Scholarlink

Background <p>Growing concern over industrial sustainability necessitates intelligent systems capable of minimizing carbon emissions across the entire product lifecycle, including raw material acquisition, manufacturing, logistics, usage, and end-of-life stages. Existing machine learning approaches provide improvements but often lack adaptability in dynamic environments, struggle with high-dimensional data, and exhibit slow convergence.</p> Methods <p>This research proposes a reinforcement learning-based optimization framework for reducing lifecycle carbon footprint while maintaining operational efficiency. A novel Multi-Agent Deep Deterministic Policy Gradient with Intelligent Cockroach Swarm Algorithm (MADDG-ICS) is developed, where MADDG enables adaptive multi-agent policy learning and ICS enhances global exploration to avoid local optima. The model operates in a simulated lifecycle environment constructed from a dataset containing manufacturing, logistics, energy, and recycling parameters. Data preprocessing includes missing value imputation and min–max normalization. Principal Component Analysis (PCA) is applied strictly as a dimensionality reduction technique to transform correlated lifecycle variables into a compact set of orthogonal components, improving computational efficiency and stabilizing the reinforcement learning state representation.</p> Results <p>The proposed model achieves a 24% reduction in carbon emissions across the product lifecycle. The proposed framework is primarily evaluated using RL metrics such as cumulative reward, convergence, and policy stability. Regression metrics MAE (4.6&#xa0;kg CO<sub>2</sub>-eq), RMSE (7.5&#xa0;kg CO<sub>2</sub>-eq), and MAPE (3.124%) are reported as secondary measures to assess the accuracy of emission outcomes produced by the learned policies, not the RL optimization process itself.</p> Conclusion <p>The MADDG-ICS framework enables real-time, emission-aware decision-making with improved convergence and robust policy learning, demonstrating strong potential for sustainable lifecycle management.</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

Optimization and decision-making model for product lifecycle carbon footprint driven by reinforcement learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月9日

查看全文
  • Qing Liu

摘要

Background

Growing concern over industrial sustainability necessitates intelligent systems capable of minimizing carbon emissions across the entire product lifecycle, including raw material acquisition, manufacturing, logistics, usage, and end-of-life stages. Existing machine learning approaches provide improvements but often lack adaptability in dynamic environments, struggle with high-dimensional data, and exhibit slow convergence.

Methods

This research proposes a reinforcement learning-based optimization framework for reducing lifecycle carbon footprint while maintaining operational efficiency. A novel Multi-Agent Deep Deterministic Policy Gradient with Intelligent Cockroach Swarm Algorithm (MADDG-ICS) is developed, where MADDG enables adaptive multi-agent policy learning and ICS enhances global exploration to avoid local optima. The model operates in a simulated lifecycle environment constructed from a dataset containing manufacturing, logistics, energy, and recycling parameters. Data preprocessing includes missing value imputation and min–max normalization. Principal Component Analysis (PCA) is applied strictly as a dimensionality reduction technique to transform correlated lifecycle variables into a compact set of orthogonal components, improving computational efficiency and stabilizing the reinforcement learning state representation.

Results

The proposed model achieves a 24% reduction in carbon emissions across the product lifecycle. The proposed framework is primarily evaluated using RL metrics such as cumulative reward, convergence, and policy stability. Regression metrics MAE (4.6 kg CO2-eq), RMSE (7.5 kg CO2-eq), and MAPE (3.124%) are reported as secondary measures to assess the accuracy of emission outcomes produced by the learned policies, not the RL optimization process itself.

Conclusion

The MADDG-ICS framework enables real-time, emission-aware decision-making with improved convergence and robust policy learning, demonstrating strong potential for sustainable lifecycle management.