Scholarlink

<p>The quality of asphalt pavement construction is directly related to the durability and service performance of roads. Among them, paving temperature control is the core link that determines the compaction degree and interlayer bonding properties of the mixture. Traditional methods rely on static decision-making models such as PID control and fuzzy logic, which are difficult to cope with the problems of sharp temperature gradient increases and insufficient local compaction caused by large temperature differences, long transportation distances, and dynamic environmental disturbances. Although existing research has introduced data-driven models and multi-agent collaborative frameworks, there are still bottlenecks such as rigid communication topology, lack of physical constraints in reward functions, and insufficient adaptation to equipment heterogeneity, leading to construction quality fluctuations and energy waste. To address these challenges, this study proposes a multi-agent reinforcement learning framework. By constructing a dynamic adjacency matrix to optimize distributed communication topology, integrating multi-source perception data such as infrared thermal imaging and embedded sensors, and designing a physically constrained layered reward mechanism, material properties and construction specifications are embedded into the agent strategy optimization process. Based on the deep deterministic policy gradient algorithm, a centralized training-distributed execution architecture is developed to achieve collaborative control of pavers, temperature control equipment, and environmental monitoring units. It dynamically adjusts the linkage parameters of vibration frequency, paving speed, and heating power, balancing the multi-objective conflicts of temperature field uniformity, compaction compliance rate, and energy efficiency. Experiments show that under extreme temperature differences, sudden wind speed changes, and equipment failures, this framework reduces temperature tracking error by 35% compared to traditional Proportional-Integral-Derivative (PID) control, decreases energy consumption per unit area by 19%, and maintains compaction stability above 90%. It also ensures minimum functional operation during communication interruptions through dynamic adjacency matrices and consensus protocols, verifying the robustness of the system. The research results provide a real-time adaptive temperature control solution for asphalt construction, and its core methods can be extended to complex engineering scenarios such as bridge paving and tunnel maintenance, promoting the evolution of the intelligent construction field towards multi-device collaboration, physical constraint embedding, and dynamic optimization decision-making.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A multi-agent reinforcement learning approach for collaborative temperature control in asphalt paving

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月13日

查看全文
  • Haiting Li,
  • Leijie Hao

摘要

The quality of asphalt pavement construction is directly related to the durability and service performance of roads. Among them, paving temperature control is the core link that determines the compaction degree and interlayer bonding properties of the mixture. Traditional methods rely on static decision-making models such as PID control and fuzzy logic, which are difficult to cope with the problems of sharp temperature gradient increases and insufficient local compaction caused by large temperature differences, long transportation distances, and dynamic environmental disturbances. Although existing research has introduced data-driven models and multi-agent collaborative frameworks, there are still bottlenecks such as rigid communication topology, lack of physical constraints in reward functions, and insufficient adaptation to equipment heterogeneity, leading to construction quality fluctuations and energy waste. To address these challenges, this study proposes a multi-agent reinforcement learning framework. By constructing a dynamic adjacency matrix to optimize distributed communication topology, integrating multi-source perception data such as infrared thermal imaging and embedded sensors, and designing a physically constrained layered reward mechanism, material properties and construction specifications are embedded into the agent strategy optimization process. Based on the deep deterministic policy gradient algorithm, a centralized training-distributed execution architecture is developed to achieve collaborative control of pavers, temperature control equipment, and environmental monitoring units. It dynamically adjusts the linkage parameters of vibration frequency, paving speed, and heating power, balancing the multi-objective conflicts of temperature field uniformity, compaction compliance rate, and energy efficiency. Experiments show that under extreme temperature differences, sudden wind speed changes, and equipment failures, this framework reduces temperature tracking error by 35% compared to traditional Proportional-Integral-Derivative (PID) control, decreases energy consumption per unit area by 19%, and maintains compaction stability above 90%. It also ensures minimum functional operation during communication interruptions through dynamic adjacency matrices and consensus protocols, verifying the robustness of the system. The research results provide a real-time adaptive temperature control solution for asphalt construction, and its core methods can be extended to complex engineering scenarios such as bridge paving and tunnel maintenance, promoting the evolution of the intelligent construction field towards multi-device collaboration, physical constraint embedding, and dynamic optimization decision-making.