The Numbers That Didn’t Add Up

The engine is a MAN B&W 10S90ME-C9.2, ten cylinders, electronically controlled. About two days after bunkering a new fuel batch, the Auto Tuning screen started flagging compression pressure red — a deviation of roughly 8 to 10.5 bar from the ordered value, putting Pcomp somewhere around 20 bar above the shop trial reference at the same load. Firing pressure, Pmax, was still tracking neatly on the shop trial curve, but only because the engine control system’s automatic Pmax control was actively correcting injection timing to hold it there, with a Pmax offset of +7 applied to keep a 15 bar margin between Pcomp and Pmax.

Scavenge air pressure was the number that worried me most: 2.6 bar at 77 percent load, against 1.95 bar at 75 percent load on the shop trial curve. The cylinder pressure and crank-angle diagrams from the PMI system showed a double-peak, almost a shoulder, near Pmax on several cylinders I checked — numbers 1, 2, 8, 9, and 10. The new fuel’s properties, as entered on the index calibration screen, showed a lower calorific value than what we’d been running, 40.79 megajoules per kilogram, with a density of 938.7 kilograms per cubic metre and a fuel temperature of 134 degrees. On top of all that, the estimated engine load on the ECS display was reading 81 percent, while the PMI system was measuring closer to 78. None of these numbers were catastrophic on their own. Together, they were a pattern I didn’t like.

The Auto Tuning system itself wasn’t giving me a clean read on any of it either. It was showing “Tuning not available,” flagged against a sensor values warning, which meant the system’s own automatic adjustment logic had effectively taken itself out of the loop rather than fight numbers it wasn’t confident in. That’s worth mentioning because it removed one layer of reassurance I might otherwise have leaned on — the system wasn’t quietly correcting for something minor in the background. It had stepped back and left the interpretation to me.

The Easy Explanation I Almost Settled For

A new fuel, a lower calorific value, elevated pressures two days later — the obvious story writes itself. Lower energy content per kilogram means the system needs to deliver more fuel for the same power, which plausibly pushes cylinder pressures up. I was ready to write that down as the answer and move on to managing it, which is exactly the point where I started drafting a technical report to send to the manufacturer’s service department, using an AI assistant to help me lay the numbers out clearly.

Laying the numbers out clearly turned out to be the thing that stopped me. Once the fuel’s calorific value was actually compared against what a difference of that size could realistically produce, the math didn’t support the story I was about to settle for. A calorific value shortfall in the range this fuel showed would reasonably account for a few percent more boost demand — not a jump of roughly a third in scavenge pressure. The fuel-quality effect was probably real. It just wasn’t, on its own, big enough to explain what I was seeing on the gauges.

What Fit and What Didn’t

The next theory worth testing was combustion timing. That double-hump shape on the PMI pressure traces is consistent with part of the fuel burning later than normal — continuing into the expansion stroke instead of finishing cleanly near top dead centre. Late-burning fuel carries extra thermal energy straight into the exhaust manifold and onto the turbine, spinning the turbocharger harder than the shop trial reference expects for that load, which raises scavenge pressure, which in turn pushes compression pressure up right behind it. It fit the shape of the pressure trace. It fit the direction of every number I had.

One thing didn’t fit. That mechanism almost always shows up as elevated exhaust gas temperature too, and across all cylinders, exhaust temperature spread was stable — no significant variation from what we’d normally expect. That single inconsistency was enough to keep the file open rather than close it. Engine sound and vibration hadn’t changed either, which ruled out anything obviously mechanical failing in a loud way. I reduced load as a precaution while the investigation continued, and kept watching exhaust temperature spread and sound as the two simplest early-warning signs available to me without stopping the engine.

I want to be honest that reducing load was itself a slightly uncomfortable decision. The ship was already running behind schedule, and cutting speed on top of an existing delay is never a popular call to make over the radio. But an inconsistency in a theory that otherwise fit the data well enough to nearly convince me was not something I was willing to sail past at full load just because the schedule was tight. The exhaust temperature spread stayed flat. If it had started climbing, I would have known within one round exactly which direction the problem was moving in.

The Detail That Changed the Diagnosis

The possibility that actually changed my thinking was the exhaust gas bypass valve. On this turbocharging arrangement, the bypass valve is meant to divert part of the exhaust gas around the turbine at higher loads, keeping the turbocharger from over-boosting the engine beyond what it’s tuned for. If that valve isn’t opening fully, more exhaust gets forced through the turbine than the system expects, the turbocharger spins harder than it should, and scavenge pressure climbs — independent of anything happening with the fuel.

