Effective Fault Diagnosis for a Quadrotor Helicopter: A Lightweight Transformer With Selective Patches and Channels Modules Method

Quadrotor helicopters have been widely applied in numerous fields and are increasingly attracting attention in many application areas. It is challenging to ensure the safety and reliability of quadrotor helicopter when faults occur in sensors or actuators, which may lead to catastrophic crashes. Recently, Transformer and its variants demonstrate powerful feature extraction capabilities, while a fault diagnosis (FD) model based on Transformer usually has a high demand for parameters and computations which limits its applications. To address this issue, this article proposes a novel lightweight Transformer with selective patches and channels modules (SPCFormer) method for quadrotor helicopter FD. First, the flight data is split into nonoverlapping patches for each channel. A selective patches module with a lightweight attention architecture is designed to extract critical local feature information from patches and mitigate multichannel coupling effects. Second, the selective channel attention is developed to form an attention vector rather than a matrix. This mechanism is integrated into the selective channels module to capture important global channel features while reducing the complexity of the model. Finally, a high-fidelity quadrotor helicopter fault simulator is developed to simulate different types of faults (i.e., actuator fault and sensor fault) under three different flight statuses and no extra sensors. The effectiveness of the proposed FD method is verified through the cross-validation on the above developed software-in-the-loop (SIL) and hardware-in-the-loop (HIL) simulators.