Scene 1

Scene 2

Predictive models developed for 11 equipment types, 2 deprioritized due to lack of data

The conveyor has been operating with certain fixed parameters to maintain a certain conveying rate (Tones per hour). Attempting to predict deviation in the parameters, through a relationship between conveyor current and conveying rate

To predict potential clogging by detecting deviation of control valve opening % from target value, as identified from historical data analysis

Linear equation of control valve % opening based on the flow rate through the valve:    For e.g.,   % opening of control valve 1 = 25.167+ 4.46 x 10-4 * Flow rate in valve 1             % opening of control valve 2 = 16.89 + 8.18 x 10-4 * Flow rate in valve 2

Linear equation of Conveyor motor current (I):     Conveyor Motor Current = 25.6 + 8.3 x 10-3 * Conveying rate

Deployed across 9 control valves across BF, Coke, SMS, HSM, Sinter Clogging incident in PCI control valve detected in March, and break-down prevented

Fans 2 To detect deviation in ID fan motor/ Mill/ Bag House by calculating difference of actual current from calculated value To predict deviations in fan parameters (like current), by identifying correlation between equipment and process parameters Linear equation of BF ID fan motor current (I):  I Fan  =  9.00 - 5.91 x 10-6 * Pressure before ID fan * Flow rate across ID fan Linear equation of ESP outlet pressure,  P ESP   = 65.39 – 1.397*fan 1_current – 1.46* Fan 2_current – 0.939 * Opacity

Deployed in ID fan motor and  Stockhouse  dedusting fan

Deployed in two conveyors in RMHS in the last 2-3 weeks On historical data, the linear model has been able to identify an instance of deviation in motor current relative to TPH, which was found to be linked to a conveyor joint maintenance activity

Linear model with flow rate gave best performance and was able to predict
failure (2/2)

Scene 3

Pump 4 To detect deviation of pump behavior, based on correlation of equipment parameters with process parameters. Linear equation of Pump flow based on current Pump Flow= 2751.7 + 13.41*Pump 1 current + 8.32*Pump 2 current (sample model developed for a 2-by-3 pump; generalizable to other pumps through if-else logic deployed)

Heat exchanger 5 To detect deviation in heat exchanger operations by monitoring the change in temperature relative to differential pressure of the system Linear equation of  Rise in temperature of combustion air = -89.49 + 0.87* Waste gas input temp – 0.862* Differential pressure of waste gas + 0.61* differential pressure of combustion air

Deployed in two pumps in BF-furnace throat stave cooling pump and hearth stave cooling pump

Deployed in the BF stock-house heat exchanger, and pending for deployment SMS heat exchanger for shutdown

Cylinder 6 To detect deviation in SMS hydraulic cylinder by monitoring the time required for completion of the process SMS cylinder- Alarm based on process time of cylinder from the time set-point is triggered

Deployed as a real-time tracking of the response time, alarm configured at 55seconds

Scene 4

Gearbox 7 To detect deviation in gearbox, based on correlation between load and current Load represented by torque and rpm for the SMS tilt gearbox Tilt Drive 1 current = Tilt Drive 1 Torque Nm *(-41.0498) + Tilt Drive 1 speed RPM * (-0.5411) + 19.7127 For BLT gearbox,  Tilting current  = ( Tilting Angle * 0.20922)  - 3.88386

Motors 8 To use equipment parameters like vibration, current, and RPM to predict abnormalities in motor operations For Sinter cooler fan, deviation in vibration acting as a leading indicator for motor break-down For SMS ID fan, Vibration = 0.0323 + 0.00025*Speed – 0.00084*Pressure

Both models deployed, but deviations observed, model logic needs to be refined

Logic developed for Sinter cooling system- On historical data, model shows the Sinter cooling system fan breakdown alarm pattern could have been detected 5+ days before actual break-down based on vibration and RPM Similar logic deployed for SMS ID fan

Shut-off valves 9 To predict abnormal valve operations, based on response time of the valve, which is the most common failure mode Tracking the time between command to close and the closed feedback, to understand the normal operating time

Valve malfunction in  mould  cooling shut-off valve detected through deviation observed in response time (3000+ sec compared to normal of under 15 seconds)

Scene 5

Lubrication system 10 To detect deviation in BLT cooling system by observing the trend of nitrogen flow vis-à-vis gearbox casing temperatures To detect anomalies in the BLT greasing system by looking at greasing cycle completion time Greasing system cycle time observed to ~40 seconds, alarm to be configured accordingly

