Machinery health indicator construction based on convolutional neural networks considering trend burr

In the study of data-driven prognostic methods of machinery, much attention has been paid to constructing health indicators (HIs). Most of the existing HIs, however, are manually constructed for a specific degradation process and need the prior knowledge of experts. Additionally, for the existing HIs, there are usually some outlier regions deviating to an expected degradation trend and reducing the performance of HIs. We refer to this phenomenon as trend burr. To deal with these problems, this paper proposes a convolutional neural network based HI construction method considering trend burr. The proposed method first learns features through convolution and pooling operations, and then these learned features are constructed into a HI through a nonlinear mapping operation. Furthermore, an outlier region correction technique is applied to detect and remove outlier regions existing in the HIs. Unlike traditional methods in which HIs are manually constructed, the proposed method aims to automatically construct HIs. Moreover, the outlier region correction technique enables the constructed HIs to be more effective. The effectiveness of the proposed method is verified using a bearing dataset. Through comparing with commonly used HI construction methods, it is demonstrated that the proposed method achieves better results in terms of trendability, monotonicity and scale similarity.