Changelog#

Version 0.3.0#

  • New features
    • Added support for Python 3.14

    • Added tsfel.__version__

  • Changes
    • TSFEL now requires Python 3.11 or newer. On Python 3.8–3.10, pip keeps installing TSFEL 0.2.0

    • Raised the minimum dependency versions: NumPy 1.26, SciPy 1.11.2, pandas 2.1.1, scikit-learn 1.3.1, PyWavelets 1.9, statsmodels 0.14 and requests 2.32.4. PyWavelets 1.9 gives identical wavelet features on NumPy 1.x and 2.x

    • setuptools is no longer installed as a runtime dependency

    • Some feature values differ from 0.2.0 because of the bugfixes below: auc, entropy, wavelet_entropy and, for synthetic signals, the spectral features based on the FFT

    • Progress bars use tqdm and show what is being counted (windows or features), and elapsed and remaining time. They are written to stderr. In Jupyter, install ipywidgets for a widget bar

    • IPython is no longer a dependency

  • Removed
    • The undocumented progress_bar_terminal, progress_bar_notebook and display_progress_bar, and tsfel.get_ipython, tsfel.HTML and tsfel.display, which were exposed by accident

  • Bugfixes
    • auc underestimated the area of signals that cross zero

    • entropy depended on floating-point round-off: computed signals with repeated values, such as sine waves, gave different results on Linux, macOS and Windows. Values that differ only by round-off are now counted as the same value

    • wavelet_entropy returned 0 for any signal whose values sum to zero, such as [1, -1, 2, -2]. It now returns 0 only for an all-zero signal

    • Spectral features ignore FFT round-off. A pure tone had a spectral skewness of about 107 and 161 spectral peaks, values that also changed between NumPy 1.x and 2.x; it now has zero spread, skewness and kurtosis, and one peak. On real signals, only spectral_entropy changes, by about 0.03%

  • Documentation
    • Migrated the documentation to the PyData Sphinx theme, with light and dark modes (#186)

    • Rewrote the personalised features guide as three steps (write, register and extract), with a working example

  • Development
    • pyproject.toml is now the only source of packaging metadata; setup.py and requirements/*.txt were removed. Set up a development environment with pip install -e . --group dev (pip 25.1 or newer)

    • CI tests Python 3.11–3.14 on Linux, macOS and Windows, the lowest supported dependency versions, the dataset loaders (in a separate job) and the pre-commit hooks

    • GitHub Actions are pinned to commit SHAs and kept up to date by Dependabot

    • The wavelet and spectral tests now check values derived from theory instead of values recorded from earlier runs

Version 0.2.0#

  • New features
    • Added support for Python 3.13 (#176)

  • Changes
    • Replaced numpy.trapz with scipy.integrate.trapezoid for consistent behavior and compatibility with NumPy < 2 (#178)

Version 0.1.9#

  • Changes
    • Replaced the Histogram feature with Histogram mode (#167)

  • Improvements
    • The correlated_features method now supports returning a filtered feature vector (#163)

  • Documentation
    • Set up a Slack community invite (#164)

    • Enhanced documentation for the time_series_feature_extraction and get_features_by_domain functions

    • Updated the main example notebook on Human Activity Recognition (#159)

Version 0.1.8#

  • New Features
    • Added a new Datasets module with convenient methods to load single-problem datasets (#156)

    • Improved the get_features_by_domain method, allowing easier selection of multiple feature domain combinations

  • Improvements
    • Significantly reduced the computational time for the LPCC feature (#153)

    • Resolved deprecation issues with SciPy Wavelets by switching to PyWavelets for features that rely on wavelets (#147)

    • Renamed the fft_mean_coefficient feature to spectrogram_mean_coefficient for descriptive correctness (#145)

  • Bugfixes
    • Fixed a bug causing a circular import issue (#154)

    • Fixed a ResourceWarning when loading the feature configuration file (#152)

    • Removed the use of eval (#150)

  • Documentation
    • Major documentation updates, including detailed explanations of the expected input and output data formats

Version 0.1.7#

  • New features
    • Implemented the Lempel-Ziv-Complexity in the temporal domain (#146)

    • Added the fractal domain with the following features (#144):
      • Detrended fluctuation analysis (DFA)

      • Higuchi fractal dimension

      • Hurst exponent

      • Maximum fractal length

      • Multiscale entropy (MSE)

      • Petrosian fractal dimension

  • Changes
    • Changed the autocorrelation logic. It now measures the first lag below (1/e) from the ACF (#142).

Version 0.1.6#

  • Changes
    • Feature total energy changed name to average power

    • Features peak to peak, absolute energy and entropy are now classified as statistical

  • Bugfixes
    • Fixed a bug on numpy bool usage (#133)

    • Fixed a bug on features’ header names

  • Improvements
    • Correlated features are now computed using absolute value

    • Unit tests improvements

    • Refactoring of some code sections and overall improved stability

Version 0.1.5#

  • Bugfixes - Fixed a bug on scipy function median_absolute_deviation to median_abs_deviation (#128) - Fixed on pandas function df.append to pd.concat (#120)

Version 0.1.4#

  • Bugfixes
    • Fixed a bug on the progress bar not being displayed if the signal is passed already divided into windows (#49)

    • Fixed a bug on the distance feature (#54)

    • Fixed a bug raising zero division in the ECDF slope feature (#57)

    • Fixed a bug when adding customised features using the JSON

    • Fixed a bug on LPC was returning inconsistent values (#58)

    • Fixed a bug on normalised autocorrelation (#64)

  • Improvements
    • Refactoring of some code sections and overall improved stability

    • The documentation has been improved and a FAQ section was created

    • The window_splitter parameter is now deprecated. If the user selected a window_size it is assumed that the signal must be divided into windows.

    • Unit tests improvements

  • New features
    • Added to return the size of the feature vector from the configuration dictionary (#50)

Version 0.1.3#

  • Bugfixes
    • Bug fixes on computational complexity calculation (#15)

    • Fixed an error on lpcc feature (#38)

    • Removed entropy warning (#38)

  • Improvements
    • Code cleaning on (TSFEL_HAR_Example.ipynb)

    • ecdf code cleaning and computational optimization

    • Overlap value is now optional and set to default as 0

    • Unit test improvements

    • Nomenclature review of peak-related features

  • New features:
    • Added new tutorials based on Jupyter notebooks (#19)

    • Added progress bar during feature extraction (#16)

    • Implemented multiprocessing. The n_jobs kwarg selects the number of CPUs to be scheduled (#30)

    • Added the neighbourhood_peaks feature

Version 0.1.1#

  • Added new features
    • Empirical cumulative distribution function

    • Empirical cumulative distribution function percentile

    • Empirical cumulative distribution function slope

    • Empirical cumulative distribution function percentile count

    • Spectral entropy

    • Wavelet entropy

    • Wavelet absolute mean

    • Wavelet standard deviation

    • Wavelet variance

    • Wavelet energy

  • Minor fixes for Google Colab

Version 0.1.0#

  • Release of TSFEL with documentation