features¶

Time-domain features.

Notation:
• $$x_i$$ : value of a signal at time index $$i$$
• $$N$$ : length of the signal
axopy.features.time.integrated_emg(x, axis=-1, keepdims=False)[source]

Sum over the rectified signal.

$\text{IEMG} = \sum_{i=1}^{N} | x_{i} |$
Parameters: x (ndarray) – Input data. Use the axis argument to specify the “time axis”. axis (int, optional) – The axis to compute the feature along. By default, it is computed along rows, so the input is assumed to be shape (n_channels, n_samples). keepdims (bool, optional) – Whether or not to keep the dimensionality of the input. That is, if the input is 2D, the output will be 2D even if a dimension collapses to size 1. y – IEMG of each channel. ndarray, shape (n_channels,)
axopy.features.time.logvar(x, axis=-1, keepdims=False)[source]

Log of the variance of the signal.

$\text{log-var} = \log \left( \frac{1}{N} \sum_{i=1}^{N} \left(x_i - \mu \right)^2 \right)$

For electrophysiological signals that are mean-zero, this is the log of the mean square value, making it similar to root_mean_square() but scaling differently (slower) with $$x$$.

For EMG data recorded from forearm muscles, log-var has been found to relate to wrist angle fairly linearly [1]_.

Note: base-10 logarithm is used, though the base is not specified in [1]_.

Parameters: x (ndarray) – Input data. Use the axis argument to specify the “time axis”. axis (int, optional) – The axis to compute the feature along. By default, it is computed along rows, so the input is assumed to be shape (n_channels, n_samples). keepdims (bool, optional) – Whether or not to keep the dimensionality of the input. That is, if the input is 2D, the output will be 2D even if a dimension collapses to size 1. y – log-var of each channel. ndarray, shape (n_channels,)

References

 [1] J. M. Hahne, F. Bießmann, N. Jiang, H. Rehbaum, D. Farina, F. C. Meinecke, K.-R. Müller, and L. C. Parra, “Linear and Nonlinear Regression Techniques for Simultaneous and Proportional Myoelectric Control,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 22, no. 2, pp. 269–279, 2014.
axopy.features.time.mean_absolute_value(x, weights='mav', axis=-1, keepdims=False)[source]

Computes the mean absolute value (MAV) of each signal.

Mean absolute value is a popular feature for obtaining amplitude information from EMG, especially in gesture classification contexts [1]_.

There is an optional windowing function applied to the rectified signal, described as MAV1 and MAV2 in some references. A custom window can also be used. The general definition is given as:

$\text{MAV} = \frac{1}{N} \sum_{i=1}^{N} w_i |x_i|$

Normal MAV does not use a windowing function, equivalent to setting all $$w_i = 1$$.

MAV1 refers to a rectangular window which de-emphasizes the beginning and ending of an input window. The first quarter of the input samples receive a weight of 0.5, the middle half of the input samples receive a weight of 1, and the final quarter recieves a weight of 0.5:

$\begin{split}w_i = \begin{cases} 1, & \frac{N}{4} \leq i \leq \frac{3N}{4} \\ 0.5, & \text{otherwise} \end{cases}\end{split}$

MAV2 uses a similar window structure to MAV1 (i.e. broken into first quarter, middle half, and final quarter), but the window is trapezoidal in shape, ramping from 0 to 1 over the first quarter and from 1 to 0 over the last quarter:

$\begin{split}w_i = \begin{cases} 1, & \frac{N}{4} \leq i \leq \frac{3N}{4} \\ \frac{4i}{N}, & i < \frac{N}{4} \\ \frac{4(i - N)}{N}, & i > \frac{3N}{4} \end{cases}\end{split}$
Parameters: x (ndarray) – Input data. Use the axis argument to specify the “time axis”. weights (str or ndarray, optional) – Weights to use. Possible values: ’mav’ : all samples in the signal are weighted equally (default). ’mav1’ : rectangular window with the middle half of the signal receiving unit weight and the first and last quarters of the signal receiving half weight. ’mav2’ : similar to ‘mav1’, but weights on the first and last quarters increase and decrease between 0 and 1 respectively, forming a trapezoidal window. [ndarray] : user-supplied weights to apply. Must be a 1D array with the same length as the signals received in the compute method. axis (int, optional) – The axis to compute the feature along. By default, it is computed along rows, so the input is assumed to be shape (n_channels, n_samples). keepdims (bool, optional) – Whether or not to keep the dimensionality of the input. That is, if the input is 2D, the output will be 2D even if a dimension collapses to size 1. y – MAV of each channel. ndarray, shape (n_channels,)

axopy.features.util.inverted_t_window()
Generates the window for MAV1
axopy.features.util.trapezoidal_window()
Generates the window for MAV2

