This python library allows detecting patterns in n-dimensional time domain signals.
Colors represent different labels over time.
The library uses the Continuous Wavelet Transform, similar to time-frequency analysis such as the fourier transform, to transform an n dimensional time domain signal into k*n dimension signals or k*n features per time sample. Then it trains a classifier using the continuous wavelet transform features.
Please cite our work!
- Numpy
- scikit-learn - optional if you are using your own classifier
For the Example (Example.ipynb)
You will also need:
- Pydblite
- Matplotlib
Time series signals are transformed into the time-frequency domain using the Continuous Wavelet Transform. This turns a n-channel signal into a k*n channel signal where k is the number of scales\frequencies of interest, providing k*n features at every sample time. The k*n features are used to describe the local behavior of the signal. These descriptions along with the class label are used to train a ML classifier.
cwt_learner is the main software package.
wavelet_feature_engineering.py provides the interface to
apply machine learning classifiers to the transformed data. The
default ML classifier is a sklearn.neural_network. MLPClassifier
but this can be changed to a different classifier.
wavelet_feature_transform.py transforms the time domain multi-channel
signal into the continuous wavelet transform. This function
interfaces with the wavelets.py library. Additional transformations
such as extracting magnitude and phase is possible. The default wavelet
is the morelet wavelet but other wavelets are also possible.