Modern Downtime Tracking: Machine Signals and Vision Beat Manual Logs
Key takeaways: a downtime record has three possible sources, and one stop record joins them.
- The operator's log gives the reason but rounds minutes and misses short stops.
- The machine signal times stops and micro-stops: run/stop state, count pulse and, where the PLC offers one, a fault code.
- A camera add-on counts output at a station with no usable signal.
- Worked example: on one puzzle press shift the supervisor reported 25 minutes, the log 55 and the signal 88; with slow running and scrap, EUR 2,440 of margin was lost while orders waited.
Downtime tracking builds a stop record from up to three sources: the operator's log, the machine signal and, at a station with no usable signal, a camera count. Joined, they give the minutes, the reason and the money of a loss.
Manual downtime tracking depends heavily on operator input. On a busy line, that means short stops are rarely written down, micro-stops are almost never recorded and root causes become guesswork. The result is a distorted picture of performance that hides real improvement opportunities.
Modern plants time stops from machine signals and count output at stations with no signal using computer vision. This guide compares the three sources, shows how to join them into one stop record and reconciles one shift from the supervisor's report to the priced loss.
Why manual downtime logs undercount losses
Manual logs made sense when equipment was simple and data collection options were limited. Today they are a bottleneck that skews OEE and misdirects improvement efforts. Common issues include:
- Missing short stops: Operators focus on major breakdowns. A 2-minute jam or 45-second sensor fault repeated many times in a shift is usually ignored, even though the accumulated loss is significant.
- Rough estimates instead of precise times: Downtime is recorded as round numbers such as 5 or 10 minutes. Over a week, this rounding error can hide hours of real loss or inflate issues that are actually minor.
- Inconsistent cause codes: The same event might be logged as "mechanical," "breakdown," or "line issue" depending on who is on shift. This inconsistency blocks accurate Pareto analysis.
- Logging delayed until the end of shift: When operators are busy, they postpone entries. By the time they update the log, details are forgotten and root causes are simplified or guessed.
- Pressure to keep numbers looking good: When downtime metrics are seen as a personal scorecard, some losses are left unrecorded or moved into "changeover" or "meeting" to avoid scrutiny.
The biggest problem is not dishonesty, it is cognitive overload. Operators are hired to run equipment safely and consistently, not to act as data-logging terminals. Any process that relies on perfect manual entry in a high-intensity environment will fail.
Automatic capture does not take the operator out of the record. A Chalmers study of automatic OEE measurement across "23 different companies and 884 machines" found an average OEE of 65%. "Almost half of the recorded OEE losses could not be classified since the loss categories were either lacking or had poor descriptions", and "90% of the stop time that was classified could be directly related to supporting activities performed by operators". The signal gives the minutes; the reason still comes from people, through a short list they trust. Stop time with no reason is unclassified downtime.
Why operators ignore downtime alerts on busy shifts. The same overload hits alerts. For alarm systems, the UK Health and Safety Executive gives a target for normal operation of "no more than one every ten minutes". An alert with no defined response and no owner gets silenced when the line is busy, so send fewer, only on priced losses, to a named person: see OEE alerts that drive action.
The third source. Where a station has no usable signal, such as a hand packing table, a camera add-on can count its output, so the station gets a count and an OEE figure. The explainer below shows how.
Explainer presented by Iris, Fabrico's AI-generated presenter: how an add-on camera counts output where a machine has no signal. Book a demo.
Plays from YouTube. Google may set cookies when you press play. Privacy policy.
Automatic downtime capture from machine signals
Machine signals solve the first half of the problem. By connecting to PLCs, sensors or existing controllers, a monitoring system reads the signals in the table below. Which states it can tell apart depends on what the PLC exposes: a press that offers only a run bit says running or stopped, not why.
| Signal | What it gives the record | Where it comes from | Watch for |
|---|---|---|---|
| Run/stop state | Start, end and minutes of a stop, micro-stops included | PLC output, relay or retrofit sensor | Idling in run mode with no product |
| Cycle or count pulse | Output and speed; slow running shows as a count below the ideal rate | PLC counter, photo eye or stroke sensor | Rejects counted as good output |
| Fault code | Which alarm fired at the stop | The PLC, where it offers one | The first fault in a cascade, not the last |
| State model | Held or suspended: an internal or an external pause | Machines built to PackML or similar | Only as good as the machine's logic |
Machines built to the PackML state model report more. The OPC Foundation's PackML specification separates pauses by where the cause sits: Suspended when "conditions external to the unit/machine require a pause in production", such as "faults to upstream or downstream equipment", and Held for internal conditions such as "low levels on materials required for production". Blocked and starved reach a downtime record only when the machine reports them; otherwise a fault code names the alarm that fired, a clue to the reason rather than the reason.
