line_tracking_of_fish_movement/eventdetection.py

1439 lines
52 KiB
Python

"""
Detect and handle peaks and troughs as well as threshold crossings in data arrays.
## Peak detection
- `detect_peaks()`: peak and trough detection with a relative threshold.
- `peak_width()`: compute width of each peak.
- `peak_size_width()`: compute for each peak its size and width.
## Threshold crossings
- `threshold_crossings()`: detect crossings of an absolute threshold.
- `threshold_crossing_times()`: compute times of threshold crossings by linear interpolation.
## Event manipulation
- `trim()`: make the list of peaks and troughs the same length.
- `trim_to_peak()`: ensure that the peak is first.
- `trim_closest()`: ensure that peaks minus troughs is smallest.
- `merge_events()`: Merge events if they are closer than a minimum distance.
- `remove_events()`: Remove events that are too short or too long.
- `widen_events()`: Enlarge events on both sides without overlap.
## Threshold estimation
- `std_threshold()`: estimate detection threshold based on the standard deviation.
- `median_std_threshold()`: estimate detection threshold based on the median standard deviation of data snippets.
- `hist_threshold()`: esimate detection threshold based on a histogram of the data.
- `minmax_threshold()`: estimate detection threshold based on maximum minus minimum value.
- `percentile_threshold()`: estimate detection threshold based on interpercentile range.
## Snippets
- `snippets()`: cut out data snippets around a list of indices.
## Peak detection with dynamic threshold:
- `detect_dynamic_peaks()`: peak and trough detection with a dynamically adapted threshold.
- `accept_peak_size_threshold()`: adapt the dection threshold to the size of the detected peaks.
"""
import sys
import numpy as np
try:
from numba import jit, int64
index_type = int64
except ImportError:
def jit(*args, **kwargs):
def decorator_jit(func):
return func
return decorator_jit
index_type = np.int
def detect_peaks(data, threshold):
"""
Detect peaks and troughs using a relative threshold according to
Bryan S. Todd and David C. Andrews (1999): The identification of peaks in physiological signals.
Computers and Biomedical Research 32, 322-335.
Parameters
----------
data: array
An 1-D array of input data where peaks are detected.
threshold: float or array
A positive number or array of numbers setting the detection threshold,
i.e. the minimum distance between peaks and troughs.
In case of an array make sure that the threshold does not change faster
than the expected intervals between peaks and troughs.
Returns
-------
peak_array: array of ints
A list of indices of detected peaks.
trough_array: array of ints
A list of indices of detected troughs.
Raises
------
ValueError: If `threshold <= 0`.
IndexError: If `data` and `threshold` arrays differ in length.
"""
if np.isscalar(threshold):
if threshold <= 0:
raise ValueError('input argument threshold must be positive!')
return detect_peaks_flat(data, threshold)
else:
if len(data) != len(threshold):
raise IndexError('input arrays data and threshold must have same length!')
if np.min(threshold) <= 0:
raise ValueError('input argument threshold must be positive!')
return detect_peaks_array(data, threshold)
@jit(nopython=True)
def detect_peaks_flat(data, thresh):
"""
Detect peaks and troughs using a fixed, relative threshold.
Parameters
----------
data: array
An 1-D array of input data where peaks are detected.
threshold: float
A positive number setting the detection threshold,
i.e. the minimum distance between peaks and troughs.
Returns
-------
peak_array: array of ints
A list of indices of detected peaks.
trough_array: array of ints
A list of indices of detected troughs.
"""
# initialize:
direction, min_inx, max_inx, pi, ti = 0, 0, 0, 0, 0
min_value, max_value = data[0], data[0]
peaks_list = np.zeros(len(data)//2,dtype=index_type)
troughs_list = np.zeros(len(data)//2,dtype=index_type)
# loop through the data:
for index, value in enumerate(data):
peaks_list, troughs_list, pi, ti, direction, min_inx, max_inx, min_value, max_value = analyse_for_peaks(index, value, thresh, peaks_list,
troughs_list, pi, ti, direction, min_inx, max_inx, min_value, max_value)
return peaks_list[:pi], troughs_list[:ti]
@jit(nopython=True)
def detect_peaks_array(data, threshold):
"""
Detect peaks and troughs using a variable relative threshold.
Parameters
----------
data: array
An 1-D array of input data where peaks are detected.
threshold: array
A positive array of numbers setting the detection threshold,
i.e. the minimum distance between peaks and troughs.
Returns
-------
peak_array: array of ints
A list of indices of detected peaks.
trough_array: array of ints
A list of indices of detected troughs.
"""
# initialize:
direction, min_inx, max_inx, pi, ti = 0, 0, 0, 0, 0
min_value, max_value = data[0], data[0]
peaks_list = np.zeros(len(data)//2,dtype=index_type)
troughs_list = np.zeros(len(data)//2,dtype=index_type)
# loop through the data:
for index, value in enumerate(data):
thresh = threshold[index]
peaks_list, troughs_list, pi, ti, direction, min_inx, max_inx, min_value, max_value = analyse_for_peaks(index, value, thresh, peaks_list,
troughs_list, pi, ti, direction, min_inx, max_inx, min_value, max_value)
return peaks_list[:pi], troughs_list[:ti]
@jit(nopython=True)
def analyse_for_peaks(index, value, thresh, peaks_list, troughs_list, pi, ti, direction, min_inx, max_inx, min_value, max_value):
"""
Detect if the current datapoint is a peak or threshold and add it to the existing peak or trough list. Method is taken from:
Bryan S. Todd and David C. Andrews (1999): The identification of peaks in physiological signals.
