Welch's method 只计算信号的多个重叠段的周期图,然后取跨段的平均值。这有效地以分辨率换取频域中的降噪。
但是,为每个小段执行大量单独的 FFT 将比为较大的段计算更少的 FFT 更昂贵。根据您的需要,您可以不使用 Welch 的方法,但将信号分成更大的片段,和/或减少它们之间的重叠(这两种方法都会减少 PSD 的方差)。
from matplotlib import pyplot as plt
# default parameters
fs1, ps1 = welch(x, sfreq, nperseg=256, noverlap=128)
# 8x the segment size, keeping the proportional overlap the same
fs2, ps2 = welch(x, sfreq, nperseg=2048, noverlap=1024)
# no overlap between the segments
fs3, ps3 = welch(x, sfreq, nperseg=2048, noverlap=0)
fig, ax1 = plt.subplots(1, 1)
ax1.hold(True)
ax1.loglog(fs1, ps1, label='Welch, defaults')
ax1.loglog(fs2, ps2, label='length=2048, overlap=1024')
ax1.loglog(fs3, ps3, label='length=2048, overlap=0')
ax1.legend(loc=2, fancybox=True)
增加分段大小并减少重叠量可以显着提高性能:
In [1]: %timeit welch(x, sfreq)
1 loops, best of 3: 262 ms per loop
In [2]: %timeit welch(x, sfreq, nperseg=2048, noverlap=1024)
10 loops, best of 3: 46.4 ms per loop
In [3]: %timeit welch(x, sfreq, nperseg=2048, noverlap=0)
10 loops, best of 3: 23.2 ms per loop
请注意,最好使用 2 的幂作为窗口大小,因为对长度为 2 的幂的信号进行 FFT 会更快。
更新
您可能会考虑尝试的另一件简单的事情是使用以 50Hz 为中心的陷波滤波器对您的信号进行带通滤波。滤波后信号的包络可以让您衡量您的信号在一段时间内包含多少 50Hz 功率。
from scipy.signal import filter_design, filtfilt
# a signal whose power at 50Hz varies over time
sfreq = 128.
nsamples = 454912
time = np.arange(nsamples) / sfreq
sinusoid = np.sin(2 * np.pi * 50 * time)
pow50hz = np.zeros(nsamples)
pow50hz[nsamples / 4: 3 * nsamples / 4] = 1
x = pow50hz * sinusoid + np.random.randn(nsamples)
# Chebyshev notch filter centered on 50Hz
nyquist = sfreq / 2.
b, a = filter_design.iirfilter(3, (49. / nyquist, 51. / nyquist), rs=10,
ftype='cheby2')
# filter the signal
xfilt = filtfilt(b, a, x)
fig, ax2 = plt.subplots(1, 1)
ax2.hold(True)
ax2.plot(time[::10], x[::10], label='Raw signal')
ax2.plot(time[::10], xfilt[::10], label='50Hz bandpass-filtered')
ax2.set_xlim(time[0], time[-1])
ax2.set_xlabel('Time')
ax2.legend(fancybox=True)
更新 2
看到@hotpaw2 的回答后,我决定尝试实现Goertzel algorithm,只是为了好玩。不幸的是,它是一种递归算法(因此不能随着时间的推移进行矢量化),所以我决定自己编写一个 Cython 版本:
# cython: boundscheck=False
# cython: wraparound=False
# cython: cdivision=True
from libc.math cimport cos, M_PI
cpdef double goertzel(double[:] x, double ft, double fs=1.):
"""
The Goertzel algorithm is an efficient method for evaluating single terms
in the Discrete Fourier Transform (DFT) of a signal. It is particularly
useful for measuring the power of individual tones.
Arguments
----------
x double array [nt,]; the signal to be decomposed
ft double scalar; the target frequency at which to evaluate the DFT
fs double scalar; the sample rate of x (same units as ft, default=1)
Returns
----------
p double scalar; the DFT coefficient corresponding to ft
See: <http://en.wikipedia.org/wiki/Goertzel_algorithm>
"""
cdef:
double s
double s_prev = 0
double s_prev2 = 0
double coeff = 2 * cos(2 * M_PI * (ft / fs))
Py_ssize_t N = x.shape[0]
Py_ssize_t ii
for ii in range(N):
s = x[ii] + (coeff * s_prev) - s_prev2
s_prev2 = s_prev
s_prev = s
return s_prev2 * s_prev2 + s_prev * s_prev - coeff * s_prev * s_prev2
它的作用如下:
freqs = np.linspace(49, 51, 1000)
pows = np.array([goertzel(x, ff, sfreq) for ff in freqs])
fig, ax = plt.subplots(1, 1)
ax.plot(freqs, pows, label='DFT coefficients')
ax.set_xlabel('Frequency (Hz)')
ax.legend(loc=1)
速度非常快:
In [1]: %timeit goertzel(x, 50, sfreq)
1000 loops, best of 3: 1.98 ms per loop
显然,这种方法只有在您只对单个频率而不是频率范围感兴趣时才有意义。