Edge recording camera for intermittent fault capture

A fault that shows up once every couple of weeks and leaves only an aftermath is a recording problem, not a sensing problem. Nobody is standing there, so the machine has to keep the evidence: continuous video from one camera, weeks of retention on a local drive, a clock you can trust, and a way to page back through it from your desk. This guide gives you the numbers that size such a set-up: the rule that turns any bitrate into days of retention, the pixel budget that decides what the camera can resolve, the NTP discipline that lines footage up with the fault log, and camera and recorder parts that suit a refrigerated room. The running example: a line that faults about every 14 days, watched at 1080p and 15 fps.

Can one camera catch a fault that happens every couple of weeks?

Yes — record continuously at the edge and prune by age: at 1080p, 15 fps and a 4 Mbps H.264 stream, a 1 TB drive holds about 23 days of video, which covers a 14-day fault interval with a 1.6 times margin. Continuous, not triggered: the PLC trigger fires when the fault bit latches, but the cause — a drifting guide, a slow leak, a part feeding crooked — starts seconds earlier, and continuous capture keeps that preamble.

If bandwidth ever worries you, hold a rolling buffer and dump the minutes before and after a trigger, but at the rates below there is rarely a reason to bother.

You also do not need a dedicated surveillance DVR for one camera. A vision controller already on the machine network records, prunes and serves the clips on the same clock as the rest of your instrumentation. A DVR earns its keep when several cameras and an existing video management system enter the picture.

How much storage does continuous recording need?

Every 1 Mbps of video stream consumes 10.8 GB of drive per day, so retention in days is drive capacity divided by bitrate times 10.8 — a calculation you can run before anyone installs anything. The constant is plain arithmetic: 1 Mbps is 1,000,000 bits per second, or 125,000 bytes per second, and 125,000 × 86,400 seconds = 10.8 GB per day.

What bitrate to plan? H.264 encoders typically land a slow industrial scene at 1080p and 15 fps between 2 and 4 Mbps, with busy motion pushing 6-8. Plan at 4 Mbps, then measure your encoder's real output over a day and redo the sums. H.265 typically saves about a third at the same picture quality.

Worked example: 4 Mbps × 10.8 = 43.2 GB/day. A 1 TB drive is 1000 GB, and 1000 ÷ 43.2 = 23.1 days. Keep about 10% for the operating system, database and file system: roughly 21 days of ring buffer, a 1.5 times margin over the 14-day gap. If the gap ever stretches to a month, the same drive at 2 Mbps holds 1000 ÷ 21.6 = 46.3 days.

Raw capacity, before the roughly 10% the operating system and file system take. Set the prune age at 1.5 times the longest gap you have actually observed, and export any clip you care about the morning after you first see it.

What resolution and frame rate show the fault clearly?

A 1.6 to 2 MP camera is enough for most fault witnessing: budget about 3 pixels across the smallest detail you must recognise, and shoot 15 fps, which resolves the timeline in 67 ms steps. Say the field is 600 mm wide and the smallest thing that matters is a 2 mm pin working loose. Three pixels across it is 1.5 px per mm, so the field needs 600 × 1.5 = 900 px across — a 1920 px wide stream covers that with more than twice the margin. If the detail were 0.5 mm, the same field would need 600 × 6 = 3600 px, which is 12 MP territory and a sign to narrow the field instead.

Frame rate is a stopwatch question. 1 ÷ 15 fps = 66.7 ms per frame, so anything faster than about 70 ms can fall between frames. A jam, a mis-feed or a door left open develops over seconds and 15 fps is plenty; a flap, a spark or a bounce needs 30 fps (33 ms steps) or a trigger-saved buffer.

Which camera do I hang on the machine?

Use a compact vision sensor with a fitted lens and locking connectors: the MV-SCB007EC-12S-WBN is a 1.6 MP colour vision sensor with a factory-installed 12 mm lens, manual focus, global shutter, M12 mount and Fast Ethernet. Set the focus once at commissioning and leave it; nothing in this job changes week to week.

Both picks below come from the smart cameras category — inspection-first devices that also hand their stream to a recorder.

Placement decides the lens before the catalogue does. The estimate is sensor width × working distance ÷ focal length. From the only free spot 3 m back, a 12 mm lens on a 1/2-inch-format sensor (6.4 mm wide) frames 6.4 × 3000 ÷ 12 = 1600 mm, about 1.6 m across: one machine bay. To cover 2.5 m from the same spot you need a wider lens or more standoff — 6.4 × 5000 ÷ 12 = about 2.7 m at 5 m.

Colour earns its place when the fault involves a printed mark, a product change or an indicator lamp; mono is fine for pure geometry. Global shutter avoids slanted edges if anything in view moves fast — at 15 fps of slow machinery it is a nice-to-have rather than a gate.

For a whole-cell view instead of one bay, the MV-SC5060XM-08M-WBN is a 6 MP mono smart camera with an 8 mm focal length on GigE. Its rolling shutter costs nothing at 15 fps on slow machinery, and the extra pixels widen your pixel-budget margin at the same field.