Computers and Biomedical Research 32, 322-335.
Parameters
----------
index: int
Index of current datapoint.
value: float
Value of current datapoint.
thresh: float
Peak/trough threshold for current datapoint.
peaks_list: numpy array
Numpy array with peak indices prior to current datapoint.
troughs_list: numpy array
Numpy array with trough indices prior to current datapoint.
pi: int
Index of next peak in peaks_list.
ti: int
Index of next trough in troughs_list.
direction: int
Direction of data derivative, e.g. +1 for up and -1 for down.
min_inx: int
Index of the minimum value in the current direction.
max_inx: int
Index of the maximum value in the current direction.
min_value: float
Minimum value in the current direction.
max_value: float
Maximum value in the current direction.
Returns
-------
peaks_list: numpy array
Numpy array with peak indices including current datapoint.
troughs_list: numpy array
Numpy array with trough indices including current datapoint.
pi: int
Index of next peak in peaks_list.
ti: int
Index of next trough in troughs_list.
direction: int
Direction of data derivative, e.g. +1 for up and -1 for down.
min_inx: int
Index of the minimum value in the current direction.
max_inx: int
Index of the maximum value in the current direction.
min_value: float
Minimum value in the current direction.
max_value: float
Maximum value in the current direction.
"""
# rising?
if direction > 0:
if value > max_value:
# update maximum element:
max_inx = index
max_value = value
# otherwise, if the new value is falling below
# the maximum value minus the threshold:
# the maximum is a peak!
elif value <= max_value - thresh:
peaks_list[pi] = max_inx
pi = pi + 1
# change direction:
direction = -1
# store minimum element:
min_inx = index
min_value = value
# falling?
elif direction < 0:
if value < min_value:
# update minimum element:
min_inx = index
min_value = value
# otherwise, if the new value is rising above
# the minimum value plus the threshold:
# the minimum is a trough!
elif value >= min_value + thresh:
troughs_list[ti] = min_inx
ti = ti + 1
# change direction:
direction = +1
# store maximum element:
max_inx = index
max_value = value
# don't know direction yet:
else:
if value <= max_value - thresh:
direction = -1 # falling
elif value >= min_value + thresh:
direction = 1 # rising
if value > max_value:
# update maximum element:
max_inx = index
max_value = value
elif value < min_value:
# update minimum element:
min_inx = index
min_value = value
return peaks_list, troughs_list, pi, ti ,direction, min_inx, max_inx,min_value, max_value
def detect_peaks_fast(data, threshold):
"""Experimental. Try to make algorithm faster.
Yeah, this is more than three times as fast as `detect_peaks()` with the for loops!
"""
peaks_list = []
troughs_list = []
# initialize:
max_value = np.maximum.accumulate(data)
min_value = np.minimum.accumulate(data)
falling_idx = np.where(data <= max_value - threshold)[0]
raising_idx = np.where(data >= min_value + threshold)[0]
direction = 1
if len(falling_idx) > 0 and (len(raising_idx) == 0 or falling_idx[0] < raising_idx[0]):
direction = -1
index = falling_idx[0]
else:
index = raising_idx[0]
# find peaks and troughs:
while index < len(data):
if direction > 0:
max_value = np.maximum.accumulate(data[index:])
idx = np.argmax(data[index:] <= max_value - threshold)
if data[index+idx] <= max_value[idx] - threshold:
indices = np.where(max_value[:idx] != max_value[idx])[0]
if len(indices) > 0:
index += indices[-1] + 1
peaks_list.append(index)
direction = -1
continue
else:
min_value = np.minimum.accumulate(data[index:])
idx = np.argmax(data[index:] >= min_value + threshold)
if data[index+idx] >= min_value[idx] + threshold:
indices = np.where(min_value[:idx] != min_value[idx])[0]
if len(indices) > 0:
index += indices[-1] + 1
troughs_list.append(index)
direction = +1
continue
break
return np.asarray(peaks_list), np.asarray(troughs_list)
def peak_width(time, data, peak_indices, trough_indices,
peak_frac=0.5, base='max'):
"""
Width of each peak.
Peak width is computed from interpolated threshold crossings at
`peak_frac` hieght of each peak.
Parameters
----------
time: array
Time, must not be `None`.
data: array
The data with the peaks.
peak_indices: array
Indices of the peaks.
trough_indices: array
Indices of corresponding troughs.
peak_frac: float
Fraction of peak height where its width is measured.
base: string
Height and width of peak is measured relative to
- 'left': trough to the left
- 'right': trough to the right
- 'min': the minimum of the two troughs to the left and to the right
- 'max': the maximum of the two troughs to the left and to the right
- 'mean': mean of the throughs to the left and to the rigth
- 'closest': trough that is closest to peak
Returns
-------
widths: array
Width at `peak_frac` height of each peak.
Raises
------
ValueError:
If an invalid value is passed to `base`.
