首先通过 join 改为原始:

First select columns by xs, apply rolling and create MultiIndex, last join to original:

df = xf.xs('bezel', axis=1, level=1).rolling(window=2).mean()
df.columns = [df.columns, ['roll2'] * len(df.columns)]

另一个使用rename的解决方案:

Another solution with rename:

df = (xf.xs('bezel', axis=1, level=1, drop_level=False).rolling(window=2).mean()
        .rename(columns={'bezel':'roll2'}))

print (df)
human           mike      dave      matt
               roll2     roll2     roll2
2018-01-31       NaN       NaN       NaN
2018-02-28  0.439297  0.756530  0.407606
2018-03-31  0.432513  0.436660  0.430393
2018-04-30  0.258736  0.469610  0.850996
2018-05-31  0.278869  0.698822  0.561285

xf = xf.join(df)
print (xf)
human           mike                          dave                      \
measure         spin      drag     bezel      spin      drag     bezel   
2018-01-31  0.811030  0.114535  0.326579  0.597781  0.194064  0.659795   
2018-02-28  0.774971  0.400888  0.552016  0.385539  0.582351  0.853266   
2018-03-31  0.794427  0.653428  0.313010  0.996514  0.524999  0.020055   
2018-04-30  0.307418  0.131451  0.204462  0.049346  0.198878  0.919165   
2018-05-31  0.196374  0.421594  0.353276  0.244024  0.930992  0.478479   

human           matt                          mike                dave  \
measure         spin      drag     bezel      roll     roll2     roll2   
2018-01-31  0.769308  0.657963  0.691395       NaN       NaN       NaN   
2018-02-28  0.564884  0.026864  0.123818  0.439297  0.439297  0.756530   
2018-03-31  0.755440  0.698443  0.736967  0.432513  0.432513  0.436660   
2018-04-30  0.782908  0.919064  0.965025  0.258736  0.258736  0.469610   
2018-05-31  0.414085  0.339771  0.157545  0.278869  0.278869  0.698822   

human           matt  
measure        roll2  
2018-01-31       NaN  
2018-02-28  0.407606  
2018-03-31  0.430393  
2018-04-30  0.850996  
2018-05-31  0.561285  

最后一次必要的排序MultiIndex:

xf = xf.join(df).sort_index(axis=1)