上面的一个答案中有一个mab的注释，它有这个方法。瓶颈有move_mean，这是一个简单的移动平均:

import numpy as np
import bottleneck as bn

a = np.arange(10) + np.random.random(10)

mva = bn.move_mean(a, window=2, min_count=1)

Min_count是一个很方便的参数，它可以取数组中该点的移动平均值。如果你不设置min_count，它将等于window，并且直到window points的所有内容都将是nan。

我的解决方案是基于维基百科上的“简单移动平均”。

from numba import jit
@jit
def sma(x, N):
    s = np.zeros_like(x)
    k = 1 / N
    s[0] = x[0] * k
    for i in range(1, N + 1):
        s[i] = s[i - 1] + x[i] * k
    for i in range(N, x.shape[0]):
        s[i] = s[i - 1] + (x[i] - x[i - N]) * k
    s = s[N - 1:]
    return s

与之前建议的解决方案相比，它比scipy最快的解决方案“uniform_filter1d”快两倍，并且具有相同的错误顺序。 速度测试:

import numpy as np    
x = np.random.random(10000000)
N = 1000

from scipy.ndimage.filters import uniform_filter1d
%timeit uniform_filter1d(x, size=N)
95.7 ms ± 9.34 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
%timeit sma(x, N)
47.3 ms ± 3.42 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

错误的比较:

np.max(np.abs(np.convolve(x, np.ones((N,))/N, mode='valid') - uniform_filter1d(x, size=N, mode='constant', origin=-(N//2))[:-(N-1)]))
8.604228440844963e-14
np.max(np.abs(np.convolve(x, np.ones((N,))/N, mode='valid') - sma(x, N)))
1.41886502547095e-13

上面有很多关于计算运行平均值的答案。我的回答增加了两个额外的特征:

忽略nan值 计算N个相邻值的平均值，不包括兴趣值本身

这第二个特征对于确定哪些值与总体趋势有一定的差异特别有用。

我使用numpy。cumsum，因为这是最省时的方法(参见上面Alleo的回答)。

N=10 # number of points to test on each side of point of interest, best if even
padded_x = np.insert(np.insert( np.insert(x, len(x), np.empty(int(N/2))*np.nan), 0, np.empty(int(N/2))*np.nan ),0,0)
n_nan = np.cumsum(np.isnan(padded_x))
cumsum = np.nancumsum(padded_x) 
window_sum = cumsum[N+1:] - cumsum[:-(N+1)] - x # subtract value of interest from sum of all values within window
window_n_nan = n_nan[N+1:] - n_nan[:-(N+1)] - np.isnan(x)
window_n_values = (N - window_n_nan)
movavg = (window_sum) / (window_n_values)

这段代码只适用于偶数n。它可以通过改变np来调整奇数。插入padded_x和n_nan。

输出示例(黑色为raw，蓝色为movavg):

这段代码可以很容易地修改，以删除从小于cutoff = 3的非nan值计算的所有移动平均值。

window_n_values = (N - window_n_nan).astype(float) # dtype must be float to set some values to nan
cutoff = 3
window_n_values[window_n_values<cutoff] = np.nan
movavg = (window_sum) / (window_n_values)

如果你选择自己生成，而不是使用现有的库，请注意浮点错误并尽量减少其影响:

class SumAccumulator:
    def __init__(self):
        self.values = [0]
        self.count = 0

    def add( self, val ):
        self.values.append( val )
        self.count = self.count + 1
        i = self.count
        while i & 0x01:
            i = i >> 1
            v0 = self.values.pop()
            v1 = self.values.pop()
            self.values.append( v0 + v1 )

    def get_total(self):
        return sum( reversed(self.values) )

    def get_size( self ):
        return self.count

如果所有的值都是大致相同的数量级，那么这将通过始终添加大致相似的数量级值来帮助保持精度。

更新:已经提出了更有效的解决方案，scipy的uniform_filter1d可能是“标准”第三方库中最好的，还有一些更新的或专门的库可用。

你可以用np。卷积得到:

np.convolve(x, np.ones(N)/N, mode='valid')

解释

The running mean is a case of the mathematical operation of convolution. For the running mean, you slide a window along the input and compute the mean of the window's contents. For discrete 1D signals, convolution is the same thing, except instead of the mean you compute an arbitrary linear combination, i.e., multiply each element by a corresponding coefficient and add up the results. Those coefficients, one for each position in the window, are sometimes called the convolution kernel. The arithmetic mean of N values is (x_1 + x_2 + ... + x_N) / N, so the corresponding kernel is (1/N, 1/N, ..., 1/N), and that's exactly what we get by using np.ones(N)/N.

边缘

np的模态参数。Convolve指定如何处理边缘。我在这里选择有效模式，因为我认为这是大多数人期望的运行方式，但您可能有其他优先级。下面是一个图表，说明了模式之间的差异:

import numpy as np
import matplotlib.pyplot as plt
modes = ['full', 'same', 'valid']
for m in modes:
    plt.plot(np.convolve(np.ones(200), np.ones(50)/50, mode=m));
plt.axis([-10, 251, -.1, 1.1]);
plt.legend(modes, loc='lower center');
plt.show()

高效的解决方案

卷积比直接的方法好得多，但(我猜)它使用FFT，因此相当慢。但是，下面的方法特别适用于计算运行平均值

def running_mean(x, N):
    cumsum = numpy.cumsum(numpy.insert(x, 0, 0)) 
    return (cumsum[N:] - cumsum[:-N]) / float(N)

要检查的代码

In[3]: x = numpy.random.random(100000)
In[4]: N = 1000
In[5]: %timeit result1 = numpy.convolve(x, numpy.ones((N,))/N, mode='valid')
10 loops, best of 3: 41.4 ms per loop
In[6]: %timeit result2 = running_mean(x, N)
1000 loops, best of 3: 1.04 ms per loop

注意numpy。allclose(result1, result2)为True，两个方法等价。 N越大，时间差异越大。

警告:虽然cumsum更快，但会增加浮点错误，这可能导致您的结果无效/不正确/不可接受

这里的评论指出了这个浮点错误问题，但我在回答中让它更明显。

# demonstrate loss of precision with only 100,000 points
np.random.seed(42)
x = np.random.randn(100000)+1e6
y1 = running_mean_convolve(x, 10)
y2 = running_mean_cumsum(x, 10)
assert np.allclose(y1, y2, rtol=1e-12, atol=0)

the more points you accumulate over the greater the floating point error (so 1e5 points is noticable, 1e6 points is more significant, more than 1e6 and you may want to resetting the accumulators) you can cheat by using np.longdouble but your floating point error still will get significant for relatively large number of points (around >1e5 but depends on your data) you can plot the error and see it increasing relatively fast the convolve solution is slower but does not have this floating point loss of precision the uniform_filter1d solution is faster than this cumsum solution AND does not have this floating point loss of precision