Mitsubishi CNC data solution: what a gateway actually reads
This page explains how a data collection gateway pulls signals out of a Mitsubishi-controlled machine tool, which values tell you something useful, and where the limits sit. It is written for process and manufacturing engineers who have to decide whether to instrument a shop floor or leave it alone.

Mitsubishi CNC data solution: where the numbers come from
A Mitsubishi CNC data solution is not one box. It is a path: the control writes values into shared memory, an interface layer reads them, and a gateway converts them into a format your software can consume. On a Mitsubishi M700 or M80 series control, that path usually runs over Ethernet through MTConnect or OPC UA, or over a serial and I/O route on older machines.
The control already knows almost everything you want. Spindle speed, feed override, axis position, program number, tool number, alarm code and cycle state all exist inside the CNC. The gateway does not measure them. It copies them, timestamps them and hands them on.
That distinction matters when something looks wrong. If a dashboard shows a spindle speed of zero while the tool is clearly turning, the fault is almost never the spindle. It sits in the read path: wrong register, stale cache, dropped packet, or a polling rate slower than the state change.
So the first engineering question is not which software to buy. It is which values exist in your control, at what refresh rate, and over which physical port. Answer that and the rest of the project becomes a wiring and naming exercise.
Ethernet, serial and I/O: picking the transport
Most shops start with Ethernet because the cable is already there. MTConnect gives you a read-only stream in a fixed dictionary, which is why integrators like it. OPC UA gives you typed values and a browse tree, which is better when you need to write back or join data across vendors.
Older Mitsubishi controls may only expose a serial port or a set of relay contacts. That is not a dead end, but it changes what you can collect. A serial link can carry cycle start, cycle end and alarm codes. It will not carry high-rate axis positions.
Digital I/O is the cheapest route and the bluntest. One input per signal tower lamp, one relay per machine. You get running, stopped and alarm. You do not get which alarm, how long the tool was in cut, or how much the feed override was dialled back.
The usual failure is mixing transports in one plant and then trying to compare the results. Keep one clock source, one naming convention and one timestamp format across every machine, or the reports will disagree with each other forever.
Which signals earn their place on the dashboard
Cycle state is the workhorse. Running, idle, alarm and setup cover most of what a planner needs, and every transport can carry them. If you only collect four things, collect these.
Spindle load is the most underrated value on a machine tool. A rising load trend on the same program and material tells you the tool is wearing. A sudden spike tells you a chip pack or a hard spot. It is a process signal, not just a maintenance one.
Program number and tool number let you tie a cycle to a job. Without them, every event is anonymous and no report can be trusted. With them, you can compare two identical machines running the same part and see which one is slower.
Axis position and feed override are useful for debugging, less so for daily reporting. They generate a lot of data, they need a fast poll, and most planning decisions do not depend on them. Add them last, after the cheap signals are stable.
What machine data changes on the shop floor
Data on its own does not hold tolerance. It changes how fast you find the cause when tolerance drifts. If a batch of aluminium housings starts running outside ±0.005 mm, spindle load and cycle time together narrow the search to tool wear, coolant or fixture clamp, usually within one shift.
It also changes quoting. When you know the real cycle time for a part family on a given machine, you stop padding estimates. At GreatLight we run 127 high-precision CNC machines, including 16 simultaneous 5-axis machining centers and 16 mill-turn centers, and a part that fits a 750 × 1,150 × 550 mm travel mill may cost very differently from one that has to go on a 4,000 mm bed.
For high-mix work the value is smaller. If every job is a one-off prototype and the operator is standing at the control, an elaborate data layer mostly records what you already saw. The payback appears when the same part runs repeatedly across several machines.
None of this replaces inspection. A 100% inspection step before shipment, with raw material checks and in-process monitoring, is what proves the parts. Machine data tells you where to look. It does not sign the report.
Control series versus what you can realistically collect
Typical field experience; confirm against your own control option list
| Control generation | Best transport | Signals you get | Watch out for |
|---|---|---|---|
| M80 / M800 (new) | Ethernet, MTConnect or OPC UA | Cycle, alarm, program, tool, load | Poll rate settings are easy to miss |
| M700 / M70 (newer) | Ethernet or OPC UA | Cycle, alarm, program, tool, load | Older firmware lacks some nodes |
| M600 / M500 (legacy) | Serial or Ethernet option | Cycle, alarm, program number | High-rate axes usually not available |
| M50 / M3 (old) | Digital I/O only | Running, stopped, alarm | One signal per wire, no detail |
| Mixed plant | One gateway per protocol | Depends on the weakest link | Clock and naming drift over time |
When to wire it up, when to leave it
If the same parts run repeatedly across several machines, build the data path and start with cycle state, alarm code and spindle load. If the shop is high-mix one-off work with the operator at the control, digital I/O for running and alarm is enough; a full gateway will cost more than it returns.
Questions engineers ask next
Do I need to write anything back to the CNC?
Most projects are read-only. MTConnect is read-only by design and that is usually enough for planning, OEE and maintenance alerts.
Writing back is a different risk class. It touches offsets, programs and interlocks, so it needs a change-control process and a rollback plan before the first write goes live.
How often should the gateway poll?
Match the rate to the decision. Cycle state every 1–5 seconds is plenty for planning. Alarm codes need to be caught within one poll or you lose the event order.
Spindle load and axis position are the ones that push rates up. If you only need trend lines, one sample per second is usually enough.
Can old machines be added later?
Yes, but expect a different data set, not a slower version of the same one. Legacy controls may only give you running, stopped and alarm through relay contacts.
Keep those machines in a separate group in your naming scheme so reports do not blend a full stream with a three-bit one.
What breaks first?
Naming, not hardware. Two integrators label the same signal differently, and a year later nobody can join the tables.
The second most common failure is the clock. Gateways that timestamp locally drift out of step with the control and with each other.
Does this replace a CMM report?
No. Machine data shows behaviour over time; metrology shows whether a part is good. They answer different questions.
Use the data to decide when to inspect more often, then let inspection decide whether the lot ships.
Who owns the data?
Agree on that before the first cable is run. Decide who hosts the gateway, who can export, and how long raw events are kept.
If parts are customer-owned designs, keep uploads and any shared data under an NDA and a defined access list.
Start with the signals you can actually use
Send us your drawing and machine list. We will quote the machining and tell you which control signals on those machines are worth collecting.
Quotation and free DFM analysis within 12 hours100% inspection before shipment