Only model transition data and valid trips should be considered
Dealing with outlier values: impute / remove the whole trip if cannot impute
Double check on mode (mode 2 are usually longer and can be checked with the propeller speed)
seperate normal and adversarial situations (i.e, rerouting, significant drop in velocity because of other craft, etc.)
Cluster with k=2, and check if modeling them separate work better or all together
To implement data loading, look for episodes training in time series (e.g. reinforcement learning approaches to get batch, use a flag, done, at the end of each episode)
@yim-fan please send today's session debrief, so I complete here.