GP2023_chirp_detection/code/chirpdetection.py

584 lines
20 KiB
Python

import os
import numpy as np
from IPython import embed
import matplotlib.pyplot as plt
from scipy.signal import find_peaks
from scipy.ndimage import gaussian_filter1d
from thunderfish.dataloader import DataLoader
from thunderfish.powerspectrum import spectrogram, decibel
from sklearn.preprocessing import normalize
from modules.filters import bandpass_filter, envelope, highpass_filter
from modules.filehandling import ConfLoader, LoadData
from modules.plotstyle import PlotStyle
ps = PlotStyle()
def instantaneos_frequency(
signal: np.ndarray, samplerate: int
) -> tuple[np.ndarray, np.ndarray]:
"""
Compute the instantaneous frequency of a signal.
Parameters
----------
signal : np.ndarray
Signal to compute the instantaneous frequency from.
samplerate : int
Samplerate of the signal.
Returns
-------
tuple[np.ndarray, np.ndarray]
"""
# calculate instantaneos frequency with zero crossings
roll_signal = np.roll(signal, shift=1)
time_signal = np.arange(len(signal)) / samplerate
period_index = np.arange(len(signal))[(
roll_signal < 0) & (signal >= 0)][1:-1]
upper_bound = np.abs(signal[period_index])
lower_bound = np.abs(signal[period_index - 1])
upper_time = np.abs(time_signal[period_index])
lower_time = np.abs(time_signal[period_index - 1])
# create ratio
lower_ratio = lower_bound / (lower_bound + upper_bound)
# appy to time delta
time_delta = upper_time - lower_time
true_zero = lower_time + lower_ratio * time_delta
# create new time array
inst_freq_time = true_zero[:-1] + 0.5 * np.diff(true_zero)
# compute frequency
inst_freq = gaussian_filter1d(1 / np.diff(true_zero), 5)
return inst_freq_time, inst_freq
def plot_spectrogram(axis, signal: np.ndarray, samplerate: float, t0: float) -> None:
"""
Plot a spectrogram of a signal.
Parameters
----------
axis : matplotlib axis
Axis to plot the spectrogram on.
signal : np.ndarray
Signal to plot the spectrogram from.
samplerate : float
Samplerate of the signal.
t0 : float
Start time of the signal.
"""
# compute spectrogram
spec_power, spec_freqs, spec_times = spectrogram(
signal,
ratetime=samplerate,
freq_resolution=50,
overlap_frac=0.2,
)
axis.pcolormesh(
spec_times + t0,
spec_freqs,
decibel(spec_power),
)
axis.set_ylim(200, 1200)
def double_bandpass(
data: DataLoader, samplerate: int, freqs: np.ndarray, search_freq: float
) -> tuple[np.ndarray, np.ndarray]:
"""
Apply a bandpass filter to the baseline of a signal and a second bandpass
filter above or below the baseline, as specified by the search frequency.
Parameters
----------
data : DataLoader
Data to apply the filter to.
samplerate : int
Samplerate of the signal.
freqs : np.ndarray
Tracked fundamental frequencies of the signal.
search_freq : float
Frequency to search for above or below the baseline.
Returns
-------
tuple[np.ndarray, np.ndarray]
"""
# compute boundaries to filter baseline
q25, q75 = np.percentile(freqs, [25, 75])
# check if percentile delta is too small
if q75 - q25 < 5:
median = np.median(freqs)
q25, q75 = median - 2.5, median + 2.5
# filter baseline
filtered_baseline = bandpass_filter(data, samplerate, lowf=q25, highf=q75)
# filter search area
filtered_search_freq = bandpass_filter(
data, samplerate, lowf=q25 + search_freq, highf=q75 + search_freq
)
return (filtered_baseline, filtered_search_freq)
def main(datapath: str) -> None:
# load raw file
data = LoadData(datapath)
# load config file
config = ConfLoader("chirpdetector_conf.yml")
# set time window
window_duration = config.window * data.raw_rate
window_overlap = config.overlap * data.raw_rate
window_edge = config.edge * data.raw_rate
# check if window duration is even
if window_duration % 2 == 0:
window_duration = int(window_duration)
else:
raise ValueError("Window duration must be even.")
