Predictive Maintenance Platform: From Condition Data to AI Agents
Unplanned downtime is the most expensive time in manufacturing — and most of it announces itself: in rising vibration levels, drifting currents, climbing temperatures. Predictive maintenance turns these warning signs into plannable service. Yet most projects fail not at the model but before it — at the infrastructure: condition data never gets captured in the first place, histories are missing, and insights never reach the maintenance team.
This guide shows how to lay the foundation for predictive maintenance with IronFlock — from sensor retrofit through data history to alarms and AI agents.
The problem: predictive maintenance rarely fails at the analysis
Four hurdles show up in almost every project:
- No condition data. Vibration, current draw, and temperature are often not measured at all on existing machines — and what the controller knows never leaves the control cabinet.
- No history. Anomalies are only recognizable by comparison with the normal state. Without months of gapless time series, there is no reference point.
- Island solutions per machine. Vendor-locked condition-monitoring systems watch one machine — but not the machine park, and certainly not across sites.
- The last mile is missing. A prediction that reaches no one prevents no failure. Insights have to arrive at the team as an alarm, a dashboard, or a work order.
Step 1: capture condition data — on existing machines too
The data foundation comes from the collector apps for factory data extraction, which run as containers on an edge PC (gateway) in the machine network:
- Retrofit with IO-Link. The IO-Link Collector connects retrofitted vibration, current, pressure, and temperature sensors (from vendors such as ifm, Balluff, or Pepperl+Fuchs) — with automatic sensor discovery and automatic IODD decoding, no account and no API key required. This is how existing machines get condition data without anyone touching the controller.
- Values from the controller. The Industrial Collector reads currents, speeds, and states from PLCs and variable-frequency drives — OPC UA and Modbus work today, Siemens S7 and Allen-Bradley are in early access. Access is read-only by design: the collectors never write into the equipment.
- CNC machines. The MTConnect Collector delivers states, programs, and axis data from machine tools.
All readings are normalized — named, with unit, timestamp, quality flag, and semantic type — and buffered locally through connection outages. The device agent keeps the gateway permanently online and remotely manageable.
Step 2: build histories — in your own database
Every project gets a dedicated TimescaleDB time-series database as its single source of truth. For predictive maintenance that means:
- Gapless time series over months and years — the reference against which anomalies become recognizable in the first place.
- Data sovereignty. The project owner is the data owner, not the app developer. Access for analytics partners is granted granularly and revocable at any time.
- Open standards. Standard TimescaleDB and open formats instead of vendor lock-in — existing analysis tools can plug in directly.
Step 3: from threshold to AI agent
Predictive maintenance does not have to start with machine learning. IronFlock allows a staged entry:
- Stage 1 — thresholds and alarms. The built-in alarm system watches live telemetry with conditions per severity level (Minor, Major, Critical) and notifies by email or SMS — for instance when vibration limits based on ISO experience values are exceeded. Auto-resolve and a comment function included.
- Stage 2 — visibility and trends. In Board Studio, trend dashboards take shape without code: vibration curves, week-over-week current draw, Split Charts per machine. Suspicious drifts become visible before thresholds break.
- Stage 3 — analytics as an app. Consuming apps build read-only on the collected histories through cross-app data access — a predictive-maintenance app that analyzes vibration, current, and temperature histories is the documented model example of this architecture. You install such apps from the App Store or build them yourself.
- Stage 4 — AI agents. With Physical AI you query your condition data in natural language; apps can bring their own app agents that explain failure predictions, put anomaly scores into context, and suggest maintenance actions. These capabilities become available in the AI assistant automatically as soon as the app is installed — no configuration.
Note: Analytics and prediction models are deliberately designed as apps — not as a black box inside the platform. That keeps it transparent which logic operates on which data, and lets specialist expertise be brought into any project as an app.
Typical condition variables — and what they reveal
Which variables are worth starting with depends on the equipment. The classics at a glance:
| Condition variable | What it typically tells you | How to capture it |
|---|---|---|
| Vibration | Bearing wear, imbalance, misalignment on rotating components | IO-Link vibration sensor, retrofitted on the bearing housing |
| Current draw | Rising mechanical load, drive wear, jamming components | Variable-frequency drive via OPC UA/Modbus or IO-Link current sensor |
| Temperature | Overload, lubrication problems, cooling failure | IO-Link temperature sensor or controller value |
| Pressure / flow | Clogged filters, leaks, pump wear in hydraulics and pneumatics | IO-Link sensors or existing process values |
| Cycle and operating hours | Basis for usage-based instead of calendar-based maintenance | Counters from PLC or CNC via collector |
Two things make this list practical. First, every variable is stored in the same normalized schema — a trend board or an alarm works for vibration exactly as it does for pressure. Second: don’t do everything at once. One variable on your most critical machines, captured cleanly, beats ten variables with no history.
For machine builders: roll out predictive maintenance as a service
Machine manufacturers can offer predictive maintenance as a digital service — it is one of the documented flagship examples for app distributors: you deliver the edge logic and the domain expertise; app distribution, updates with rollback, secure remote access, messaging, and data storage come from the platform. Through the deployment options, the same service runs in the cloud, on an appliance at the customer’s site, or in their private cloud — without code changes.
Frequently asked questions
What data do I need to get started?
Less than commonly assumed. A single continuously captured variable per machine — current draw, vibration, or temperature — delivers reliable trends and sensible alarm thresholds after just a few weeks of history. You can expand at any time: the collectors use a shared table schema, and dashboards and analyses grow with you.
Do I have to develop my own machine-learning models?
No. The staged entry starts with threshold alarms and trend dashboards — no models at all. Analytics arrives as an app: from the App Store, from specialists, or built in-house. AI agents from such apps then become available in the assistant automatically, answering questions about anomalies and maintenance recommendations.
Does this work with existing machines that lack modern interfaces?
Yes — that is the most common case. Retrofitted IO-Link sensors (vibration, current, temperature) are discovered and decoded automatically, without touching the controller. Where controller data is available, it comes in additionally via OPC UA, Modbus, or MTConnect.
Where do collection and analysis run — cloud or edge?
Collection always runs at the edge, in containers on a gateway in the machine network, with local buffering through connection outages. History and analysis live in the project database — in the cloud (EU data centers), on an appliance in the local network, or in your private cloud, as you choose. The feature set is identical across all models.
Where to go next
The pragmatic start: use the getting started guide to connect a gateway, install the IO-Link or Industrial Collector, and bring the first condition variable onto a board — in demo mode that works even without hardware. The building blocks in detail: factory data extraction, IoT alarms, and Physical AI.