摘要
针对铜冶炼转炉在生产过程中因熔体喷溅、摇炉操作等动态工况导致炉壁温度出现周期性剧烈波动,传统静态温度监测方法难以准确预测的问题,本文提出一种融合LSTM神经网络与图像匹配技术的智能监测方法。通过部署于炉腹、风眼区、端盖东、端盖西四部位的红外热像仪采集时序温度数据,创新性地采用模板区域提取与灰度差异分析算法对摇炉遮挡等异常图像进行预处理,有效提升数据质量。在此基础上,构建LSTM预测模型,利用其门控机制捕捉温度序列的长期依赖关系,实现对未来温度趋势的精准预测。工业验证结果表明,该模型在炉腹和端盖西的预测平均绝对误差(MAE)为1.35~1.44℃,风眼区等复杂工况下MAE控制在3.66~4.20℃,显著优于传统方法。该方法能够可靠识别炉衬蚀损引起的温度上升趋势,为转炉预测性维护提供数据支撑,对保障安全生产、延长炉寿及推动冶炼智能化具有重要工程价值。
This study presents an integrated framework addressing the critical industrial challenge of predicting periodic thermal fluctuations in PS converter sidewalls during copper smelting operations,where conventional static temperature monitoring methods prove inadequate under dynamic operational conditions such as molten material splashing and converter tilting.To overcome these limitations,we have developed an intelligent monitoring system that synergistically integrates Long Short-Term Memory(LSTM)neural networks with advanced image matching technology,establishing a robust platform for thermal trend prediction.The methodological implementation commences with the strategic deployment of infrared thermal cameras at four critical locations on the converter vessel-specifically the furnace belly,tuyere zone,east end cover,and west end cover-enabling comprehensive sequential temperature data acquisition with high temporal resolution.A fundamental technological contribution resides in the novel image preprocessing algorithm based on template region extraction and grayscale difference analysis,which effectively identifies and filters abnormal images resulting from operational occlusions,thereby substantially enhancing input data quality and reliability through sophisticated computational techniques that maintain data integrity under challenging industrial environments.Building upon this refined dataset,we construct an optimized LSTM prediction model that leverages the architecture's inherent gating mechanisms to capture complex long-term dependencies within multivariate temperature time series,enabling highly accurate forecasting of future thermal trends through advanced pattern recognition capabilities.The model undergoes rigorous training employing a structured methodology incorporating multiple crucial stages:data normalization to ensure feature consistency,careful partitioning into training and test sets following an 80∶20 ratio to validate generalization performance,and implementation of sophisticated batch training strategies with input sequences specifically designed to cover complete 8-hour smelting cycles,ensuring comprehensive learning of thermal behavior patterns throughout operational phases.Industrial validation results demonstrate exceptional performance metrics across diverse operational conditions,with the model achieving remarkably low Mean Absolute Error(MAE)of 1.35-1.44℃at relatively stable locations such as the furnace belly and west end cover,while maintaining robust performance with MAE of 3.66-4.20℃under the more challenging conditions of the tuyere zone where thermal fluctuations are most pronounced.These results represent a substantial improvement over conventional monitoring methods that typically exhibit errors exceeding 5-7℃under comparable conditions.The system's demonstrated capability to reliably identify temperature rise trends indicative of lining erosion provides crucial data support for transitioning from traditional scheduled maintenance to predictive maintenance paradigms,thereby offering substantial engineering value through multiple dimensions:enhance operational safety by preventing potential breakdowns,extend furnace service life through optimized maintenance scheduling,reduce unplanned downtime resulting in improved production efficiency,and optimize refractory material consumption leading to significant cost reductions.Furthermore,while specifically optimized for converter smelting operations,the core"data preprocessing-deep learning modeling"framework demonstrates considerable potential for adaptation to other high-sensor industrial processes requiring accurate prediction of key parameter evolution,including but not limited to blast furnace operations,chemical reactor monitoring,and power generation systems.This research establishes a new benchmark for intelligent monitoring systems in heavy industries,bridging the gap between theoretical research and practical industrial applications,with promising implications for operational excellence across multiple industrial sectors where precise temperature monitoring and prediction are critical for ensuring safety,optimizing efficiency,and promoting sustainable practices.
作者
陈习堂
孙鼎然
张鑫
高荣
王恩志
徐建新
CHEN Xitang;SUN Dingran;ZHANG Xin;GAO Rong;WANG Enzhi;XU Jianxin(Chuxiong Dianzhong Nonferrous Metal Smelting Co.,Ltd.,Chuxiong 675000,Yunnan,China;School of Metallurgical and Energy Engineering,Kunming University of Science and Technology,Kunming 650093,China)
出处
《有色金属(冶炼部分)》
北大核心
2026年第1期9-19,共11页
Nonferrous Metals(Extractive Metallurgy)
基金
云南省科技厅科技计划-基础研究计划资助项目(202501AS070131)。
关键词
PS转炉
LSTM神经网络
温度预测
预测性维护
图像匹配
PS converter
LSTM neural network
temperature prediction
predictive maintenance
image matching
