Retrofit Predictive Maintenance: Bring Old Machines Online

In one line: You don't need a new machine to get predictive maintenance; you need a few external sensors, a read-only gateway that speaks OPC-UA or MQTT, and a model that learns this machine's normal from the day the sensors go on. A 30-year-old pump with no PLC and no network is a first-class candidate for condition monitoring; the machine learning doesn't care how old the asset is, only what its sensors say.

The machine that has run for 30 years and never been watched

Every plant has one. A pump, a motor, a compressor, a gearbox that has run since before the current maintenance team was hired. It's reliable enough that nobody has touched the capital budget to replace it, critical enough that the line stops when it stops, and old enough that it has no built-in sensors, no fieldbus, and no way to tell anyone how it's doing. So it runs to failure. And when it fails, it does so at 3 a.m., with no warning, on a part nobody stocks.

That machine is exactly where predictive maintenance pays off the most. And the reason plants skip it is a myth: that condition monitoring needs a modern, connected, sensor-equipped asset. It doesn't. It needs a signal. You can add the signal.

What "retrofit predictive maintenance" actually means

Retrofit predictive maintenance is the practice of adding condition monitoring to an existing machine that wasn't built for it: by mounting external sensors on the outside of the asset, collecting their data through a small gateway, and feeding it to a predictive model. Nothing inside the machine changes. You're not modifying the control system, re-validating the process, or replacing the equipment. You're attaching a nervous system to a machine that never had one.

It's the opposite of the greenfield case, where a new asset ships with embedded sensors and a network stack. The industrial world is overwhelmingly brownfield: a well-maintained machine tool, pump, or motor routinely runs for 20, 30, even 40 years, and most of the installed base predates the idea of IIoT connectivity entirely. Retrofit is how that installed base gets monitored without a rip-and-replace capital project.

Why old machines are the best PdM candidates, not the worst

Here's the counterintuitive part: the oldest, least-connected machines usually have the highest return on condition monitoring, because they're the ones nobody is watching.

New assets often arrive with some built-in monitoring and land on a preventive schedule. Old brownfield machines tend to sit on run-to-failure or a fixed calendar interval, which is precisely the strategy predictive maintenance beats by the widest margin. The U.S. Department of Energy's O&M Best Practices work at PNNL puts the savings of a predictive program at roughly 8–12% over preventive maintenance and 30–40% or more over reactive, run-to-failure maintenance. The bigger delta is against exactly the strategy your oldest machines are running.

Stack that on the cost of the failure itself. Deloitte's analysis of predictive technologies for asset maintenance estimates unplanned downtime costs industrial manufacturers on the order of $50 billion a year, with predictive approaches lifting equipment uptime by 10–20%. A retrofit that gives you three weeks of warning on a critical 25-year-old machine pays for a handful of sensors many times over the first time it turns a seizure into a scheduled swap.

The capital-avoidance argument seals it. Replacing a working machine to "make it smart" is the most expensive way to get a signal that a handful of external sensors provides. Retrofit lets you keep decades of proven mechanical investment and add only the part that was missing.

The retrofit stack: how you actually wire it up

A retrofit monitoring path has three layers, and none of them touch the machine's control logic:

1. External sensors, mounted non-invasively. The workhorses bolt or clamp onto the outside of the asset:

Vibration: accelerometers (IEPE or 4–20 mA) on the bearing housings, the single richest signal for rotating equipment. Two axes on a critical bearing beat one.
Temperature: RTDs or thermocouples on bearings, windings, or casings.
Current and voltage: clamp-on current transformers on the motor leads, a non-contact window into mechanical load and electrical faults.
Pressure, flow, RPM, and acoustic where the failure mode calls for them.

You don't instrument everything. You pick the two or three signals that reveal the machine's dominant failure modes: for a pump, that's usually vibration on the bearings plus temperature plus motor current. (Which sensor reveals which failure mode is the whole selection problem.)

2. A retrofit gateway or edge PLC. The sensors wire into a small industrial gateway that digitizes the analog channels and publishes them over a standard protocol. A gateway that exposes an OPC-UA server or publishes MQTT drops straight into a modern ingestion path; one that only speaks Modbus can be bridged up to either. This box is the machine's new network interface: the thing it never had.

3. Read-only ingestion into an on-premise pipeline. The monitoring layer subscribes to the gateway's tags read-only and pushes them into a local time-series pipeline that computes features (rolling statistics, FFT, bearing frequencies) and runs the models. Because the gateway sits beside the machine rather than inside its control loop, there is no write path back to anything, a property that matters enormously when the OT-security team reviews it. If you already run a SCADA platform, the same principle applies through Ignition's OPC-UA server; retrofit is the case where there's no SCADA to read from, so you add the gateway instead.

The result is a machine that, from the pipeline's point of view, looks identical to a modern connected asset. The models never learn that the pump is old. They only see the data.

The honest part: what age does and doesn't change

Retrofit works because a good anomaly model learns behavior, not specifications. An LSTM autoencoder trained on a specific machine learns to reconstruct that machine's normal vibration, temperature, and current patterns; when a bearing starts to degrade, reconstruction error climbs and the anomaly surfaces, regardless of the machine's age, make, or model. There's no equipment library to match against and no assumption that the asset is new.

