Early bearing fault diagnosis for imbalanced data in offshore wind turbine using improved deep learning based on scaled minimum unscented kalman filter

The development of low-speed fault diagnosis methods especially in offshore wind turbines is considered of utmost importance for mainly solving two challenges. These include diagnosis based on imbalanced data with low signal to noise ratio and invariant features acquired from multi-sensors. To effectively address these issues, in this work, an improved deep belief network, termed Scaled-minimum Unscented Kalman Filter-aided DBN, was proposed for processing imbalanced data under low-speed. First, a Gramian Angular Summation Field was designed to preserve absolute temporal relation in time-series for 2-D feature maps. Second, the traditional deep belief network was improved by using a Scaled-minimum Unscented Kalman Filter to enhance the nonlinear tracking ability. The latter can make the feature representation of 2-D feature maps dynamically adapt its configuration and enhance the anti-noise ability of the diagnosis model. Wherein, minimum sigma set and scaled unscented transform were introduced to improve the ability of discriminative fault features in imbalanced data with low-speed, making the diagnostic model more efficient. Two different low-speed experimental cases were conducted to analyse the performance of the proposed method. From the extracted results, the anti-noise ability to diagnose the fault in imbalanced data was demonstrated.