Temporal Machine Learning Payload Prediction for DJI Matrice 100 Quadcopter Drone Based on Tracking Data

This paper presents three different machine-learning techniques to predict the payloads of DJI Matrice 100 quadcopter drones. The tracking data is based on real-life experimentation that is provided as open source. The payloads are 0.0, 250, 500, and 750 grams. The Machine Learning techniques are LSTM, TCN, and GRU. The values of the tracks’ kinematics come from several different flights for the different loads. The tracks’ kinematics in addition to some flight parameters are used in training the three models. The temporal nature of the data values triggers the need for machine learning methods that use history/memory as part of inputs. The input variables went through the data reduction stage as some variables turned out to be less relevant than others for the prediction operation. The original data has more than 1400 records for every flight with more than 22 variable values. More than 270 flights were conducted with the 4 payloads. The flight/track data is accessed in parallel at every time stamp by the learning models and the model converges after a few epochs to the payload label. The training/validation/testing values show that the three models captured the predicted load efficiently. Comparisons between the three models as predictors for the carried payloads are presented and discussed in the paper. The TCN showed slight superiority over the other models