"""
def left_base(data, left_inx, right_inx, peak_inx):
return data[left_inx]
def right_base(data, left_inx, right_inx, peak_inx):
return data[right_inx]
def min_base(data, left_inx, right_inx, peak_inx):
return min(data[left_inx], data[right_inx])
def max_base(data, left_inx, right_inx, peak_inx):
return max(data[left_inx], data[right_inx])
def mean_base(data, left_inx, right_inx, peak_inx):
return np.mean((data[left_inx], data[right_inx]))
def closest_base(data, left_inx, right_inx, peak_inx):
return data[left_inx] if peak_inx-left_inx <= right_inx-peak_inx else data[right_inx]
widths = np.zeros(len(peak_indices))
if len(peak_indices) == 0:
return widths
# we need a trough before and after each peak:
peak_inx = np.asarray(peak_indices, dtype=int)
trough_inx = np.asarray(trough_indices, dtype=int)
if len(trough_inx) == 0 or peak_inx[0] < trough_inx[0]:
trough_inx = np.hstack((0, trough_inx))
if peak_inx[-1] > trough_inx[-1]:
trough_inx = np.hstack((trough_inx, len(data)-1))
# base for size of peaks:
base_func = closest_base
if base == 'left':
base_func = left_base
elif base == 'right':
base_func = right_base
elif base == 'min':
base_func = min_base
elif base == 'max':
base_func = max_base
elif base == 'mean':
base_func = mean_base
elif base == 'closest':
base_func = closest_base
else:
raise ValueError('Invalid value for base (%s)' % base)
# width of peaks:
for j in range(len(peak_inx)):
li = trough_inx[j]
ri = trough_inx[j+1]
baseval = base_func(data, li, ri, peak_inx[j])
thresh = baseval*(1.0-peak_frac) + data[peak_inx[j]]*peak_frac
inx = li + np.argmax(data[li:ri] > thresh)
if inx > 0:
ti0 = np.interp(thresh, data[inx-1:inx+1], time[inx-1:inx+1])
else:
ti0 = time[0]
inx = ri - np.argmax(data[ri:li:-1] > thresh)
if inx+1 < len(data):
ti1 = np.interp(thresh, data[inx+1:inx-1:-1], time[inx+1:inx-1:-1])
else:
ti1 = time[-1]
widths[j] = ti1 - ti0
return widths
def peak_size_width(time, data, peak_indices, trough_indices,
peak_frac=0.75, base='closest'):
"""
Compute for each peak its size and width.
Parameters
----------
time: array
Time, must not be `None`.
data: array
The data with the peaks.
peak_indices: array
Indices of the peaks.
trough_indices: array
Indices of the troughs.
peak_frac: float
Fraction of peak height where its width is measured.
base: string
Height and width of peak is measured relative to
- 'left': trough to the left
- 'right': trough to the right
- 'min': the minimum of the two troughs to the left and to the right
- 'max': the maximum of the two troughs to the left and to the right
- 'mean': mean of the throughs to the left and to the rigth
- 'closest': trough that is closest to peak
Returns
-------
peaks: 2-D array
First dimension is the peak index. Second dimension is
time, height (value of data at the peak),
size (peak height minus height of closest trough),
width (at `peak_frac` size), 0.0 (count) of the peak. See `peak_width()`.
Raises
------
ValueError:
If an invalid value is passed to `base`.
"""
def left_base(data, left_inx, right_inx, peak_inx):
return data[left_inx]
def right_base(data, left_inx, right_inx, peak_inx):
return data[right_inx]
def min_base(data, left_inx, right_inx, peak_inx):
return min(data[left_inx], data[right_inx])
def max_base(data, left_inx, right_inx, peak_inx):
return max(data[left_inx], data[right_inx])
def mean_base(data, left_inx, right_inx, peak_inx):
return np.mean((data[left_inx], data[right_inx]))
def closest_base(data, left_inx, right_inx, peak_inx):
return data[left_inx] if peak_inx-left_inx <= right_inx-peak_inx else data[right_inx]
peaks = np.zeros((len(peak_indices), 5))
if len(peak_indices) == 0:
return peaks
# time point of peaks:
peaks[:, 0] = time[peak_indices]
# height of peaks:
peaks[:, 1] = data[peak_indices]
# we need a trough before and after each peak:
peak_inx = np.asarray(peak_indices, dtype=int)
trough_inx = np.asarray(trough_indices, dtype=int)
if len(trough_inx) == 0 or peak_inx[0] < trough_inx[0]:
trough_inx = np.hstack((0, trough_inx))
if peak_inx[-1] > trough_inx[-1]:
trough_inx = np.hstack((trough_inx, len(data)-1))
# base for size of peaks:
base_func = closest_base
if base == 'left':
base_func = left_base
elif base == 'right':
base_func = right_base
elif base == 'min':
base_func = min_base
elif base == 'max':
base_func = max_base
elif base == 'mean':
base_func = mean_base
elif base == 'closest':
base_func = closest_base
else:
raise ValueError('Invalid value for base (%s)' % base)
# size and width of peaks:
for j, pi in enumerate(peak_inx):
li = trough_inx[j]
ri = trough_inx[j+1]
baseval = base_func(data, li, ri, pi)
thresh = baseval*(1.0-peak_frac) + data[pi]*peak_frac
inx = li + np.argmax(data[li:ri] > thresh)
if inx > 0:
ti0 = np.interp(thresh, data[inx-1:inx+1], time[inx-1:inx+1])
else:
ti0 = time[0]
inx = ri - np.argmax(data[ri:li:-1] > thresh)
if inx+1 < len(data):
ti1 = np.interp(thresh, data[inx+1:inx-1:-1], time[inx+1:inx-1:-1])
else:
ti1 = time[-1]
if np.any(np.isfinite((data[pi], baseval))):
peaks[j, 2] = data[pi] - baseval
peaks[j, 3] = ti1 - ti0
return peaks
def threshold_crossings(data, threshold):
"""
Detect crossings of a threshold with positive and negative slope.
Parameters
----------
data: array
An 1-D array of input data where threshold crossings are detected.
threshold: float or array
A number or array of numbers setting the threshold
that needs to be crossed.
Returns
-------
up_indices: array of ints
A list of indices where the threshold is crossed with positive slope.
down_indices: array of ints
A list of indices where the threshold is crossed with negative slope.
Raises
------
IndexError: If `data` and `threshold` arrays differ in length.
