AlarmIQ Engineering Guide • March 28, 2026 • 10 min read

AI Predictive Maintenance for PLCs: From Reactive to Proactive

Your PLC already knows your equipment is failing. The data is there — in cycle times, motor currents, temperature trends, and alarm patterns. AI just connects the dots before the breakdown happens.

The Maintenance Maturity Spectrum

Most manufacturing plants operate somewhere on a spectrum of maintenance strategies, and the position determines both downtime and cost:

Reactive (run-to-failure) — fix it when it breaks. The cheapest approach until something expensive fails catastrophically at 2 AM on a Friday. Common in plants with tight budgets and no maintenance planning.
Preventive (time-based) — replace parts on a fixed schedule regardless of condition. Better than reactive, but wasteful. You replace a bearing at 6 months even if it had 2 years of life left, and miss the one that fails at 5 months.
Condition-based — monitor equipment health indicators and act when degradation is detected. More efficient than time-based, but still reactive to degradation rather than predictive of failure.
Predictive (AI-driven) — use machine learning to forecast when failure will occur based on patterns in operating data. Maintenance is scheduled at the optimal time — late enough to maximize component life, early enough to prevent unplanned downtime.

The jump from preventive to predictive is where AI enters the picture. But "AI predictive maintenance" as marketed by most vendors is often vaporware. Here is what actually works and what does not.

What PLC Data Feeds Predictive Models

Every PLC continuously collects data that contains maintenance-relevant signals. The challenge is knowing which signals matter and how to interpret their trends. The most valuable PLC signals for predictive maintenance fall into several categories:

Motor Current and Power

Motor current draw is one of the most reliable indicators of mechanical health. A motor driving a healthy conveyor draws a consistent current at a given speed and load. When bearings start to fail, friction increases and current rises. When a belt begins to slip, the current pattern changes. When a gearbox loses lubrication, the current during acceleration increases.

Most variable frequency drives (VFDs) make motor current available as a PLC tag. If your PLC reads Drive_01.OutputCurrent, you already have the data — you just need to trend it over weeks and months rather than viewing it as a point-in-time value.

Cycle Time and Throughput

Cycle time is the simplest and most universally available signal. Nearly every machine tracks it. Healthy equipment produces consistent cycle times. When a hydraulic system leaks, cycle times gradually increase. When a pneumatic cylinder slows, the cycle extends. When a tool wears, either the cycle lengthens (if the process compensates) or quality degrades (if it does not).

The key insight is not the absolute cycle time but the variance. A machine that runs 4.2-second cycles with 0.05-second variance is healthy. The same machine running 4.2-second cycles with 0.3-second variance is telling you something is wrong, even though the average has not changed.

Temperature Trends

Temperature sensors are everywhere in industrial equipment — on motors, bearings, hydraulic reservoirs, spindles, gearboxes, and enclosures. Temperature alone is a lagging indicator (by the time a bearing is hot, damage is advanced), but the rate of temperature rise under load is a leading indicator.

A healthy motor might reach 65C after 30 minutes of operation. If that same motor reaches 65C after 20 minutes, with no change in ambient conditions or load, the cooling system or bearing condition has degraded. An AI model can detect this drift long before the temperature reaches an alarm setpoint.

Vibration Data

Vibration analysis is the gold standard for rotating equipment health. Traditionally, this required dedicated vibration sensors and specialized analysis software. Increasingly, vibration sensors are connected directly to PLC I/O modules or edge devices that make the data available as PLC tags.

The useful vibration signals for predictive maintenance are:

Overall RMS velocity — general indicator of vibration severity (ISO 10816 classification)
Frequency spectrum peaks — specific frequencies correspond to specific failure modes (1x = imbalance, 2x = misalignment, bearing frequencies = bearing defects)
Crest factor — ratio of peak to RMS, increases as bearing defects develop

Alarm Patterns

This is the most overlooked predictive signal and the one most directly relevant to PLC alarm management. Equipment does not fail suddenly in most cases. It degenerates. And that degeneration shows up in alarm patterns:

An intermittent sensor alarm that fires once a week, then twice a week, then daily
A high-temperature alarm that starts activating in the afternoon when it never did before
A drive fault that clears on reset but recurs with increasing frequency
A sequence timeout that grows from rare to occasional to common

Each of these patterns is a predictive signal. A maintenance technician who reviews alarm logs might notice the trend manually. An AI system detects it automatically and quantifies the rate of change.

