Damage identification for multi-rotor drone using convolutional neural network technique

In contemporary society, multi-rotor drone have found extensive usage in various fields, such as agriculture, cargo transportation, and aerial photography. Damage to multi-rotor drone can compromise their safety and reduce efficiency. Therefore, early damage detection is crucial as it can prevent catastrophic failures and decrease the associated economic and human costs. At present, visual inspection is the primary method used for detecting damage in multi- rotor drone. However, this technique may not be entirely reliable in identifying minor faults that are difficult to discern with the naked eye. This study focuses on three experimental parts; firstly,to fabricate a multi-rotor drone as the research subject; secondly, to develop a vibration data acquisition device with MPU6050 and STM32 micro-controller,and thirdly,to identify the damage using machine learning techniques. Damage scenarios were set by releasing the bolts at different conjunction points of the multi-rotor body frame.Three damaged cases were set by releasing one bolt at arm conjunction, two bolts at arm conjunction and one bolt at motor conjunction, respectively. The first case (undamaged) is considered as the reference. Any change in structure can reflect in a vibration signal. Three axes vibration data were acquired under different conditions,for the sake of safety,the UAV was conducted under the ground with a idol motor speed. After the data collection,the data preprocessing techniques linear interpolation method Laida criterion were adopted to process the missing data and inconsistent data.For damage identification,three machine learning techniques, including decision tree, random forest, K-Nearest- Neighbours (KNN) were adopted to identify the damage for multi-rotor drone and finally with the accuracy of 68.74%, 67.96%, 91.71%, respectively. Then, Convolutional Neural Networks (CNN), as the state-of-the-art machine learning technique also called deep learning was proposed and achieved outstanding success with 100% accuracy for damage identification. It is important to consider the parameter used in the CNN,so,in this research,the parameter used in the CNN,including sample length, convolution kernel, number of convolutional layer,activation function,batch-size,dropout,learning rate were analyzed by Python platform and the best parameter were selected.In summary, machine learning techniques can effectively detect damage for multi- rotor drone, however,CNN technique convolutional neural network possesses superior feature extraction capability and classification accuracy compared to traditional machine learning techniques.

Text FK 2023 3 IR.pdf Download (1MB)

Actions (login required)

View Item