UAV Fault Diagnosis for Limited Data and Multi Task Scenarios

With the increasing system complexity and accident rate, Unmanned Aerial Vehicle (UAV) fault diagnosis has received widespread attention. Compared with the traditional neural network-based classification problems, UAV fault diagnosis is characterized by limited fault data and different fault scenarios that are difficult to obtain in advance, which brings great challenges to the construction of diagnosis models. To address the above problems, this article proposes a UAV fault diagnosis algorithm (UAV-FDA) from the perspective of limited data and model adaption. First, for limited data, multi-channel representation learning satisfying the bijection property is proposed for data enhancement. Second, for different fault diagnosis scenarios, a model adaptive algorithm without human involvement is proposed to automatically adjust the model according to different task characteristics. Based on real UAV flight data, in six diagnosis scenarios, UAV-FDA is compared with other algorithms in terms of diagnosis accuracy, confusion matrix, parameter number and floating-point operations. Based on independent repeated experiments, its accuracy ranked 1/42 in 5 scenarios (96.85%, 97.02%, 96.39%, 94.34% and 95.93%) and 3/42 in one scenario (92.67%). The standard deviation of the accuracy does not exceed 1.12%. The parameter number and operations are ranked as 7/42 and 16/42. It shows the advantages in terms of accuracy, stability, parameter number and operations.