电子测量与仪器学报

王昱,张哲成,杨晓庆.基于MTCN和双重注意力的航空发动机RUL预测[J].电子测量与仪器学报,2025,39(11):142-151

基于MTCN和双重注意力的航空发动机RUL预测

Remaining useful life prediction for aircraft engine basedon MTCN and dual attention

DOI：

英文关键词:aircraft engine remaining useful life multi-scale time convolutional network dual attention GeLU

基金项目:国家自然科学基金 (61906125，62373261)、辽宁省科技计划联合计划项目(2025-MSLH-572)、辽宁省高校基本科研业务费项目（LJ232410143020，LJ212410143047）资助

作者	单位
王昱	沈阳航空航天大学自动化学院沈阳110136
张哲成	沈阳航空航天大学自动化学院沈阳110136
杨晓庆	沈阳航空航天大学自动化学院沈阳110136

Author	Institution
Wang Yu	School of Automation, Shenyang Aerospace University, Shenyang 110136, China
Zhang Zhecheng	School of Automation, Shenyang Aerospace University, Shenyang 110136, China
Yang Xiaoqing	School of Automation, Shenyang Aerospace University, Shenyang 110136, China

摘要点击次数: 409

全文下载次数: 331

中文摘要:

当前航空发动机剩余使用寿命预测常局限于多源传感器数据的整体分析，采用单一时间尺度或以空间维度聚焦退化特征，忽视了不同传感器在不同时刻所呈现的关键特征差异，导致特征提取不充分。为此，首先将各传感器信息视为整体，设计了多尺度时间卷积网络（MTCN），以全面提取其长期与短期特征。在此基础上，引入了包含“通道注意力”和“自注意力”的双重注意力机制，通过自适应的权重分配，不仅显著增强了空间特征的提取，还成功补充了对各传感器信道在关键时间点信息的精准聚焦。通过MTCN与双重注意力机制有效协作，实现了时空特征的全面且高效融合。此外，采用高斯误差线性单元（GeLU）作为激活函数，进一步提升模型对航空发动机非线性数据的处理能力。在美国航天局C-MAPSS数据集上的实验验证结果表明，该方法应对复杂工况及多样故障模式时，预测精度和鲁棒性均得到大幅提升，与现有先进方法相比，其整体预测指标均方根误差（RMSE）和Score分别平均下降了7%和13.1%。

英文摘要:

Current aircraft engine remaining useful life prediction methods often rely on a holistic analysis of multi-source sensor data, typically using a single time scale or focusing on spatial features, which neglects key differences in sensor data at different time points. To address these limitations, a novel multi-scale temporal convolutional network (MTCN) is proposed to comprehensively extract both long-term and short-term temporal features from multi-source sensor data. Additionally, a dual attention mechanism, integrating channel attention and self-attention, is designed to enhance spatial feature representation and selectively focus on critical sensor measurements at key time points. The collaborative integration of MTCN and the dual attention mechanism facilitates effective spatiotemporal feature fusion, improving the model’s capacity to capture complex degradation patterns. Moreover, the Gaussian error linear unit (GeLU) activation function is employed to enhance the network’s nonlinear fitting capability. Experimental evaluations conducted on the NASA C-MAPSS benchmark dataset demonstrate that the proposed method significantly outperforms state-of-the-art approaches, achieving average reductions of 7% in root mean square error (RMSE) and 13.1% in Score, thereby verifying its superior prediction accuracy and robustness.

查看全文查看/发表评论下载PDF阅读器