def beatsfilter(self, data):
self.all_data.append(data)
scaled_csi = self.all_data
no_frames = len(scaled_csi)
if no_frames < 256:
return None
Fs = 10
no_subcarriers = scaled_csi[0]["csi"].shape[0]
finalEntries = self.getCSI(scaled_csi)
sigs = []
for x in range(no_subcarriers):
finalEntry = finalEntries[x].flatten()
filtData = bandpass(7, 1, 1.5, Fs, finalEntry)
for i in range(0, 70):
filtData[i] = 0
sigs.append(filtData)
pxxs = []
for data in sigs:
f, Pxx_den = signal.welch(data, Fs)
pxxs.append(Pxx_den)
meanPsd = np.mean(pxxs, axis=0)
print("Beats: %.2f" % float(f[np.argmax(meanPsd)] * 60))
self.all_data = []
def beatsfilter(scaled_csi):
xax = getTimestamps(scaled_csi)
csi, no_frames, no_subcarriers = getCSI(scaled_csi)
Fs = 1 / np.mean(np.diff(xax))
stabilities = []
for x in range(no_subcarriers):
finalEntry = csi[x].flatten()
variation = stats.variation(finalEntry)
if not np.isnan(variation) and variation < 0.5:
hampelData = hampel(finalEntry, 5)
smoothedData = running_mean(hampelData, 5)
filtData = bandpass(5, 1, 1.3, Fs, smoothedData)
wLength = 1 * Fs
N = np.mod(len(filtData), wLength)
windows = np.array_split(filtData, N)
psds = []
for window in windows:
f, Pxx_den = signal.welch(window, Fs)
fMax = f[np.argmax(Pxx_den)]
psds.append(fMax)
specStab = 1 / np.var(psds)
print("Sub: {} has spectral stability: {}".format(x, specStab))
# plt.plot(xax, filtData, alpha=0.5)
stabilities.append({
"sub": x,
"stability": specStab,
"data": filtData
})
stabilities.sort(key=lambda x: x["stability"], reverse=True)
bestStab = stabilities[0]["stability"]
news = []
for stab in stabilities:
if stab["stability"] >= (bestStab / 100) * 20:
news.append(stab)
print("Using {} subcarriers.".format(len(news)))
pxxs = []
for sub in news:
filtData = sub["data"]
f, Pxx_den = signal.welch(filtData, Fs)
pxxs.append(Pxx_den)
meanPsd = np.mean(pxxs, axis=0)
plt.plot(f * 60, meanPsd, alpha=0.5)
plt.xlabel("Estimated Heart Rate [bpm]")
plt.ylabel("PSD [V**2/Hz]")
plt.legend(loc="upper right")
plt.show()
def heatmap(scaled_csi):
timestamps = getTimestamps(scaled_csi)
csi, no_frames, no_subcarriers = getCSI(scaled_csi)
Fs = 1 / np.mean(np.diff(timestamps))
print("Sampling Rate: " + str(Fs))
limits = [0, timestamps[-1], 1, no_subcarriers]
for x in range(no_subcarriers):
hampelData = hampel(csi[x].flatten(), 5)
smoothedData = running_mean(hampelData, 5)
butteredData = bandpass(7, 0.2, 0.3, Fs, smoothedData)
csi[x] = butteredData
fig, ax = plt.subplots()
im = ax.imshow(csi, cmap="jet", extent=limits, aspect="auto")
cbar = ax.figure.colorbar(im, ax=ax)
cbar.ax.set_ylabel("Amplitude (dBm)")
plt.xlabel("Time (s)")
plt.ylabel("Subcarrier Index")
plt.show()
def shorttime(reader, subcarrier):
scaled_csi = reader.csi_trace
#Replace complex CSI with amplitude.
no_frames, no_subcarriers, finalEntry = getCSI(scaled_csi)
# hampelData = hampel(finalEntry, 30)
# smoothedData = running_mean(hampelData, 20)
y = finalEntry[subcarrier]
n = no_frames
Fs = 20
fmin = 1
fmax = 1.7
y = bandpass(9, fmin, fmax, Fs, y)
# f, t, Zxx = signal.stft(y, Fs, nperseg=2000, noverlap=1750)
f, t, Zxx = signal.stft(y, Fs)
#bin indices for the observed amplitude peak in each segment.
