基于集成神经网络的剩余寿命预测

针对机器或设备的剩余寿命（Remaining useful life, RUL）预测精度低的问题，提出基于一维卷积神经网络（Convolutional neural network, CNN）和双向长短期记忆（Bidirectional long short-term memory, BD-LSTM）的集成神经网络模型。为了更好地抽取时间序列上的特征，以及产生更多的训练样本，采用滑动窗口对数据进行处理，同时采用卡尔曼滤波对数据进行降噪处理，将数据标准化以及设置RUL标签。与人工提取特征不同，利用一维CNN对数据进行特征提取，并舍弃了CNN中的池化层。然后将提取到的高维特征输入到BD-LSTM进行回归预测，并采用Bagging的方式对此神经网络进行集成来预测RUL。最后通过在NASA的数据集上验证该模型的有效性，以及相比于其他机器学习或者深度学习模型的优越性，实验表明所提模型在RUL预测方面更加准确。 

Remaining useful life prediction based on an integrated neural network

Unexpected failures and unscheduled maintenance activities of mechanical systems might incur considerable waste of resources and high investment costs. Thus, in recent years, prognostics and health management (PHM) has received a lot of attention because of its importance in maintenance cost reduction and machine fault prognostics. The remaining useful life (RUL) of machinery is defined as the length from the current time to the end of its useful life, which is the core technology of PHM. During the operation of machines and equipment, a large amount of data generated by different sensors in the system is collected using various methods. These data often characterize the health status of machinery to a certain extent. By applying the systematic approach to these data, valuable information for strategic decision-making can be obtained. However, traditional machine learning algorithms are usually not efficient enough to handle the complex and nonlinear characteristics of the system and deal with big data. With the rapid development of modern computational hardware and theory, deep learning algorithms show unique advantages in characterizing the system complexity and processing big data. Because of the low-accuracy prediction of the RUL of machines or equipment, a neural network integrating the one-dimensional convolutional neural network (1D CNN) and the bidirectional long short-term memory (BD-LSTM) was proposed. To extract the features of the time series and generate more training samples, the sliding window algorithm was used to process the data and the Kalman filter was applied to denoise the data. Then, the dataset was standardized and the RUL labels were set. Instead of artificial feature extraction, this study used 1D CNN to extract features from the data and discarded the pooling layer of CNN. The extracted high-dimensional features were inputted into the BD-LSTM for regression prediction, and the neural network was integrated by bagging to predict the RUL. Finally, the effectiveness and superiority of the model compared with the machine or deep learning model were verified using the National Aeronautics and Space Administration dataset. Results showed that the proposed model can more accurately predict the RUL than the machine or deep learning model.