Artificial Intelligence Techniques and Cloud Computing for Wind Turbine Pitch Bearing Fault Detection

Blade bearings are critical rotating units for maximizing wind power yield. It is essential to detect the blade bearing faults at an early stage and prevent their catastrophic failure. One major challenge lies in the signal denoising under the time-varying operating conditions. This time-varying condition is often treated as a series of piece-wise time invariant conditions when filtering the collected signal. The duration of the time invariant period, also referred to as the window length or comprehensive period, is often determined by trial-and-error, which could lead to improper separating the time varying signals and poor fault detection performance. In this paper, to find a suitable window length, a novel method called the Temporal Convolutional Augmented Bayesian Search (TCABS) algorithm is used to search for a ‘comprehensive period’ for the unknown signal. After estimating the window length, the Split Bayesian Augmented Lagrangian Algorithm (SBAL) was used based on split window techniques to construct time-varying models. The proposed TCABS and SBAL are validated by real signals collected from an industrial-scale wind turbine in operation for over 15 years.