秦子轩,张晓东,白广芝,任先聪.基于多尺度特征融合的航空发动机剩余寿命预测[J].航空发动机,2024,50(4):114-120
基于多尺度特征融合的航空发动机剩余寿命预测
Aeroengine Remaining Useful Life Prediction Based on Multi-scale Feature Fusion
  
DOI:
中文关键词:  深度学习  多头注意力机制  多尺度卷积双向长短期记忆网络  剩余可用寿命  航空发动机
英文关键词:deep learning  multi-head attention mechanism  multi-scale convolutional bidirectional long short-term memory network  remaining useful life  aeroengine
基金项目:国家自然科学基金(61801517)、中央高校基本科研业务专项经费(19CX02029A,19CX02027A)资助
作者单位
秦子轩,张晓东,白广芝,任先聪 中国石油大学(华东) 计算机科学与技术学院山东青岛 266580 
摘要点击次数: 490
全文下载次数: 498
中文摘要:
      针对航空发动机原始数据中存在多样化退化信息及大量噪声干扰的问题,建立了一种基于多尺度特征融合的发动机剩 余可用寿命(RUL)预测模型。构建了一种基于统计量的方法来降低原始数据中的噪声干扰;基于卷积双向长短期记忆网络 (ConvBiLSTM)和多头注意力机制(Multi-Attention)设计了加权时空特征提取模块(WSTFEM);采用多尺度学习策略,构建多尺度 卷积双向长短期记忆网络(MCBLSTM)提取数据在不同时间尺度下的加权时空特征;提取数据手工特征为RUL预测提供具有针对 性和解释性的退化信息;将上述特征进行特征融合后输入至全连接网络获得RUL预测值。以FD004子集为例,使用C-MAPSS数 据集对模型进行仿真试验验证。结果表明:MCBLSTM模型在4个子数据集上RUL预测精度更高。相较于BiLSTM,均方根误差减 小了20.35%,非对称评分函数下降了54.76%。
英文摘要:
      Aiming at the issues of diversified degradation information and large amounts of noise interference in the raw data of aeroengine, an engine remaining useful life (RUL) prediction model of based on multi-scale feature fusion was established. A statistic- based method is constructed to reduce noise interference in the raw data. The weighted spatial-temporal feature extraction module (WSTFEM) was designed based on ConvBiLSTM and Multi-Attention. The multi-scale learning strategy was utilized to construct a multi- scale convolutional bidirectional long short-term memory network (MCBLSTM) to extract weighted temporal and spatial features of data at different temporal scales. The handcrafted features of data were extracted to provide targeted and explanatory degradation information for RUL prediction. The above features were fused and fed into the fully connected network to obtain the predicted RUL value. Taking the FD004 subset as an example, the C-MAPSS dataset was used to verify the model. The results show that the MCBLSTM model has higher RUL prediction accuracy on the four sub-datasets. Compared with BiLSTM, the root mean square error is reduced by 20.35%, and the asymmetric scoring function is reduced by 54.76%.
查看全文  查看/发表评论  下载PDF阅读器