Loss-of-control (LOC) is the main cause of crashes for drones. On-board prevention systems should be designed that require low computing power and memory. Data-driven techniques serve as a solution. This study proposes the use of recurrent neural networks (RNN) for LOC prediction. The aim is to identify which RNN model is most suitable and if this model can predict loss-of-control for changing aerodynamic characteristics, wind conditions, quadcopter types and LOC events. Real-life flight tests using a Tiny Whoop quadcopter were performed
where LOC was initiated by demanding a too high yaw rate of ±2000 deg/s. Using data from these failure runs, four RNN networks were trained: long short-term memory (LSTM), bidirectional LSTM (BiLSTM), LSTM preceded by a convolutional neural network (CNN-LSTM) and gated recurrent unit (GRU).
Only on-board sensor measurements are necessary for LOC prediction. The commanded rotor values provide clearest early warning signals for loss-of-control, as these values show saturation before LOC. Beside this, for optimal performance, an additional parameter should be used as well, which is either the
gyroscopic measurement or acceleration. All networks could predict LOC correctly and equally well, leading to no preference for one specific model type. Next to
this, when compensating for expected deviations in the prediction, the models can still be used for change in mass and different propellers. To draw conclusions
for flying in wind conditions and using different quadcopters, more research is necessary. Finally, the models can only be used to predict loss-of-control due to a too high yaw rate and the time to loss-of-control should be similar to those from the training set.