Speed–Load Insensitive Fault Diagnosis Method of Wind Turbine Gearbox Based on Adversarial Training

The rotational speed and load torque of wind turbine gearboxes can vary widely during operation, which has an obvious impact on the gearbox fault diagnosis carried out based on vibration signals. To address this problem, this paper proposes a fault diagnosis method that introduces an adversarial training mechanism and designs a game learning strategy among the feature extractor, fault recognizer, rotational speed estimator, and load estimator. In this way, the network tends to acquire fault features with weaker correlation with rotational speed and load and thus improves the performance of the fault diagnosis network in the face of the samples from the rotational speed and load ranges that are not covered by the training set. At the same time, in order to verify the effectiveness of the proposed method, in this paper, we have designed an experimental platform for wind turbine gearbox scaling, carried out simulation experiments of variable speed and torque faults, collected experimental data, and constructed a variable speed and load fault dataset. Comparing the proposed method with the baseline model, when confronted with data from RPMs or load ranges not covered by the training set, the accuracy of the baseline model drops by anywhere from 10.54% to 16.46%, while the accuracy of the method drops by only 1.39%. The results show that the method can effectively improve the performance of the fault diagnosis network when facing a variation of speed and load.