"""
if np.isscalar(threshold):
up_indices = np.nonzero((data[1:]>threshold) & (data[:-1]<=threshold))[0]
down_indices = np.nonzero((data[1:]<=threshold) & (data[:-1]>threshold))[0]
else:
if len(data) != len(threshold):
raise IndexError('input arrays data and threshold must have same length!')
up_indices = np.nonzero((data[1:]>threshold[1:]) & (data[:-1]<=threshold[:-1]))[0]
down_indices = np.nonzero((data[1:]<=threshold[1:]) & (data[:-1]>threshold[:-1]))[0]
return up_indices, down_indices
def threshold_crossing_times(time, data, threshold, up_indices, down_indices):
"""
Compute times of threshold crossings by linear interpolation.
Parameters
----------
time: array
Time, must not be `None`.
data: array
The data.
up_indices: array of ints
A list of indices where the threshold is crossed with positive slope.
down_indices: array of ints
A list of indices where the threshold is crossed with negative slope.
Returns
-------
up_times: array of floats
Interpolated times where the threshold is crossed with positive slope.
down_times: array of floats
Interpolated times where the threshold is crossed with negative slope.
"""
up_times = np.zeros(len(up_indices))
for k, inx in enumerate(up_indices):
up_times[k] = np.interp(threshold, data[inx:inx+2], time[inx:inx+2])
down_times = np.zeros(len(down_indices))
for k, inx in enumerate(down_indices):
down_times[k] = np.interp(-threshold, -data[inx:inx+2], time[inx:inx+2])
return up_times, down_times
def trim(peaks, troughs):
"""
Trims the peaks and troughs arrays such that they have the same length.
Parameters
----------
peaks: array
List of peak indices or times.
troughs: array
List of trough indices or times.
Returns
-------
peaks: array
List of peak indices or times.
troughs: array
List of trough indices or times.
"""
# common len:
n = min(len(peaks), len(troughs))
# align arrays:
return peaks[:n], troughs[:n]
def trim_to_peak(peaks, troughs):
"""
Trims the peaks and troughs arrays such that they have the same length
and the first peak comes first.
Parameters
----------
peaks: array
List of peak indices or times.
troughs: array
List of trough indices or times.
Returns
-------
peaks: array
List of peak indices or times.
troughs: array
List of trough indices or times.
"""
# start index for troughs:
tidx = 0
if len(peaks) > 0 and len(troughs) > 0 and troughs[0] < peaks[0]:
tidx = 1
# common len:
n = min(len(peaks), len(troughs[tidx:]))
# align arrays:
return peaks[:n], troughs[tidx:tidx + n]
def trim_closest(peaks, troughs):
"""
Trims the peaks and troughs arrays such that they have the same length
and that peaks-troughs is on average as small as possible.
Parameters
----------
peaks: array
List of peak indices or times.
troughs: array
List of trough indices or times.
Returns
-------
peaks: array
List of peak indices or times.
troughs: array
List of trough indices or times.
"""
pidx = 0
tidx = 0
nn = min(len(peaks), len(troughs))
if nn == 0:
return np.array([]), np.array([])
dist = np.abs(np.mean(peaks[:nn] - troughs[:nn]))
if len(peaks) == 0 or len(troughs) == 0:
nn = 0
else:
if peaks[0] < troughs[0]:
nnp = min(len(peaks[1:]), len(troughs))
distp = np.abs(np.mean(peaks[1:nnp] - troughs[:nnp - 1]))
if distp < dist:
pidx = 1
nn = nnp
else:
nnt = min(len(peaks), len(troughs[1:]))
distt = np.abs(np.mean(peaks[:nnt - 1] - troughs[1:nnt]))
if distt < dist:
tidx = 1
nn = nnt
# align arrays:
return peaks[pidx:pidx + nn], troughs[tidx:tidx + nn]
def merge_events(onsets, offsets, min_distance):
"""Merge events if they are closer than a minimum distance.
If the beginning of an event (onset, peak, or positive threshold crossing,
is too close to the end of the previous event (offset, trough, or negative
threshold crossing) the two events are merged into a single one that begins
with the first one and ends with the second one.
Parameters
----------
onsets: 1-D array
The onsets (peaks, or positive threshold crossings) of the events
as indices or times.
offsets: 1-D array
The offsets (troughs, or negative threshold crossings) of the events
as indices or times.
min_distance: int or float
The minimum distance between events. If the beginning of an event is separated
from the end of the previous event by less than this distance then the two events
are merged into one. If the event onsets and offsets are given in indices than
min_distance is also in indices.
Returns
-------
merged_onsets: 1-D array
The onsets (peaks, or positive threshold crossings) of the merged events
as indices or times according to onsets.
merged_offsets: 1-D array
The offsets (troughs, or negative threshold crossings) of the merged events
as indices or times according to offsets.
"""
onsets, offsets = trim_to_peak(onsets, offsets)
if len(onsets) == 0 or len(offsets) == 0:
return np.array([]), np.array([])
else:
diff = onsets[1:] - offsets[:-1]
indices = diff > min_distance
merged_onsets = onsets[np.hstack([True, indices])]
merged_offsets = offsets[np.hstack([indices, True])]
return merged_onsets, merged_offsets
def remove_events(onsets, offsets, min_duration, max_duration=None):
"""Remove events that are too short or too long.
If the length of an event, i.e. `offset` (offset, trough, or negative
threshold crossing) minus `onset` (onset, peak, or positive threshold crossing),
is shorter than `min_duration` or longer than `max_duration`, then this event is
removed.