Zxx = np.abs(Zxx)
maxAx = np.argmax(Zxx, axis=0)
freqs = []
for i in maxAx:
arc = f[i]
if arc >= fmin and arc <= fmax:
arc *= 60
freqs.append(arc)
else:
freqs.append(None)
bpms = []
for freq in freqs:
if freq != None:
bpms.append(freq)
print(bpms)
print("Average HR: " + str(np.average(bpms)))
#We're only interested in frequencies between 0.5 and 2.
freq_slice = np.where((f >= fmin) & (f <= fmax))
f = f[freq_slice]
Zxx = Zxx[freq_slice, :][0]
plt.pcolormesh(t, f * 60, np.abs(Zxx))
plt.plot(t, freqs, color="red", marker="x")
plt.title('STFT Magnitude')
# plt.ylabel('Frequency [Hz]')
plt.ylabel('Heart Rate [BPM]')
plt.xlabel('Time [sec]')
plt.show()
def updateHeat2(self, data):
self.all_data.append(data)
if not self.updateTimestamps():
self.all_data = self.all_data[:-1]
return None
scaled_csi = self.all_data
no_frames = len(scaled_csi)
no_subcarriers = scaled_csi[0]["csi"].shape[0]
ylimit = scaled_csi[no_frames - 1]["timestamp"]
if no_frames < 80:
return None
# limits = [1, no_subcarriers, 0, ylimit]
limits = [0, ylimit, 1, no_subcarriers]
finalEntry = np.zeros((no_subcarriers, no_frames))
#Replace complex CSI with amplitude.
for y in range(no_subcarriers):
for x in range(no_frames):
scaled_entry = scaled_csi[x]["csi"]
finalEntry[y][x] = db(abs(scaled_entry[y][0][0]))
for j in range(no_subcarriers):
sig = finalEntry[j]
#hampelData = hampel(sig, 10)
#smoothedData = running_mean(sig, 30)
y = sig.flatten()
y = bandpass(5, 1.0, 1.3, 20, y)
for x in range(70):
y[x] = 0
finalEntry[j] = y
#x = subcarrier index
#y = time (s)
#z = amplitude (cBm)
if not hasattr(self, "im"):
self.im = self.ax.imshow(finalEntry,
cmap="jet",
extent=limits,
aspect="auto")
cbar = self.ax.figure.colorbar(self.im, ax=self.ax)
cbar.ax.set_ylabel("Amplitude (dBm)", rotation=-91, va="bottom")
else:
self.im.set_array(finalEntry)
self.im.set_extent(limits)
self.ax.relim()
self.ax.autoscale_view()
self.fig.canvas.draw()
self.fig.canvas.flush_events()
def updateButterworth(self, data):
self.all_data.append(data)
self.updateTimestamps()
if not self.updateTimestamps():
self.all_data = self.all_data[:-1]
return None
scaled_csi = self.all_data
no_frames = len(scaled_csi)
no_subcarriers = scaled_csi[0]["csi"].shape[0]
if no_frames < 50:
return None
#Replace complex CSI with amplitude.
finalEntry = [
db(abs(scaled_csi[x]["csi"][15][0][0])) for x in range(no_frames)
]
hampelData = hampel(finalEntry, 10)
smoothedData = running_mean(hampelData, 30)
y = smoothedData
x = list([x["timestamp"] for x in scaled_csi])
tdelta = (x[-1] - x[0]) / len(x)
Fs = 1 / tdelta
n = no_frames
y = bandpass(5, 1.0, 1.3, Fs, y)
ffty = np.fft.rfft(y, len(y))
freq = np.fft.rfftfreq(len(y), tdelta)
freqX = [((i * Fs) / n) * 60 for i in range(len(freq))]
self.plotButt.set_xdata(freqX)
self.plotButt.set_ydata(np.abs(ffty))
self.ax.relim()
self.ax.autoscale_view()
self.fig.canvas.draw()
self.fig.canvas.flush_events()
def fft(reader):
scaled_csi = reader.csi_trace
Fs = 20
no_frames, no_subcarriers, finalEntry = getCSI(scaled_csi)
y = bandpass(7, 1, 1.5, Fs, finalEntry[15])
#rfft is a real Fast Fourier Transform, as we aren't interested in the imaginary parts.
#as half of the points lie in the imaginary/negative domain, we get an n/2 point FFT.
#rfftfreq produces a list of frequency bins, which requires calculation to produce
#interpretable frequencies for BPM tracking. by plotting calculated bpm values,
#the graph can be more easily read.
#despite rfft producing an n/2 point fft, the calculation still uses n,
#because the point spacing is still based on n points.
