Gas Turbine Rotor Fault Diagnosis Based on Domain Adversarial Adaptation Transfer Learning for Small Samples

Fault diagnosis for gas turbine rotor system is of great importance, which can effectively improve its operation safety and reliability. In the fault diagnosis process, the labeled fault data is insufficient. A gas turbine domain adversarial adaptation transfer learning method combining with multidimensional feature fusion and domain adversarial adaptation transfer learning is proposed. Firstly, fault samples signals are input to the Deep Convolution Neural Network and Residual Network (ResNet) for multidimensional feature extraction. Secondly, the obtained features are effectively fused by Discriminant Correlation Analysis (DCA) method to obtain more rich and discriminate feature presentation. Then, the obtained model network is trained by supervised learning and transferred to target domain. Finally, domain adversarial adaption based on Wasserstein distance under different working conditions is adopted to fine-tune the network to realize the domain adaptation and improve the fault diagnosis accuracy under the condition of few label samples. The gas turbine test bench experiments under different working conditions are verified the method. Test results prove that the method obtains good fault diagnosis accuracy and adaptability for gas turbine.