A Multiscale Cross-Channel Attention Network for Remaining Useful Life Prediction With Variable Sensors

The remaining useful life (RUL) prediction of machinery based on deep learning (DL) represents a crucial component in the field of prognostics and health management (PHM). However, these DL-based methods for RUL prediction tend to become unreliable during online inference when only partial signals are available. To address this issue, we introduce the variable sensors scenario and propose a multiscale cross-channel attention network (MSCCAN) specifically designed for RUL prediction with variable sensors. For each input sample with variable sensors, the embedding layer is utilized to transform the dimensions of the input tensor. The multiscale cross-channel attention (MSCCA) layer is employed to extract and fuse multichannel degradation information, where the multiscale convolutional attention (MSCA) blocks extract the multiscale degradation features, and the anti-missing cross-channel attention (AMCCA) block effectively integrates feature information while mitigating interference from missing sensors. A mask global average (MGA) layer is used to compress high-dimensional features without being affected by the channels from missing sensors. Moreover, a new data augmentation method is used to improve the robustness of the model to the inputs with variable sensors. Finally, the experiments on the commercial modular aero-propulsion system simulation (CMAPSS) dataset and the New CMAPSS (N-CMAPSS) dataset validate the effectiveness of MSCCAN under both normal and variable sensor scenarios. Experimental results demonstrate that the proposed method can reduce the prediction error by more than 40% compared with the comparison methods under the worst scenario with variable-sensor inputs.