Marine condition monitoring rarely fails at the sensor. It fails at the last link in the chain, and the meta-description of this page names it precisely: the data has to trigger real work orders, not just dashboards. The usual sequence runs sensor, reading, threshold, alert, dashboard — and then stops. Somebody ashore looks at a screen, forms an impression, and nothing is assigned to anybody with a date attached. Six months later the fleet has a monitoring subscription, a wall of trend charts, and a maintenance programme running on exactly the same intervals it used before any of it was installed. The discipline exists and it is well standardised. ISO 17359 sets out the general principles of condition-based maintenance from establishing the monitoring programme through data analysis to decision-making; ISO 13374 covers data processing and communication; and the ISO 20816 series governs vibration measurement and evaluation, with ISO 20816-3 providing a starting chart for alarm settings across machinery groups defined by power rating and whether the base is rigid or flexible. None of that describes what happens after the alarm. That part is your integration, and it is the part that determines whether any of the rest was worth buying. Start a free trial of Marine Inspection and specify the last link first.
Sensor
Vibration, temperature, pressure, oil condition. Rarely the problem, and the part everyone spends the evaluation on.
Reading
A measured value in a defined quantity and unit, against a known machine, at a known moment.
Threshold
A limit somebody chose. The most consequential decision in the whole system and the one given the least attention.
Alert
Something crossed the limit. Now a person has to decide whether it matters and what to do about it.
Dashboard
And this is where most implementations end. Awareness without accountability, reviewed by whoever remembers to look.
Work order
Assigned to a named person, with a date, appearing in the same list as every other job that vessel owes. The only step that produces maintenance.
Sensors produce data. Dashboards produce awareness. Only work orders produce maintenance. If the integration specification stops at the dashboard, you have bought monitoring rather than condition-based maintenance, and the two are separated by exactly one link.
The Threshold Is the Project
Everything downstream inherits this decision, and getting it wrong in either direction destroys the programme in a way that better sensors cannot rescue. Book a Marine Inspection demo and plan how thresholds will be set and revised before installing anything.
Set too tight
Alarms fire on normal operating variation. Crews investigate, find nothing, and within a few weeks learn that alarms from this system do not mean anything. Alarm fatigue is worse than no monitoring, because it degrades trust in every alert the platform produces including the ones that matter.
Set too loose
Nothing fires until damage is already advanced, which delivers a monitoring system that confirms failures rather than anticipating them. Less visible than the first failure mode and equally expensive, because it looks like a healthy programme with no alarms.
Where to start
ISO 20816-3 provides a chart establishing basic alarm settings for machinery groups defined by power rating and base type, rigid or flexible — which gives a defensible starting point rather than a guess. Treat it as the opening position and expect to revise it against your own machines, because a published chart cannot know that this particular pump has always run slightly rough.
The revision process matters as much as the initial setting. Decide in advance who can change a threshold, what evidence justifies a change, and whether the change is recorded — because a limit quietly widened after a run of nuisance alarms is indistinguishable, six months later, from a limit that was correct all along. That record is also what a surveyor will want to see if the monitoring underpins a survey arrangement.
The Unit Trap
A silent failure mode that produces plausible-looking data and wrong answers, and one that is entirely preventable at specification stage. Sign up for Marine Inspection and pin the quantity and unit down per measurement point before anything is connected.
Which quantity
Acceleration, velocity or displacement. These describe the same physical event and produce entirely different numbers, and choosing between them is one of the questions that makes vibration alarm setup genuinely difficult rather than merely fiddly.
Which statistic
Root mean square, peak, or peak to peak. A threshold expressed in one and a reading delivered in another will breach or fail to breach for reasons that have nothing to do with the machine's condition.
Which unit
Millimetres per second RMS is common in metric practice, inches per second peak in the Americas. Both serve the same purpose in different units, which is exactly why a mismatch between a sensor's output and a system's expectation produces numbers that look reasonable and are wrong.
The specification line
For every measurement point: the quantity, the statistic, the unit, the sampling condition, and the machine it belongs to. Written down, agreed by both the monitoring supplier and the maintenance platform, and checked against a known reading during commissioning rather than assumed.
Specify the work order before you specify the sensor
When this threshold is breached on this machine, what job is created, against which component in the equipment register, at what priority, assigned to which role, with what acceptance criterion for closing it. Write that sentence for every monitored point before choosing hardware. It takes an afternoon, it forces every ambiguous decision into the open, and it is the difference between condition-based maintenance and a subscription to charts.
What Class Actually Asks For
Where condition monitoring supports a class-approved maintenance arrangement, the survey requirement is more specific than most monitoring installations are set up to satisfy. Schedule a walkthrough and check your position against each of these before an audit does.
