Transformers have achieved promising results on aeroengine remaining useful life(RUL)prediction,but they still have several limitations:1)Aeroengine domain knowledge,which contains rich information that can reflect th...Transformers have achieved promising results on aeroengine remaining useful life(RUL)prediction,but they still have several limitations:1)Aeroengine domain knowledge,which contains rich information that can reflect the aeroengine’s health statue,is largely ignored in modeling process;2)Traditional transformer ignores the valuable degradation information from other time scales.To address these issues,a novel domain knowledge-augmented multiscale transformer(DKAMFormer)is developed by integrating domain knowledge and multiscale learning to improve the prognostic performance and reliability.First,to obtain rich and professional aeroengine domain knowledge,multiple detail and complete knowledge graphs(KGs)are established based on the working principle of aeroengine,including aeroengine structure,components working characteristics and sensor parameters.Second,the domain knowledge contained in KGs is convert to embedded vector by KG representative learning,which are then utilized to strengthen and enrich the original multidimensional time-series(MTS)monitoring data,aiming to intergrade domain knowledge and monitoring data to train DKAMFormer.Third,to learn rich and complementary degradation features,a novel multiscale time scale-guided self-attention(MTSGSA)mechanism is designed,which maps original MTS into different time-scale feature spaces,and then employs multiple independent self-attention head to extract the degradation features from different time-scale spaces.Finally,through a series of comparative experiments on the public CMAPSS and NCMAPSS datasets and compared with 17 SOTA methods,the developed DKAMFormer significantly improves the RUL prediction performance under multiple operation conditions and degradation modes.展开更多
基金supported in part by the National Natural Science Foundation of China(52305570)the National Natural Science Foundation of China Key Support Project(U2133202)+2 种基金China Postdoctoral Science Foundation(2022M720955)Postdoctoral Science Foundation of Heilongjiang Province(LBH-Z22187)Outstanding Doctoral Dissertation Funding Project of Heilongjiang Province(LJYXL2022-011).
文摘Transformers have achieved promising results on aeroengine remaining useful life(RUL)prediction,but they still have several limitations:1)Aeroengine domain knowledge,which contains rich information that can reflect the aeroengine’s health statue,is largely ignored in modeling process;2)Traditional transformer ignores the valuable degradation information from other time scales.To address these issues,a novel domain knowledge-augmented multiscale transformer(DKAMFormer)is developed by integrating domain knowledge and multiscale learning to improve the prognostic performance and reliability.First,to obtain rich and professional aeroengine domain knowledge,multiple detail and complete knowledge graphs(KGs)are established based on the working principle of aeroengine,including aeroengine structure,components working characteristics and sensor parameters.Second,the domain knowledge contained in KGs is convert to embedded vector by KG representative learning,which are then utilized to strengthen and enrich the original multidimensional time-series(MTS)monitoring data,aiming to intergrade domain knowledge and monitoring data to train DKAMFormer.Third,to learn rich and complementary degradation features,a novel multiscale time scale-guided self-attention(MTSGSA)mechanism is designed,which maps original MTS into different time-scale feature spaces,and then employs multiple independent self-attention head to extract the degradation features from different time-scale spaces.Finally,through a series of comparative experiments on the public CMAPSS and NCMAPSS datasets and compared with 17 SOTA methods,the developed DKAMFormer significantly improves the RUL prediction performance under multiple operation conditions and degradation modes.
