Source code for bruges.attribute.spectrogram

# -*- coding: utf-8 -*-

:copyright: 2015 Agile Geoscience
:license: Apache 2.0
import numpy as np
from scipy.fftpack import fft
from scipy.signal import get_window

from bruges.util import next_pow2

[docs]def spectrogram(data, window_length, dt=1.0, window_type='boxcar', overlap=0.5, normalize=False, zero_padding=0): """ Calculates a spectrogram using windowed STFTs. Args: data (array): 1D numpy array to process into spectra. window_length (float): The length of the window in seconds if dt is set, otherwise in samples. Will zero pad to the closest power of two. window_type: A string specifying the type of window to use for the STFT. The same input as scipy.signal.get_window dt (float or int): The time sample interval of the trace. Defaults to 1, which allows window_length to be in seconds. overlap (float): The fractional overlap for each window. A value of 0 uses no redudant data, a value of 1 slides the STFT window one sample at a time. Defaults to 0.5 normalize (bool): Normalizes the each spectral slice to have unity energy. zero_padding (float or int): The amount of zero padding to when creating the spectra. Returns ------- out : A spectrogram of the data ([time, freq]). (2D array for 1D input) See Also -------- spectraldecomp : Spectral decomposition """ # Make the base window window_n = int(np.floor(window_length / dt)) pad = int(np.floor(zero_padding / dt)) window = get_window(window_type, window_n) # Calculate how many STFTs we need to do. stride = int(window.size * (1 - overlap) + 1) n_windows = int(np.ceil((data.size - window.size) / stride) + 1) # Pad window to power of 2 padded_window = np.zeros(next_pow2(window.size+pad)) # Init the output array output = np.zeros([n_windows, int(padded_window.size // 2)]) # Do the loop for i in range(int(n_windows)): start = int(i * stride) end = start + int(window.size) # Do not compute last samples if we don't have a full window if (end > data.size-1): break padded_window[0:window.size] = window*data[start:end] spect = (2. * np.absolute(fft(padded_window)) / window.size)[0:int(padded_window.size // 2)] if normalize: output[i, :] = spect / np.sum(spect**2) else: output[i, :] = spect return output