I couldn’t confirm the valve’s actual position without stopping the vessel — from the control room, all I had was the indication, not a direct view of how far it was actually travelling. But the theory fit everything I had: the pressure numbers, the stable exhaust temperatures, the unchanged sound. And it changed the nature of the problem in a way that mattered immediately. This wasn’t an active combustion abnormality that might be getting worse by the hour. It was a mechanical restriction, contained and reasonably well understood, that I could manage carefully until we reached port.

Managing It Until We Could Stop

The vessel was already behind schedule, and the valve itself was hard to inspect properly, let alone repair, without stopping to open it up — something we couldn’t do until the next port, still a few days off. In the meantime, the plan came down to a handful of concrete steps. I found the load threshold where the bypass valve is meant to start opening on this engine, typically somewhere in the 75 to 90 percent range, and stayed below it operationally wherever I could, since a stuck valve matters far less if the system doesn’t need it engaged in the first place.

I also watched whether scavenge pressure tracked proportionally better at reduced load than the normal load curve would predict — if it did, that was a useful confirming sign I had the right cause rather than a coincidence. Exhaust temperature spread and engine sound stayed my two early-warning checks throughout. And since Pmax was already being actively managed by the engine control system, and the mechanism now looked mechanical rather than an active runaway condition, continuing at a reduced, sub-threshold load for the transit felt like a reasonable way to manage the problem rather than something that justified an emergency stop at sea. I flagged the suspected cause to the superintendent ahead of arrival too, so a technician and the right actuator or seal spares for the bypass valve would already be on hand, instead of losing more time troubleshooting after we’d already stopped.

What the AI Conversation Actually Did For Me

I want to be precise about this, because it would be easy to overstate it. The AI assistant did not diagnose the bypass valve. That possibility came from my own experience with this engine type, once a mechanical explanation was on the table as something worth considering alongside the fuel theory. What the conversation actually did was narrower and, in its own way, more useful: it forced me to quantify the fuel-calorific-value argument instead of accepting it because it sounded reasonable, and it kept pointing back at the one number that didn’t fit — the stable exhaust temperature spread — instead of letting me close the combustion-timing theory out on a partial match.

It also suggested two specific checks I hadn’t thought to run yet: comparing turbocharger rpm against its own reference curve for that scavenge pressure, which would show whether the turbocharger itself was behaving abnormally or simply responding to more exhaust energy than expected, and checking the air cooler differential for fouling, as a separate possible contributor to elevated boost that had nothing to do with either the fuel or the valve. I hadn’t reached for either of those on my own yet. And writing the technical report itself, with the numbers organized into an actual argument rather than a list of readings, is what surfaced the calorific-value math in the first place. None of that replaced my own judgment. It gave the judgment better material to work with.

I think there’s a real distinction worth naming here, between an AI tool acting as an answer machine and one acting as a disciplined second reader. The first would have been comforting and probably wrong — the fuel-quality theory was plausible enough that a system optimizing for a fast, confident-sounding answer could easily have handed it back to me dressed up as a conclusion. What actually happened was closer to being interviewed by someone who kept asking “how do you know that” and “what would tell you this is wrong” until the weak points in my own reasoning became impossible to ignore. That’s a different kind of usefulness than I expected going in, and it’s the reason I keep reaching for it during a live investigation now, rather than only after the fact when writing something up.

What I Changed Since

We reached port a few days later, stopped, and opened the bypass valve assembly for inspection. The actuator linkage was stiff, partly seized from a buildup of exhaust deposits restricting its full travel — enough to keep the valve from ever reaching its intended open position at higher loads, even though the control signal calling for it looked correct the entire time. We freed and serviced the linkage, and scavenge pressure at a comparable load afterward came back close to the shop trial reference.

• After any fuel change, I now check whether the size of a parameter shift actually matches what the fuel-property difference alone could produce — not just whether the direction of the shift makes sense.
• I check the exhaust gas bypass valve’s position and travel as a standard step whenever scavenge pressure deviates materially from the shop trial curve, not only once fuel quality is already the leading suspect.
• I write a short, structured technical summary of unusual readings before reporting them anywhere, even briefly, because organizing scattered numbers into an actual argument keeps surfacing gaps I would otherwise miss.
• I treat “explains most of the gap” as different from “explains all of the gap,” and keep looking for a second contributor whenever a gap remains after the first cause checks out.
• I now use an AI assistant as a genuine second opinion during live troubleshooting — not for the answer, but specifically to be pushed on what doesn’t fit the current theory yet, and what hasn’t been checked at all.