This type of machine monitoring system transforms the quality of downtime data:
- Stops are captured: If a connected machine stops, it is recorded. Micro-stops, frequent short jams and low-speed losses appear in the event stream; starved and blocked appear only where the machine reports them.
- Precise timestamps: Start and end times are captured directly from the machine. OEE metrics such as Availability and Performance are built from objective data, not memory.
- Standardized state definitions: States are configured once, then applied consistently across shifts, lines and plants, which makes benchmarking possible.
- Minimal operator burden: Operators pick the reason for a timed stop from a short list instead of starting and stopping timers; see operator downtime reason entry.
With this foundation, operations and continuous improvement managers can trust their production monitoring data. The conversation shifts from "Are these numbers correct?" to "What is the best next action?"
Where machine signals fall short
Machine signals show when the asset is not producing, but they do not always explain why. Typical gaps include:
- Ambiguous states: The PLC may show the machine is stopped, but the true cause could be material, quality, a blocked conveyor or an upstream machine.
- Complex lines: In lines with multiple interconnected machines, one stop can ripple across the line. It is hard to see which machine is the initiator and which are just blocked or starved.
- Human-driven issues: Operator absence, incorrect setup, slow manual packing or visual quality checks are rarely visible in PLC data.
- Manual workstations: Assembly cells or inspection stations without sophisticated controls can remain blind spots if you rely on signals alone.
A camera covers part of that within its frame, and the operator covers the rest with the reason.
Computer vision and stops: what a camera adds and what it cannot see
Computer vision uses cameras and AI models to interpret what is happening on and around the machine. In general, a vision system can identify motion, material flow, accumulation, operator presence, and in some cases visual defects or misalignments. Vision products on the market are sold for these jobs around a stop:
- Detecting line-level blockers and starvation: A camera over the line can show a full upstream conveyor or an empty infeed, so a person can judge whether a stop started at the machine or at another asset.
- Distinguishing mechanical from material issues: If the PLC shows a stop and the camera sees an empty pallet, missing boxes or no operator, a person reviewing the stop can tell a material wait from a machine fault.
- Tracking manual micro-stops: In semi-automatic cells, a camera count shows small hesitations and repeated manual interventions as output below the ideal rate, a performance loss OEE can carry.
- Supporting setup and changeover analysis: Some products cut changeover video into tasks to show where manual steps lose time, which helps SMED and standard work projects.
Vision does not replace machine signals, it adds context. Combined with the signal and the operator's reason, it gives a fuller account of a stop than any manual log.
What a camera cannot see. A 2026 peer-reviewed review of industrial object counting notes that on crowded conveyors, occlusion can make a model "identify multiple overlapping targets as a single one or miss detection completely", and that "some stations have backlight and low-light shooting conditions". Causes inside a machine or outside the frame need the signal, a sensor or a person.
What Fabrico's camera add-on does today. It counts output where a station has no usable signal, and where a camera add-on covers the station, it shows the video for the minute of a stop, streamed on request and not stored in Fabrico. It does not classify stops or find causes; the operator gives the reason. On Fabrico's roadmap, not live today: the camera add-on detecting stops by itself, blocked, starved or material calls from the image, and the camera add-on suggesting a likely reason (see how a camera could suggest a stop reason).
Connecting line cameras to MES for downtime codes. Keep the jobs apart: the signal times the stop, the camera add-on counts where there is no signal, and the code comes from the operator's short reason list. The steps are in connect line cameras to MES for downtime codes.
