The fast and efficient fault diagnosis is the key to guarantee uninterrupted working of facilities, which is more frugal and trustworthy than scheduled upkeep. At present, data acquisition and fault diagnosis based on a variety of sensors have become an indispensable means for manufacturing enterprises. However, through the independent analysis of all kinds of sensor data, the traditional analysis method fails to make full use of the interrelationship between data sources. A new feature fusion approach that is based on Convolutional Neural Network (CNN) is put forward in this study for rotating machinery fault diagnosis. For multi-source data, some data sources are extracted with empirical features and others are extracted with hidden features. CNN is adopted to obtain the recessive features of complex signal waveform, such as acceleration, displacement, etc. The fusion of statistical features and recessive features is a new set of features and is input into Light Gradient Boosting Machine (LightGBM) model. The stator and rotor fault experiment is designed and implemented to verify the advantages of the proposed method. Compared with the traditional approaches, this method is 3% more accurate or at least 4 times faster than the traditional method under the same conditions.

快速有效的故障诊断是保证设备不间断工作的关键，这比计划的维护更节俭和可信赖。目前，基于各种传感器的数据采集和故障诊断已成为制造企业不可或缺的手段。但是，通过对各种传感器数据的独立分析，传统的分析方法无法充分利用数据源之间的相互关系。提出了一种基于卷积神经网络（CNN）的特征融合新方法，用于旋转机械故障诊断。对于多源数据，某些数据源具有经验特征，而另一些则具有隐藏特征。采用CNN获得复杂信号波形的隐性特征，例如加速度，统计特征和隐性特征的融合是一组新特征，并输入到“光梯度增强机（LightGBM）”模型中。设计并实施了定子和转子故障实验，以验证所提方法的优点。与传统方法相比，该方法在相同条件下的准确度要比传统方法高3％，或至少快4倍。