Remaining useful life prediction of aero-engine via temporal convolutional network with gated convolution and channel selection unit

As a hot spot in the prognostics and health management (PHM), predicting the remaining useful life (RUL) is significant to ensure the availability and reliability of industrial systems. Data-driven methods, such as the traditional temporal convolutional networks (TCNs), have been widely used for RUL prediction. However, traditional TCNs lack effective means to sufficiently extract both the spatial and temporal features of multisensory data. Furthermore, the prediction ability of traditional TCNs is constrained by their inflexible network structure. Here, an integrated model called the improved temporal convolutional network with gated convolution and channel selection unit (GCSU-ITCN) is proposed as a solution to the above problems. First, "the loss boundary to mapping ability" (LM) method is introduced to select sensors with superior mapping ability to the RUL. It contributes to the extraction of features from multi-sensory data. Then, the gated convolution and channel selection unit (GCSU) is designed to extract both local and global features of multi-sensory data, encompassing the spatial and temporal dimensions. Last, the improved temporal convolutional network (ITCN) is developed to capture long temporal correlations within the extracted features. The ITCN has a flexible structure that enables it to learn deep temporal features more effectively than traditional TCNs. A series of comparative experiments is conducted on the C-MAPSS dataset to evaluate the prediction capability of the GCSU-ITCN. Furthermore, ablation experiments are also conducted to assess the contribution of each component within the model.