Pipes 11 To detect potential clogging or leakages in pipelines by checking the correlation between flow and pressure Logic based on ratio of Flow and pressure to be maintained

Cooling system deployed; model to be retrained to improve performance with guidance from team Logic for L1 deployment of greasing cycle time monitoring to be  finalised

Currently deployed in BF scrubber pipes To be developed for Coke COG pipes

ROT motor (Table rolls) 12 To detect deviation in SMS hydraulic cylinder by monitoring the time required for completion of the process SMS cylinder- Alarm based on process time of cylinder from the time set-point is triggered

Gas  Analyser 13 To predict deviations in flow, as insufficient flow can throw incorrect gas composition results Most common failure mode linked to flow of gas to  analyser . However, there is no sensor to track the flow of gases to the  analyser  and therefore modeling not possible

Scene 6

Two part approach towards development of advanced analytics backed models

Critical equipment for cross deployment for each type to be identified – BF & Caster to be  prioritised Reliablility  squad, supported by McKinsey, to identify and train sub-teams for cross deployment Reliability squad to identify potential to leverage the integrated under-development IT infrastructure by TCS for additional infrastructure needs

Cross-deployment across critical equipment's across plants (week 13 onwards)

Model development & deployment to be  lead  by 7 member reliability squad supported by designated plant teams

POC  development for generic equipment types (12 weeks)

Model development lead by McKinsey team & supported by the 7-member reliability squad

Control valves Fans Conveyors Pumps Heat exchanges Table rolls Lubrication system Hydraulic systems Compressor Chiller Motors Transformers Cylinders Switchgears Shears Mandrels Spindle

Gearbox Pully Shut-off valves U-seal Storage bins Pipelines Analysers M/C WT, Stackers

Criticality for operations (e.g., equipment from BF & SMS  prioritised ) Data availability for all critical parameters Ease of data extraction (e.g.,  IBA  system based restrictions on  HSM , SMS) Cross deployment potential

Scene 7

Identification of equipment's  where understanding of critical parameters and  predictive model can help monitor equipment health Identifying the  types of failure  that need  to be modelled  e.g., bearing failure, winding failure

Final modelling  exercise with  critical parameters

Building of AA  model to identify  critical parameters

Identification  of parameters  and data  engineering

Understand process  from Operations team and  identify all process parameters using L1 HMI Identify tags of all critical parameters and  extract data for analysis Structure data by identifying time lags and  granularity  of data

Machine learning model to be tested on data to  identify critical parameters  impacting the output variable (as per failure to be studied) Discussion with process experts to help  shortlist 1-2 critical parameters for further modelling

Test linear model  to identify a relation of output and critical parameters Study historical data  using the identified equation to check if  model is able to predict instances  of failure

Models would be  productionized in L1 system;  Two columns added in L1 system – predicted value and delta of actual from predicted   Model will raise  alarm to operator  when  deviation from predicted value  is high

Scene 8

Valve operation: Predictive model built to detect clogging in injection lines and allow for timely corrective action

Historical data to be collected from L1 system  Predictive model to help determine target % opening based on operating parameters Deviation from this target % will result in an alarm for checking

To predict clogging in the injection line by calculating deviation of control valve opening % from target value, as identified from historical data analysis

Deployed across 8 control valves: BF: 1  HSM: 3 SMS: 2 Coke: 1 Sinter: 1

Scene 9

Linking process and equipment parameters for PCI control valve

IV-1 IV-2 IV-3 Pressure in vessel (bar) Weight in vessel (kg) Pressure in vessel (bar) Weight in vessel (kg) Pressure in vessel (bar) Weight in vessel (kg) Additional nitrogen 1 Flow rate Additional nitrogen 2 Flow rate Control valve 1 Control valve 2 To 19  Tuyeres   To 19  Tuyeres   Core fluidization Flow rate % opening bottom fluidization Flow rate % opening Pressure probes across injection lines CV1 – flow rate and % opening  CV2 – flow rate and % opening

Scene 10

Linear models were constructed to identify most suitable model

X = Flow rate, pressure in injection line

X = Flow rate, pressure in injection line, Core Fluidization

X = Flow rate, pressure in injection line, Core and bottom fluidization

Scene 11

Linear model with flow rate gave best performance and was able to predict failure (1/2)

