A helper library to perform essential manipulations on time-series EMG data of NinaPro DB2
This library aims at easing the preprocessing, amd training of the EMG signals from the NinaPro DB2 database. The working of all the functions is showcased in NinaPro_Utility.ipynb file.
The library provides following functions :
Download the data from the NinaPro official website and save the data in a folder. The get_data() requires the folder path and the file name as the input arguements.
It returns a dataframe with the EMG signals, labels and the repetition number.
normalise() scales the data using the normalise() uses the StandardScaler() from scikit-learn to normalize the data. It fits on training reps only and then transforms the whole data (excluding stimulus and repetition ofcourse).
Before normalising data :
After normalization :
Notice the scale
Sometimes, it is required that the signal is filtered with low noise or high noise frequencies. filter_data uses Butterworth filter to filter the data. It requires the cutoff frequency, the butterworth order, and the type of filter (btype is one of lowpass, highpass, bandpass).
The bandpass filter requires the f value to be a tuple or a list containing lower cutoff frequency and higher cutoff frequency.
After applying lowpass filter:
After applying bandpass filter
After applying highpass filter
This function rectifies the signals and converts all the negative values to positive values by simply using the absolute value.
The rectifies signal looks like this :
windowing() is used to augment the data. The function requires the following arguements : data, reps, gestures, win_len, win_stride.
data= Pandas dataframe just like returned by any of the above functionsreps= Repetitions that you want to use for windowinggestures= The gesture movements that you wish to classifywin_len= (Length of window in milisecond) x 2. For example, for a window of 300ms, use 600 as thewin_lensince the sampling frequency of signal is 2000Hz.win_stride= (Length of stride in milisecond) x 2. For example, for a stride of 10ms, use 20 as thewin_stridesince the sampling frequency of signal is 2000Hz.
To understand what windowing function does, please refer to below image (source : Google)
For multiclass classification, we need the labels to be represented in one-hot representation.
get_categorical() helps in converting the integer labels to one-hot representation.
It takes the following arguements:
saved_model= The model you have already trainedX_testy_test
returns a confision matrix.