References

 [1] B. Hudgins, P. Parker, and R. N. Scott, “A New Strategy for Multifunction Myoelectric Control,” IEEE Transactions on Biomedical Engineering, vol. 40, no. 1, pp. 82-94, 1993.
 [2] A. Phinyomark, P. Phukpattaranont, and C. Limsakul, “Feature Reduction and Selection for EMG Signal Classification,” Expert Systems with Applications, vol. 39, no. 8, pp. 7420-7431, 2012.
axopy.features.time.root_mean_square(x, axis=-1, keepdims=False)[source]

Computes the root mean square of each signal.

RMS is a commonly used feature for extracting amplitude information from physiological signals.

$\text{RMS} = \sqrt{\frac{1}{N} \sum_{i=1}^N x_i^2}$
Parameters: x (ndarray) – Input data. Use the axis argument to specify the “time axis”. axis (int, optional) – The axis to compute the feature along. By default, it is computed along rows, so the input is assumed to be shape (n_channels, n_samples). keepdims (bool, optional) – Whether or not to keep the dimensionality of the input. That is, if the input is 2D, the output will be 2D even if a dimension collapses to size 1. y – RMS of each channel. ndarray, shape (n_channels,)
axopy.features.time.slope_sign_changes(x, threshold=0, axis=-1, keepdims=False)[source]

Computes the number of slope sign changes (SSC) of each signal.

A slope sign change occurs when the middle value of a group of three adjacent values in the signal is either greater than or less than both of the other two.

Parameters: x (ndarray) – Input data. Use the axis argument to specify the “time axis”. threshold (float, optional) – A threshold for discriminating true slope sign changes from those caused by low-level noise fluctuating about a specific value. By default, no threshold is used, so every slope sign change in the signal is counted. axis (int, optional) – The axis to compute the feature along. By default, it is computed along rows, so the input is assumed to be shape (n_channels, n_samples). keepdims (bool, optional) – Whether or not to keep the dimensionality of the input. That is, if the input is 2D, the output will be 2D even if a dimension collapses to size 1. y – SSC of each channel. ndarray, shape (n_channels,)

References

 [1] B. Hudgins, P. Parker, and R. N. Scott, “A New Strategy for Multifunction Myoelectric Control,” IEEE Transactions on Biomedical Engineering, vol. 40, no. 1, pp. 82-94, 1993.
axopy.features.time.waveform_length(x, axis=-1, keepdims=False)[source]

Computes the waveform length (WL) of each signal.

Waveform length is the sum of the absolute value of the deltas between adjacent values (in time) of the signal:

$\text{WL} = \sum_{i=1}^{N-1} | x_{i+1} - x_i |$
Parameters: x (ndarray) – Input data. Use the axis argument to specify the “time axis”. axis (int, optional) – The axis to compute the feature along. By default, it is computed along rows, so the input is assumed to be shape (n_channels, n_samples). keepdims (bool, optional) – Whether or not to keep the dimensionality of the input. That is, if the input is 2D, the output will be 2D even if a dimension collapses to size 1. y – WL of each channel. ndarray, shape (n_channels,)

References

 [1] B. Hudgins, P. Parker, and R. N. Scott, “A New Strategy for Multifunction Myoelectric Control,” IEEE Transactions on Biomedical Engineering, vol. 40, no. 1, pp. 82-94, 1993.
axopy.features.time.zero_crossings(x, threshold=0, axis=-1, keepdims=False)[source]

Computes the number of zero crossings (ZC) of each signal.

A zero crossing occurs when two adjacent values (in time) of the signal have opposite sign. A threshold is used to mitigate the effect of noise around zero. It is used as a measure of frequency information.

Parameters: x (ndarray) – Input data. Use the axis argument to specify the “time axis”. threshold (float, optional) – A threshold for discriminating true zero crossings from those caused by low-level noise situated about zero. By default, no threshold is used, so every sign change in the signal is counted. axis (int, optional) – The axis to compute the feature along. By default, it is computed along rows, so the input is assumed to be shape (n_channels, n_samples). keepdims (bool, optional) – Whether or not to keep the dimensionality of the input. That is, if the input is 2D, the output will be 2D even if a dimension collapses to size 1. y – ZC of each channel. ndarray, shape (n_channels,)

References

 [1] B. Hudgins, P. Parker, and R. N. Scott, “A New Strategy for Multifunction Myoelectric Control,” IEEE Transactions on Biomedical Engineering, vol. 40, no. 1, pp. 82-94, 1993.