Manual logs, machine signals and camera counts: one stop record
The four totals of one shift in the worked example below, 25, 55, 88 and 122 lost minutes, come from three sources, and each answers a different question about a stop. For the general case for replacing the log, see manual vs automatic downtime tracking.
| Source | What it records | What it misses | Effort | Who owns it |
|---|---|---|---|---|
| Operator's log | The reason, and stops long enough to write down | Short stops, slow running, exact minutes; reasons drift between shifts | Writing on shift, then reconciling | Operator and shift lead |
| Machine signal | Start, end and minutes of stops and micro-stops; count and speed; fault codes where offered | Why the machine stopped; stations with no signal | One connection per machine | Production, with OT |
| Camera add-on count | Output at a station with no usable signal; with Fabrico, the video for the minute of a stop, streamed on request and not stored in Fabrico | Anything outside its frame or light; the reason | One camera over a count line, proved against a reference count | Production manager |
| Joined stop record | Minutes, count, reason and money in one row | What none of the three sees | A short reason list and a price per constraint minute | CI manager |
How to join them into one stop record:
- The signal owns the clock: start, end and minutes come from the machine; nobody edits them by hand.
- The count owns the output: a PLC counter where one exists, a camera add-on count where none does, a hand tally only as a referee.
- The operator owns the reason: one pick from a short list, at the line, while the stop is fresh.
- The plan owns the price: planned time, the ideal rate and the margin turn minutes into money.
The fields of a stop record. For a Pareto by reason and a ranking by money: line and station, start and end, minutes, machine state, fault code where the PLC offers one, reason code, shift, SKU, count and the price of a constraint minute.
Why downtime logs do not match what supervisors report
When the signal goes live, three numbers disagree: the supervisor's report, the operator's log and the machine. Nobody is lying. Five things pull them apart:
- Definitions. One person books a die change or a cleaning as planned time, another as downtime.
- Estimates. People write 10 or 15 minutes; the machine counts 14 or 21.
- Thresholds. A paper log asks only for stops above a threshold, so 40-second misfeeds never appear.
- Speed and scrap. Slow running and scrapped output are OEE losses too, and only a count shows them.
- Timing. The supervisor writes the report after the shift, from the log and from memory.
A reconciliation walks from the smallest number to the largest and names the gap at every step, as the example below does for one shift.
Worked example: one shift reconciled, from manual logs to machine signals
Example inputs, not customer figures; replace them with the plant's.
The line. A jigsaw puzzle plant runs a laminator, a die-cutting press, a breaking and bagging machine and a hand boxing table. The press is the constraint at an ideal 10 sheets a minute; one sheet is one puzzle, so a lost press minute is 10 puzzles. Its PLC gives a run/stop state, a stroke count and a fault code when the double-sheet detector trips. The boxing table has no signal, so a camera add-on counts the boxes.
One early shift in the autumn peak. 480 minutes less a 30-minute break and a 5-minute handover leave 445 planned minutes. The signal timed 88 minutes in 39 stops: one die change of 34 minutes, 3 waits for laminated board totalling 21, 9 stripping jams totalling 14, and 26 sheet feeder misfeeds totalling 19, none over 2 minutes, all with the double-sheet fault code. In 357 run minutes the press counted 3,310 sheets against 3,570 at the ideal rate: 26 minutes of slow running. The camera add-on counted 3,230 boxed puzzles, so 80 sheets were pulled as mis-cut: 8 minutes.
OEE. Availability 357 / 445 = 80.2%, performance 3,310 / 3,570 = 92.7%, quality 3,230 / 3,310 = 97.6%: OEE 72.6%, the same as 323 good minutes / 445 planned. The minutes add up: 323 good + 88 stopped + 26 slow + 8 scrap = 445.
Three reports of the same shift. The operator's paper log (stops of 2 minutes or more) lists die change 30, waiting for board 15 and jams 10: 55 minutes. The supervisor's report says 25, with the die change booked as planned. The signal says 88.
| Step | Minutes | Running total | What explains the gap |
|---|---|---|---|
| Supervisor's shift report | 25 | 25 | Unplanned stops only; the die change booked as planned |
| Die change, as logged | +30 | 55 | The operator's log counts it; the supervisor's report does not |
| Estimates on the three logged stops | +14 | 69 | Die change 34 not 30, board waits 21 not 15, jams 14 not 10 |
| 26 sheet feeder misfeeds | +19 | 88 | Under the log's 2-minute threshold; the signal timed them |
| Slow running | +26 | 114 | Stroke count below the ideal rate; in no downtime log |
| Mis-cut sheets | +8 | 122 | Press count minus the camera add-on box count |
The money. Example price EUR 8.00 a puzzle ex works at an example 25% margin: EUR 2.00 a puzzle, EUR 20 a press minute. NYU Stern's margins by sector ("Data used is as of January 2026") give Recreation, 49 firms, a gross margin of 39.79% and a pre-tax unadjusted operating margin of 9.69%; 25% sits between. In the autumn peak, when orders wait, the 122 lost minutes are 1,220 puzzles: EUR 2,440 a shift, EUR 102,480 a month over 42 shifts. The log shows EUR 1,100 of it (45%), the supervisor's report EUR 500 (20%). When orders do not wait, the minutes are capacity, priced at the overtime or extra shift they would save.
