摘要
铝粉自燃是工业生产中的重大安全隐患,早期预警对降低风险至关重要。传统方法依赖固定温度阈值,难以适应动态变化的实际情况,预警精度受限。为了解决这一问题,通过试验构建了铝粉自燃早期痕迹气体体积分数数据集,系统记录氢气、一氧化碳等关键指标的时序变化,为模型开发奠定了数据基础。在试验基础上,提出了一种基于长短期记忆(Long Short-Term Memory,LSTM)网络的预测模型,利用时序特征分析可提前检测自燃风险。结果表明,该模型在测试集上不同粒径的铝粉样本中,能比温度突变提前平均11~57 s发出预警,显著优于传统方法。研究将深度学习方法与安全预警结合,不仅为铝粉自燃监测提供了新的技术手段,也为工业生产中的其他安全隐患预警提供了参考,具有广泛的应用前景和社会经济效益。
Spontaneous combustion of aluminum powder presents a significant safety risk in industrial production environments,making effective early warning systems essential for risk reduction.Traditional detection methods,based on fixed threshold approaches,are limited in their adaptability to dynamic changes,warning accuracy,and response effectiveness.To overcome these limitations,this study experimentally established a comprehensive dataset of trace gas concentrations during the early stages of aluminum powder spontaneous combustion.The experimental framework systematically recorded the temporal evolution of key parameters,such as hydrogen and carbon monoxide concentrations,providing a data-driven foundation for the development of predictive models.Building on the experimental data foundation,this study develops a Long Short-Term Memory(LSTM)-based predictive model.In contrast to conventional temperature-threshold methods,the LSTM model leverages temporal feature analysis to identify safety risks associated with the spontaneous combustion of aluminum powder.The experimental validation process involves multiple aluminum powder samples with varying particle sizes,moisture contents,and accumulation thicknesses,all tested under controlled laboratory conditions.This study demonstrates that the LSTM-based predictive model provides safety warnings 1157 s in advance of temperature anomalies in aluminum powder samples with varying particle sizes,significantly outperforming conventional methods.Additionally,the LSTM model exhibits stable predictive capability for hydrogen and carbon monoxide concentrations,effectively capturing the corresponding concentration ranges throughout the combustion process.The time advance for safety warnings is closely related to the particle size of the aluminum powder samples.For the different particle sizes,the time advance varies as follows:a minimum of 20 s(5μm),11 s(45μm),and 19 s(75μm);a maximum of 39 s(5μm),32 s(45μm),and 57 s(75μm);and an average of 30 s(5μm),22 s(45μm),and 41 s(75μm).These differences in time advance are attributed to the varying reactivity of aluminum powder depending on particle size.Through comparative analysis with traditional temperature-threshold methods,this study quantitatively demonstrates the superior early warning performance of the LSTM-based approach.The integration of deep learning techniques with safety monitoring systems marks a significant advancement in industrial safety protocols,particularly for facilities processing combustible metal powders.The developed methodology presents a novel technical approach for monitoring aluminum powder combustion and offers valuable reference frameworks applicable to early warning systems aimed at preventing safety hazards.This contribution enhances both the theoretical understanding and practical applications within industrial safety management.
作者
钟圣俊
苏星雨
刘春旺
王博雅
蒋关宇
ZHONG Shengjun;SU Xingyu;LIU Chunwang;WANG Boya;JIANG Guanyu(College of Metallurgy,Northeastern University,Shenyang 110819,China)
出处
《安全与环境学报》
北大核心
2026年第3期896-904,共9页
Journal of Safety and Environment
基金
国家重点研发计划项目(2023YFC3010603)。
关键词
安全工程
堆积铝粉
自燃预警
深度学习
时序预测
safety engineering
accumulated aluminum powder
spontaneous combustion warning
deep learning
time series prediction