% Opening = 45.85 + 0.0004384 *Flow rate – 8.32 * Inlet Pressure + 9.23 * Outlet Pressure

Scene 12

Performance metric Actual vs Predicted opening plot SD R2 RMSE MAE 0.74 6.09 4.83 12.01 Percentage of opening of control valve 2 =  47.45 +(0.0009814* Flow-(10.85*Inlet pressure)+(10.12*Outlet pressure)

Linear model with flow rate gave best performance and was able to predict failure (2/2)

Scene 13

A clogging incident was faced on 28 th  November, 2018 As can be seen from the above graph, the model would have indicated the onset of clogging as percentage opening of control valve was 15% more than the predicted opening

Delta = % actual opening - % predicted opening

The model was tested on historical data to identify periods of clogging

Scene 14

Alarm generated due to high deviation between predicted opening and actual opening, investigation reflected a break-down and valve replaced timely

Actual vs predicted value of % opening for the PCI control valve in April

28-04-2019 8.06201
39.3574
28-04-2019
2.67 days
40.2489
3(E04-2019
30-04-2019

Scene 15

Mixed gas line control valve: Across HSM furnace, data was extracted of parameters impacting the % opening of control valves

Furnace pressure  -  not considered as a model parameter as multiple lines use same furnace and pressure is measured at the furnace level

Scene 16

Linear model with flow rate gave best performance and was able to predict deviations with high accuracy

Performance metric Variable Importance Plot SD R2 0.86 RMSE 2.34 MAE 2.06 6.28 Percentage of opening of control valve = 4.3469936 + 0.0155574 * Flow set point - 0.0083060 * actual flow

Scene 17

SMS control valve data was extracted across the BOF system

OXYGEN
TROGEN
OXYGEN
NITROGEN
LANCE-I START RELEASE
2094 'bar
bar
20 66 'bar
19
LANCE-2 START RELEASE
Lance Not Active
- msrnuuegr
602 bar
N rn%rnu bar
INSTRUMEUT
6 46 fbar
CONTROL
Uanual
PV
0.0
, urn•nnüi
Nrn'/min
Nm'lmin
LANCE 1
Real
12628 mm
Lance 2
LANCE 2
Real H.qht
22698 mm
I-ance 1
GÅBLER VALVE
VALVE OPENNG
15701 mm
LANCE HELHT
RESPONSE
16776 mm
23
SP Nrn'"nin
PV
nrn'/min
00
o o m inun

Scene 18

% opening of control valve = 43.59-0.124* Lance 2 O 2  line flow + 10.23 * O 2  line back pressure + 0.02* Lance 2 O 2  set point

Scene 19

The model was tested on historical data to  identify deviations

Delta = % actual opening - % predicted opening Time

Due to challenges with data extraction (long lead time, high granularity, non-integrated system), 10 days data has been considered Working with OEM to simplify extraction (scheduling overnight in IBA system); model to be strengthened with data for 3 months

Scene 20

Predictive model in ID fan motor operations and in-process equipment (Mill & Bag house)

Problem statement Approach Historical data to be collected from L1 system  Exploratory analysis to identify important parameters Predictive model to help determine target ID fan motor current based on operating parameters Deviation from this target current will result in an alarm for checking  To predict issues in ID fan motor/ Mill/ Bag House by calculating deviation of motor current from target value, as identified from historical data analysis

Scene 21

Data across the mill and bag house equipment is collected for building predictive model

Conveyor current (Directly proportional to grinding rate)

Drag chain conveyor for transferring coal to mill

Scene 22

Linear models with different parameters was constructed to identify most suitable parameters

X = Pressure before ID fan * Flow rate across ID fan Y = ID fan motor current

Current in ID fan motor = 9.00 - 5.91 x 10-6 * Pressure before ID fan * Flow rate across ID fan

Scene 23

Delta = Actual current – Predicted current

As per model prediction, the current in ID fan motor was 23 A greater than the predicted current at 3:30 AM The model would have raised an alarm at 3:30 AM on 1st Jan indicating a blockage which is resulting in higher current consumption

The model was applied on historical data to identify periods  of clogging

Scene 24

The current rise in the ID fan motor was due to increase in pressure in the mill; A model was built to help predict clogging  in mill