Parameters
----------
onsets: 1-D array
The onsets (peaks, or positive threshold crossings) of the events
as indices or times.
offsets: 1-D array
The offsets (troughs, or negative threshold crossings) of the events
as indices or times.
min_duration: int, float, or None
The minimum duration of events. If the event offset minus the event onset
is less than `min_duration`, then the event is removed from the lists.
If the event onsets and offsets are given in indices than
`min_duration` is also in indices. If `None` then this test is skipped.
max_duration: int, float, or None
The maximum duration of events. If the event offset minus the event onset
is larger than `max_duration`, then the event is removed from the lists.
If the event onsets and offsets are given in indices than
`max_duration` is also in indices. If `None` then this test is skipped.
Returns
-------
onsets: 1-D array
The onsets (peaks, or positive threshold crossings) of the events
with too short and too long events removed as indices or times according to onsets.
offsets: 1-D array
The offsets (troughs, or negative threshold crossings) of the events
with too short and too long events removed as indices or times according to offsets.
"""
onsets, offsets = trim_to_peak(onsets, offsets)
if len(onsets) == 0 or len(offsets) == 0:
return np.array([]), np.array([])
elif min_duration is not None or max_duration is not None:
diff = offsets - onsets
if min_duration is not None and max_duration is not None:
indices = (diff > min_duration) & (diff < max_duration)
elif min_duration is not None:
indices = diff > min_duration
else:
indices = diff < max_duration
onsets = onsets[indices]
offsets = offsets[indices]
return onsets, offsets
def widen_events(onsets, offsets, max_time, duration):
"""Enlarge events on both sides without overlap.
Subtracts `duration` from the `onsets` and adds `duration` to the offsets.
If two succeeding events are separated by less than two times the `duration`,
then the offset of the previous event and the onset of the following event are
set at the center between the two events.
Parameters
----------
onsets: 1-D array
The onsets (peaks, or positive threshold crossings) of the events
as indices or times.
offsets: 1-D array
The offsets (troughs, or negative threshold crossings) of the events
as indices or times.
max_time: int or float
The maximum value for the end of the last event.
If the event onsets and offsets are given in indices than
max_time is the maximum possible index, i.e. the len of the
data array on which the events where detected.
duration: int or float
The number of indices or the time by which the events should be enlarged.
If the event onsets and offsets are given in indices than
duration is also in indices.
Returns
-------
onsets: 1-D array
The onsets (peaks, or positive threshold crossings) of the enlarged events.
offsets: 1-D array
The offsets (troughs, or negative threshold crossings) of the enlarged events.
"""
new_onsets = []
new_offsets = []
if len(onsets) > 0:
on_idx = onsets[0]
new_onsets.append( on_idx - duration if on_idx >= duration else 0 )
for off_idx, on_idx in zip(offsets[:-1], onsets[1:]):
if on_idx - off_idx < 2*duration:
mid_idx = (on_idx + off_idx)//2
new_offsets.append(mid_idx)
new_onsets.append(mid_idx)
else:
new_offsets.append(off_idx + duration)
new_onsets.append(on_idx - duration)
if len(offsets) > 0:
off_idx = offsets[-1]
new_offsets.append(off_idx + duration if off_idx + duration < max_time else max_time)
return new_onsets, new_offsets
def std_threshold(data, samplerate=None, win_size=None, thresh_fac=5.):
"""Estimates a threshold for peak detection based on the standard deviation of the data.
The threshold is computed as the standard deviation of the data
multiplied with `thresh_fac`.
In case of Gaussian distributed data, setting `thresh_fac=2.0` (two standard deviations)
captures 68% of the data, `thresh_fac=4.0` captures 95%, and `thresh_fac=6.0` 99.7%.
If `samplerate` and `win_size` is given, then the threshold is computed for
each half-overlapping window of duration `win_size` separately.
In this case the returned threshold is an array of the same size as data.
Without a `samplerate` and `win_size` a single threshold value determined from
the whole data array is returned.
Parameters
----------
data: 1-D array
The data to be analyzed.
samplerate: float or None
Sampling rate of the data in Hz.
win_size: float or None
Size of window in which a threshold value is computed.
thresh_fac: float
Factor by which the standard deviation is multiplied to set the threshold.
Returns
-------
threshold: float or 1-D array
The computed threshold.
"""
if samplerate and win_size:
threshold = np.zeros(len(data))
win_size_indices = int(win_size * samplerate)
for inx0 in range(0, len(data), win_size_indices//2):
inx1 = inx0 + win_size_indices
std = np.std(data[inx0:inx1], ddof=1)
threshold[inx0:inx1] = std * thresh_fac
return threshold
else:
return np.std(data, ddof=1) * thresh_fac
@jit(nopython=True)
def median_std_threshold(data, samplerate, win_size=0.0005, n_snippets=1000, thresh_fac=6.0):
"""Estimate a threshold for peak detection based on the median standard deviation of data snippets.
On `n_snippets` snippets of `win_size` duration the standard
deviation of the data is estimated. The returned threshold is the
median of these standard deviations multiplied by `thresh_fac`.
Parameters
----------
data: 1-D array of float
The data to be analysed.
samplerate: int or float
Sampling rate of the data
win_size: float
Duration of windows on which standarad deviations are computed in seconds.
n_snippets: int
Number of snippets on which the standard deviations are estimated.
thresh_fac: float
Factor by which the median standard deviation is multiplied to set the threshold.
Returns
-------
threshold: float
The computed threshold.
