I'll take you through the entire thing from the very beginning to the live alert, in deep

detail,
with no step skipped. I'm going to assume the person knows nothing and build every idea
on the one before it. This is long because you asked for full depth — read it as a complete
walkthrough you can absorb once and then teach from.

Let me set the structure first so you know where we're going: I'll cover (A) the feed and why
it needs heating, (B) what an exchanger is and the physics of how heat actually moves,
(C) the eight measurements and exactly what each one tells you, (D) how the three
calculated quantities are built and what each means physically, (E) what fouling is,
mechanism by mechanism, (F) how fouling reveals itself in the data, (G) why a naive
alarm fails and how the model fixes it, and (H) the full alert pipeline from raw reading
to "Clean Now." Each section assumes you've absorbed the one before.

A. The feed, and why it has to be heated at all

A1. Where the feed comes from. Crude oil is a soup of thousands of different molecules,
from very light (gases, petrol) to very heavy (tar-like). A refinery separates them by boiling
point. First the crude goes to the atmospheric distillation column — heat it, the lighter
things boil off and are collected, the heavy stuff sinks to the bottom. That bottom is sent to
a second column held under vacuum (low pressure makes things boil at lower
temperature, so you can pull off more without overheating). Whatever still refuses to boil
even under that vacuum is the heaviest material in the whole refinery. That leftover is
Vacuum Residue (VR) — and it's your DCU feed.

A2. What VR is like physically. Think of it as cold molten tar. It's extremely viscous (barely
flows when cool), very dense, and packed with the biggest, most complex molecules —
including asphaltenes, which will matter enormously later. It also carries sulphur, metals,
and a lot of "carbon-forming" potential. Nobody upstream could turn it into anything
valuable, so it lands at the DCU as a last chance to extract value.

A3. What the DCU does with it. The Delayed Coker Unit's job is to heat VR so hot that the
big molecules break apart ("crack") into smaller, lighter, sellable products — gases,
naphtha, gas oils — leaving behind solid petroleum coke. To make this cracking happen,
the feed has to be driven up to roughly 330 °C entering the furnace, and then the furnace
pushes it further to the coil-outlet temperature where cracking occurs. There's a hard
ceiling: above 330 °C before the furnace, the feed starts coking inside the wrong place —
the equipment itself — which is damaging. So 330 °C is a wall you approach but never
cross.
A4. Why we don't just heat it all in the furnace. You could take cold VR and heat the
entire way with the furnace burning fuel gas. But fuel gas costs money, every hour, forever.
Heating heavy oil from near-ambient all the way to furnace conditions purely on fuel would
be enormously expensive. So refineries do something clever: they pre-warm the feed for
free before it ever reaches the furnace. Every degree of free pre-warming is a degree the
furnace doesn't pay to supply. This is the entire economic reason your project exists —
squeeze the maximum free pre-warm out of the system so the furnace burns the minimum
gas.

B. What a heat exchanger is, and how heat actually crosses

B1. The core trick: two streams, one wall, no mixing. A heat exchanger lets a hot fluid
and a cold fluid pass close to each other, separated by a metal wall, so heat flows from hot
to cold through the wall — but the two fluids never touch or blend. You keep your feed pure
and you keep the hot stream pure; only the heat moves.

B2. The shell-and-tube design specifically. Yours are shell-and-tube exchangers. Picture
a fat horizontal pipe — the shell. Running through it lengthwise is a bundle of many thin
tubes. One fluid flows inside the tubes; the other flows inside the shell but outside the
tubes, weaving around the bundle. In your train, the cold VR feed runs through the tubes,
and a hot leftover product stream runs through the shell around them. The tube walls are
the meeting surface — hot on the outside, cold on the inside, heat soaking through the
metal.

B3. Why heat moves at all — the only law you need. Heat always flows from hotter to
colder, never the reverse, and it flows faster the bigger the temperature difference. Hold a
metal spoon: one end in boiling water, the other in your hand — heat marches up the
spoon because one end is much hotter than the other. Shrink that difference and the flow
slows. This single idea — heat flows down a temperature gap, and a bigger gap means
more heat — is the foundation of everything that follows.

