Hello,
I am working on a machine learning project that maps EMG / IMU (input) to predict joint angles (output).
However, as I split my data into training and testing sets, my X_train and Y_train need to have the same number of rows to map inputs to outputs. Since EMG was sampled at 2000 Hz whereas motion capture was sampled at 120 Hz, should I upsample my joint angle data or should I downsample the EMG data? My intuition is that it is 'better' to downsample because in real applications, you would not have 'fake' data added to the system. Is that a correct assumption or does upsampling vs. downsampling not matter in this situation?
For example, if I have 10 seconds of walking data, I would want my training shapes as:
Shape = (# rows, # columns)
Downsample scenario:
X_train.shape = (120Hz * 10sec, # of EMG sensors = 5) = (1200, 6)
Y_train.shape = (1200, 1)
OR
Upsample scenario:
X_train.shape = (2000Hz * 10 sec, # of EMG sensors = 5) = (20000, 5)
Y_train.shape = (20000, 1)
Hi David,
It actually depends on what you want to do after up/down-sampling. I guess that as you are actually working with physio/physical signals. You don't need to look at high freq info. So, I would go for downsampling. I always prefer to downsample when I have such high-frequency sampling rates. That would alleviate some time complexity post-processing, and also, as you mentioned, it would be closer to that of wearable/constrained applications. That said without knowing exactly the particularities of your experiment. By the way, when you upsample the data, you are not faking anything, upsampling is actually applied with physio signals in real applications (for instance when working with HRV extraction, etc).
Hi David,
I will go for upsampling. EMG signals are acquired with such high sampling frequencies for a reason, so you may lose some important EMG features by downsampling,
Hakim
Thank you Hakim Mecheri for your reply. In this case, since EMG information is important for predicting joint angles, I will upsample the target labels (joint angles). In fact, when I originally downsampled the EMG inputs (X_train and X_test), the RMSE for predicting joint angles (Y_train and Y_test) increased considerably because the relevant EMG information was lost. Therefore, preserving relevant EMG features seems to do a better job at mapping EMG signals to joint angles.
- Badges
- Science topic
- Similar topics
- Geoscience
- Asia
- India