scripts/impute_expand.py

import pandas as pd
import numpy as np
from matplotlib import pyplot as plt
import json
import os
from tqdm import tqdm
from concurrent.futures import ThreadPoolExecutor

# disable the warning from pandas
pd.options.mode.chained_assignment = None

id_info = json.load(open('expanded_id_info.json'))
id_list = list(id_info.keys())

print('Reading data...')
dfs = {}

def read_parquet(id):
    return id, pd.read_parquet(f'./slim_data/id_{id}.parquet')

with ThreadPoolExecutor(max_workers=100) as executor:
    futures = [executor.submit(read_parquet, id) for id in id_list]
    for future in tqdm(futures):
        id, df = future.result()
        dfs[id] = df

sensor_downtimes = {}
for id in tqdm(id_list):

    tqdm.write(f"Preparing ID: {id}")

    df = dfs[str(id)].copy()
    df['DATA_AS_OF'] = pd.to_datetime(df['DATA_AS_OF'])

    # round the DATA_AS_OF to the nearest 5 minutes
    df['DATA_AS_OF'] = df['DATA_AS_OF'].dt.round('5min') 
    tqdm.write(f"De-duplicating ID: {id}")
    #find duplicate timestamps
    missing_gap0 = df['DATA_AS_OF'].diff().dt.total_seconds() == 0
    # missing_gap5 = df['DATA_AS_OF'].diff().dt.total_seconds() > (60*5)
    missing_gap0 = missing_gap0[missing_gap0].index

    to_remove=[]
    for ind in missing_gap0:
        # if the DATA_AS_OF of ind is 10min smaller than ind+1 then add 5 min to ind
        if ind+1 == len(df):
            to_remove.append(ind)
        elif df['DATA_AS_OF'].iloc[ind] + pd.Timedelta('10min') == df['DATA_AS_OF'].iloc[ind+1]:
            # print(f'adjust {ind}')
            df['DATA_AS_OF'].iloc[ind] = df['DATA_AS_OF'].iloc[ind] + pd.Timedelta('5min')
        else:
            # remove the row of ind
            to_remove.append(ind)
    df = df.drop(to_remove)
    df = df.reset_index(drop=True)
    # check missing_gap0 again
    missing_gap0 = df['DATA_AS_OF'].diff().dt.total_seconds() == 0
    missing_gap0 = missing_gap0[missing_gap0].index
    assert len(missing_gap0) == 0, 'There are still duplicate timestamps'

    tqdm.write(f"Small gaps ID: {id}")
    # get the missing gaps of 5~2h
    threshold_time = 120 # in min

    missing_gap15 = df['DATA_AS_OF'].diff().dt.total_seconds() <= (60*threshold_time) # make it 2h
    missing_gap5 = df['DATA_AS_OF'].diff().dt.total_seconds() > (60*5)
    missing_gap = missing_gap15 & missing_gap5

    missing_gap = missing_gap[missing_gap].index

    def linear_impute(start_idx, end_idx):
        start_time = df['DATA_AS_OF'][start_idx]
        end_time = df['DATA_AS_OF'][end_idx]
        start_speed = df['SPEED'][start_idx]
        end_speed = df['SPEED'][end_idx]
        start_travel_time = df['TRAVEL_TIME'][start_idx]
        end_travel_time = df['TRAVEL_TIME'][end_idx]
        gap = end_time - start_time
        gap = gap.total_seconds()
        new_rows = []
        for j in range(1, int(gap // 300)):
            new_rows.append({
                'DATA_AS_OF': start_time + pd.Timedelta(f'{j*5}min'),
                'SPEED': start_speed + (end_speed - start_speed) * j / (gap // 300),
                'TRAVEL_TIME': start_travel_time + (end_travel_time - start_travel_time) * j / (gap // 300)
            })
        return new_rows

    with ThreadPoolExecutor(max_workers=100) as executor:
        futures = [executor.submit(linear_impute, i - 1, i) for i in missing_gap]
        results = [future.result() for future in tqdm(futures)]

    # Flatten the list of lists
    new_rows = [item for sublist in results for item in sublist]

    # Create a DataFrame from the new rows and concatenate it with the original DataFrame
    new_df = pd.DataFrame(new_rows)
    df = pd.concat([df, new_df], ignore_index=True)

