基于多尺度排列熵和长短时记忆神经网络的航空发动机剩余寿命预测

针对航空发动机性能退化失效的变点和多状态参数的时间序列预测, 构建了基于多尺度排列熵算法和长短时记忆神经网络的剩余寿命预测模型; 使用多尺度排列熵算法对时间序列进行变点分析, 求解出性能退化过程中的突变点, 得到了有故障征兆的性能退化起始点; 构建了包含多变量的长短时记忆神经网络模型, 将多个状态参数代入到模型中得到对应的剩余寿命; 将变点后的航空发动机多状态参数和剩余寿命作为样本, 代入到长短时记忆神经网络模型中进行多步和多变量的时间序列预测; 通过综合航空发动机状态参数变点分析方法和时间序列预测模型, 得到最终的剩余寿命预测结果。研究结果表明: 多尺度排列熵算法能够及时监控各个状态参数的变化, 当发现状态参数异常时, 排列熵的值会发生跳变, 从而有助于及时发现故障征兆; 长短时记忆神经网络模型通过门控单元对长时间序列数据进行信息筛选, 充分保留了有效信息用于时间序列预测; 多变量长短时记忆神经网络能够对多状态参数进行同步分析, 并且将状态参数直接与剩余寿命相对应, 提高了模型效率; 通过多尺度排列熵算法和长短时记忆神经网络模型的结合, 能够考虑到航空发动机的多退化模式, 得到更符合实际退化过程的剩余寿命预测结果; 经过算例分析, 提出方法的剩余寿命预测的均方根误差为5.3, 与长短时记忆神经网络、反向传播神经网络和支持向量机相比, 误差分别降低了63%、72%和78%。 

Residual life prediction of aeroengine based on multi-scale permutation entropy and LSTM neural network

Aiming at the change point of aeroengine performance degradation failure and the time series prediction of multi-state parameters, the residual life prediction model based on the multi-scale permutation entropy(MPE) algorithm and long-short term memory(LSTM) neural network was constructed. The change points in time series were analyzed by the MPE algorithm, and the mutation points in the process of performance degradation were solved. The starting point of performance degradation with fault symptoms was obtained. The LSTM neural network model with multi-variables was constructed, and the corresponding residual life was obtained by introducing the multi-state parameter data into the model.The aeroengine multi-state parameters and residual life after the change point were taken as samples and substituted into the LSTM neural network model, the multi-step and multi-variable time series prediction was carried out.The final residual life prediction results were obtained by integrating the state parameter change point analysis method and time series prediction model of aeroengine. Research result shows that the MPE algorithm can monitor the changes of state parameters in time. When abnormal state parameters are found, the value of permutation entropy will jump, which is helpful to discover the fault symptoms in time. The LSTM neural network model selects the information of long time series data through the gated units, and the effective information can be fully reserved for the time series prediction. The multi-variable LSTM neural network can synchronously analyze the multi-state parameters, and directly correspond to the residual life, which improves the efficiency of the model. The combination of MPE algorithm and LSTM neural network model can take the multiple degradation modes of aeroengine into account, and the residual life prediction results of aeroengine are more in line with the actual degradation process. After an example analysis, the root mean square error of the proposed residual life prediction method is 5.3, which is 63%, 72% and 78% lower than that of LSTM neural network, back-propagation neural network and support vector machine, respectively.