How Machine Learning Models Work on PLC Data

The most practical ML approach for PLC-based predictive maintenance is anomaly detection, not failure prediction. The distinction matters.

Failure prediction requires labeled training data: examples of normal operation paired with examples of specific failure modes. This data is expensive to collect because failures are (hopefully) rare, and each failure mode requires separate training examples. For most plants, this data simply does not exist.

Anomaly detection requires only normal operating data. The model learns what "normal" looks like across multiple signals simultaneously, then flags deviations from normal. This approach works because:

Normal operating data is abundant — every day of production generates training examples
The model does not need to know what specific failure looks like, only that current behavior differs from baseline
It catches unexpected failure modes that would not be in a labeled training set

A practical example: A packaging machine runs with motor current of 12.3A, cycle time of 3.8s, and bearing temperature of 52C during normal production. An anomaly detection model learns these relationships, including that current increases slightly as temperature rises. One day, current is 13.1A but temperature is only 48C. No single value is alarming, but the relationship between them has changed. The model flags the anomaly. Investigation reveals a worn belt that is increasing mechanical load while reduced friction is lowering temperature. A traditional alarm on current alone would not trigger until 15A.

Practical Implementation Steps

Implementing AI predictive maintenance does not require a massive IT project. A pragmatic approach:

Start with data collection. Before any AI, you need historical time-series data. Configure your PLC or edge gateway to log key signals (motor currents, temperatures, cycle times, alarm events) to a database or historian. Minimum useful history: 3 months of production data.
Pick one machine. Do not try to monitor the entire plant at once. Choose a machine that is critical, has a history of unplanned downtime, and has accessible PLC data. Success on one machine justifies expansion.
Define your signals. Select 5-15 signals per machine that are relevant to equipment health. More is not better — irrelevant signals add noise that degrades model performance.
Establish a baseline. Run the anomaly detection model on 2-3 months of normal operating data. Tune the sensitivity so it does not alert on routine process variation (product changeovers, startup/shutdown sequences, shift changes).
Validate with maintenance records. Compare anomaly detections against historical maintenance events. Did the model detect the bearing replacement that happened in Month 4? Did it flag the drive failure in Month 6? If not, adjust signals or sensitivity.
Integrate into workflows. An AI prediction has zero value if nobody acts on it. Connect anomaly alerts to your CMMS or maintenance scheduling system so that they generate work orders, not just dashboard notifications.

The Current State of AI in Manufacturing Maintenance

Honest assessment: AI predictive maintenance is real but overhyped. Here is what works today and what does not:

What Works

Anomaly detection on time-series PLC data — well-established, practical, and delivers measurable results when properly implemented
Alarm pattern analysis — detecting increasing alarm frequency and correlating alarm clusters with impending failures
Remaining useful life (RUL) estimation for specific, well-studied components like bearings with dedicated vibration sensors
Automated report generation — AI summarizing equipment health trends and highlighting machines that need attention

What Does Not Work (Yet)

Plug-and-play predictive maintenance — no product works out of the box without domain knowledge about your specific equipment and process
Predicting specific failure modes without labeled failure data from your equipment
Replacing maintenance expertise — AI augments experienced maintenance teams, it does not replace them. A model can flag an anomaly, but a technician still needs to diagnose and repair.
Generic models across different machines — a model trained on one extruder does not transfer to a different extruder without retraining

AI-Powered Alarm Diagnostics

AlarmIQ uses AI to analyze PLC alarm patterns, detect increasing failure signatures, and trace every alarm to its root cause. Predictive insights from the alarm data your PLC already collects.

Learn About AlarmIQ

Signals You Should Be Collecting Today

Even if you are not ready for full AI predictive maintenance, start collecting these signals now. Future you will thank present you for having the data:

Every VFD output current and speed — the single most valuable predictive signal for motor-driven equipment
All temperature sensors — motors, bearings, hydraulic reservoirs, spindle housings
Cycle times with timestamps — not just averages, but individual cycle durations
Alarm events with timestamps — every alarm activation, acknowledgment, and return-to-normal
Hydraulic and pneumatic pressures — trends reveal pump wear, leaks, and valve degradation
Runtime hours per equipment — basic but essential for correlating degradation with usage