【摘要】 在故障监测信号中，不同时间尺度的时间序列片段会呈现辨别性退化特征。为了全面捕捉这些差异化的多尺度退化信息，提出了多时间尺度趋势注意力卷积网络故障预测方法。该方法旨在聚焦关键信号，提取表征潜在故障的多尺度信息，实现精确的设备故障预测。可解释性分析实验揭示了该方法在故障预测过程中的部分逻辑过程。该网络通过趋势注意力机制提取信号趋势信息，以此计算不同信号的注意力权重；采用多时间尺度卷积核对加权多元时间序列进行特征提取与融合，将融合特征输入全卷积网络以提取深度退化特征，并预测故障剩余时间。在C-MAPSS涡扇发动机数据集上与先进模型进行对比实验，证实了本方法在故障预测任务中的有效性与先进性。更多还原

【Abstract】 In fault monitoring signals, time series fragments across various time scales exhibit distinctive degradation characteristics. To comprehensively capture these multi-scale degradation features, a multi-time scale trend attention convolutional network for fault prediction is proposed. This method is designed to precisely focus on extracting multi-scale information from critical signals that indicate potential faults, thereby enabling accurate equipment fault prediction. An interpretability analysis experiment sheds light on the logical processes underlying this method during fault prediction. The network employs a trend attention mechanism to compute attention weights for different signals, utilizing multi-time scale convolution kernels to extract and fuse these weighted multivariate time series. Subsequently, the fused features are fed into a fully convolutional network to extract deep degradation features and estimate the remaining time until a fault occurs.Comparative experiments with state-of-the-art models on the C-MAPSS turbofan engine dataset have validated the effectiveness and superiority of this method in fault prediction tasks.更多还原