# check if window ovelap is even
if window_overlap % 2 == 0:
window_overlap = int(window_overlap)
else:
raise ValueError("Window overlap must be even.")
# make time array for raw data
raw_time = np.arange(data.raw.shape[0]) / data.raw_rate
# good chirp times for data: 2022-06-02-10_00
t0 = (3 * 60 * 60 + 6 * 60 + 43.5) * data.raw_rate
dt = 60 * data.raw_rate
# generate starting points of rolling window
window_starts = np.arange(
t0,
t0 + dt,
window_duration - (window_overlap + 2 * window_edge),
dtype=int
)
# ask how many windows should be calulated
nwindows = int(
input("How many windows should be calculated (integer number)? "))
# ititialize lists to store data
chirps = []
fish_ids = []
for st, start_index in enumerate(window_starts[: nwindows]):
# make t0 and dt
t0 = start_index / data.raw_rate
dt = window_duration / data.raw_rate
# set index window
stop_index = start_index + window_duration
# calucate median of fish frequencies in window
median_freq = []
track_ids = []
for _, track_id in enumerate(np.unique(data.ident[~np.isnan(data.ident)])):
window_idx = np.arange(len(data.idx))[
(data.ident == track_id) & (data.time[data.idx] >= t0) & (
data.time[data.idx] <= (t0 + dt))
]
median_freq.append(np.median(data.freq[window_idx]))
track_ids.append(track_id)
# convert to numpy array
median_freq = np.asarray(median_freq)
track_ids = np.asarray(track_ids)
# iterate through all fish
for tr, track_id in enumerate(np.unique(data.ident[~np.isnan(data.ident)])):
print(f"Track ID: {track_id}")
# get index of track data in this time window
window_idx = np.arange(len(data.idx))[
(data.ident == track_id) & (data.time[data.idx] >= t0) & (
data.time[data.idx] <= (t0 + dt))
]
# get tracked frequencies and their times
freq_temp = data.freq[window_idx]
powers_temp = data.powers[window_idx, :]
# approximate sampling rate to compute expected durations if there
# is data available for this time window for this fish id
track_samplerate = np.mean(1 / np.diff(data.time))
expected_duration = ((t0 + dt) - t0) * track_samplerate
# check if tracked data available in this window
if len(freq_temp) < expected_duration * 0.9:
continue
fig, axs = plt.subplots(
7,
config.number_electrodes,
figsize=(20 / 2.54, 12 / 2.54),
constrained_layout=True,
sharex=True,
sharey='row',
)
# get best electrode
best_electrodes = np.argsort(np.nanmean(
powers_temp, axis=0))[-config.number_electrodes:]
# frequency where second filter filters
search_window = np.arange(
np.median(freq_temp)+config.search_df_lower, np.median(
freq_temp)+config.search_df_upper, config.search_res)
# search window in boolean
search_window_bool = np.ones(len(search_window), dtype=bool)
# get tracks that fall into search window
check_track_ids = track_ids[(median_freq > search_window[0]) & (
median_freq < search_window[-1])]
# iterate through theses tracks
if check_track_ids.size != 0:
for j, check_track_id in enumerate(check_track_ids):
q1, q2 = np.percentile(
data.freq[data.ident == check_track_id],
config.search_freq_percentiles
)
search_window_bool[(search_window > q1) & (
search_window < q2)] = False
# find gaps in search window
search_window_indices = np.arange(len(search_window))
# get search window gaps
search_window_gaps = np.diff(search_window_bool, append=np.nan)
nonzeros = search_window_gaps[np.nonzero(
search_window_gaps)[0]]
nonzeros = nonzeros[~np.isnan(nonzeros)]
# if the first value is -1, the array starst with true, so a gap
if nonzeros[0] == -1:
stops = search_window_indices[search_window_gaps == -1]
starts = np.append(
0, search_window_indices[search_window_gaps == 1])
# if the last value is -1, the array ends with true, so a gap
if nonzeros[-1] == 1:
stops = np.append(
search_window_indices[search_window_gaps == -1],
