dissipation during each braking process.

[0173] Table 1 Drum brake heat dissipation performance comparison data table

[0174]

[0175]

[0176] As can be seen from the data in Table 1, the heat dissipation control system driven by reinforcement learning proposed in the present invention significantly outperforms the traditional fixed-threshold control method in multiple key performance indicators. In terms of the average maximum temperature, the brake drum temperature of the experimental group was effectively controlled at 187.3 °C, while that of the control group reached 226.8 °C, with a reduction amplitude of 17.4%. This result indicates that the system of the present invention can effectively suppress the temperature peak and avoid performance degradation or thermal fade problems of the braking system due to overheating. Further analyzing the number of times the temperature exceeded 220 °C, no temperature overlimit event occurred in the experimental group during continuous multi-day and multiple downhill tests, while the traditional control system had 12 high-temperature situations where the temperature exceeded 220 °C under the same conditions. This result proves that the system has a strong ability to prevent high-temperature risks by predicting the temperature rise trend in advance and dynamically adjusting the heat dissipation strategy, significantly improving the safety and reliability of the system. In terms of control response, the average response delay of the fan is also an important indicator to measure the intelligence level of the system. The response time of the experimental group was 0.46 seconds, much faster than 2.13 seconds of the control group, with an advance response amplitude of 78.4%. This shows that the system intervenes in control when the temperature trend begins to show signs through the real-time policy generation mechanism, avoiding the lag of the traditional "after-the-fact remedy" type of control.

[0177] From the perspective of energy consumption, the average fan on-time of the experimental group was 148 seconds, which was more than 50 seconds less than 201 seconds of the control group, saving approximately 26.4% of the heat dissipation energy consumption. This shows that on the premise of achieving a better level of temperature control, the present invention can also reduce the fan workload, improve the overall energy efficiency, and extend the service life of the equipment.

[0178] Finally, in terms of heat dissipation efficiency, the time for the temperature to drop from the high point to the set safe range is an indicator that comprehensively reflects the responsiveness of the model and the system regulation ability. The experimental group only needed an average of 176 seconds to complete the drop process, while the control group needed 292 seconds, with an improvement amplitude as high as 40%. This indicates that in the closed-loop execution of heat dissipation control, the system not only responds faster and controls more accurately, but also is more efficient in restoring cooling.

[0179] In summary, the data in the table fully reflects the technical advantages of the system of the present invention in the field of heat dissipation control of drum brakes. By integrating reinforcement learning, gated recurrent unit network, and improved Harris hawk optimization algorithm, the system realizes intelligent closed-loop management from prediction to control, successfully solving the problems existing in the traditional heat dissipation control methods such as response lag, high-temperature out-of-control, and high energy consumption, providing a feasible and efficient engineering path for thermal management intelligence.

[0180] The above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, shall be covered by the protection scope of the present invention.

Claims

1. A dynamic control system and method for a heat dissipation drum brake based on reinforcement learning, characterized in that, It includes the following steps: S1. Collect the operation data of the drum brake system, preprocess the operation data, and form a thermal state data sequence; S2. Use a gated recurrent unit neural network to train the thermal state data sequence to obtain a thermal state prediction model, and output temperature rise trend data; S3. Based on the thermal state data and the temperature rise trend data, use the proximal policy optimization algorithm to train and obtain a control policy network. The control policy network constructs a reward function with the goals of temperature stability, minimum energy consumption, and shortest response time on the basis of the defined state space and action space, and outputs a heat dissipation control strategy at each moment; S4. Use the Harris hawk optimization algorithm to jointly optimize the key hyperparameters of the thermal state prediction model and the control policy network to obtain an optimal parameter combination, and form an optimized thermal state prediction model and control policy network; S5. Generate a heat dissipation control instruction based on the optimized thermal state prediction model and control policy network, and control the heat dissipation actuator to perform corresponding operations; S6. Continuously collect the temperature feedback data of the drum brake, compare it with the temperature rise trend data, and dynamically adjust the parameters of the thermal state prediction model and the control policy network according to the prediction error, and update the training data samples.

2. The dynamic control system and method of a heat dissipation drum brake based on reinforcement learning according to claim 1, characterized in that The operation data of the drum brake system specifically includes brake temperature, braking duration, brake pedal opening, vehicle speed, ambient temperature, and friction frequency, which are used to generate thermal state data and train the thermal state prediction model and the control policy network.

3. The dynamic control system and method of a heat dissipation drum brake based on reinforcement learning according to claim 1, characterized in that, The preprocessing of the operation data specifically includes data cleaning, outlier removal, and normalization processing, which are used to eliminate noise and unify the data scale, and provide a standardized input for generating thermal state data.