"""
win_size_indices = int(win_size * samplerate)
if win_size_indices < 10:
win_size_indices = 10
step = len(data)//n_snippets
if step < win_size_indices//2:
step = win_size_indices//2
stds = np.array([np.std(data[i:i+win_size_indices])
for i in range(0, len(data)-win_size_indices, step)])
return np.median(stds)*thresh_fac
def hist_threshold(data, samplerate=None, win_size=None, thresh_fac=5.,
nbins=100, hist_height=1.0/np.sqrt(np.e)):
"""Estimate a threshold for peak detection based on a histogram of the data.
The standard deviation of the data is estimated from half the
width of the histogram of the data at `hist_height` relative height.
This estimates the data's standard deviation by ignoring tails of the distribution.
However, you need enough data to robustly estimate the histogram.
If `samplerate` and `win_size` is given, then the threshold is computed for
each half-overlapping window of duration `win_size` separately.
In this case the returned threshold is an array of the same size as data.
Without a samplerate and win_size a single threshold value determined from
the whole data array is returned.
Parameters
----------
data: 1-D array
The data to be analyzed.
samplerate: float or None
Sampling rate of the data in Hz.
win_size: float or None
Size of window in which a threshold value is computed in sec.
thresh_fac: float
Factor by which the width of the histogram is multiplied to set the threshold.
nbins: int or list of floats
Number of bins or the bins for computing the histogram.
hist_height: float
Height between 0 and 1 at which the width of the histogram is computed.
Returns
-------
threshold: float or 1-D array
The computed threshold.
center: float or 1-D array
The center (mean) of the width of the histogram.
"""
if samplerate and win_size:
threshold = np.zeros(len(data))
centers = np.zeros(len(data))
win_size_indices = int(win_size * samplerate)
for inx0 in range(0, len(data), win_size_indices//2):
inx1 = inx0 + win_size_indices
std, center = hist_threshold(data[inx0:inx1], samplerate=None, win_size=None,
thresh_fac=thresh_fac, nbins=nbins,
hist_height=hist_height)
threshold[inx0:inx1] = std
centers[inx0:inx1] = center
return threshold, centers
else:
maxd = np.max(data)
mind = np.min(data)
contrast = np.abs((maxd - mind)/(maxd + mind))
if contrast > 1e-8:
hist, bins = np.histogram(data, nbins, density=False)
inx = hist > np.max(hist) * hist_height
lower = bins[0:-1][inx][0]
upper = bins[1:][inx][-1] # needs to return the next bin
center = 0.5 * (lower + upper)
std = 0.5 * (upper - lower)
else:
std = np.std(data)
center = np.mean(data)
return std * thresh_fac, center
def minmax_threshold(data, samplerate=None, win_size=None, thresh_fac=0.8):
"""Estimate a threshold for peak detection based on minimum and maximum values of the data.
The threshold is computed as the difference between maximum and
minimum value of the data multiplied with `thresh_fac`.
If `samplerate` and `win_size` is given, then the threshold is computed for
each half-overlapping window of duration `win_size` separately.
In this case the returned threshold is an array of the same size as data.
Without a samplerate and win_size a single threshold value determined from
the whole data array is returned.
Parameters
----------
data: 1-D array
The data to be analyzed.
samplerate: float or None
Sampling rate of the data in Hz.
win_size: float or None
Size of window in which a threshold value is computed.
thresh_fac: float
Factor by which the difference between minimum and maximum data value
is multiplied to set the threshold.
Returns
-------
threshold: float or 1-D array
The computed threshold.
"""
if samplerate and win_size:
threshold = np.zeros(len(data))
win_size_indices = int(win_size * samplerate)
for inx0 in range(0, len(data), win_size_indices//2):
inx1 = inx0 + win_size_indices
window_min = np.min(data[inx0:inx1])
window_max = np.max(data[inx0:inx1])
threshold[inx0:inx1] = (window_max - window_min) * thresh_fac
return threshold
else:
return (np.max(data) - np.min(data)) * thresh_fac
def percentile_threshold(data, samplerate=None, win_size=None, thresh_fac=1.0, percentile=1.0):
"""Estimate a threshold for peak detection based on an inter-percentile range of the data.
The threshold is computed as the range between the percentile and
100.0-percentile percentiles of the data multiplied with
thresh_fac.
For very small values of `percentile` the estimated threshold
approaches the one returned by `minmax_threshold()` (for same values
of `thresh_fac`). For `percentile=16.0` and Gaussian distributed data,
the returned theshold is twice the one returned by `std_threshold()`
or `hist_threshold()`, i.e. twice the standard deviation.
If you have knowledge about how many data points are in the tails of
the distribution, then this method is preferred over
`hist_threhsold()`. For example, if you expect peaks that you want
to detect using `detect_peaks()` at an average rate of 10Hz and
these peaks are about 1ms wide, then you have a 1ms peak per 100ms
period, i.e. the peaks make up 1% of the distribution. So you should
set `percentile=1.0` or lower. For much lower percentile values, you
might choose `thresh_fac` slightly smaller than one to capture also
smaller peaks. Setting `percentile` slightly higher might not change
the estimated threshold too much, since you start incorporating the
noise floor with a large density, but you may want to set
`thresh_fac` larger than one to reduce false detections.
If `samplerate` and `win_size` is given, then the threshold is computed for
each half-overlapping window of duration `win_size` separately.
In this case the returned threshold is an array of the same size as data.
Without a samplerate and win_size a single threshold value determined from
the whole data array is returned.
Parameters
----------
data: 1-D array
The data to be analyzed.
samplerate: float or None
Sampling rate of the data in Hz.
win_size: float or None
Size of window in which a threshold value is computed.
percentile: float
The interpercentile range is computed at percentile and 100.0-percentile.