    # sort by the DATA_AS_OF
    df = df.sort_values('DATA_AS_OF')
    df = df.reset_index(drop=True)

    # check again
    missing_gap15 = df['DATA_AS_OF'].diff().dt.total_seconds() <= (60*threshold_time)
    missing_gap5 = df['DATA_AS_OF'].diff().dt.total_seconds() > (60*5)
    missing_gap = missing_gap15 & missing_gap5

    missing_gap = missing_gap[missing_gap].index
    assert len(missing_gap) == 0, 'There are still missing gaps'

    tqdm.write(f"Large gaps ID: {id}")
    missing_gaplarge = df['DATA_AS_OF'].diff().dt.total_seconds() > (60*threshold_time) # make it 2h #30 min is ok

    missing_gaplarge = missing_gaplarge[missing_gaplarge].index

    def zero_impute(start_idx, end_idx):
        start_time = df['DATA_AS_OF'][start_idx]
        end_time = df['DATA_AS_OF'][end_idx]
        gap = end_time - start_time
        gap = gap.total_seconds()
        new_rows = []
        for j in range(1, int(gap // 300)):
            new_rows.append({
                'DATA_AS_OF': start_time + pd.Timedelta(f'{j*5}min'),
                'SPEED': 0,
                'TRAVEL_TIME': 0
            })
        return new_rows

    with ThreadPoolExecutor(max_workers=100) as executor:
        futures = [executor.submit(zero_impute, i - 1, i) for i in missing_gaplarge]
        results = [future.result() for future in tqdm(futures)]

    # Flatten the list of lists
    new_rows = [item for sublist in results for item in sublist]

    # Create a DataFrame from the new rows and concatenate it with the original DataFrame
    new_df = pd.DataFrame(new_rows)
    df = pd.concat([df, new_df], ignore_index=True)

    # sort by the DATA_AS_OF
    df = df.sort_values('DATA_AS_OF')
    df = df.reset_index(drop=True)

    # check again
    missing_anygap = df['DATA_AS_OF'].diff().dt.total_seconds() > (60*5)
    missing_anygap = missing_anygap[missing_anygap].index
    assert len(missing_anygap) == 0, 'There are still missing gaps'

    tqdm.write(f"Sensor downtime ID: {id}")
    # get sensor downtime
    # get all the SPEED=0
    zero_speed = df['SPEED']==0
    speed_goes_down = df['SPEED'].diff() < 0
    speed_goes_up = df['SPEED'].diff(-1) < 0

    speed_goto_zero = zero_speed & speed_goes_down
    speed_goto_zero = speed_goto_zero[speed_goto_zero].index

    speed_gofrom_zero = zero_speed & speed_goes_up
    speed_gofrom_zero = speed_gofrom_zero[speed_gofrom_zero].index

    threshold_step = threshold_time//5
    sensor_downtime = {}
    i=0
    for start, end in zip(speed_goto_zero, speed_gofrom_zero):
        if end - start > threshold_step:
            sensor_downtime[i] = {'time':(df['DATA_AS_OF'][start], df['DATA_AS_OF'][end]), 'index':(start, end)} 
            i+=1


    # if the downtime is between 0:00 to 6:00 then remove the downtime from dictionary
    def check_22_6(time):
        if time.hour >= 0 and time.hour < 6:
            return True
        elif time.hour >=22:
            return True
        else:
            return False
    for key in list(sensor_downtime.keys()):
        if check_22_6(sensor_downtime[key]['time'][0]) and check_22_6(sensor_downtime[key]['time'][1]):
            del sensor_downtime[key]

    # convert the 'time' segment to string
    for key in sensor_downtime.keys():
        sensor_downtime[key]['time'] = (str(sensor_downtime[key]['time'][0]), str(sensor_downtime[key]['time'][1]))

    sensor_downtime = dict(enumerate(sensor_downtime.values()))

    df.to_parquet(f'./expanded_impute_data/id_{id}.parquet')
    sensor_downtimes[id] = sensor_downtime
    id_info[str(id)]['sensor_downtime'] = sensor_downtime
    id_info[str(id)]['len'] = len(df)

json.dump(id_info, open('./expanded_impute_data/id_info_imputed.json', 'w'), indent=4)

json.dump(sensor_downtimes, open('./expanded_impute_data/sensor_downtimes.json', 'w'), indent=4)
print('Done!')