B4. The three hurdles heat must clear. To get from the hot stream into the cold feed, heat
crosses three barriers in series: first it has to move from the bulk of the hot fluid to the
tube's outer surface; then conduct through the metal tube wall; then move from the tube's
inner surface into the bulk of the cold feed. Each barrier resists a little. The metal is a good
conductor so it resists little; the fluid films on each side resist more. Remember this
"barriers in series" picture — fouling will later add a fourth barrier, and that's the whole
story.
B5. Where the hot stream comes from. The beautiful part of the design: the plant heats
its own feed with its own waste heat. The DCU produces hot products that have to be
cooled down anyway before storage. Instead of throwing that heat away to cooling water or
air, those hot products are routed through the shells of the preheat exchangers, dumping
their heat into the incoming cold feed. The feed gets warmed; the products get cooled;
nothing is wasted. So every exchanger has a hot product giving up heat on the shell side
and your cold feed receiving it on the tube side.

C. The eight measurements — what each sensor is actually telling you

For every exchanger you instrument exactly eight points. They split into three pairs of "in vs
out" plus flows. Here's what each one physically means, not just its name.

C1. T_ci — cold-side inlet temperature. How warm the feed already is when it arrives at
this exchanger. In a series train this is whatever the previous exchanger left it at. It's your
starting line for this stage.

C2. T_co — cold-side outlet temperature. How warm the feed is when it leaves this
exchanger, heading to the next one. The difference between T_co and T_ci is how much this
single exchanger warmed the feed — its visible contribution. If T_co is barely above T_ci,
this exchanger is doing little; if it's much higher, it's pulling its weight.

C3. T_hi — hot-side inlet temperature. How hot the leftover product stream is when it
enters the shell. This sets the top of your temperature gap — the hotter this is, the more
potential there is to heat the feed.

C4. T_ho — hot-side outlet temperature. How hot that product is when it leaves, now
cooler. The drop from T_hi to T_ho is how much heat the hot stream gave away in this
exchanger.

C5. m_c — cold-side mass flow. How much feed (mass per unit time) is passing through
the tubes. This matters because warming a little feed by 20° is a small amount of heat, but
warming a lot of feed by the same 20° is a large amount of heat. Temperature change alone
doesn't tell you the heat moved — you need to know how much stuff was warmed.

C6. m_h — hot-side mass flow. Same idea for the hot product: how much of it is flowing
and therefore how much heat it's capable of carrying and dumping.

C7. ΔP_c — cold-side pressure drop. How much the feed's pressure falls as it travels
through the tubes. A clean, open tube offers little resistance, so the pressure barely drops.
As we'll see, this number is a quiet early traitor that reveals fouling.
C8. ΔP_h — hot-side pressure drop. The same for the shell side. Resistance to the hot
stream's passage.

C9. The cross-check hidden in these eight. Notice you can describe the heat moved two
ways: from the cold side (how much feed × how much it warmed) and from the hot side
(how much product × how much it cooled). These two descriptions of the same heat must
agree. If they don't, a sensor is wrong or something is leaking or bypassing. This built-in
agreement test is your honesty check, and it's why the system can flag bad instruments
instead of trusting them blindly.

D. The three calculated quantities — built one on top of the next

You never measure the exchanger's "health" directly. You compute it from those eight
readings, in three layered steps. Each step is meaningless without the one before, so
follow the chain.

D1. First quantity — Heat Duty (call it Q): "how much heat actually crossed." You take
how much feed flowed and how much warmer it got, and that tells you the total heat that
entered the feed. You can equally take how much hot product flowed and how much it
cooled — same heat, counted from the other side. Conceptually: Q is the real tonnage of
heat moved per hour. It's the honest accounting of energy that crossed the wall. On its
own, though, Q doesn't tell you if the exchanger is healthy — because Q depends on how
hard you're driving it (a big temperature gap can produce high Q even through a somewhat
dirty exchanger). So Q is necessary but not sufficient. We need to normalize it.