If zero, compute maximum minus minimum data value as the interpercentile range.
thresh_fac: float
Factor by which the inter-percentile range of the data is multiplied to set the threshold.
Returns
-------
threshold: float or 1-D array
The computed threshold.
"""
if percentile < 1e-8:
return minmax_threshold(data, samplerate=samplerate, win_size=win_size,
thresh_fac=thresh_fac)
if samplerate and win_size:
threshold = np.zeros(len(data))
win_size_indices = int(win_size * samplerate)
for inx0 in range(0, len(data), win_size_indices//2):
inx1 = inx0 + win_size_indices
threshold[inx0:inx1] = np.squeeze(np.abs(np.diff(
np.percentile(data[inx0:inx1], [100.0 - percentile, percentile])))) * thresh_fac
return threshold
else:
return np.squeeze(np.abs(np.diff(
np.percentile(data, [100.0 - percentile, percentile])))) * thresh_fac
def snippets(data, indices, start=-10, stop=10):
"""
Cut out data arround each position given in indices.
Parameters
----------
data: 1-D array
Data array from which snippets are extracted.
indices: list of int
Indices around which snippets are cut out.
start: int
Each snippet starts at index + start.
stop: int
Each snippet ends at index + stop.
Returns
-------
snippet_data: 2-D array
The snippets: first index number of snippet, second index time.
"""
idxs = indices[(indices>=-start) & (indices<len(data)-stop)]
snippet_data = np.empty((len(idxs), stop-start))
for k, idx in enumerate(idxs):
snippet_data[k] = data[idx+start:idx+stop]
# XXX alternative: check speed and behavior for empty idxs
# snippets = np.vstack([data[idx+start:idx+stop] for idx in idxs])
return snippet_data
def detect_dynamic_peaks(data, threshold, min_thresh, tau, time=None,
check_peak_func=None, check_trough_func=None, **kwargs):
"""
Detect peaks and troughs using a relative threshold according to
Bryan S. Todd and David C. Andrews (1999): The identification of peaks in physiological signals.
Computers and Biomedical Research 32, 322-335.
The threshold decays dynamically towards min_thresh with time constant tau.
Use `check_peak_func` or `check_trough_func` to reset the threshold to an appropriate size.
Parameters
----------
data: array
An 1-D array of input data where peaks are detected.
threshold: float
A positive number setting the minimum distance between peaks and troughs.
min_thresh: float
The minimum value the threshold is allowed to assume.
tau: float
The time constant of the the decay of the threshold value
given in indices (`time` is None) or time units (`time` is not `None`).
time: array
The (optional) 1-D array with the time corresponding to the data values.
check_peak_func: function
An optional function to be used for further evaluating and analysing a peak.
The signature of the function is
```
r, th = check_peak_func(time, data, peak_inx, index, min_inx, threshold, **kwargs)
```
with
time (array): the full time array that might be None
data (array): the full data array
peak_inx (int): the index of the detected peak
index (int): the current index
min_inx (int): the index of the trough preceeding the peak (might be 0)
threshold (float): the threshold value
min_thresh (float): the minimum value the threshold is allowed to assume.
tau (float): the time constant of the the decay of the threshold value
given in indices (time is None) or time units (time is not None)
**kwargs: further keyword arguments provided by the user
r (scalar or np.array): a single number or an array with properties of the peak or None to skip the peak
th (float): a new value for the threshold or None (to keep the original value)
check_trough_func: function
An optional function to be used for further evaluating and analysing a trough.
The signature of the function is
```
r, th = check_trough_func(time, data, trough_inx, index, max_inx, threshold, **kwargs)
```
with
time (array): the full time array that might be None
data (array): the full data array
trough_inx (int): the index of the detected trough
index (int): the current index
max_inx (int): the index of the peak preceeding the trough (might be 0)
threshold (float): the threshold value
min_thresh (float): the minimum value the threshold is allowed to assume.
tau (float): the time constant of the the decay of the threshold value
given in indices (time is None) or time units (time is not None)
**kwargs: further keyword arguments provided by the user
r (scalar or np.array): a single number or an array with properties of the trough or None to skip the trough
th (float): a new value for the threshold or None (to keep the original value)
kwargs: key-word arguments
Arguments passed on to `check_peak_func` and `check_trough_func`.
Returns
-------
peak_list: np.array
A list of peaks.
trough_list: np.array
A list of troughs.
If time is `None` and no `check_peak_func` is given,
then these are lists of the indices where the peaks/troughs occur.
If `time` is given and no `check_peak_func`/`check_trough_func` is given,
then these are lists of the times where the peaks/troughs occur.
If `check_peak_func` or `check_trough_func` is given,
then these are lists of whatever `check_peak_func`/`check_trough_func` return.
Raises
------
ValueError: If `threshold <= 0` or `min_thresh <= 0` or `tau <= 0`.
IndexError: If `data` and `time` arrays differ in length.
