Predictive maintenance means using data from the machinery to tell you that something is going wrong before it breaks, so you can fix it at a time you choose. On a ship that usually means the generators, the main engine, pumps, compressors, purifiers and cooling systems. AI is the part that reads thousands of sensor values and says "this one is not behaving like last month".
This guide is written from the ship side. It covers what these systems actually watch, which fleets use them today, which numbers you can trust, and what to ask before you sign.
Three kinds of maintenance, in plain words
Most ships run a mix of three.
- Planned (time-based) maintenance. Change the part every 8,000 running hours, because the maker says so. This is what your planned maintenance system (PMS) is built around. It is simple and class understands it. But it replaces some parts that were fine and misses some that fail early.
- Condition-based maintenance. Measure the machine (vibration, temperature, oil analysis) and act when a value crosses a limit. An engineer still decides what the numbers mean.
- Predictive maintenance. The system learns what "normal" looks like for each machine and warns when the pattern starts to drift, often days or weeks before a limit is crossed. The best ones also say which part is the likely cause.
The difference between the last two matters when you read a brochure. Many products sold as "predictive" are really condition monitoring with a nicer screen. That is not bad. But you should know which one you are paying for.
What it watches on a ship
The systems work best on rotating machinery with good sensors and a clear failure history:
- diesel generators and their turbochargers
- main engine (cylinder pressures, exhaust temperatures, bearing temperatures)
- seawater and freshwater cooling pumps
- compressors, purifiers, fans
- batteries on hybrid and electric ships, where the system tracks cell temperature, voltage and wear
- coatings and corrosion, using photos taken by the crew rather than sensors
A good example of the last one: Ardmore Shipping gave its crews an app to photograph hull and tank coatings. The AI reads the photos and builds a picture of where corrosion is spreading across the fleet. No new sensors, no new wiring.
Who uses it today
These are the cases we have reported on. Notice how few of them publish a result.
Equinor is the strongest evidence, though it is offshore, not shipping. It monitors more than 700 rotating machines with more than 24,000 sensors, and says predictive maintenance has delivered $120 million in value since 2020. These are Equinor's own figures, but they come from the operator, not a supplier, and they cover years, not a pilot. Pumps and compressors on a platform are close cousins of the ones in your engine room.
OPearl LNG put Wärtsilä's Expert Insight on all 14 of its LNG carriers. A whole fleet is a real decision. No savings have been published yet.
Saipem started on the drillship Saipem 12000 in February 2026, and is testing a second system on the diesel generators of the crane vessel Saipem 7000. Starting with generators is a sensible choice, as we explain below.
K Line and Kawasaki have worked on an AI machinery support system with failure prediction since 2021, built with the Japanese AI company Preferred Networks. It is now being joined to their voyage planning service.
Groundup.ai works with Tidewater, which runs more than 200 offshore support vessels. It claims 22 months with zero unplanned downtime for some customers and about $220,000 of value per vessel per year. These are the supplier's own numbers. Ask to speak to the customer.
AST Reygar and Seacat Services ran a seven-month trial on crew transfer vessels in the UK, from September 2025 to April 2026, on diesel, electric and hybrid boats. Small vessels, but a real fleet and a fixed test period.
And from research: a study on a tanker's central seawater cooling system used 41 measured values to tell normal running from early faults, and fed the result into the ship's PMS. That last step is the one that matters on board.
Be careful with the big numbers
You will see claims like "20 to 30% less downtime" and "billions saved every year". Most of them come from market reports or supplier marketing, not from a ship. Before you believe a number, ask three things:
- Which ships, and how many? One ship for three months proves very little.
- Compared with what? "Less downtime" than the same ships last year, or than an industry average nobody can check?
- Who measured it? The owner, the class society, or the supplier?
The Equinor figure passes most of these tests. Most others do not, yet.
What it needs from you
This is the part the brochure leaves out.
Good data. Sensors must work and be calibrated. If the bearing temperature sensor on No. 2 generator has been reading 0 °C for a year, the AI learns that 0 °C is normal.
A clean maintenance history. The system learns from past failures. If your PMS says "done" against jobs that were never done, or the failure was written as "repaired" with no cause, the AI learns nothing.
A link to shore. Most systems send data to the supplier's cloud. Check how much data, how often, and what happens when the link is down.
A person who reads the alerts. Even the robot dogs now doing inspection rounds at a steel mill need a person in a control room to read what they find. On a ship that is the chief engineer, or a superintendent ashore. If nobody owns the alerts, they become noise within a month.
Crew time. A World Maritime University study found that 72% of seafarers have no time on board to learn new systems. An engine monitoring system needs training, and the training needs hours in the schedule. Plan them before installation, not after.
Where to start
Do not start with the whole ship. Start where the money and the data are.
- Pick one or two machines with a history of failures. Diesel generators are the usual choice: there are several of them, they fail in known ways, and a failure costs real money. Purifiers and cooling pumps are also good candidates.
- One ship, six months. Write down the baseline first: unplanned stops, repair costs and spare parts used in the last two years.
- Decide who reacts to an alert, and how fast. Write it in the procedure.
- Connect it to the PMS. An alert that does not become a work order is a wasted alert.
- Judge it at six months on your own numbers, not the supplier's report.
Questions to ask the supplier
- Is this predictive (it forecasts a fault) or condition monitoring (it shows values against limits)?
- Which sensors does it need, and which do we already have? Who installs and calibrates them?
- How many alerts per ship per month should we expect, and how many were real faults on other ships?
- Can we speak to a customer who has used it for more than a year?
- Does it write work orders into our PMS, or do we copy them by hand?
- How much data goes ashore, and does it work when the satellite link is down?
- Who owns the data, and can we take it with us if we leave?
- How does it fit with our class maintenance scheme? Will class accept its records?
- What is the total cost per ship per year, including sensors, installation, data and training hours?
Common questions
Is predictive maintenance the same as condition monitoring? No. Condition monitoring shows you the values. Predictive maintenance tries to forecast the fault from how the values change. Many products do the first and call it the second.
Does it replace the planned maintenance system? No. It should feed the PMS, so jobs move earlier or later based on the machine's real condition. The PMS stays the record that the office, class and auditors look at.
Can an older ship use it? Yes, if the main machines have working sensors or can be fitted with them. Some systems work from photos or portable vibration readings taken by the crew.
How long before it is useful? The system needs time to learn what normal looks like on your ship, usually some weeks of clean data. Judge it after months, not days.
The ship-side view
Predictive maintenance works. The best evidence, from Equinor, is real, and fleets like OPearl and Tidewater are betting on it. But the AI is the easy part. The hard part is the data, the maintenance history and the person who has time to act on the alert. Get those right on one ship first. Then the AI has something to work with.
We update this guide when new evidence appears. The latest stories on predictive maintenance from our weekly digest are listed below.