Remaining Useful Life (RUL) estimation is a fundamental task in the prognostic and health management (PHM) of industrial equipment and systems. To this end, we propose a novel approach for RUL estimation in this paper, based on deep neural architecture due to its great success in sequence learning. Specifically, we take the Transformer encoder as the backbone of our model to capture short- and long-term dependencies in a time sequence. Compared with convolutional neural network based methods, there is no limitation from the kernel size for a complete receptive field of all time steps. While compared with recurrent neural network based methods, we develop our model based on dot-product self-attention, enabling it to fully exploit parallel computation. Moreover, we further propose a gated convolutional unit to facilitate the model’s ability of incorporating local contexts at each time step, for the attention mechanism used in the Transformer encoder makes the output high-level features insensitive to local contexts. We conduct experiments on the C-MAPSS datasets and show that, the performance of our model is superior or comparable to those of other existing methods. We also carry out ablation studies and demonstrate the necessity and effectiveness of each component used in the proposed model.

剩余使用寿命（RUL）估算是工业设备和系统的预后和健康管理（PHM）的基本任务。为此，由于深度学习在序列学习中的巨大成功，我们在本文中提出了一种基于深度神经结构的RUL估计新方法。具体来说，我们将Transformer编码器作为模型的主干，以捕获时间序列中的短期和长期依赖性。与基于卷积神经网络的方法相比，对于所有时间步长的完整接收域，内核大小没有限制。与基于递归神经网络的方法相比，我们基于点积自注意力开发了模型，从而使其能够充分利用并行计算。而且，我们进一步提出了门控卷积单元，以促进模型在每个时间步合并局部上下文的能力，因为在Transformer编码器中使用的注意力机制使输出高级特征对局部上下文不敏感。我们对C-MAPSS数据集进行了实验，结果表明，我们模型的性能优于或优于其他现有方法。我们还进行了消融研究，并证明了所提出模型中使用的每个组件的必要性和有效性。我们模型的性能优于或优于其他现有方法。我们还进行了消融研究，并证明了所提出模型中使用的每个组件的必要性和有效性。我们模型的性能优于或优于其他现有方法。我们还进行了消融研究，并证明了所提出模型中使用的每个组件的必要性和有效性。