len(search_window) - 1
)
# else it starts with false, so no gap
if nonzeros[0] == 1:
stops = search_window_indices[search_window_gaps == -1]
starts = search_window_indices[search_window_gaps == 1]
# if the last value is -1, the array ends with true, so a gap
if nonzeros[-1] == 1:
stops = np.append(
search_window_indices[search_window_gaps == -1],
len(search_window)
)
# get the frequency ranges of the gaps
search_windows = [search_window[x:y]
for x, y in zip(starts, stops)]
search_windows_lens = [len(x) for x in search_windows]
longest_search_window = search_windows[np.argmax(
search_windows_lens)]
search_freq = (
longest_search_window[1] - longest_search_window[0]) / 2
else:
search_freq = config.default_search_freq
print(f"Search frequency: {search_freq}")
# iterate through electrodes
for el, electrode in enumerate(best_electrodes):
print(el)
# load region of interest of raw data file
data_oi = data.raw[start_index:stop_index, :]
time_oi = raw_time[start_index:stop_index]
# filter baseline and above
baseline, search = double_bandpass(
data_oi[:, electrode],
data.raw_rate,
freq_temp,
search_freq
)
# compute instantaneous frequency on broad signal
broad_baseline = bandpass_filter(
data_oi[:, electrode],
data.raw_rate,
lowf=np.mean(freq_temp)-5,
highf=np.mean(freq_temp)+100
)
# compute instantaneous frequency on narrow signal
baseline_freq_time, baseline_freq = instantaneos_frequency(
baseline, data.raw_rate
)
# compute envelopes
baseline_envelope_unfiltered = envelope(
baseline, data.raw_rate, config.envelope_cutoff)
search_envelope = envelope(
search, data.raw_rate, config.envelope_cutoff)
# highpass filter envelopes
baseline_envelope = highpass_filter(
baseline_envelope_unfiltered,
data.raw_rate,
config.envelope_highpass_cutoff
)
# envelopes of filtered envelope of filtered baseline
baseline_envelope = envelope(
np.abs(baseline_envelope),
data.raw_rate,
config.envelope_envelope_cutoff
)
# bandpass filter the instantaneous
inst_freq_filtered = bandpass_filter(
baseline_freq,
data.raw_rate,
lowf=config.instantaneous_lowf,
highf=config.instantaneous_highf
)
# CUT OFF OVERLAP ---------------------------------------------
# cut off first and last 0.5 * overlap at start and end
valid = np.arange(
int(window_edge), len(baseline_envelope) -
int(window_edge)
)
baseline_envelope_unfiltered = baseline_envelope_unfiltered[valid]
baseline_envelope = baseline_envelope[valid]
search_envelope = search_envelope[valid]
# get inst freq valid snippet
valid_t0 = int(window_edge) / data.raw_rate
valid_t1 = baseline_freq_time[-1] - \
(int(window_edge) / data.raw_rate)
inst_freq_filtered = inst_freq_filtered[
(baseline_freq_time >= valid_t0) & (
baseline_freq_time <= valid_t1)
]
baseline_freq = baseline_freq[
(baseline_freq_time >= valid_t0) & (
baseline_freq_time <= valid_t1)
]
baseline_freq_time = baseline_freq_time[
(baseline_freq_time >= valid_t0) & (
baseline_freq_time <= valid_t1)
] + t0
# overwrite raw time to valid region
time_oi = time_oi[valid]
baseline = baseline[valid]
broad_baseline = broad_baseline[valid]
search = search[valid]
# NORMALIZE ---------------------------------------------------
baseline_envelope = normalize([baseline_envelope])[0]
search_envelope = normalize([search_envelope])[0]
inst_freq_filtered = normalize([inst_freq_filtered])[0]
# PEAK DETECTION ----------------------------------------------
# detect peaks baseline_enelope
prominence = np.percentile(
baseline_envelope, config.baseline_prominence_percentile)
baseline_peaks, _ = find_peaks(
np.abs(baseline_envelope), prominence=prominence)
# detect peaks search_envelope
prominence = np.percentile(
search_envelope, config.search_prominence_percentile)
search_peaks, _ = find_peaks(