D2. Second quantity — the driving gap (LMTD): "how hard the exchanger is being
pushed." Recall heat flows down a temperature difference. But the difference isn't a single
number — it's different at each end of the exchanger (big gap where hot-in meets cold-out,
smaller gap where hot-out meets cold-in). So we compute one representative "average
driving gap" across the whole exchanger. Conceptually: this is the size of the push that's
forcing heat across. A large gap is a strong push; a small gap is a weak push. We need this
because the same exchanger will move more heat when pushed harder — and we want to
separate "moved a lot of heat because it's good" from "moved a lot of heat because it was
pushed hard."

D3. Third quantity — Performance (U): "how good the exchanger is right now." Here's
the crucial combination. Take the heat actually moved (Q) and ask: given how hard it was
being pushed (the driving gap) and how big the exchanger is (its surface area), how much
heat did it move per unit of push, per unit of area? That ratio is U, the overall heat transfer
coefficient — the true scorecard of the exchanger. Think of it as fuel efficiency for a car:
not "how far did you go" (that's Q) but "how far per litre" (that's U). A high U means the
exchanger transfers heat brilliantly for the push it's given; a low U means it's sluggish — it's
getting plenty of push but failing to move heat. U is the headline number. It strips out how
hard you're pushing and how big the box is, leaving only the intrinsic quality of the heat
transfer right now.

D4. Why the surface area matters in that ratio. A bigger exchanger (more tubes, more
wall) naturally moves more heat, just like a bigger radiator. So to judge quality rather than
size, you divide out the area. This is why the project needs the datasheets (tube count,
length, diameter) — and why plugged tubes matter: a plugged tube is dead area, so the
effective area shrinks, and you must account for that or you'll misjudge U.

D5. The benchmark — clean U. When the exchanger was brand new and spotless, it had a
particular, high U — its "clean" performance, recorded on the original datasheet. That's
your reference line, the "this is as good as it gets" mark. Everything about fouling is
measured as how far today's U has fallen below clean U.

D6. Fourth quantity — Fouling resistance (Rf): "how thick the blanket is." Now connect
back to the "barriers in series" idea from B4. A clean exchanger has three barriers (hot film,
metal, cold film). Fouling adds a fourth — a layer of deposit on the wall that heat must also
fight through. Rf is a number that captures the size of that extra barrier. When the
exchanger is clean, Rf is essentially zero. As gunk builds, Rf grows. The way we get it:
compare today's U against clean U; the shortfall is the extra resistance, the fouling blanket,
expressed as a number. So Rf rising over time = the deposit thickening. Rf suddenly
snapping back to near-zero = somebody cleaned the exchanger (which, conveniently, is
also how the system learns what a cleaning event looks like in the data).

To recap the chain so it's locked in: eight readings → Q (heat moved) → combine with the
driving gap and area → U (quality now) → compare to clean U → Rf (the fouling blanket).
Each rests on the one before.

E. What fouling actually is — the mechanisms, in plain physical terms

You can't explain detection convincingly without explaining the thing being detected.
Fouling isn't one process; it's several, all leaving a deposit on the tube wall.

E1. Asphaltene deposition — the main villain. Remember asphaltenes from A2 — the
biggest, heaviest molecules in the VR. In the flowing feed they're normally held
dissolved/suspended in the surrounding lighter oil, like sugar dissolved in tea. But they're
fragile in solution. When conditions shift — local heating, a change in feed composition,
shear — they fall out of suspension, clump together, and stick to the tube wall. Picture
sugar crystallizing out of over-sweet tea and caking on the cup. This is the dominant fouling
mechanism in a VR preheat train.

E2. Coking / carbon laydown. VR has high "carbon-forming potential." Where the wall is
hottest, some of this material can thermally degrade and lay down a hard carbonaceous
deposit — essentially baked-on residue, like the burnt crust at the bottom of a pan left on
the heat too long. This tends to happen on the higher-temperature exchangers near the
furnace end.

E3. Corrosion products. Sulphur and other species can corrode metal surfaces; the
corrosion products themselves form a layer and also roughen the surface, giving
asphaltenes more to grab onto. So corrosion and deposition feed each other.

E4. Why running 18% over design makes it worse. Your unit runs at 410 against a 346.5
design feed rate. Two consequences: first, the feed spends less time in each exchanger
(less residence time), and the flow is more forceful (more shear), both of which can knock
asphaltenes out of suspension faster; second, you're simply pushing a lot more of this
asphaltene-rich material through per hour, so there's more raw material available to
deposit. Over-design throughput is a fouling accelerant — and, usefully, it also means the
fouling signal is strong, which helps detection.