"""
if threshold <= 0:
raise ValueError('input argument threshold must be positive!')
if min_thresh <= 0:
raise ValueError('input argument min_thresh must be positive!')
if tau <= 0:
raise ValueError('input argument tau must be positive!')
if time is not None and len(data) != len(time):
raise IndexError('input arrays time and data must have same length!')
peaks_list = list()
troughs_list = list()
# initialize:
direction = 0
min_inx = 0
max_inx = 0
min_value = data[0]
max_value = min_value
# loop through the data:
for index, value in enumerate(data):
# decaying threshold (first order low pass filter):
if time is None:
threshold += (min_thresh - threshold) / tau
else:
idx = index
if idx + 1 >= len(data):
idx = len(data) - 2
threshold += (min_thresh - threshold) * (time[idx + 1] - time[idx]) / tau
# rising?
if direction > 0:
# if the new value is bigger than the old maximum: set it as new maximum:
if value > max_value:
max_inx = index # maximum element
max_value = value
# otherwise, if the new value is falling below the maximum value minus the threshold:
# the maximum is a peak!
elif max_value >= value + threshold:
# check and update peak with the check_peak_func function:
if check_peak_func:
r, th = check_peak_func(time, data, max_inx, index,
min_inx, threshold,
min_thresh=min_thresh, tau=tau, **kwargs)
if r is not None:
# this really is a peak:
peaks_list.append(r)
if th is not None:
threshold = th
if threshold < min_thresh:
threshold = min_thresh
else:
# this is a peak:
if time is None:
peaks_list.append(max_inx)
else:
peaks_list.append(time[max_inx])
# change direction:
min_inx = index # minimum element
min_value = value
direction = -1
# falling?
elif direction < 0:
if value < min_value:
min_inx = index # minimum element
min_value = value
elif value >= min_value + threshold:
# there was a trough:
# check and update trough with the check_trough function:
if check_trough_func:
r, th = check_trough_func(time, data, min_inx, index,
max_inx, threshold,
min_thresh=min_thresh, tau=tau, **kwargs)
if r is not None:
# this really is a trough:
troughs_list.append(r)
if th is not None:
threshold = th
if threshold < min_thresh:
threshold = min_thresh
else:
# this is a trough:
if time is None:
troughs_list.append(min_inx)
else:
troughs_list.append(time[min_inx])
# change direction:
max_inx = index # maximum element
max_value = value
direction = 1
# don't know direction yet:
else:
if max_value >= value + threshold:
direction = -1 # falling
elif value >= min_value + threshold:
direction = 1 # rising
if max_value < value:
max_inx = index # maximum element
max_value = value
elif value < min_value:
min_inx = index # minimum element
min_value = value
return np.asarray(peaks_list), np.asarray(troughs_list)
def accept_peak_size_threshold(time, data, event_inx, index, min_inx, threshold,
min_thresh, tau, thresh_ampl_fac=0.75, thresh_weight=0.02):
"""Accept each detected peak/trough and return its index or time.
Adjust the threshold to the size of the detected peak.
To be passed to the `detect_dynamic_peaks()` function.
Parameters
----------
time: array
Time values, can be `None`.
data: array
The data in wich peaks and troughs are detected.
event_inx: int
Index of the current peak/trough.
index: int
Current index.
min_inx: int
Index of the previous trough/peak.
threshold: float
Threshold value.
min_thresh: float
The minimum value the threshold is allowed to assume..
tau: float
The time constant of the the decay of the threshold value
given in indices (`time` is `None`) or time units (`time` is not `None`).
thresh_ampl_fac: float
The new threshold is `thresh_ampl_fac` times the size of the current peak.
thresh_weight: float
New threshold is weighted against current threshold with `thresh_weight`.
Returns
-------
index: int
Index of the peak/trough if `time` is `None`.
time: int
Time of the peak/trough if `time` is not `None`.
threshold: float
The new threshold to be used.
"""
size = data[event_inx] - data[min_inx]
threshold += thresh_weight * (thresh_ampl_fac * size - threshold)
if time is None:
return event_inx, threshold
else:
return time[event_inx], threshold
if __name__ == "__main__":
import matplotlib.pyplot as plt
print("Checking eventetection module ...")
print('')
# generate data:
dt = 0.001
time = np.arange(0.0, 10.0, dt)
f = 2.0
data = (0.5 * np.sin(2.0 * np.pi * f * time) + 0.5) ** 4.0
data += -0.1 * time * (time - 10.0)
data += 0.1 * np.random.randn(len(data))
print("generated waveform with %d peaks" % int(np.round(time[-1] * f)))
plt.plot(time, data)
print('')
print('check detect_peaks(data, 1.0)...')
peaks, troughs = detect_peaks(data, 1.0)
print(peaks)
print(troughs)
p, t = detect_peaks_fast(data, 1.0)
print(p)
print(t)
# print peaks:
print('detected %d peaks with period %g that differs from the real frequency by %g' % (
len(peaks), np.mean(np.diff(peaks)), f - 1.0 / np.mean(np.diff(peaks)) / np.mean(np.diff(time))))
# print troughs:
print('detected %d troughs with period %g that differs from the real frequency by %g' % (
len(troughs), np.mean(np.diff(troughs)), f - 1.0 / np.mean(np.diff(troughs)) / np.mean(np.diff(time))))
# plot peaks and troughs:
plt.plot(time[peaks], data[peaks], '.r', ms=20)
plt.plot(time[troughs], data[troughs], '.g', ms=20)
# detect threshold crossings:
onsets, offsets = threshold_crossings(data, 3.0)
onsets, offsets = merge_events(onsets, offsets, int(0.5/f/dt))
plt.plot(time, 3.0*np.ones(len(time)), 'k')
plt.plot(time[onsets], data[onsets], '.c', ms=20)
plt.plot(time[offsets], data[offsets], '.b', ms=20)
plt.ylim(-0.5, 4.0)
plt.show()
# check the faster algorithm:
import timeit
def wrapper(func, *args, **kwargs):
def wrapped():
return func(*args, **kwargs)
return wrapped
wrapped = wrapper(detect_peaks, data, 1.0)
t1 = timeit.timeit(wrapped, number=200)
print(t1)
wrapped = wrapper(detect_peaks_fast, data, 1.0)
t2 = timeit.timeit(wrapped, number=200)
print(t2)
print('new algorithm takes %.0f%% of old algorithm' % (100.0*t2/t1))