基于并联CNN-SE-Bi-LSTM的轴承剩余使用寿命预测

滚动轴承作为一种机械标准件,广泛应用于各类旋转机械设备,其健康状况对机器设备的正常运行至关重要,掌握其剩余使用寿命(RUL)可以更好地保证生产活动安全有效的进行.针对目前基于深度学习的机器RUL预测方法普遍存在:a)预测性能很大程度依赖手工特征设计;b)模型不能够充分提取数据中的有用特征;c)学习过程中没有明确考虑多传感器数据等缺点,提出了一种新的深度预测网络——并联多个带有压缩激励机制的卷积神经网络和双向长短期记忆网络集成网络(CNN-SE-Bi-LSTM),用于设备的RUL预测.在该预测网络中,不同传感器采集的监测数据直接作为预测网络的输入.然后,在改进的压缩激励卷积网络(CNN-SE-Net)提取空间特征的基础上进一步通过双向长短期记忆网络(Bi-LSTM)提取时序特征,建立起多个独立的可以自动从输入数据中学习高级表示的RU L预测模型分支.最后,将各独立分支学习到的特征通过全连接层并联获得最终的RU L预测模型.通过滚动轴承加速退化实验的数据,验证了所提网络的有效性并与现有的一些改进算法进行了对比实验.结果表明,面对原始多传感器数据,该算法能够自适应地提供准确的RU L预测结果,且预测表现优于现有一些预测方法.

Prediction of bearing remaining useful life based on parallel CNN-SE-Bi-LSTM

As a kind of mechanical standard parts, rolling bearing is widely used in all kinds of rotating machinery. Its health condition is very important for the normal operation of equipments. Mastering its remaining useful life(RUL) can better ensure the security and efficiency of production activities. Aiming at the common problems of current machine RUL prediction methods based on deep learning: a) the prediction performance largely depends on manual feature design, b) the models cannot fully extract the useful features from the data, c) the multi-sensor data is not explicitly considered in the learning process. This paper proposed a new equipment RUL prediction network: parallel a set of integrated network comprised of CNN network with SE mechanism(CNN-SE-net) and Bi-LSTM network(CNN-SE-Bi-LSTM). It directly used data from different sensors as inputs to the prediction network, and extract spacial features with the improved CNN-SE-net, extracted the temporal features with Bi-LSTM to establish several independent branches of RUL prediction model which can automatically learn high-level representation from input data. It obtained the final RUL prediction model by paralleling the features learned from the branches with the fully connection layer. The effectiveness of the proposed network is verified by the data of accelerated degradation test of rolling bearing, and comparing experiments with some existing improved algorithms. The results show that, face with the original multi-sensor data, the algorithm can adaptively provide accurate RUL prediction results, and the prediction performance is better than some existing prediction methods.