Trends, not values
Where machinery is on condition-based maintenance under an approved scheme, the annual audit reviews vibration and lubricating oil trend analyses. A system holding current readings and recent alarms without retained history cannot produce a trend, and therefore cannot satisfy the thing being asked for.
Calibration status
Also reviewed at the annual audit. A sensor outside its calibration period casts doubt on every reading it produced during that period, which makes calibration tracking part of the monitoring record rather than a separate maintenance chore.
Per-item selection
Condition-based maintenance can be selected on an item-by-item basis rather than as an all-or-nothing scheme, so an approved arrangement can cover specific machinery while the rest remains on intervals. That makes a staged programme legitimate rather than a compromise.
What this means for the integration
Retention and calibration are requirements, not features. Ask how long readings are kept, whether a trend can be produced for a specific point across a full survey cycle, and whether sensor calibration dates live alongside the readings rather than in a separate spreadsheet somebody maintains.
Bandwidth Decides the Architecture
Vibration data is high-frequency by nature and satellite capacity is metered, which forces a design decision that shore-based condition monitoring never has to make. Start a free trial and establish what actually crosses the link.
The problem
Raw vibration waveforms are large. Sending them continuously from every monitored point on every vessel to a shore analytics platform is a bandwidth commitment that changes the economics of the whole programme, and published shipowner guidance already identifies poorly optimised synchronisation as a source of additional satellite traffic.
The answer
Process at the edge and transmit features rather than raw signal — overall values, band energies, derived indicators, alarm states — with raw data retained aboard and pulled only when somebody needs to diagnose something specific. This edge-processing pattern is standard in industrial deployments and is close to mandatory at sea.
The question to ask
How many megabytes per vessel per day does this system transmit in normal operation, and what happens on reconnection after a week with no link — does it send the backlog at once. A supplier who has deployed at sea will answer immediately. One who has only deployed ashore will not have considered that the network is metered.
Sensors produce data. Dashboards produce awareness. Only work orders produce maintenance.
Be Honest About Which Programme You Are Buying
There are three distinct things sold under adjacent language, with very different costs and very different prerequisites. Choosing the wrong one is the most expensive mistake available here. Book a walkthrough and be specific about which you are scoping.
Level 1
Instrumented monitoring
Sensors, readings, thresholds and alarms, with the alarm producing a work order. Modest cost, immediate benefit, and the level at which most fleets should start. Requires disciplined threshold setting and an equipment register good enough for alarms to attach to the right component.
Prerequisite: a maintenance system that can receive a triggered job
Level 2
Condition-based maintenance
Intervals replaced by condition for selected items, with trends retained, calibration tracked and, where relevant, the arrangement approved by class. Formalised in ISO 17359 and supported by ISO 13374 for data processing. This is the level that changes the maintenance programme rather than supplementing it.
Prerequisite: retained history, calibration records, and class engagement
Level 3
Predictive, with machine learning
Models trained on time-series sensor data, typically with edge inference and cloud training. Published 2026 figures for industrial deployments put pilot investment in the region of fifty to two hundred thousand euros and production deployment at five hundred thousand to two million, with the strongest fit in capital-intensive industries where downtime exceeds a hundred thousand dollars an hour.
Prerequisite: years of labelled data, and an honest look at whether your downtime economics justify it
Most fleets asking about IoT are describing level one and being quoted level three. The reported benefits at the top level are substantial — availability gains, maintenance cost reduction in the region of twenty to forty percent, unplanned downtime reduction of thirty to fifty percent — and those figures come from process industries with continuous production and enormous hourly downtime costs. They are not a forecast for a twelve-vessel fleet, and a supplier quoting them without asking about your downtime economics is selling a model built for somebody else's problem.
The Specification, Point by Point
Complete this per monitored measurement point before any hardware is ordered. It is tedious and it prevents almost everything that goes wrong. Start a free trial and use the equipment register as the anchor for every row.
