Pandas resample：非精确时间戳重采样全为NaN的原因问询

我有一个从设备获取的DataFrame，其timestamp为非精确秒级，示例数据如下：

hr                                                                                                                       
timestamp                                                                                                                                            
2022-11-02 20:23:20.850611  72                                                                                                                       
2022-11-02 20:23:21.868609  71                                                                                                                       
2022-11-02 20:23:22.932606  71                                                                                                                       
2022-11-02 20:23:24.057612  72                                                                                                                       
2022-11-02 20:23:25.182701  71                                                                                                                       
2022-11-02 20:23:25.932692  74                                                                                                                       
2022-11-02 20:23:27.057694  72                                                                                                                       
2022-11-02 20:23:28.182689  72                                                                                                                       
2022-11-02 20:23:28.932730  72                                                                                                                       
2022-11-02 20:23:30.057686  71                                                                                                                       
2022-11-02 20:23:31.182692  69                                                                                                                       
2022-11-02 20:23:31.932689  68                                                                                                                       
2022-11-02 20:23:33.057890  67                                                                                                                       
2022-11-02 20:23:34.182879  68                                                                                                                       
2022-11-02 20:23:34.932871  68                                                                                                                       
2022-11-02 20:23:36.057870  66                                                                                                                       
2022-11-02 20:23:37.182873  68                                                                                                                       
2022-11-02 20:23:37.933040  72                                                                                                                       
2022-11-02 20:23:39.058044  73                                                                                                                       
2022-11-02 20:23:40.183046  78                                                                                                                       
2022-11-02 20:23:40.959045  84                                                                                                                       
2022-11-02 20:23:42.058044  84                                                                                                                       
2022-11-02 20:23:43.183052  75                                                                                                                       
2022-11-02 20:23:43.936050  73                                                                                                                       
2022-11-02 20:23:45.058047  81  

我尝试使用以下代码进行重采样，获取精确秒级、250ms分辨率的估计值：

df_resamp = df.resample('250ms').interpolate('cubic')
print(df_resamp.head(30))

但返回结果全为NaN：

timestamp                                                                                                                                            
2022-11-02 20:23:20.750 NaN                                                                                                                          
2022-11-02 20:23:21.000 NaN                                                                                                                          
2022-11-02 20:23:21.250 NaN                                                                                                                          
2022-11-02 20:23:21.500 NaN                                                                                                                          
2022-11-02 20:23:21.750 NaN                                                                                                                          
2022-11-02 20:23:22.000 NaN                                                                                                                          
2022-11-02 20:23:22.250 NaN                                                                                                                          
2022-11-02 20:23:22.500 NaN                                                                                                                          
2022-11-02 20:23:22.750 NaN                                                                                                                          
2022-11-02 20:23:23.000 NaN                                                                                                                          
2022-11-02 20:23:23.250 NaN                                                                                                                          
2022-11-02 20:23:23.500 NaN                                                                                                                          
2022-11-02 20:23:23.750 NaN                                                                                                                          
2022-11-02 20:23:24.000 NaN                                                                                                                          
2022-11-02 20:23:24.250 NaN                                                                                                                          
2022-11-02 20:23:24.500 NaN                                                                                                                          
2022-11-02 20:23:24.750 NaN                                                                                                                          
2022-11-02 20:23:25.000 NaN                                                                                                                          
2022-11-02 20:23:25.250 NaN                                                                                                                          
2022-11-02 20:23:25.500 NaN                                                                                                                          
2022-11-02 20:23:25.750 NaN                                                                                                                          
2022-11-02 20:23:26.000 NaN                                                                                                                          
2022-11-02 20:23:26.250 NaN                                                                                                                          
2022-11-02 20:23:26.500 NaN                                                                                                                          
2022-11-02 20:23:26.750 NaN                                                                                                                          
2022-11-02 20:23:27.000 NaN                                                                                                                          
2022-11-02 20:23:27.250 NaN                                                                                                                          
2022-11-02 20:23:27.500 NaN                                                                                                                          
2022-11-02 20:23:27.750 NaN                                                                                                                          
2022-11-02 20:23:28.000 NaN 

请问为何会出现全NaN的情况？

问题原因 #

• 重采样锚点不匹配：resample('250ms')默认以整点（如分钟的0秒）为锚点生成时间序列，你的原始数据起始时间是2022-11-02 20:23:20.850611，而重采样生成的第一个时间点是2022-11-02 20:23:20.750，早于原始数据起始时间，且后续所有重采样时间点和原始数据的时间戳完全无重叠，没有可用的非NaN值启动插值。
• 插值依赖已有数据：interpolate必须基于已存在的非NaN值计算，若重采样后的序列中没有保留原始数据的时间点对应值，就无法进行插值计算。

解决方案 #

方式1：锚定重采样起始点为原始数据开头 #

让重采样序列从原始数据的第一个时间戳开始对齐，确保原始数据点被包含在内：

# 以原始数据起始时间为锚点重采样
df_resamp = df.resample('250ms', origin='start').interpolate('cubic')
print(df_resamp.head(30))

方式2：先保留原始数据点再插值 #

先生成250ms间隔的空序列，填充原始数据值后再插值：

# 生成250ms间隔序列
df_resamp = df.resample('250ms').asfreq()
# 将原始数据的值填充到对应时间点
df_resamp.loc[df.index, 'hr'] = df['hr']
# 执行三次样条插值
df_resamp = df_resamp.interpolate('cubic')
print(df_resamp.head(30))

补充：强制从精确秒级开始 #

如果需要严格从整点秒（如2022-11-02 20:23:21.000）作为起始点，可以手动指定锚点：

from pandas import Timestamp
# 指定精确秒作为重采样锚点
df_resamp = df.resample('250ms', origin=Timestamp('2022-11-02 20:23:21')).interpolate('cubic')

内容的提问来源于stack exchange，提问作者Jean Nobrega