python里去除极值的三种方法 (rolling_median)

3 ways to remove outliers from your data

https://ocefpaf.github.io/python4oceanographers/blog/2015/03/16/outlier_detection/

According to Google Analytics, my post "Dealing
with spiky data", is by far the most visited on the blog. I think that the reasons are: it is one of the oldest posts, and it is a real problem that people have to deal everyday.

Recently I found an amazing series of post writing by Bugra on how to perform outlier detection using FFT,
median filtering, Gaussian
processes, and MCMC

I will test out the low hanging fruit (FFT and median filtering) using the same data from my original
post.

In [2]:

from datetime import datetime
from pandas import read_table

fname = './data/spikey_v.dat'
cols = ['j', 'u', 'v', 'temp', 'sal', 'y', 'mn', 'd', 'h', 'mi']

df = read_table(fname , delim_whitespace=True, names=cols)

df.index = [datetime(*x) for x in zip(df['y'], df['mn'], df['d'], df['h'], df['mi'])]
df = df.drop(['y', 'mn', 'd', 'h', 'mi'], axis=1)

df.head()

Out[2]:

	j	u	v	temp	sal
1994-11-22 04:00:00	1056.1667	-0.1	0.0	23.9	34.6
1994-11-22 05:00:00	1056.2083	-0.2	0.7	23.9	34.6
1994-11-22 06:00:00	1056.2500	-0.1	2.0	23.9	34.6
1994-11-22 07:00:00	1056.2917	0.0	3.1	23.9	34.6
1994-11-22 08:00:00	1056.3333	-0.1	2.7	23.9	34.6

Starting with Bugra's

get_median_filtered()

we
have:

In [3]:

import numpy as np

def get_median_filtered(signal, threshold=3):
signal = signal.copy()
difference = np.abs(signal - np.median(signal))
median_difference = np.median(difference)
if median_difference == 0:
s = 0
else:
s = difference / float(median_difference)
mask = s > threshold
signal[mask] = np.median(signal)
return signal

In [4]:

import matplotlib.pyplot as plt

figsize = (7, 2.75)
kw = dict(marker='o', linestyle='none', color='r', alpha=0.3)

df['u_medf'] = get_median_filtered(df['u'].values, threshold=3)

outlier_idx = np.where(df['u_medf'].values != df['u'].values)[0]

fig, ax = plt.subplots(figsize=figsize)
df['u'].plot()
df['u'][outlier_idx].plot(**kw)
_ = ax.set_ylim(-50, 50)

Not bad. Still, it missed the two lower spikes. Let's go for the

detect_outlier_position_by_fft()

.

In [5]:

def detect_outlier_position_by_fft(signal, threshold_freq=0.1,
frequency_amplitude=.001):
signal = signal.copy()
fft_of_signal = np.fft.fft(signal)
outlier = np.max(signal) if abs(np.max(signal)) > abs(np.min(signal)) else np.min(signal)
if np.any(np.abs(fft_of_signal[threshold_freq:]) > frequency_amplitude):
index_of_outlier = np.where(signal == outlier)
return index_of_outlier[0]
else:
return None

In [6]:

outlier_idx = []

y = df['u'].values

opt = dict(threshold_freq=0.01, frequency_amplitude=0.001)

win = 20
for k in range(win*2, y.size, win):
idx = detect_outlier_position_by_fft(y[k-win:k+win], **opt)
if idx is not None:
outlier_idx.append(k + idx[0] - win)
outlier_idx = list(set(outlier_idx))

fig, ax = plt.subplots(figsize=(7, 2.75))

df['u'].plot()
df['u'][outlier_idx].plot(**kw)
_ = ax.set_ylim(-50, 50)

Not sure if this method is the best here... Maybe if the signal was contaminated by high frequency noise this method would perform better.

Inspired by Bugra's median filter let's try a

rolling_median

filter
using pandas.

In [7]:

from pandas import rolling_median

threshold = 3
df['pandas'] = rolling_median(df['u'], window=3, center=True).fillna(method='bfill').fillna(method='ffill')

difference = np.abs(df['u'] - df['pandas'])
outlier_idx = difference > threshold

fig, ax = plt.subplots(figsize=figsize)
df['u'].plot()
df['u'][outlier_idx].plot(**kw)
_ = ax.set_ylim(-50, 50)

Update: A friend, that knows this data, challenged me to use the same
technique on $v$. Here it is:

In [8]:

df['pandas'] = rolling_median(df['v'], window=3, center=True).fillna(method='bfill').fillna(method='ffill')

difference = np.abs(df['v'] - df['pandas'])
outlier_2 = difference > 2
outlier_3 = difference > 3

fig, ax = plt.subplots(figsize=figsize)
df['v'].plot()
df['v'][outlier_2].plot(**kw)
kw.update(color='g', marker='*', alpha=1)
df['v'][outlier_3].plot(**kw)
_ = ax.set_ylim(-50, 60)

Note that I had to reduce the threshold from 3 -> 2 to get them all.