Delta Pressure across the mill is largely dependent on the grinding rate in the mill. Current in the drag chain conveyor can be used as a proxy for the grinding rate Delta pressure in mill = 7.368 * Current in conveyor belt Delta pressure Current in conveyor belt

Scene 25

Deviation from predicted across time Zoomed in view of Dec 28 th  to Jan 1 st , 2019 Delta pressure in mill had been gradually increasing beyond the predicted delta pressure. The increase was observed starting on 29 th  December and it was severe on the night of 31 st  December Similar increase has been observed in other periods – Oct 1 st  and Dec 20 th Delta = Actual Pressure across mill – Predicted Pressure across mill Delta Time Delta Time

The model was applied on historical data to identify periods of clogging

Scene 26

Delta Pressure across the bag house is largely dependent on the grinding rate. Current in the drag chain conveyor can be used as a proxy for the grinding rate Delta pressure in bag house = 1.188 * Current in conveyor belt Baghouse Delta pressure Current in conveyor belt

A similar model is built for delta pressure in bag house and grinding rate

Scene 27

The model was applied on historical data of bag house to identify periods of clogging in  bag house

Difference = Actual Delta Pressure across bag house – Predicted Delta Pressure across bag house

Delta pressure in the bag house has been gradually increasing as the bag house gets older There have been few drops in pressure indicating a hole/ gap but there have been no instances of blockage in the bag house The pressure difference in the bag house has been gradually increasing and needs to be monitored to decide future action

Scene 28

For  Stockhouse  dedusting system fan, 4 months of data was extracted across several parameters

Voltage and current of ESP fields Motor current being measured  Chimney opacity ESP outlet temperature ESP outlet pressure Assuming opacity accounts for ESP outlet temperature and ESP field current

Scene 29

The fan performance, measured through outlet pressure, has been found to be linearly related to the fan system currents and the chimney opacity

Outlet pressure = 65.39 – 1.397*fan 1_current – 1.46* Fan 2_current – 0.939 * Opacity

Scene 30

The model was applied to historical data to identify periods of deviation of outlet pressure from the predicted trend

Delta = Actual outlet pressure – predicted outlet pressure

Currently the outlet pressure is being modeled on fan current and chimney opacity To improve model accuracy, can potentially consider ESP field current and temperature, if not highly correlated with opacity

Scene 31

For the coal conveyor CC2, data was extracted for past 1 year

RMHS COAL STACKING AND RECLAIMING CIRCUIT
ORE CIRCUIT
From
cc-l
Current
cc-g
Current
12 A
Belt scale
146128t
Totalizer -
76.50 T
cc-2
Current
26 A
Belt scale
Totalizer
1 OOT
COAL YARD
cc.10
PCICIRCUT
cc-3
Current
27 28 A
Current:
1749A
Current -
TATA
COAL PECLAmn.'G TO CORE MIS
ccs
Current
Current
20 20 A
cc-7
27 63 A
Current
2710 T
C&12

Conveyor motor current, Conveying rate (Tons/Hr)

Other parameters Belt speed not measured separately  Vibration data not being archived

Scene 32

Model was tried out on complete data for past 1 year (Feb 2018 to Feb 2019)

Data was identified as of 2 motors – original motor (red) in place and new motor (green) installed after breakdown New motor was seen requiring lower current to transport same tons/hour compared to old motor

Scene 33

Linear model for new motor gave best performance and was able  to predict failure

Performance metric Model fit RMSE R 2 0.68 Standard deviation 5.33 MAE 1.97 10.2 Conveyer Motor Current = 25.6 + 8.3 x 10-3 * Conveying rate Current Conveying rate- Tons/hour

Scene 34

CC2 conveyor – trends observed for the new motor backing  the model predictions

During its operation, there was a period where the new motor consumed lesser current than it typically does for the same tons/hour Analyzing data for this motor in a time series shows that in mid-Jan 2019, the new motor consumed lower current than usual Investigation suggests that conveyor joint maintenance was carried out at this time

Scene 35

Coal stacking conveyors analyzed as a system to check for system health

In the steady state, conveyors from CC1 – CC5 follow the same pattern of current consumption with a slight delay For a change in conveying rate TPH, current values for all the conveyers change in sync