Table 1: Per-Point Specification
| Item | What to define | Why it matters | The failure if it is skipped |
|---|---|---|---|
| The component | Which entry in the equipment register this point measures | Alarms must attach to a component, not to a sensor name | Work orders raised against a device rather than the machine |
| Quantity and statistic | Acceleration, velocity or displacement; RMS, peak or peak to peak | The same event produces very different numbers in each | Thresholds and readings expressed differently, silently |
| Unit | Stated explicitly, on both sides of the interface | Metric and imperial conventions coexist in this field | Plausible numbers that are wrong by a constant factor |
| Operating condition | The load or state at which the measurement is valid | Vibration monitoring is defined for normal operation | Readings taken at idle compared against loaded thresholds |
| Alert threshold | The value, its basis, and who set it | ISO guidance gives a starting point, not a final answer | Alarm fatigue or silence, both of which end the programme |
| Threshold governance | Who may change it, on what evidence, and where the change is recorded | A quietly widened limit is undetectable later | No way to distinguish a tuned threshold from a suppressed one |
| The resulting job | What work order is created, at what priority, for which role | This is the link that turns monitoring into maintenance | A dashboard, and a maintenance programme unchanged |
| Acceptance criterion | What condition allows the resulting job to be closed | Otherwise closure is somebody's judgement with no record | Jobs closed on the reading falling back below the line by itself |
| Retention | How long readings are kept and at what resolution | Trend analysis over a survey cycle requires retained history | Unable to produce the trend an annual audit reviews |
| Calibration | Sensor calibration interval, status and where it is recorded | Calibration status is reviewed alongside the trends | Readings of uncertain validity across an unknown period |
| Bandwidth | What crosses the satellite link and at what volume | Raw waveform transmission is a recurring cost per vessel | A communications bill nobody attributed to the monitoring project |
| Failure behaviour | What happens when the sensor, the gateway or the link fails | A silent sensor looks identical to a healthy machine | Monitoring that stopped months ago and nobody noticed |
The Failure Nobody Designs For
A dead sensor and a healthy machine produce the same thing on a dashboard: no alarms. That equivalence is the most dangerous property of any monitoring system. Schedule a demo and ask specifically how the absence of data is detected.
Heartbeat monitoring
The system should expect a reading from each point on a defined cadence and raise something when one does not arrive. Silence has to be an event rather than an absence of events, or a failed sensor becomes indistinguishable from a machine in perfect condition.
Plausibility checks
A reading that is identical every hour, or zero, or physically impossible for that machine, is a sensor fault rather than a measurement. Flat-lined data is a common failure and it looks like stability to anybody scanning a chart.
Coverage reporting
A periodic statement of how many monitored points reported successfully in the last period, per vessel. Without it, coverage erodes gradually as sensors fail and nobody notices until somebody asks for a trend and finds a gap.
And it becomes a work order too
A sensor that stops reporting should raise a job to check it, exactly like any other equipment fault, assigned to a person with a date. The monitoring hardware is machinery on the vessel and belongs in the maintenance system alongside everything else it watches.
Questions Before You Commit
Ask these of the monitoring supplier and of the maintenance platform, because this only works if both ends agree. Start a free trial and put them to us as readily as to any sensor vendor.
Table 2: Questions for Both Sides
| Area | The question | A real answer | What should worry you |
|---|---|---|---|
| The last link | Show me a threshold breach producing an assigned work order, live | Demonstrated end to end on real equipment | A dashboard, and an explanation of how you would then raise a job |
| Component mapping | Does the alarm attach to our equipment register entry or to a sensor ID? | To the component, so history accumulates in the right place | A parallel asset list maintained separately from the maintenance one |
| Threshold setting | How do you arrive at initial thresholds, and how are they revised? | A standards-based starting point with a defined revision process | Defaults nobody can explain, or a promise to tune it later |
| Trend retention | Can you produce a trend for one point across a full survey cycle? | Yes, with the retention period stated | Recent data only, which fails the audit question |
| Calibration | Where does sensor calibration status live? | Alongside the readings, with expiry visible | A separate spreadsheet, which is where calibration goes to die |
| Bandwidth | Megabytes per vessel per day, and behaviour after a week offline? | A measured figure, because they have deployed at sea | Never having been asked, which tells you about their deployments |
| Dead sensor | How do we find out a point has stopped reporting? | Heartbeat alerting plus periodic coverage reporting | Silence, which is indistinguishable from good health |
| Programme level | Are you selling monitoring, condition-based maintenance, or prediction? | A clear answer with the prerequisites for each stated | Predictive language attached to threshold alarms |
| Evidence | Which fleet of our size runs this pairing, and can we speak to them? | A contactable reference with both systems in production | Industrial case studies from land-based process plants |
CONDITION MONITORING REALITY
Standards references here are indicative and you should work from the current published texts. ISO 17359 covers general principles of condition-based maintenance, ISO 13374 addresses data processing and communication, the ISO 20816 and ISO 10816 series govern vibration measurement and evaluation with alarm-setting guidance by machinery group, and lubricant analysis has its own standards. Applicability, current editions and the specific criteria for your machinery should be confirmed rather than assumed from any summary. Class requirements are vessel and society specific. Where condition-based maintenance supports an approved survey arrangement, the audit expectations described — trend analyses and calibration status — reflect general practice; confirm the position with your own society before relying on a monitoring installation for survey credit. The predictive maintenance investment and benefit figures come from industrial deployments outside shipping, principally in capital-intensive process industries with very high hourly downtime costs, and should not be read as a forecast for a commercial fleet. Marine Inspection connects IoT sensor data and remote telemetry into inspection and maintenance records; it is not a vibration analysis platform, and a serious condition monitoring programme pairs a specialist monitoring supplier with a maintenance system that can act on what it produces.
