张玉彦, 李文浩, 闫玲娣,等.基于多源信息融合的矿山辊磨机故障诊断方法[J].河南理工大学学报（自然科学版）,2026,45(5):40-47.

ZHANG Y Y, LI W H, YAN L D ,et al.Fault diagnosis method for mine roller mill based on multi-source information fusion[J].Journal of Henan Polytechnic University(Natural Science) ,2026,45(5):40-47.

基于多源信息融合的矿山辊磨机故障诊断方法

张玉彦1, 李文浩1, 闫玲娣2, 李浩1, 何文斌1, 明五一1, 文笑雨1, 王昊琪1

1.郑州轻工业大学 河南省机械装备智能制造重点实验室，河南 郑州  450000；2.中信重工 矿山机械研究院，河南 洛阳 471000

摘要: 目的针对传统故障诊断方法往往只依赖于单一数据源，如振动数据或温度数据，在诊断多故障激励源的矿用辊磨机时准确率低的问题，提出多源信息融合的故障诊断方法。  方法首先，采用箱形图方法进行统计分析，对所有样本标记是否异常，进一步对多源数据进行t-SNE降维预处理，挖掘最关键的特征量；其次，将降维后的数据输入神经网络模型中进行训练，另外，提取3种单源数据样本（电流源、温度源、压力源）；最后，将多源信息融合数据和单一信息源数据分别输入神经网络中得到诊断结果。结果 以中信重工机械股份有限公司的高压辊磨机实际生产作业多源传感器数据为例，对多源融合数据和单源数据下辊磨机故障诊断准确率进行对比试验。实验对比结果显示，电流源数据训练迭代218次，训练时间为466 s，测试准确率为90.4%；压力源数据训练迭代672次，训练时间为1 426 s，测试准确率为94.1%；温度源数据训练迭代658次，训练时间为1 247 s，测试准确率为86.1%；多源信息融合数据只训练迭代36次便达到目标阈值，停止迭代，训练时间为77 s，测试准确率为98.9%。多源信息融合方法明显优于仅用单一数据源的方法，诊断误差显著降低。  结论基于多源信息融合的故障诊断方法可以显著提高故障诊断的准确率，是一种非常有效的故障诊断方法。

关键词:多源信息融合;故障诊断;箱形图;t-SNE降维;神经网络

doi:10.16186/j.cnki.1673-9787.2025100055

基金项目:国家自然科学基金资助项目（52105536）；广东省基础与应用基础研究基金资助项目（2022A1515140066）；河南省重点研发专项项目（221111240200）；河南省重点研发与推广专项项目（科技攻关）（232102221009）

收稿日期:2025/10/25

修回日期:2025/12/20

出版日期:2026-07-24

Fault diagnosis method for mine roller mill based on multi-source information fusion

Zhang Yuyan1, Li Wenhao1, Yan Lingdi2, Li Hao1, He Wenbin1, Ming Wuyi1, Wen Xiaoyu1, Wang Haoqi1

1.Henan Key Laboratory of Mechanical Equipment Intelligent Manufacturing， Zhengzhou University of Light Industry， Zhengzhou  450000， Henan， China；2.Mining Machinery Research Institute， CITIC Heavy Industry， Luoyang  471000， Henan， China

Abstract: Objectives Traditional fault diagnosis methods often rely on a single data source， such as vibration data or temperature data， resulting in low accuracy when applied to mine roller mills with multiple fault excitation sources. To address this issue， a fault diagnosis method based on multi source information fusion is proposed.  Methods First， the boxplot method is used for statistical analysis to label all samples as normal or abnormal. Then， multi source data are preprocessed using t-SNE dimensionality reduction to extract the most critical features. Subsequently， the reduced dimension data are fed into a neural network for training. Meanwhile， three types of single source data samples （current， temperature， and pressure） are also extracted. Finally， the multi source fused data and the single source data are input separately into the neural network to obtain the diagnostic results.  Results Taking the multi sensor data from the high pressure roller mill of CITIC Heavy Industries Co.， Ltd. as an example， comparative experiments on fault diagnosis accuracy are conducted using both multi source fused data and single source data. The experimental results show that： for the current source data， the training converges after 218 iterations， with a training time of 466 s and a test accuracy of 90.4%； for the pressure source data， training converges after 672 iterations， with a training time of 1 426 s and a test accuracy of 94.1%； for the temperature source data， training converges after 658 iterations， with a training time of 1 247 s and a test accuracy of 86.1%. In contrast， the multi source fused data reaches the target threshold after only 36 iterations， with a training time of 77 s and a test accuracy of 98.9%. The multi source information fusion method significantly outperforms the single source methods， markedly reducing the diagnosis error.  Conclusions The fault diagnosis method based on multi source information fusion can substantially improve fault diagnosis accuracy and is a highly effective approach.

Key words:multi-source information fusion;fault diagnosis;boxplot;t-SNE dimensionality reduction;neural network