对，经过很多的折腾，这似乎起作用了....我很想听听有没有更好的方法来做这件事。我的方法很乱！！

#Set the gap length to detect
    gap_length = 40
        
    #Create a boolean mask to identify where signal changes occur
    mask = df['change'].ne(df['change'].shift()) | (df['change'].rolling(gap_length+1).sum() == 0)
        
    #Generate a grouping identifier for each contiguous sequence of "ON" signals
    group_id = mask.cumsum()
        
    #Use groupby and cumcount to generate a count of "ON" signals up to each row
    df['mask'] = df.groupby(group_id).cumcount() + 1
        
    #Replace the 0's with NaN values
    df['mask'][df['mask'] > 1] = 0
        
    #Apply rolling window - signal inverted, so *1+1 to flip it and center it 
    df['mask'] = (df['mask'].rolling(window=3, center=True).min()*-1)+1

    # pick out where the filled in mask changes
    df['on_number'] = d['mask'].diff()

    # Keep only the first "ON" signal
    df['on_number'][df['on_number'] < 0] = 0

    # Sum on signals to get a rolling count of true "ONs"
    df['on_number'] = df['on_number'].cumsum()