To address the insufficient generalization ability of abnormal state assessment caused by cross-domain data distribution shifts in wind turbine blades under complex operating conditions
this paper proposes a Dynamic Domain Adaptive Residual Attention Network (DDA-RAN) based cross-domain anomaly assessment method. This approach integrates transfer learning with a dynamic attention mechanism
achieving collaborative optimization of domain shift through a bidirectional dynamic calibration strategy. Specifically
a Dynamic Domain Adaptive Attention Residual Block (DDAAR-Block) is designed. It utilizes a dynamic domain adaptive attention mechanism combined with switchable domain identifiers to adaptively activate either the source or target domain branch
balancing the representation of domain-invariant and domain-specific features. A cross-domain alignment module is constructed
implementing phased dynamic training via gradient truncation and a gating mechanism to promote cross-domain feature interaction and distribution alignment. An adaptive multi-scale residual block is adopted to extract deep and shallow temporal features
enhanced by global adaptive pooling for feature robustness. Furthermore
a regularized Multi-Layer Perceptron (MLP) classifier is introduced to improve model generalization. Experimental results demonstrate that in cross-domain assessment tasks for blade cracking and icing
the proposed method significantly outperforms mainstream comparison models (e.g.
ADDA
DANN) in terms of target domain accuracy (up to 99.36%)
F1-score (99.35%)
and MCC (98.71%). The accuracy shows an average improvement of 2.1%-3.8%. This provides a valuable reference for cross-domain anomaly assessment in industrial scenarios.
JIN X H, LYU S Y, KONG Z Q, et al.Graph spatio-temporal networks for condition monitoring of wind turbine[J]. IEEE Transactions on Sustainable Energy, 2024, 15(4): 2276-2286.
Yang J Y, An W Z, Yan C C, et al.Context-aware domain adaptation in semantic segmentation[C]//2021 IEEE Winter Conference on Applications of Computer Vision (WACV). Waikoloa, HI, USA, 2021: 514-524.
LIU F X, LI Z J, LIANG M L, et al.Prediction method of non-stationary random vibration fatigue reliability of turbine runner blade based on transfer learning[J]. Reliability Engineering & System Safety, 2023, 235: 109215.
Yue R X,Jiang G Q, Jin X H, et al.Spatio-temporal feature alignment transfer learning for cross-turbine blade icing detection of wind turbines[J]. IEEE Transactions on Instrumentation and Measurement, 2024, 73:1-17.
Yang W Q, Yu G.Federated multi-model transfer learning-based fault diagnosis with peer-to-peer network for wind turbine cluster[J]. Machines, 2022, 10(11): 972.