- PLCs are the richest source of machine-level data in any plant — cycle counts, fault codes, current draws, servo positions — but almost none of this data is analysed because there's no layer designed to read it for intelligence.
- AI+PLC integration is always read-only at the PLC layer — AI models consume PLC data from a historian or OPC server; they never write to PLC registers or modify control logic.
- The value unlock in AI+PLC is converting PLC fault codes and parameter drift from reactive alarms into predictive signals — weeks earlier than traditional threshold alerts.
PLCs know more about your machines than any other system in your plant. Almost none of it gets used.
A typical PLC running a CNC machine or packaging line captures cycle time, servo load, axis position, fault code history, and dozens of process-specific parameters — at scan rates measured in milliseconds. This is the most granular operational data in the plant. It almost never gets analysed.
The reason: PLCs were designed for control, not analytics. SCADA provides a visibility layer above them, but typically stores only the tags explicitly configured for trending. The bulk of PLC data — fault histories, servo torque traces, cycle time micro-variations — stays in the PLC’s local buffer and rolls over every few hours.
AI integration changes the value equation: it doesn’t change what PLCs do. It reads what they’re already recording and extracts the failure signatures buried in that data.
AI+PLC integration must never modify PLC program logic or write to PLC control registers without explicit engineering authorisation and a formal change management process. The integration layer reads data only. Automating setpoint changes based on AI recommendations requires a separate, safety-reviewed closed-loop control design — it is not a default feature of predictive maintenance AI.
What PLC data AI actually analyses
Cycle time drift: A press cycle that should take 4.2 seconds and is now averaging 4.4 seconds — the 5% drift is invisible on an HMI screen but visible to a model that knows the baseline.
Fault code frequency patterns: A fault code that fires once a week versus once a month versus once a day — the acceleration tells you more than any single occurrence.
Servo and drive current signatures: Motor current draw contains bearing and gear degradation signatures detectable weeks before a mechanical failure would show up as a threshold breach.
Synchronisation drift in multi-axis systems: Small timing misalignments between axes that individually look within spec but whose combination pattern is a known precursor to a specific failure mode.
The integration architecture
FactoLake connects to PLCs via OPC-UA gateway or direct PLC protocol (Siemens S7, Allen-Bradley Ethernet/IP, Mitsubishi MELSEC) and stores the data stream in a time-series database where MaintIQ can build and run predictive models.
Frequently Asked Questions
No program changes are required for basic data extraction. An OPC-UA server (hardware gateway or software layer) reads the PLC’s existing data table without modifying the control program. For access to data not currently mapped to OPC tags, minor PLC configuration changes (not program logic changes) may be needed.
FactoLake supports Siemens (S7 family), Allen-Bradley/Rockwell (Logix family), Mitsubishi, Omron, Schneider, and Beckhoff via OPC-UA. For legacy PLCs without OPC support, protocol-specific gateways are available.
6-12 months of historical data that includes at least a few past failure events for each equipment type. For equipment that fails rarely, the model learns from similar equipment’s failure patterns and transfers that knowledge.
Related: Connecting SCADA to MES · How to Connect PLCs to AI · Using OPC-UA for AI Integration