E5. The physical effect of the deposit — back to barriers. Whatever the mechanism, the
result is the same: a layer of low-conductivity gunk coating the wall. Deposits are poor heat
conductors — they insulate. So heat that used to cross three barriers now has to fight
through four, and the extra barrier (Rf) keeps growing as the layer thickens. That's why U
falls. And separately, the deposit physically narrows the channel the fluid flows through —
which is where the second clue comes from.

F. How fouling reveals itself in the data — the two clues, in detail
F1. Clue one — the heat-transfer quality (U) drifts down. As the insulating blanket
thickens (E5), the exchanger transfers less heat for the same push. In the readings, you see
it as the feed coming out less warm than it used to under the same conditions — the gap
between T_co and T_ci shrinking relative to the push available. Computed, U sags lower
and lower, and Rf climbs. This is the primary, direct indicator of fouling. But — and this is
important — it's a slow, gradual drift over days and weeks, and it can be partly disguised by
other things changing (covered in G), so you rarely act on U alone.
F2. Clue two — the flow gets choked (ΔP rises). The deposit narrows the bore of the
tubes (or the gaps in the shell). To push the same amount of fluid through a narrower
space, you need more pressure — so the pressure drop across the exchanger climbs.
Picture a slowly clogging shower head: same water supply, but the spray weakens and
back-pressure builds. Crucially, this clue often shows up earlier and more sharply than the
thermal clue, because even a thin deposit measurably restricts flow before it has badly
hurt heat transfer. So ΔP is your early tripwire.

F3. Why you watch both together. Each clue alone can be fooled. U can dip for reasons
that aren't fouling (next section). ΔP can jump because someone moved a valve. But
fouling is essentially the only thing that makes both move adversely at the same time,
gradually, in the same exchanger. So when U is sagging and ΔP is climbing together on one
exchanger, you have a high-confidence fouling call. Two independent symptoms agreeing
— exactly how a careful doctor diagnoses rather than reacting to a single reading.

F4. The series-train amplifier — why one fouled exchanger matters to all. Because the
exchangers are in a line, the feed leaving one is the feed entering the next. If exchanger #4
fouls and under-delivers, every exchanger after it receives feed that's colder than
designed. The downstream ones now have a slightly bigger temperature gap to work with,
so they transfer a bit more and partially mask the shortfall — but never fully. The leftover
deficit lands at the far end: the feed reaches the furnace below target, and the furnace
burns extra gas to close the gap. So one fouled exchanger raises the fuel bill for the whole
train — and, because of that masking, you cannot diagnose the problem by watching only
the final outlet temperature. You must look inside, exchanger by exchanger. This is the
core justification for per-exchanger monitoring.

G. Why a naive alarm fails, and how the model fixes it

G1. The trap: U moves for innocent reasons. Here's the subtlety that separates a credible
system from a noisy one. The exchanger's apparent performance shifts even when it's
perfectly clean — because:

• Throughput changes. Run more feed through and the flow scrubs the walls more
vigorously, which actually improves heat transfer; run less and it worsens. So U
naturally rises and falls with feed rate.
• Feed quality changes. A different VR batch — heavier, lighter, different viscosity —
transfers heat differently. So U shifts with whatever's in the feed today.
• Hot-stream conditions change. If the leftover product feeding the shell is hotter or
flowing faster on a given day, the numbers shift.
If you set a dumb fixed threshold — "alarm if U drops below X" — you'd get false alarms
every time the feed rate dipped or the feed batch changed, and operators would quickly
learn to ignore the system. That's the death of any monitoring tool.

G2. The fix — predict the expected U, then watch the deviation. Instead of comparing U
to a fixed line, you first predict what U should be right now, given today's throughput,
today's feed quality, today's hot-stream conditions. This expected value comes from the
physics (the first-principles backbone) plus a learned correction. Then you compare reality
against that moving, condition-aware expectation. The gap between "what U should be
today" and "what U actually is today" is the real fouling signal — with all the innocent
variation stripped out. A clean exchanger on a high-throughput day and the same
exchanger on a low-throughput day will both sit near their expected U, so neither raises a
flag. Only a genuine deposit pulls actual U persistently below expected. This is precisely
why the solution is a hybrid model — physics computes the expected behaviour, machine
learning sharpens the prediction by learning the residual patterns, and the deviation is
what you alarm on.

