End-to-end wind turbine damage detection model based on multi-branch feature sensing and contextual information reuse in harsh environments

Large wind turbines work in harsh environments for long periods of time, and blade damage is a frequent problem. Accurate detection of blade damage is particularly important for the safe and economic operation of wind turbines. Traditional target detection algorithms are unable to integrate global features and form a long-term memory of features when facing large-scale multi-category datasets such as wind turbine damage, and are prone to feature loss problems as the depth of the network increases. In this paper, we propose an end-to-end lightweight damage detection model to solve the above problem. Efficient feature encoders and decoders are first used to enhance the model’s ability to memorize features over time. Subsequently, a multi-branch reparameterized feature extraction network is designed in reducing the computational complexity of the model and improving the dynamic splicing and cross-layer fusion ability of the model. To enhance the ability of multi-scale feature perception and contextual information utilization, sparse parallel feature pyramid networks are designed to improve the enhancement of deep and shallow features in terms of coarse- and fine-grained aspects and to reduce the inter-channel dependency of features. The proposed detection model has the best detection performance in the designed wind turbine dataset.