| 陈俊英，席月芸，徐 琳，殷春武.基于MLP集成随机子空间决策树的航空发动机
剩余使用寿命预测[J].航空发动机,2024,50(6):81-87 |
| 基于MLP集成随机子空间决策树的航空发动机
剩余使用寿命预测 |
| Remaining Useful Life Prediction of Aeroengines Based on MLP Integrated RandomSubspace Decision Trees |
| |
| DOI： |
| 中文关键词: 剩余使用寿命预测 航空发动机 随机子空间 决策树 集成方法 多层感知器 |
| 英文关键词:remaining useful life prediction aeroengine random subspace decision tree integrated method multi-layer perceptron |
| 基金项目:陕西省自然科学基金（2023-JC-YB-562）资助 |
|
| 摘要点击次数: 124 |
| 全文下载次数: 74 |
| 中文摘要: |
| 为了预测航空发动机的剩余使用寿命（RUL），针对从众多发动机状态参数中选择特定特征组合进行预测的问题，构建
了基于多层感知器（MLP）集成随机子空间决策树的剩余寿命预测模型；随机选取抽样样本的特征子空间构建决策回归树；构建
MLP模型结构和损失函数，通过适应性矩估计（Adam）算法优化MLP模型参数，基于MLP模型集成多棵决策树的预测结果，得到
发动机的剩余使用寿命；在C-MAPSS数据集上进行的消融试验验证了预测模型中随机特征子空间、决策回归树和MLP集成模块
均有益于改善平均绝对误差（MAE）、均方根误差（RMSE）、惩罚得分、拟合优度以及准确率等预测指标值。结果表明：当真实剩余
寿命周期小于30时，预测准确率比循环神经网络（RNN）的预测结果提高了7.46%；与其他几种预测方法相比，该方法在多个指标
综合评价下具有较好的性能，为多参数航空发动机剩余寿命的预测提供了一种有效方案。 |
| 英文摘要: |
| To predict the Remaining Useful Life (RUL) of aeroengines, a prediction model based on the integration of Multi-Layer
Perceptron (MLP) with random subspace decision trees was constructed to address the issue of selecting a specific feature combination from
numerous engine state parameters. This model randomly selects feature subspaces from sampled data to build decision regression trees. The
MLP model structure and loss function were established, and the parameters of the MLP model were optimized using the Adaptive Moment
Estimation (Adam) algorithm. By integrating the prediction results from multiple decision trees based on the MLP model, the RUL of the
engine is obtained. Ablation experiments conducted on the C-MAPSS dataset demonstrate that the random feature subspaces, decision
regression trees, and MLP integration modules in the prediction model all contribute to improving prediction metrics such as Mean
Absolute Error (MAE), Root Mean Square Error (RMSE), Penalty Score, Goodness of Fit and Accuracy. The results show that when the true
RUL cycle is less than 30, the prediction accuracy is improved by 7.46% compared to the predictions from Recurrent Neural Networks
(RNN). Compared with other prediction methods, this method exhibits better performance under comprehensive evaluations across multiple
metrics, providing an effective solution for multi-parameter prediction of the RUL of aeroengines. |
| 查看全文 查看/发表评论 下载PDF阅读器 |