基于K-Means聚类的粒子群优化CNN-BiGRU-HAM发动机剩余使用寿命预测方法  

A Particle Swarm Optimization CNN-BiGRU-HAM Engine Remaining Service Life Prediction Method Based on K-Means Clustering

作　　者：王晓鹏 王磊 韩小伟 张鹏超 徐浩然[1,2] WANG Xiaopeng;WANG Lei;HAN Xiaowei;ZHANG Pengchao;XU Haoran(School of Mechanical Engineering,Shaanxi University of Technology,Hanzhong Shaanxi 723001,China;Shaanxi Key Laboratory of Industrial Automation,Shaanxi University of Technology,Hanzhong Shaanxi 723001,China)

机构地区：[1]陕西理工大学机械工程学院,陕西汉中723001 [2]陕西理工大学陕西省工业自动化重点实验室,陕西汉中723001

出　　处：《机床与液压》2024年第20期239-247,共9页Machine Tool & Hydraulics

基　　金：国家自然科学基金面上项目(62176146)。

摘　　要：飞机在多种工况条件下运行时,发动机退化特征复杂性不断增加,导致发动机剩余寿命预测精度低。针对此问题,提出一种基于聚类分析的端到端剩余寿命(RUL)预测方法。采用K-Means聚类方法对发动机的多种工况和运行条件进行分组;再利用卷积神经网络(CNN)提取反映剩余寿命复杂动态变化的高维特征,将结果输入到双向门控循环单元(BiGRU)中学习特征之间的变化规律,设计并引入了新的混合注意力机制(HAM),充分考虑变量之间的关系,对重要特征信息赋予更大的权重,同时抑制冗余信息的影响;然后进行非线性变换,获得RUL预测结果;最后使用粒子群优化算法对神经网络的超参数进行调优。采用美国航天局NASA研究中心提供的涡轮发动机模拟数据集验证所提网络模型的有效性。结果表明:对于多工况运行条件,所提方法的均方根误差相比于CNN、LSTM、BiLSTM、CNN-LSTM分别降低了49.2%、37.1%、33.6%、24.8%,有效提升了模型的预测精度。When the aircraft operates under multiple operating conditions,the complexity of engine degradation features is increasing,resulting in low accuracy of engine remaining life prediction.Aiming at the problem,an end-to-end remaining useful life(RUL)prediction method was proposed based on clustering analysis.A K-Means clustering method was used to group multiple working conditions and operating conditions of the engine.A convolutional neural network(CNN)was used to extract high-dimensional features reflecting the complex dynamic changes of the remaining life,and the results were inputted into the bidirectional gated recurrent unit(BiGRU)to learn the changing law between features.A new hybrid attention mechanism(HAM)was designed and introduced,the relationship between variables were fully considered and more weight was given to important feature information,the influence of redundant information was suppressed.A nonlinear transformation was performed to obtain the RUL prediction results.Finally,the hyperparameters of the neural network were tuned using the particle swarm optimization algorithm.The turbine engine simulation dataset provided by NASA Research Centre of the United States space agency NASA was used to verify the effectiveness of the proposed network model.The results show that for multiple operating conditions,the root-mean-square error of the method is reduced by 49.2%,37.1%,33.6%and 24.8%compared to CNN,LSTM,BiLSTM and CNN-LSTM,respectively,which effectively improves the prediction accuracy of the model.

关 键 词：多工况聚类 卷积神经网络(CNN) 双向门控循环神经网络 混合注意力机制(HAM) 粒子群优化算法 

分 类 号：V263.5[航空宇航科学与技术—航空宇航制造工程] TP181[自动化与计算机技术—控制理论与控制工程]