| Loss on the shift (example) | Minutes, signal and count | Minutes in the log | EUR a shift, orders waiting |
|---|---|---|---|
| Die change | 34 | 30 | 680 |
| Slow running | 26 | No | 520 |
| Waiting for laminated board | 21 | 15 | 420 |
| Sheet feeder misfeeds, 26 stops | 19 | No | 380 |
| Stripping jams, 9 stops | 14 | 10 | 280 |
| Mis-cut sheets, 80 | 8 | No | 160 |
| Total | 122 | 55 | 2,440 |
What the reconciliation changes. The log ranks the die change first, then waiting for board, then jams. Priced from the signal and the counts, slow running (EUR 520) is second and the misfeeds (EUR 380) are fourth, and neither appears in the log at all. With the estimate gaps and the mis-cut sheets, EUR 1,340 a shift sits outside the log, EUR 56,280 a month in the peak.
The business case for a signal over a manual log. The paperwork is the small part: the log and the reconciliation take an example 35 minutes of a shift lead's time, and Eurostat puts 2025 "average hourly labour costs in the whole economy" at EUR 38.2 in the euro area, so EUR 22.28 a shift. The large part is the loss the log cannot see. If a feeder fix removes 10 of the 19 misfeed minutes (an example target), that is EUR 200 a shift, EUR 8,400 a month while orders wait. Set that against the quoted cost of the connection, the camera add-on and the software.
Production loss tracking: from stop minutes to money and actions
Capturing stops is only valuable if teams act on the information. Production loss tracking puts minutes and money on the losses: how long, how much output, and what that cost while orders waited. In Fabrico, the financial impact module prices each loss from SKU price and margin and ranks stops by money.
Typical workflows include:
- Production actions: When recurring short stops are linked to changeover, cleaning or material handling, production teams can standardize work, adjust staffing or re-balance tasks directly from a prioritized list of losses.
- Maintenance actions: When a stop needs a technician, a maintenance work order opens from the logged stop in one tap, and chosen stop reasons can create it automatically when the reason is logged. Maintenance managers can plan inspections or replacements based on real failure patterns.
- Shared view for operations and maintenance: Both teams work from the same timeline of stops, causes and actions. There is less debate about "Is this downtime really maintenance or operations" because the data is granular and transparent.
- Structured improvement cycles: Pareto charts of downtime and speed loss, built from the signal, the camera add-on count and the operators' reasons, point CI teams to the next most valuable improvement project.
The result is a production environment where OEE improvements are built on evidence, not on partial or biased logs.
Where AI fits into modern downtime tracking
AI adds another layer of intelligence on top of raw signals and images. In the market, AI is sold to classify events, suggest causes and find patterns that are hard to see in spreadsheets. Each claim deserves the same two questions: what data does it read, and who confirms the answer?
Examples of AI-driven value in AI production monitoring include:
- Automatic cause suggestions: Combining PLC fault codes, machine state patterns and vision input, AI can propose a likely cause for a person to confirm. On Fabrico's roadmap, not live today: the camera add-on suggesting a likely reason. Today the operator picks the reason from a short list.
- Pattern detection across shifts and products: AI can highlight that minor stops spike on specific SKUs or during certain shift combinations, pointing to training or setup issues.
- Prioritization by impact: Rather than a long list of small problems, a ranking shows which recurring stops and speed losses cost the most on key products. In Fabrico that ranking is the financial impact module, which prices losses from SKU price and margin: arithmetic, not AI.
AI does not replace engineers. It can shorten the search for an answer. Fabrico's AI Assistant answers questions from manuals, maintenance history and recorded stops with their causes, so a technician can ask how a similar stop was fixed before.