4. A dynamic control method for a heat dissipation drum brake based on reinforcement learning according to claim 1, characterized in that, The specific content of S2 includes: S21. Reorganize the operation data of the drum brake system in the order of collection time to construct a thermal state data sequence X = {x1, x2, …, x N}, where the i-th thermal state vector is x i , i ∈ [1, N], and N is the total length of the thermal state data sequence; S22. Construct a gated recurrent unit neural network, which includes an input layer, at least one GRU hidden layer, and an output layer, and is used to extract the time series features in the thermal state data sequence and perform temperature prediction modeling; S23. In the GRU hidden layer, iteratively update the hidden state at each moment; S24. Use the thermal state data sequence X to train the gated recurrent unit neural network, and optimize it using a mean square error loss function containing a thermal trend consistency term: Among them, y i represents the true temperature value at the i-th moment, represents the predicted temperature value at the i-th moment, λ is the thermal trend consistency penalty coefficient, L represents the total loss function value, N represents the total number of data points in the sample sequence, y i-1 represents the true temperature value at the (i - 1)-th moment, represents the predicted temperature value at the (i - 1)-th moment; S25. After the training is completed, obtain a thermal state prediction model M, which is used to model the future temperature evolution law of the drum brake; S26. Input the thermal state data sequence X into the thermal state prediction model to obtain the temperature rise trend data within a preset future time period.

5. A dynamic control method for a heat dissipation drum brake based on reinforcement learning according to claim 1, characterized in that, The specific content of S3 includes: S31. Combine the hot state data x t with the temperature rise trend data T' t and the driving mode data M t to form an extended state vector: Among them, x t represents the thermal state vector at time t, T' t represents the temperature rise trend data at time t, M t represents the driving mode state vector at time t; S32. Construct a heat dissipation control action space \(A=\{a_1,a_2,\cdots,a\) M \}, where each action \(a\) j represents a heat dissipation control instruction, and \(M\) is the total number of actions; S33. Train the control policy network π through the proximal policy optimization algorithm to establish the mapping relationship between the extended state vector s t and the action a t . π: s t → a t ; S34. Define the reward function R t for evaluating the performance of the control strategy: R t = -α|T t -T set |-βE t -γτ t -δφ(M t ); Among them, T t represents the actual temperature at time t, T set represents the predetermined temperature, E t represents the energy consumption at time t, τ t represents the response time at time t, φ(M t ) represents the driving mode deviation value, and α, β, γ, and δ are the weight coefficients of temperature deviation, energy consumption, response time, and driving mode deviation respectively; S35. During the process of training the control policy network π, define the training objective function J(θ) through the proximal policy optimization algorithm: where, r t (θ) represents the policy ratio, represents the advantage function at time t, ∈ represents the clipping parameter, μ represents the weight coefficient of the driving mode consistency penalty term, Δ(M t ) represents the driving mode deviation at time t, θ represents the parameter of the control policy network π, represents the calculation of the expected value for all time steps t, clip() represents the clipping function, and min represents the minimum operation; S36. After the training is completed, control the policy network π to output the corresponding heat dissipation control policy a for each extended state vector s t t .​ 6. The dynamic control method of a heat dissipation drum brake based on reinforcement learning according to claim 1, wherein The specific content of S4 includes: S41. Determine the key hyperparameter set Θ of the thermal state prediction model GRU ; S42. Determine the key hyperparameter set Θ of the control policy network PPO ; S43. Construct a combined hyperparameter set Θ = Θ GRU ∪ Θ PPO , which is used to uniformly represent all the parameters to be optimized for the thermal state prediction model and the control strategy network; S44. Construct a temperature control physical model through the principles of heat conduction and energy conservation: Collect physical parameters related to the heat conduction, heat dissipation efficiency and frictional heat generation of the brake to form a parameter set θ p ={k, h, A, m, c}, where k is the thermal conductivity, h is the convective heat transfer coefficient, A is the heat dissipation area, m is the mass of the brake, and c is the specific heat capacity; Based on the principles of energy conservation and heat conduction, a temperature control physical model is constructed, and the temperature evolution function T exp (Θ): T exp (Θ) = f(T0, θ p ); Among them, T0 is the initial temperature, and the function f() is obtained from the heat conduction and convective heat transfer formulas; Construct a working condition index C according to the vehicle operation data, where the working condition index is composed of the braking duration, vehicle speed, and ambient temperature; S45. By fusing the training error and physical deviation terms of the temperature control physical model and the control policy network, define the fitness function F(Θ) as the objective evaluation function of the Harris hawk optimization algorithm: F(Θ) = L GRU (Θ GRU ) + L PPO (Θ PPO ) + α o |T exp (Θ) - T obs | + κ|C - C opt |; Among them, L GRU (Θ GRU ) represents the mean square prediction error during the training process of the control policy network, and L PPO (Θ PPO ) represents the policy error of the control policy network during the training process of the proximal policy optimization algorithm. T obs represents the observed actual temperature, and C opt represents the preset ideal working condition index. α o and κ are the weight factors of the temperature control physical constraint term and the working condition deviation term respectively; S46. Combine the temperature rise trend prediction data and the driving mode information to construct an adaptive dynamic step factor λ through a thermal risk estimation mechanism k : Among them, λ0 represents the initial step size, ΔT is the temperature rise trend, and T thr is the temperature warning threshold, and T set represents the predetermined temperature, φ(M t ) represents the driving mode deviation value, φ ref is the reference driving mode complexity, ζ is the scaling coefficient affected by the driving mode, and exp is the exponential function; S47. When the temperature rise trend ΔT exceeds the warning threshold T thr a boosted raid update is performed on the candidate hyperparameter solution through the dynamic step factor λ k : Among them, X (k) represents the candidate hyperparameter combination at the k-th iteration, and U (k) represents the reference solution obtained based on the current local optimal information. is an indicator function that takes the value of 1 when ΔT exceeds T thr and 0 otherwise, and X (k+1) represents the candidate hyperparameter combination at the (k + 1)-th iteration; S48. Through the improved Harris hawk optimization algorithm, under the guidance of the fitness function F(Θ), the candidate hyperparameter set Θ is iteratively searched and updated until the preset termination condition is met, so as to obtain the optimal hyperparameter combination Θ * ; S49. Update the parameters of the hot state prediction model and the control strategy network respectively according to the optimal hyperparameter combination Θ * ​ S410. Construct an optimized thermal state prediction model and a control strategy network using the updated hyperparameters to form an optimized model jointly guaranteed by temperature control physical constraints, adaptive driving conditions, and an improved surprise mechanism.