Scene 36

For the ore stacking conveyor CO9, data was extracted for 3 months

RMHS ORE STACKING CIRCUIT
COAL CIRCUIT
From WT-2
From WT-3
CA From WT-4
From Ground Hopper 1
From Ground Hopper 2
From Ground Hopper 3
co-l
Current:- 12.36 A
co-a
co-9
Current 37.08 A
Belt scale:-1107964Yh
Totalizer -
Current -
Belt scale
Totalizer -
Current
Belt scale
Totalizer 0 T
co-13
Current
Current -
1811 A
Current --
OOO A
co-10
1927 T
000 t
00 t
co-14
Current :- C A
SuS FEEDLNG MIS
co-15
Current
co-16
Current 046
co.17
Current 026
Current 0 43 A
Belt 381682 tph
co-11
Current:- 087
To SMS
LC-18
co-34
TATA
P-V-BB FEEDING
To ORE and FLUX yard
To ORE and FLUX yard
To ORE and FLUX yard
BF FEEDLXG
20-18
Current
055 A
From SMS yard
Current --
184 A
Belt scale:- 0.00 tph
Totalizer
OT

Scene 37

A linear curve based on the conveying rate is able to predict the conveyor motor current with a reasonable accuracy

Conveyer Motor Current = 32.24 + 0.0052 * Conveying rate

Scene 38

While the model has been able to predict deviation, further improvement has to be explored through data aggregation

Model performance to be improved further through data aggregation and by including further data-from the current 6 months to 1 year

Delta = actual current - predicted current

Scene 39

For the furnace hearth stave cooling pump (pump 1552), data was extracted across the cooling circuit

Scene 40

Model can’t be cross-deployed, only applicable for one-by-two systems

Granular data used- 1 minute intervals- failed to offer satisfactory prediction of flow

Rolling avg flow and current, at 10-minute intervals used for modeling; higher correlation observed

Scene 41

Three conditional linear equations were developed for the “two-by-three” pump

Pump discharge flow =(0.3592 * pump_2 current) - (1.39635* pump_3 current + (43.8983 * pressure) + 3406.5995          R 2 =0.37

Pump discharge flow =(5.6308 * pump_1 current) - (0.1432 * pump_3 current + (24.5652* pressure) + 3261.7209           R 2 =0.22

Model suitable for deployment, based on 5 months of data

Limited data available; needs to be strengthened with more data

Pump discharge flow = (8.2254 * pump_1 current) + (11.0686 * pump_2 current) + (49.7919 * pressure) + 2579.5614 R 2 = 0.66

Scene 42

Pump Flow= 2579.56 + 8.22*Pump 1 current + 11.06*Pump 2 current+ 49.79*suction pressure Performance metric RMSE R2 0.66 Standard deviation 20.1 MAE 6.32 12.8 Variable Importance Plot

3690 -
3660 -
3630-
3600 -
3570 -
Feb
colour
— pred
Mar
May
Jun
date time

Scene 43

The model has been able to predict deviations in pump flow

Data gap observed because of low quality data and due to change in combination of operating pump The deviation was marked by a sudden drop in the flow, pressure and current on 11 th  April; team to investigate the cause for change

Delta = actual flow – predicted flow for Condition 1- pump 1, 2 running, 3 on standby

Scene 44

For the furnace throat stave secondary cooling system pump (Pump 1330), data was extracted across the system

Scene 45

The model was tested on historical data to identify deviations in pump flow based on current; pressure to be included in next iteration

Pump Flow = -597.92+ 84.31*pump_1_current + 42.05*pump_2_current Performance metric RMSE R2 0.51 Standard deviation 68.3 MAE 35.4 47.5 Variable Importance Plot

2400 -
2100 -
Mar 04
Mar 11
Mar 18
date_tirne
colour
Mar 25

Scene 46

Currently high delta observed based on 1 month of data for pump motor currents and flow Model to be strengthened to include data for 3 months, pressure reading and inspection of past deviations Need to explore risk of high correlation (and hence bias), between flow and pressure

100 -
-100
Mar 04
Mar 11
Mar 18
date time
Mar 25
Apr 01

Scene 47

Heat exchanger: Across BF stock-house, data was extracted of parameters impacting the temperature

