Quadrotor UAV blade damage detection based on improved Transformer

Due to the execution of various flying maneuvers of the Quadrotor Unmanned Aerial Vehicles (UAV), the variations in these actions lead to inconsistencies in local features within the flight monitoring data over the time series. And traditional methods are difficult to extract the fault features that diminishes the effectiveness of the UAV blade damage detection. In this paper, an UAV blade detection method based on the improved time series Transformer is proposed. A new network named Kolmogorov-Arnold Networks (KAN) is employed to replace the traditional Multi-Layer Perceptron (MLP) layer within the Transformer encoder layer. A fully connected module is then introduced for additional feature extraction, enhancing the model's ability to capture damage-related features more effectively. The experimental results demonstrate that the approach enables proficiently extraction of informative patterns from flight data, and the fault diagnosis accuracy reaches 98.9%, which verify how well the method performs.