Randomization-based neural networks for image-based wind turbine fault diagnosis

As the development of wind energy industry, the safe production of wind farms has become an urgent problem. To avoid serious faults and deterioration, building effective diagnostic model for wind turbine (WT) has raised increasing attentions in wind-power industry. However, the challenges like big data of sensors and model construction exist still. In this paper, to achieve better performance and suitable framework, a three channel broad learning system (3-BLS) is proposed for image-based fault diagnosis (FD) on overall WT system. First, multiple sensor series are collected and converted into interpretable RGB images via right-sized sliding window for broader information and grabbing relations; Next, features are extracted in respective RGB channels, and a manual feature layer is added in the 3-BLS, where the structure is temporary non-specific; Finally, with the help of an optimizer, the concrete 3-BLS is auto-built with its structure configured reasonably and the manual features binary-coded and enabled selectively. In addition, an inter-channel attention scheme is formed during 3-BLS dynamic updating process, and several BLS prototypes different in projections are studied. In experiments, the optimized 3-BLS with less parameters got over 10% accuracy gain than adjusted single BLS and achieved over 98% fault detection on actual collected WT data.