Which source to add first, and who decides
Start from the reconciliation, not the technology: the largest priced gap points to the source to add.
| If the reconciliation shows | Add first | Effort | Who picks |
|---|---|---|---|
| Stop minutes far above the log at the constraint | The machine signal: run/stop state and count | One connection per machine | Production manager with OT |
| A station with no usable signal that holds the line back | A camera add-on count, proved against a reference count | One camera over a count line, light by day and night | Production manager |
| Stop time with no reason or a vague one | A shorter reason list, picked at the line | A workshop with operators | CI manager |
| Reasons that hinge on what happened out of sight | The video for the minute of a stop, streamed on request and not stored in Fabrico, where a camera add-on covers the station | A camera with the stop zone in frame | CI manager |
| Faults inside the machine | PLC fault codes or a sensor | A mapping, or a sensor per machine | Maintenance lead |
Three manager roles.
- The plant manager rules that losses are priced at the constraint margin while orders wait, and picks the first line.
- The production manager owns the stop record day to day: the signal on the constraint, camera add-on counts where there is no signal, the short reason list.
- The CI manager runs the reconciliation, ranks losses by money and gives the top one a project and an owner.
OT or IT owns the connection; the maintenance lead takes stops that need a technician, through the work order.
Want to see the three sources joined on one OEE timeline before choosing what to add? Book a demo.
Key evaluation questions for MES and OEE buyers
When you evaluate MES and OEE platforms, especially those that claim to track downtime automatically, it is useful to ask:
- How do you connect to machines and what events can you capture from PLCs or controllers?
- How do you handle semi-automatic or manual workstations where signals are limited?
- If a camera is offered, what does it do today, what is still roadmap, and where does its video live?
- How is a camera count proved against a reference count, by day and by night?
- Can a logged stop open a maintenance work order, and which stop reasons create one automatically?
- How does an operator give the reason for a stop, and how short is the reason list?
- How do you present downtime and OEE metrics to different roles: operators, team leaders, CI managers and maintenance?
Where Fabrico fits. Fabrico is OEE software and light MES in one cloud platform: the machine signal times stops and micro-stops, the operator picks the reason from a short list, the camera add-on counts output where a station has no usable signal, OEE is calculated live, and the financial impact module ranks stops by money. Maintenance is built in: a work order opens from a logged stop in one tap.
A quarter checklist for moving beyond manual logs
For mid-size and large manufacturers, the limitations of manual downtime logs have become a strategic issue. As product mixes grow and supply chains tighten, unmeasured losses translate directly into missed orders, overtime and capital spending that may not be necessary.
The steps below are for a plant, production or CI manager, on one line first; the plant sets its own pace.
- Pick the constraint line, where a lost minute costs the most.
- List the signals the machine offers (run/stop, count, fault codes, a state model) before buying sensors.
- Mark the stations with no usable signal and prove any camera add-on count against a reference count.
- Cut the reason list to what operators will use at the line, and agree which stops count as planned.
- Reconcile one shift as in the example, then price the top loss and give it an owner.
- Prune the alerts to those with a defined response and an owner, and reconcile again at quarter end.
At the end of the quarter, the plant knows where its capacity goes, in minutes and in money. Manual logs can still play a role as a backup or for special cases, but they no longer define the data.
Start with one line and one reconciled shift.
Frequently asked questions
What machine signals does a downtime record use?
A run/stop state for the start, end and minutes of a stop, a cycle or count pulse for output and speed, and a fault code where the PLC offers one. PackML machines can also report held and suspended states.
Why do downtime events not match what supervisors report?
Definitions, estimates, thresholds and timing: a changeover booked as planned, rounded minutes, short stops below the log's threshold, and slow running and scrap that no downtime log lists. Reconcile one shift step by step.
How do you connect line cameras to MES for accurate downtime codes?
Keep the jobs apart: the machine signal times the stop, the camera add-on counts output where there is no signal, and the code comes from the operator's short reason list. On Fabrico's roadmap, not live today: a camera add-on that suggests the code.
Why do operators ignore downtime alerts during busy shifts?
Too many alerts with no clear response. The UK Health and Safety Executive's alarm target for normal operation is no more than one every ten minutes. Send fewer, only on priced losses, to a named owner.
What is the business case for machine signals instead of manual OEE logs?
The losses the log cannot see. In the example shift the log showed EUR 1,100 of a EUR 2,440 loss; EUR 1,340 a shift sat outside it while orders waited, against EUR 22.28 of paperwork.
Want to see one shift's manual log, machine signals and camera add-on counts reconciled in minutes and money? Book a demo.
Related articles

Real Fabrico tiles, sample data.