# ((binNumber*samplingRate)/n)*60 produces a per minute rate.
ffty = np.fft.rfft(y, len(y))
freq = np.fft.rfftfreq(len(y), 0.05)
freqX = [((i * Fs) / no_frames) * 60 for i in range(len(freq))]
plt.subplot(211)
plt.xlabel("Time (s)")
plt.ylabel("Amplitude (Hz)")
plt.plot(y)
plt.subplot(212)
plt.xlabel("Breaths Per Minute (BPM)")
plt.ylabel("Amplitude (dB/Hz)")
axes = plt.gca()
axes.set_xlim([50, 100])
plt.plot(freqX, np.abs(ffty))
plt.show()
def beatsfilter(reader, Fs):
scaled_csi = reader.csi_trace
no_frames, no_subcarriers, finalEntries = getCSI(scaled_csi,
metric="phasediff")
sigs = []
for x in range(no_subcarriers):
finalEntry = finalEntries[x]
# hampelData = hampel(finalEntry, 10, 5)
# # detrended = dynamic_detrend(hampelData, 5, 3, 1.2, Fs)
# # rehampeledData = hampel(detrended, 10, 0.1)
# # hampelData = hampel(finalEntry, 10)
# # runningMeanData = running_mean(hampelData, 20)
# filtData = bandpass(7, 1, 1.5, Fs, hampelData)
# hampelData = hampel(finalEntries[x], 5, 3)
# runningMeanData = running_mean(hampelData, 20)
# smoothedData = dynamic_detrend(runningMeanData, 5, 3, 1.2, 10)
# doubleHampel = hampel(smoothedData, 10, 3)
filtData = bandpass(7, 1.1, 1.4, Fs, finalEntry)
for i in range(0, 70):
filtData[i] = 0
sigs.append(filtData)
pxxs = []
for data in sigs:
f, Pxx_den = signal.welch(data, Fs)
pxxs.append(Pxx_den)
meanPsd = np.mean(pxxs, axis=0)
# plt.plot(f*60, meanPsd)
# plt.xlim([40, 100])
# plt.show()
return f[np.argmax(meanPsd)] * 60
def filter_seismo(sismo, freqs=[0.1, 10], ftype='bandpass', dt=0.01):
"""
filter seismograms.
Inputs:
-freqs[tuple][(0.1,10)] : corner frequencies of filter response
-ftype[str][default=bandpass] : filter type
Return:
-Updates self.velocity array.
"""
filtered_s = np.zeros(sismo.shape)
if ftype == 'bandpass':
for i in range(sismo.shape[-1]):
filtered_s[:, i] = bandpass(sismo[:, i],
freqs[0],
freqs[1],
dt=dt,
corners=4,
zerophase=True)
return filtered_s
def filter_signal(self, signal):
filtered_signal = filters.bandpass(signal, self.sample_rate,
MORSE_FREQUENCY-self.options['bandwidth'],
MORSE_FREQUENCY+self.options['bandwidth'])
return filtered_signal
def specstabfilter(reader, Fs):
scaled_csi = reader.csi_trace
no_frames, no_subcarriers, finalEntries = getCSI(scaled_csi)
#Using CSI phase difference data.
stabilities = []
for x in range(no_subcarriers):
finalEntry = finalEntries[x]
#CardioFi paper uses 100Hz uniformly sampled data, which may mean we need to use different
#filter parameters. For now, we will copy them verbatim.
#Currently the initial hampel filter is run using T1=0.5s t1=0.4
# dynamic_detrending is run with c=5 l=3 alpha=1.2 Fs=Fs
# the second hampel filter is run using T1=0.5s t1=0.1
# the bandpass filter is run at 5th order with a range of 1-2Hz.
hampelData = hampel(finalEntry, 20)
detrended = dynamic_detrend(hampelData, 5, 3, 1.2, Fs)
rehampeledData = hampel(detrended, 20)
filtData = bandpass(7, 1, 1.5, Fs, rehampeledData)
#Spectral Stability section.
#The spectral stability metric aims to score subcarriers based on the
#variance of their heart rate estimations.
#To do this, both the length of the data being used for estimations,
#and a sliding window length must be selected.
#While the dataLength should ideally be around 40s to mirror that used
#in the CardioFi paper, this is difficult to replicate using the data
#which was produced over the last few months.
#Shorter data and window lengths are being experimented with but do not
#appear to be working well. This is evidenced by the number of subcarriers
#being selected using the metric, which is typically lower than that seen
#in the CardioFi paper.
#Notably, it is also difficult to take consistent ground truth heart rate
#measurements over a 40 second period.
wLength = 5 * Fs
dataLength = 10 * Fs
windows = []
position = 0
#This sliding window system splits filtData into wLength-sized windows,
#except for the final window which contains the remainder of the data.
#Each sliding window has an overlap of 1 second (or Fs).
while position < dataLength:
if (position + wLength) > dataLength:
window = filtData[position:]
else:
window = filtData[position:position + wLength]
windows.append(window)
position += wLength - Fs
#PSD measurements are then taken for each window, and the variance of
#these measurements are then collected to produce a final score for
#this subcarrier.
psds = []
for window in windows:
f, Pxx_den = signal.welch(window, Fs)
fMax = f[np.argmax(Pxx_den)]
psds.append(fMax)
specStab = 1 / np.var(psds)
if specStab == np.inf:
specStab = 0
stabilities.append([x, specStab, filtData])
# print("Sub: {} has spectral stability: {}".format(x, specStab))
# print("Sub: {} mean pred: {}".format(x, np.mean(psds)*60))
stabilities.sort(key=lambda x: x[1], reverse=True)
bestStab = stabilities[0][1]
news = [x for x in stabilities if x[1] > bestStab * 0.2]
print("Using {} subcarriers.".format(len(news)))
pxxs = []
for sub in news:
data = sub[2]
f, Pxx_den = signal.welch(data, Fs)
pxxs.append(Pxx_den)
#The mean PSD of the top subcarriers is then selected, with the
#peak taken as the final heart rate estimation.
meanPsd = np.mean(pxxs, axis=0)
return f[np.argmax(meanPsd)] * 60