7. A dynamic control method for a heat dissipation drum brake based on reinforcement learning according to claim 6, characterized in that, The key hyperparameter set of the thermal state prediction model specifically includes the number of units in the GRU hidden layer, the time window length, and the learning rate of the GRU network, which are used to adjust the structural complexity of the thermal state prediction model and the range of time series data processing.

8. A dynamic control method for a heat dissipation drum brake based on reinforcement learning according to claim 6, characterized in that, The key hyperparameter set of the control strategy network specifically includes the discount factor, the learning rate of the control strategy network, and the weight coefficients of temperature deviation, energy consumption, response time, and driving mode deviation in the reward function, which balance and adjust the learning rate of the control strategy network and the reward weights.

9. A dynamic control method for a heat dissipation drum brake based on reinforcement learning according to claim 1, characterized in that The S5 specifically includes: S51. Obtain real-time thermal state data and input it into the optimized thermal state prediction model, and use the thermal state prediction model to predict the temperature rise trend of the drum brake in the short term in the future, which is used to judge the temperature change trend and control requirements. S52. Input the thermal state data at the current moment, the predicted temperature rise trend result, and the current driving mode state into the optimized control strategy network. S53. The control strategy network outputs a heat dissipation control strategy and control parameters according to the input information. The control strategy includes the selection of the heat dissipation mode, and the control parameters include the start signal of the heat dissipation device, the wind speed level, and the flow rate of the cooling medium. S54. Generate a control instruction according to the control parameters output by the control strategy network, and send the control instruction to the corresponding heat dissipation actuator through the communication interface. The heat dissipation actuator includes a fan, a heat dissipation air duct, or a liquid cooling device. S55. After receiving the control instruction, the heat dissipation actuator performs heat dissipation operations according to the specified parameters, including starting the heat dissipation device, adjusting the wind speed and air volume, and regulating the flow rate of the cooling liquid. S56. After the heat dissipation actuator completes the control task, it transmits the execution status information and real-time temperature feedback data to the system.

10. A dynamic control system for a heat-dissipating drum brake based on reinforcement learning, a dynamic control method for a heat-dissipating drum brake based on reinforcement learning according to any one of claims 1 to 9, characterized in that, It includes the following modules: The data acquisition and preprocessing module is used to collect the operation data of the drum brake in real time and preprocess the operation data. The temperature rise prediction module is used to process the input thermal state data and predict the temperature change trend of the drum brake in the future time period. The control strategy generation module is used to output a heat dissipation control strategy and control parameters according to the thermal state data, the temperature rise trend, and the driving mode state. The heat dissipation execution module is used to receive the control instruction issued by the control strategy generation module and control the operation status of the heat dissipation actuator. The feedback and update module is used to receive the execution status and temperature feedback information returned by the heat dissipation actuator, and dynamically update the parameters of the thermal state prediction model and the control strategy network in combination with the prediction error. The parameter optimization module is used to jointly optimize the key hyperparameters in the temperature rise prediction module and the control strategy generation module.

Images