What records the stream and prunes it?

A vision controller next to the panel does the whole job: it pulls the stream continuously — the same pattern as continuous camera acquisition mode — writes segmented files, deletes anything past the prune age and serves the rest back over the network. Its drive size, chosen with the 10.8 GB/day rule, sets your retention.

Vision controllers publish their storage in the model name. The MV-VC2140-128G40-1T runs an Intel N-series N97 with 8 GB of RAM, boots from a 128 GB SSD and records to a 1 TB HDD across its 3-4 LAN ports. At the 43.2 GB/day planning rate, that 1 TB drive holds 23 days between prunes — your 21-day target with margin.

Need power and video on one cable to the camera? The MV-VC5525-256G70-NN adds PoE camera ports on a Core i5 with 32 GB of RAM and a 256 GB SSD. For the longest window, the MV-VC4712-256G10-2TH pairs a 256 GB SSD with a 2 TB HDD and keeps PoE camera ports and PCIe expansion, shipping without a vision software licence key so it stays a pure recorder. At 43.2 GB/day a 2 TB drive holds 46.3 days, so even a month between events stays covered.

What runs on the box is ordinary: a capture process that pulls the camera's stream — acquiring an RTSP stream in your vision program describes the pull, and keeping the RTSP buffer empty keeps it from lagging — writing 5-10 minute segments, pruning by file age and keeping an index. Put the ring buffer on the HDD and leave the SSD for the system and exported clips.

Pick the drive for the retention window you computed; the cards line the three up.

How do I keep the timestamp honest?

Point the camera and the controller at the same NTP server and leave them to it: on a switched plant LAN, NTP typically holds devices within tens of milliseconds, tight enough to line footage up with PLC log entries frame by frame. The vision sensor NTP timestamp fix walks through the settings that get a device there.

The fortnight cadence is exactly why this matters: clocks that drift apart by the time you need them turn your evidence into two unverifiable stories. So make the recorder write timestamps into segment file names and into a one-line log for every segment it opens. When the PLC latches the fault, have it send the event with its own timestamp to the recorder; the review list then shows tagged events, and the export is one or two segments either side.

How do I review and export footage from my desk?

Keep the recorder on the machine's existing Ethernet, browse to it and page through segments in the browser; for a camera in another building, a VPN route reaches it without exposing anything to the internet. Three pieces make the review fast. Segmented files of 5-10 minutes mean exporting the fault is copying one or two files, not a 40 GB day file. The tagged event list lands you on the fault, not near it. And a live view matters during commissioning week — you will aim and focus once, then never again.

For the live picture in a browser tab, the machine vision camera livestream workflow and restreaming a GigE Vision camera over RTSP cover the two standard routes, and the VPN remote subnet route sorts out reaching a device across sites.

Export before the prune age passes; the ring buffer will not wait for your Friday review. At the 4 Mbps planning rate a 5-minute segment is 4 Mbps × 300 s ÷ 8 = 150 MB, so one exported fault is a few hundred megabytes — keep those next to the fault report.

Will the set-up cope with a refrigerated room?

Yes — 5 °C is comfortable territory for industrial control hardware, which is normally rated down to 0 °C or below; the engineering risk in a cold room is condensation, not cold. Condensation follows the dew point: air from the 20 °C side of the door carries moisture that condenses on any surface cooled near 5 °C. Let equipment acclimate to the room powered off for an hour or two before energising, keep cabinet doors closed so warm air does not bathe cold surfaces, and favour filtered air if the cabinet breathes.

If the recorder sits in a cabinet at the machine rather than inside the cold store, check what its enclosure rating actually claims. Do not read an IP code as a NEMA type — immersion and jet protection are different claims — and NEMA 4X versus IP66 ratings sets out how the two scales compare.

What do I check before ordering?

Six checks, fifteen minutes at the machine, and the order writes itself: measure the bay, count the pixels, do the drive arithmetic, name the NTP server, plan the network path, decide the mount.

• Field of view. Stand where the camera will mount and measure to the extremes of what you must see; the estimate above turns that distance into a lens choice.
• Pixel budget. Field width in mm ÷ (smallest detail in mm ÷ 3) gives the pixels you need across; over 1920 means narrow the field or step the sensor up.
• Retention. Prune age = 1.5 × longest observed gap; drive size = prune age × 10.8 × planned Mbps, plus 10% overhead.
• Clock. Name the NTP server now — panel PC or plant server — and put both devices on it at commissioning.
• Network. A spare port on the machine switch, PoE if you want one cable to the camera, a VPN path if the site is remote.
• Mount and environment. Rigid fixing, no vibration in frame, and the condensation routine above if the room cycles temperature.

If you send us the bay width, the mounting distance and the prune target, we will spec the camera and recorder against them, or start with the camera finder and the configurator below; anything that does not fit a form goes through the contact page.