But be honest about two constraints, because a vendor who isn't will burn you:

The model doesn't get the machine's 30-year history: the sensors weren't there. It learns "normal" from the moment your new sensors start producing data. That's fine: cold-start anomaly detection works from day one (starting from a conservative, general baseline), and the model's picture of this specific machine sharpens over the following weeks as it sees more normal operation across speeds and loads. You get early, coarse coverage immediately and precise, per-machine coverage soon after, not a magic read of decades you never measured.

A retrofit signal is only as good as sensor quality and placement. A loosely mounted accelerometer or a sensor on the wrong side of a coupling produces data that no model can rescue. Mounting, orientation, and sample rate are where retrofit projects live or die, which is why the two or three well-placed sensors beat a dozen sloppy ones.

And the question every reliability engineer asks next: what happens when a retrofit sensor fails? It will: outdoor clamps come loose, cables chafe. A monitoring system built for the real world has to distinguish a sensor fault from a machine fault, and keep running on what's left. Prevly validates every incoming reading, stamps it with a quality score, and treats a dead or out-of-range sensor as a flagged sensor problem, not a phantom machine anomaly. The models that don't depend on the missing channel keep scoring; the ones that do fail safe and say "sensor offline" rather than emitting garbage. For measurements you can't afford to lose, the fix is deliberate redundancy (a second accelerometer on the critical bearing), decided up front, not after the outage.

A 25-year-old pump, retrofitted

Here's what a retrofit looks like in practice, illustrative of the pattern, not a specific customer result. Take a 25-year-old centrifugal pump: no sensors, no network, running to failure on a "replace it when it dies" plan.

You clamp two IEPE accelerometers on the drive-end and non-drive-end bearing housings, an RTD on the drive-end bearing, and a current transformer on the motor supply. All four wire into a small gateway exposing an OPC-UA server. Prevly Edge subscribes read-only and starts ingesting. From day one, cold-start anomaly detection is watching; within a couple of weeks it has a tight baseline for the pump at its normal operating point.

Six weeks later, the model flags an anomaly. And because it's explainable, the alert carries the evidence: elevated radial vibration (vibration_x_rms: +0.34), bearing temperature rising faster than the housing (temperature_delta: +0.21), and a subtle roughness in motor current (current_kurtosis: +0.12). A reliability engineer reads that as a developing outer-race bearing defect, confirms it with a targeted envelope measurement, and schedules the swap for the next planned downtime. A machine that had never in 25 years told anyone anything just bought three weeks of warning, from sensors that cost less than one hour of the downtime they prevented.

What this looks like with Prevly

Prevly is built for the brownfield reality, not just the greenfield demo:

Read-only ingestion from your gateway. Prevly Edge subscribes to the retrofit gateway over read-only OPC-UA (or ingests MQTT), a proven ingestion path, with no write path to any control system. It uses whatever sensors you mount; there's no proprietary hardware requirement.
Runs on-premise. The pipeline, models, and storage stay inside the plant network. A retrofit doesn't drag your old machine's data to someone else's cloud: see on-premise vs cloud PdM.
Detection from day one, RUL with honest uncertainty. Cold-start anomaly detection covers the machine immediately; remaining-useful-life estimates come as conformal prediction intervals, validated on real NASA C-MAPSS data (an honest headline of RMSE 14.33 on the standard FD001 test), so you plan around a defensible window instead of a false-precision date.
Explainable, and it becomes a work order. Every alert reports which sensors drove it, and a prediction becomes a drafted work order (asset, likely fault, recommended action, sensor evidence attached) that coexists with your existing CMMS rather than replacing it.

Scoping which sensors go where on your specific machine is part of setting up a pilot, not a self-service checkout: we walk through the failure modes and the sensor placement with you before anything gets mounted.

Frequently asked questions

Can you do predictive maintenance on old machines without built-in sensors? Yes. That's what a retrofit is: you mount external sensors (vibration, temperature, current) on the outside of the machine, collect their data through a small gateway, and feed it to a predictive model. The machine's age doesn't matter to the model: it learns that specific machine's normal behavior from the data the new sensors produce.

Do I have to replace my PLC or the machine itself? No. A retrofit adds a monitoring layer alongside the machine; it doesn't modify the control system or the equipment. Many brownfield machines have no PLC at all; the retrofit gateway becomes the machine's first network interface, and it reads sensors, it doesn't control anything.

What sensors do I need to retrofit a machine? Start with the signals that reveal the dominant failure modes. For most rotating equipment that's vibration on the bearing housings, plus temperature, plus motor current. Add pressure, flow, RPM, or acoustic where the asset warrants it. Two or three well-placed, well-mounted sensors outperform a dozen poorly installed ones.

Will the sensors or gateway interfere with the machine's operation? No. Sensors mount non-invasively on the outside of the asset, and a monitoring gateway with no write path is a passive observer: it reads values and never sends commands. Done correctly, the machine runs exactly as it did before; you've only added something that listens.

What if a retrofit sensor fails after installation? A well-built system distinguishes a sensor fault from a machine fault. Prevly validates each reading, stamps a quality score, and flags a dead or out-of-range sensor as a sensor problem rather than a machine anomaly. Models that don't need the missing channel keep scoring; the ones that do fail safe. For critical measurements, add a redundant sensor up front so one failure can't blind you.

Start with one machine

You don't have to digitize the whole plant to prove this. Pick the oldest, most critical, least-monitored machine you have (the one that scares you) and retrofit condition monitoring onto it. Read-only, on-premise, on your existing mechanical asset. Within days of the sensors going on, you'll see what a machine that has run silently for decades has actually been doing.

Request a Prevly demo and we'll scope the sensor kit for your first machine with you.