Frequently Asked Questions
Why do condition monitoring projects fail?
Almost never at the sensor, and almost always at the last link. The chain runs sensor, reading, threshold, alert, dashboard — and stops there. Somebody ashore looks at a screen, forms an impression, and nothing is assigned to anybody with a date. The result is a monitoring subscription, a wall of trend charts, and a maintenance programme still running on the intervals it used before. Sensors produce data and dashboards produce awareness, but only work orders produce maintenance. Specify what job is created, against which component, at what priority and for which role, before choosing any hardware.
How should thresholds be set?
From a standards-based starting point, then revised against your own machines under a defined process. The ISO 20816 series provides alarm-setting guidance for machinery groups defined by power rating and base type, rigid or flexible, which gives a defensible opening position rather than a guess. Both failure directions are serious: too tight produces alarms on normal operating variation, and alarm fatigue is worse than no monitoring because it destroys trust in every alert the platform generates; too loose means nothing fires until damage is advanced, which looks like a healthy programme with no alarms. Decide in advance who may change a threshold, on what evidence, and where the change is recorded.
What does a unit mismatch actually do?
Produces plausible numbers that are wrong, which is the most dangerous kind of error because nothing looks broken. Vibration can be expressed as acceleration, velocity or displacement, measured as root mean square, peak or peak to peak, in millimetres per second or inches per second depending on regional convention — and both metric and imperial systems serve the same purpose in different units. A threshold expressed one way against a reading delivered another will breach or fail to breach for reasons unconnected to the machine. Define the quantity, statistic, unit and operating condition per measurement point, agree them on both sides of the interface, and verify against a known reading at commissioning.
What does class want from a condition monitoring installation?
Trends and calibration, rather than current values. Where machinery runs on condition-based maintenance under an approved scheme, the annual audit reviews vibration and lubricating oil trend analyses along with calibration status — so a system holding recent readings without retained history cannot produce what is being asked for, and a sensor outside its calibration period casts doubt on everything it recorded. Condition-based maintenance can also be selected item by item rather than as an all-or-nothing scheme, which makes a staged programme legitimate. Confirm the specific expectations with your own society before relying on monitoring for survey credit.
How much data can we actually send from a vessel?
Less than a shore-based architecture assumes, which forces edge processing. Raw vibration waveforms are large, and transmitting them continuously from every monitored point on every vessel is a bandwidth commitment that changes the economics of the programme — published shipowner guidance already lists poorly optimised synchronisation among the hidden costs of maritime software. The workable pattern is processing at the edge and transmitting derived features, overall values and alarm states, with raw signal retained aboard and retrieved only for specific diagnosis. Ask any supplier for megabytes per vessel per day in normal operation and what happens on reconnection after a week offline.
Do we need machine learning?
Probably not yet, and being clear about the level you are buying prevents an expensive mismatch. Instrumented monitoring with sensible thresholds feeding work orders is where most fleets should start and delivers benefit immediately. Condition-based maintenance, formalised in ISO 17359, replaces intervals with condition for selected items and requires retained trends, calibration records and class engagement. Predictive modelling on time-series data sits above both, and published 2026 industrial figures put pilot investment around fifty to two hundred thousand euros with production deployment between five hundred thousand and two million — with the strongest fit in capital-intensive industries where downtime exceeds a hundred thousand dollars an hour. Most fleets asking about IoT are describing level one and being quoted level three.
The link that makes it worth buying
Threshold Breach to Assigned Work Order, With a Name and a Date
Write the sentence for every monitored point before ordering hardware: when this limit is crossed on this machine, this job is created against this component, at this priority, for this role, closable on this criterion. Define the quantity, statistic and unit on both sides. Decide who may move a threshold and where that is recorded. Retain readings long enough to produce a trend across a survey cycle, keep calibration status alongside them, process at the edge so the satellite link survives it, and treat a silent sensor as an event rather than as good news. Do that and condition monitoring becomes maintenance. Skip the last link and it stays a dashboard.
.png)
.png)
