def wpi(dataset, n_channels, seq_length):
"""
function to calculate a wavelet packet energy image
# Ding.2017: Energy-Fluctuated Multiscale Feature Learning With Deep ConvNet for Intelligent Spindle Bearing Fault Diagnosis
:param dataset: the raw data read from .csv or .npy file
:type dataset: ndarray
:param n_channels: number of channels
:type n_channels: int
:param seq_length: length of one data stream of a single sensor
:type seq_length: int
:return: the flattened images, tupel holding the image size
"""
level = 10 # choose an even number, paper uses 10
wavelet = 'db8' # Daubechies 8 used in paper
# wavelet = 'coif3' # Daubechies 8 used in paper
order = "natural" # other option is "freq"
clip_energy = 1 # threshold to clip the calculated energy features (negative and positive)
num_samples = dataset.shape[0]
img_size = np.power(2, level // 2)
size_flat = img_size * img_size
pics = np.zeros([num_samples, n_channels * size_flat]) # initialize array
wp_image = np.zeros([img_size, img_size])
for sample in range(num_samples): # loop over all samples
for ch in range(n_channels): # loop over all channels
# Construct wavelet packet tree for signal of one channel
wp = WaveletPacket(dataset[sample][ch * seq_length:(ch + 1) *
seq_length],
wavelet,
'symmetric',
maxlevel=level)
nodes = wp.get_level(
level, order=order
) # !required! access the tree to populate it (might be a bug of the library?)
i = 0
# loop through the tree (ordered from aaa..a to ddd..d)
for node in wp.get_leaf_nodes():
# use only the coefficients from node (i, p) with p = 0..(2^i-1), i.e. set all other coefficients to zero
new_wp = WaveletPacket(None, wavelet, 'symmetric', level)
new_wp[node.path] = wp[node.path].data
# get the reconstruction coefficients (length 2^i)
reconst = np.asarray(new_wp.reconstruct(update=False))
# phase shift --> arrange energy features calculated as the squared sum over the reconstruction coefficients
wp_image[i % img_size,
i // img_size] = np.sum(np.multiply(reconst, reconst),
axis=-1)
# remove node from wp tree
del new_wp[node.path]
i += 1
# (!) THP modification (!), clip the wpi to fixed range: especially the approximation coefficients hold a lot of energy which
# scales very differently
wp_image = np.clip(wp_image, -clip_energy, clip_energy)
# collect all pictures, shape (samples, ch1+ch2..)
pics[sample][ch * size_flat:(ch + 1) * size_flat] = np.reshape(
wp_image, [1, size_flat])
return pics, [img_size, img_size]
#!/usr/bin/env python
# -*- coding: utf-8 -*-
from pywt import WaveletPacket
wp = WaveletPacket(range(16), 'db2', maxlevel=3)
print [node.path for node in wp.get_leaf_nodes(decompose=False)]
print [node.path for node in wp.get_leaf_nodes(decompose=True)]
coeffs = [(node.path, node.data) for node in wp.get_leaf_nodes(decompose=True)]
print coeffs
wp2 = WaveletPacket(None, 'db2', maxlevel=3)
for path, data in coeffs:
wp2[path] = data
#print wp["a"]
print [node.path for node in wp2.get_leaf_nodes(decompose=False)]
print wp2.reconstruct()
#!/usr/bin/env python
# -*- coding: utf-8 -*-
from pywt import WaveletPacket
wp = WaveletPacket(range(16), 'db2', maxlevel=3)
print[node.path for node in wp.get_leaf_nodes(decompose=False)]
print[node.path for node in wp.get_leaf_nodes(decompose=True)]
coeffs = [(node.path, node.data) for node in wp.get_leaf_nodes(decompose=True)]
print coeffs
wp2 = WaveletPacket(None, 'db2', maxlevel=3)
for path, data in coeffs:
wp2[path] = data
#print wp["a"]
print[node.path for node in wp2.get_leaf_nodes(decompose=False)]
print wp2.reconstruct()
