Machine Monitoring 101: What It Is and What It Actually Tells You

Machine monitoring is the practice of collecting signals from production equipment (run state, cycle counts, alarms, and sometimes process values like temperature or current) and turning them into a live, timestamped record of what each machine did and when. At minimum it tells you whether machines are running, how many parts they made, and how long they sat idle. It does not, by itself, tell you why anything happened. That distinction is where most monitoring projects succeed or quietly fail.

If you run a shop with ten or thirty machines, you already have machine monitoring. It's called walking the floor. It works, sort of, during first shift, when you're there, on the machines you happen to pass. The pitch for doing it with sensors and software is simple: the same visibility, on every machine, every minute, every shift, with a memory that goes back months instead of to lunchtime.

What Is Machine Monitoring, Exactly?

Strip away the vendor language and machine monitoring is three layers.

The first layer is signal capture. Something on or near each machine produces a data stream: a current transformer on the motor feed, a proximity sensor counting cycles, a tap on the stack light, or a direct connection to the machine's PLC or CNC over Ethernet/IP, Modbus TCP, or OPC UA. Modern CNCs often expose rich data through MTConnect or FANUC's FOCAS library; a 1990s machine might give you nothing but an amp reading. Both are fine starting points. If your equipment is older, our guide on getting production data out of old machines covers the retrofit side in detail.

The second layer is state interpretation. Raw signals get converted into states with rules: current above a threshold means running, a cycle sensor pulse increments the part count, no pulse for four minutes during a scheduled shift means idle. This layer sounds trivial and is not. A rule that misreads a warm-up cycle as production, or counts a double-fed part twice, poisons everything downstream. Expect to tune it per machine.

The third layer is storage and display. States and counts land in a time-series database and show up on dashboards, shift reports, and alerts. This is the part vendors demo. It's also, honestly, the easiest part.

What Signals Does Machine Monitoring Actually Capture?

The core set, roughly in order of how often shops start with them: machine state (off, idle, running, faulted), part counts (total, good, reject where you can detect it), cycle times (the timestamps between count pulses, which quietly become one of your most valuable datasets), and alarms or fault codes where a controller exposes them.

Beyond the core set live process signals: spindle load, motor current, hydraulic pressure, barrel temperatures, vacuum levels. These matter for quality and predictive maintenance work, but they're a second chapter. Plenty of shops get twelve months of value from nothing but states and counts.

One signal deserves special mention: the operator. A machine can report that it stopped at 10:42 for eleven minutes. Only a human can say it stopped because the material ran out and the forklift was on the other dock. Good monitoring systems make it fast for operators to attach a reason code to a stop, usually one tap on a floor tablet. Systems that skip this end up with beautiful charts of unexplained downtime.

What the Data Actually Tells You (and What It Doesn't)

Here's what typically falls out of the first few weeks of monitoring a small shop.

Utilization is lower than everyone believed. Shops that estimate 75 percent utilization routinely measure 45 to 55 percent once the counting is honest. This is normal, it's not an indictment of anyone, and it's the single most useful number the system will ever produce, because it reframes the capacity conversation. That machine you were about to buy? You may already own it.

Small stops dominate. A filler line that stops twelve times a shift for ninety seconds each loses more time than the one dramatic two-hour breakdown everyone remembers. Clipboards never catch micro-stops. Sensors catch all of them, and the pattern (always after a changeover, always on the night shift, always lane 3) usually points straight at the fix.

Cycle times drift. The quoted cycle was 34 seconds; the measured median is 41. Multiply seven seconds by 800 parts a day and you've found a phantom machine's worth of capacity hiding in feeds, speeds, and habits.

Now the limits. Monitoring tells you what, when, and how often. It does not tell you why (that needs reason codes and human eyes), it does not fix anything (a dashboard has never repaired a bearing), and it does not automatically improve OEE, a number you should understand before you chase it. Our piece on OEE for small manufacturers covers that math and its traps.

The failure mode we see most often isn't technical. It's a system that gets installed, admired for a month, and then ignored because nobody's morning routine changed. Data with no meeting attached is decoration.

How Much Does Machine Monitoring Cost?

For hardware, plan on $500 to $1,500 per machine for a retrofit (sensors, edge device, wiring) or a few hundred dollars for a gateway if the machine's controller is already networkable. Installation is often half a day to a day per machine.

For software, the market splits in two. Subscription platforms typically run $50 to $150 per machine per month, which is painless at five machines and real money at forty. A custom or self-hosted stack (an MQTT broker, a time-series database, purpose-built dashboards) costs more up front, typically in the tens of thousands including engineering, and then very little per month. The crossover point usually lands somewhere between ten and twenty-five machines, depending on how much of the plumbing you already have.

The honest budgeting advice: whatever you spend on software, reserve time and attention for the human side. The shops that win with monitoring are the ones where the 7 a.m. production meeting starts with yesterday's numbers on a screen and someone asking "what happened at 2:15?"

Where Should a Small Shop Start?

Pick three to five machines, not the whole floor. Choose your constraint machine, your most suspect machine, and one that's easy to instrument so you get a quick win. Capture states and counts only. Put one clear screen where the shift lead can see it, and add reason codes for stops within the first month, while the habit is forming.

Run that for a quarter. You'll learn what your floor actually does, what your team actually looks at, and what the full rollout should look like, and you'll learn it for a fraction of the cost of guessing. Scaling up after a good pilot is easy. Recovering from a forty-machine rollout nobody uses is not.

This is also the stage where architecture choices matter more than they appear to. Pick protocols and storage that won't box you in: MQTT for transport, a real time-series database, and dashboards you can change without a vendor ticket. The industrial automation work we do usually starts exactly here, at the pilot, where the decisions are still cheap.

FAQ

What's the difference between machine monitoring and a full MES?

Machine monitoring answers "what are my machines doing?" An MES (manufacturing execution system) also manages orders, routings, scheduling, quality workflows, and traceability. Monitoring is a component you can run standalone in weeks; an MES is an enterprise project measured in quarters. Most small shops need monitoring long before they need an MES.

Do operators react badly to being monitored?

Sometimes, at first, and the framing decides everything. Monitor machines, not people: the useful findings are almost always material starvation, changeover friction, and maintenance issues rather than operator effort. Shops that share the data openly and use it to remove operators' daily annoyances typically see the skepticism fade within a few weeks.

Can machine monitoring predict failures?

Basic state-and-count monitoring can't, but it builds the foundation. Once you add process signals like vibration, current signatures, or temperatures, trend-based and model-based predictions become possible. Walk before running: a shop that can't yet explain last week's downtime has no business buying a predictive maintenance platform.

How long until monitoring pays for itself?

Typically within three to nine months, and usually from one unglamorous finding: a recurring stop nobody had quantified, a cycle time that drifted, or the discovery that a planned equipment purchase wasn't needed. A single avoided machine purchase can pay for monitoring across the whole floor.

If you're weighing platforms, retrofits, or a pilot on your toughest machines, we're happy to talk through it with no obligation. Willowark's industrial automation team builds monitoring systems across the messy boundary between old iron and modern software. Reach out and tell us what your floor looks like.