Detection of quadcopter propeller failure by machine learning methods

The paper presents a study of options for detecting a failure or defect in the propeller of an unmanned aircraft system (quadcopter) using machine learning methods. An original accuracy evaluation of the known algorithms using in practice the data obtained from the quadcopter in its flight conditions is performed. The proposed method is based on the classification of three propeller states (serviceable propellers, one propeller artificially deformed, one propeller broken) using machine learning algorithms. The input information is the data obtained from the quadcopter measuring system in real time: speed, acceleration and rotation angle relative to three axes. For the correct work of the presented algorithm, data was preprocessed with division into time intervals and applying to the obtained intervals the fast Fourier transform. Based on the processed data, machine learning algorithms were trained using the reference vector method, k-nearest neighbor algorithm, decision tree algorithm, and multilayer perceptron. The obtained accuracy values of the proposed methods are compared. It is shown that the application of machine learning methods can detect and classify the propeller states with an accuracy of up to 96 %. The best result is achieved using the decision tree algorithm. The results of the study can be of practical importance for real-time systems to detect propeller defect and breakage for unmanned aerial vehicles. It is possible to predict with high accuracy the propeller wear; it is possible to improve the stability and safety of the flight.