BF Gas Daiw Consumption
Gas Monttw Consumption
STOVE PREHEATER
IS 12042 0
24638080 o
924
59451
Strrve 11 +Waulic
Waste Gas
12
PT4030_2
oust
1117 mgrrn3
1813
Force Stan
39
Stow Com Press.
50
44 NmYh
283.8
CA FAN CONTROL
Pr PV
144.2
AIR PREHEATER
891 mm
FanlCAFan1Gr
2 m rnWC
13 mm
6 mWC
so •c
135 •c
60 %
62.4 •
Stove 13 '-Waulic
VT 4000_0' select
Calc. CV tKCall
1052B
Concentration
131.94 ba
Stwell Y-vd
130.28 ba
stTve12
163.41 ba
stcrve13
7 SO
mmWC
1 mmWC
7 SO
mmWC
891 mmwc
Maint
au
750
mmWC
891 mmwc
10 %
52.8
rss-
CA
TTØS3_02 Select
516336
Nma,m
Stove I
Stove 2
stove 3
Trend
Pressure Trend
w,'AnayzerTrend
Combustion Air
Mixed Gas
Calorific value ot COG
Calorific value of BFG
Desired calorific value
Ot mixed gas
C O Consumption
COKE OVEN GAS
CO Gas Daibi Consumption
co Gas MonthV Consumption
Start

Delta Pressure combustion air ( mmwc ) (PDT4043)

Scene 48

The change in temperature of combustion air was explained through a linear model based on waste gas temp and heat exchanger differential pressure

Rise in temperature of combustion air = -89.49 + 0.87* Waste gas input temp – 0.862* Differential pressure of waste gas + 0.61* differential pressure of combustion air

Scene 49

The model was applied to historical data to identify periods of deviation of differential temperature from the predicted trend

Delta = % actual temp change - % predicted temp change Time

Scene 50

The model was deployed to live track predicted and actual values

2506/19
Slag Gran- Plant
EMG. STATUS
17 09 479 'o
MCG E-mgCond1
E um•ce lop
Electrical
"Nm3/h RAFT
30g/Nm3 PCl
. CA now below preset value
Gas C'eaning
Furnace UMIitoring I
S3
C PRIMETALS
CondJAlarr
33353' Nn,3'h TGP
3.31 bar DP -
bar
bar
TGT
HBT-
g.30.oo P"
6/202019
7
degCl Hum
It nooo PM
LOWER
UPPER K
1-37
BFGN Pr
2, 15
880 mmwc Radar 3
2.00 m
.67
'.30.00AM
02
GBT
•SSO.OOAN
snsnot 9
66deg'C 2nd Lvi
320.00 AM
61251201 9
53000
Name
Measured Jemo.
calculated To
13000 PM
6,r24J201 9
fi value
ait a
2.3100B*'
6425420 9
ggonoau
11 :3000AV
6050019
6050019
nooo PM
6tW2ti19
X valuemme "amp

Reliability

Scene 1 (0s)

Scene 2 (7s)

Scene 3 (1m 25s)

Scene 4 (2m 14s)

Scene 5 (3m 10s)

Scene 6 (4m 0s)

Scene 7 (4m 46s)

Scene 8 (5m 37s)

Scene 9 (6m 3s)

Scene 10 (6m 25s)

Scene 11 (7m 2s)

Scene 12 (7m 21s)

Scene 13 (7m 39s)

Scene 14 (7m 59s)

Scene 15 (8m 16s)

Scene 16 (8m 36s)

Scene 17 (8m 53s)

Scene 18 (9m 17s)

Scene 19 (9m 39s)

Scene 20 (9m 57s)

Scene 21 (10m 20s)

Scene 22 (10m 47s)

Scene 23 (11m 10s)

Scene 24 (12m 7s)

Scene 25 (12m 28s)

Scene 26 (12m 52s)

Scene 27 (13m 10s)

Scene 28 (13m 36s)

Scene 29 (13m 50s)

Scene 30 (14m 12s)

Scene 31 (14m 31s)

Scene 32 (14m 55s)

Scene 33 (15m 20s)

Scene 34 (15m 36s)

Scene 35 (15m 57s)

Scene 36 (16m 12s)

Scene 37 (16m 47s)

Scene 38 (17m 6s)

Scene 39 (17m 22s)

Scene 40 (17m 49s)

Scene 41 (18m 18s)

Scene 42 (18m 50s)

Scene 43 (19m 9s)

Scene 44 (19m 30s)

Scene 45 (19m 45s)

Scene 46 (20m 5s)

Scene 47 (20m 28s)

Scene 48 (21m 6s)

Scene 49 (21m 29s)

Scene 50 (21m 41s)