摘要
牵引电机温度预测在动车组牵引电机状态评估和日常维护中具有重要作用。针对现有时序预测模型提取牵引电机时序数据的特征不充分,导致模型预测精度不高的问题,提出一种基于MultiCNN-GRU-ITA的动车组牵引电机温度预测模型,通过更深层次地提取数据的时空特征来预测牵引电机的温度。该模型提出了多通道卷积神经网络(multi-channel convolutional neural networks, MultiCNN)的空间特征提取模块,多尺度地获取牵引电机数据的空间特征,增强特征的表征能力;设计了GRU(gated recurrent unit, GRU)堆叠的时间特征提取模块,采用门控循环单元捕捉数据的长期依赖关系,提取牵引电机数据的时间特征,更准确地预测温度的动态变化;引入改进的时序注意力机制模块(improved temporal attention,ITA),聚焦时空特征中的关键信息,进一步提升模型对重要特征的识别能力。利用动车组实际运行数据制作了数据集,并在多种预测场景下进行了实验测试。实验结果表明,在预测输出步长为5、10、15、20 min的4种场景下,MultiCNN-GRUITA模型在MAE和MSE方面均表现出明显的优势,相比于LSTM、GRU、SVR、ARIMA模型,MAE和MSE指标降低了41.03%和65.32%以上;在不同预测步长下,MultiCNN-GRU-ITA模型的温度预测曲线与实际值具有很高的拟合度,该模型能有效捕捉牵引电机的温度变化趋势,可为构建高精确性的牵引电机故障预测与健康评估系统提供模型支撑。
The temperature prediction is crucial for the condition assessment and daily maintenance of traction motors in high-speed trains.Aiming at the problem of insufficient feature extraction of traction motor timing data in existing timing prediction models,resulting in low prediction accuracy,a MultiCNN-GRU-ITA based temperature prediction model for high-speed train traction motors was proposed.The model could predict the temperature by deeply extracting spatiotemporal features of the extra data.This model proposed a multi-channel convolutional neural networks(MultiCNN)module,which obtained the spatial features of traction motor data at multiple scales and enhances the representation ability of features.It designed a GRU stack module,which uses Gated Recurrent Unit(GRU)to capture long-term dependencies,extracted the temporal features from the traction motor data,and more accurately predicted dynamic temperature changes.The Improved Attention Mechanism(ITA)module was introduced to focus on key information in spatiotemporal features,further enhancing the model's ability to recognize important features.The dataset used in this study was created using actual operational data from high-speed trains,and experiments were conducted in various prediction scenarios.The experimental results show that the MultiCNN-GRU-ITA model exhibits significant advantages in MAE and MSE in four scenarios with predicted output step sizes of 5 minutes,10 minutes,15 minutes,and 20 minutes.Compared with LSTM,GRU,SVR,and ARIMA models,the MAE and MSE indicators are reduced by more than 41.03%and 65.32%,respectively.Under different prediction intervals,the temperature prediction curves of the MultiCNNGRU-ITA model exhibit a high degree of fit with the actual values.This model can effectively capture the temperature change trend of the traction motor and provide model support for constructing a high-precision traction motor fault prediction and health assessment system.
作者
王运明
李明阳
陈梦华
常振臣
WANG Yunming;LI Mingyang;CHEN Menghua;CHANG Zhenchen(School of Automation and Electrical Engineering,Dalian Jiaotong University,Dalian 116028,China;State Engineering Technology Center,CRRC Changchun Railway Vehicle Co.,Ltd.,Changchun 130000,China)
出处
《铁道科学与工程学报》
北大核心
2025年第5期2367-2379,共13页
Journal of Railway Science and Engineering
基金
国家自然科学基金资助项目(62103074)
辽宁省教育厅基本科研项目(LJKMZ20220857)
辽宁省交通科技项目(202345)。
关键词
牵引电机
温度预测
多通道卷积神经网络
门控循环单元
注意力机制
traction motor
temperature prediction
multi-channel convolutional neural network
gated recurrent unit
attention mechanism