G3. Why the cleaning history matters here. To learn what "fouling that's worth acting on"
looks like, the model is shown past cleaning events — the dates exchangers were cleaned
and how bad they'd gotten just before. Those past episodes are the labelled examples that
teach the model the shape of a developing fouling problem versus normal scatter. Without
that history, the system can still detect drift (something is changing) but is weaker at
judging how serious and how soon.

H. The full alert pipeline — from raw reading to "Clean Now"

Now everything assembles into the live system. Six stages, each building on the prior.

H1. Ingest and validate. Every minute (the historian's recording cadence), the eight
readings for all twelve exchangers stream in. First they're checked for sanity: stuck
sensors, impossible values, the hot-side-vs-cold-side heat agreement test from C9. Bad
data is flagged rather than trusted, so the rest of the chain isn't poisoned by a faulty
thermocouple. (This matters more here because there's no redundant instrumentation to
cross-check against — so the energy-balance check and outlier handling do that job.)

H2. Compute the live scorecard per exchanger. For each exchanger, the system
computes Q, then U, then Rf — the chain from section D — and also tracks ΔP. So at any
instant you have, for all twelve, a live picture: how much heat each is moving, how good
each is right now, how thick each one's fouling blanket is, and how choked each one's flow
is.
H3. Compare against expectation (the smart part). Using the approach in G2, the system
also computes what each exchanger's U should be under the current conditions, and looks
at the deviation. A small deviation is normal scatter. A deviation that is growing,
persistently, in the adverse direction — especially with ΔP also creeping up — is flagged as
genuine fouling beginning. This is the early catch, well before the exchanger becomes a
"red" problem.

H4. Forecast forward — fouling rate and remaining life. It's not enough to say "this is
fouling." The system projects the trend: at the current rate of deterioration, how many more
days until this exchanger reaches the point where it's worth cleaning? This converts a
present measurement into a prediction — "exchanger E/F will hit the cleaning threshold in
roughly N days" — which is what lets people plan rather than react.

H5. Translate into money, with lead time built in. Fouling costs you because the furnace
burns extra gas to make up the lost preheat. Using the value of each degree of lost preheat
(the currency-per-°C figure the project needs from the economics team), the system
continuously asks: is the accumulating extra-fuel cost about to exceed the cost of just
cleaning this exchanger? And critically, because arranging a cleaning takes one to two
weeks to mobilize, it raises the warning that much earlier — so the recommendation lands
before you start bleeding money, not after. When that economic crossover is reached, the
system raises a "Clean Now" advisory on that specific exchanger.

H6. Present it — and keep a human in charge. The advisory isn't a vague "something's
wrong." It names the exchanger, shows the evidence (U sagging below expected, ΔP rising,
the forecast, the projected cost), and — within the 330 °C ceiling and the flow/pressure
limits — can suggest an interim move like repositioning a bypass to recover some preheat
while cleaning is arranged. It's routed to the right person on a role-based dashboard, with
an email alert and a full audit trail. And here is the line that matters most for management
and for safety: the system only advises. It never actuates anything. No automatic valve
moves, no write-back to the control system in this phase. It recommends; an authorized
human decides and acts. A very smart assistant — not an autopilot.

That's the entire chain end to end, every link explained and resting on the one before it: the
feed exists because it's the heaviest leftover and must be heated → we heat it for free in
exchangers to save furnace fuel → an exchanger moves heat through tube walls down a
temperature gap → we watch eight points to capture that → from them we build Q, then U,
then Rf to score health → fouling is a growing insulating, channel-narrowing deposit → it
shows up as U sagging and ΔP rising together → we judge it against condition-aware
expectations to avoid false alarms → and we forecast it, price it, and surface a human-
owned "Clean Now" advisory at the right moment.

If it would help, I can now lay this same end-to-end logic out as a single flow diagram — the
full pipeline from the eight sensors through to the advisory — so you have one picture that
holds the entire story for the briefing. Want that?