search_envelope, prominence=prominence)
# detect peaks inst_freq_filtered
prominence = np.percentile(
inst_freq_filtered,
config.instantaneous_prominence_percentile
)
inst_freq_peaks, _ = find_peaks(
np.abs(inst_freq_filtered),
prominence=prominence
)
# # SAVE DATA ---------------------------------------------------
# PLOT --------------------------------------------------------
# plot spectrogram
plot_spectrogram(
axs[0, el], data_oi[:, electrode], data.raw_rate, t0)
# plot baseline instantaneos frequency
axs[1, el].plot(baseline_freq_time, baseline_freq -
np.median(baseline_freq))
# plot waveform of filtered signal
axs[2, el].plot(time_oi, baseline, c=ps.green)
# plot broad filtered baseline
axs[2, el].plot(
time_oi,
broad_baseline,
)
# plot narrow filtered baseline envelope
axs[2, el].plot(
time_oi,
baseline_envelope_unfiltered,
c=ps.red
)
# plot waveform of filtered search signal
axs[3, el].plot(time_oi, search)
# plot envelope of search signal
axs[3, el].plot(
time_oi,
search_envelope,
c=ps.red
)
# plot filtered and rectified envelope
axs[4, el].plot(time_oi, baseline_envelope)
axs[4, el].scatter(
(time_oi)[baseline_peaks],
baseline_envelope[baseline_peaks],
c=ps.red,
)
# plot envelope of search signal
axs[5, el].plot(time_oi, search_envelope)
axs[5, el].scatter(
(time_oi)[search_peaks],
search_envelope[search_peaks],
c=ps.red,
)
# plot filtered instantaneous frequency
axs[6, el].plot(baseline_freq_time, np.abs(inst_freq_filtered))
axs[6, el].scatter(
baseline_freq_time[inst_freq_peaks],
np.abs(inst_freq_filtered)[inst_freq_peaks],
c=ps.red,
)
axs[6, el].set_xlabel("Time [s]")
axs[0, el].set_title("Spectrogram")
axs[1, el].set_title("Fitered baseline instanenous frequency")
axs[2, el].set_title("Fitered baseline")
axs[3, el].set_title("Fitered above")
axs[4, el].set_title("Filtered envelope of baseline envelope")
axs[5, el].set_title("Search envelope")
axs[6, el].set_title(
"Filtered absolute instantaneous frequency")
print(el)
# DETECT CHIRPS IN SEARCH WINDOW -------------------------------
baseline_ts = time_oi[baseline_peaks]
search_ts = time_oi[search_peaks]
freq_ts = baseline_freq_time[inst_freq_peaks]
# check if one list is empty
if len(baseline_ts) == 0 or len(search_ts) == 0 or len(freq_ts) == 0:
continue
# get index for each feature
baseline_idx = np.zeros_like(baseline_ts)
search_idx = np.ones_like(search_ts)
freq_idx = np.ones_like(freq_ts) * 2
timestamps_features = np.hstack(
[baseline_idx, search_idx, freq_idx])
timestamps = np.hstack([baseline_ts, search_ts, freq_ts])
# sort timestamps
timestamps_idx = np.arange(len(timestamps))
timestamps_features = timestamps_features[np.argsort(
timestamps)]
timestamps = timestamps[np.argsort(timestamps)]
# # get chirps
# diff = np.empty(timestamps.shape)
# diff[0] = np.inf # always retain the 1st element
# diff[1:] = np.diff(timestamps)
# mask = diff < config.chirp_window_threshold
# shared_peak_indices = timestamp_idx[mask]
current_chirps = []
for tt in timestamps:
cm = timestamps_idx[(timestamps >= tt) & (
timestamps <= tt + config.chirp_window_threshold)]
if set([0, 1, 2]).issubset(timestamps_features[cm]):
chirps.append(np.mean(timestamps[cm]))
current_chirps.append(np.mean(timestamps[cm]))
fish_ids.append(track_id)
for ct in current_chirps:
axs[0, el].axvline(ct, color='r', lw=1)
axs[0, el].scatter(
baseline_freq_time[inst_freq_peaks],
np.ones_like(baseline_freq_time[inst_freq_peaks]) * 600,
c=ps.red,
)
axs[0, el].scatter(
(time_oi)[search_peaks],
np.ones_like((time_oi)[search_peaks]) * 600,
c=ps.red,
)
axs[0, el].scatter(
(time_oi)[baseline_peaks],
np.ones_like((time_oi)[baseline_peaks]) * 600,
c=ps.red,
)
plt.show()
if __name__ == "__main__":
datapath = "../data/2022-06-02-10_00/"
main(datapath)