Skip to content
GREEN FUTURETECH.
Explore
Keep local inspection operatingIllustrative default: 8 cameras, 2 edge nodes, 4 hours of combined outage, 720 Wh nominal battery.INTERFACEOperator decisions stay localANALYTICSDeferred central reportingCOREStore and forwardVISIONInspection stays on the LANEDGEProtected power and storagePROTECTED LANVARIABLE UPLINK

GFT instruments / Edge to dashboard

Keep the decision local.
Let the signal travel later.

Keep the factory decision local. Compare streaming and event traffic, expose a recurring queue deficit, and size an illustrative local buffer and battery.

Configure the failure boundary
Operate the instrumentProcedural visualisation. Schematic, not a facility.

An instrument for the decision

What must run at the edge when power and links fail?

Change the inputs. Read the result. Keep the assumptions visible.

Sizing study / all defaults illustrativeOne contiguous daily outage of mains and uplink. Not a measured GFT installation.

Define the failure envelope.

Each slider has an exact numeric partner. These are planning assumptions, not programme defaults.

Set payload, power and battery assumptions

LOCAL CONTINUITY / WAN RECOVERY

2 edge nodes · 4 h outage · daily queue clears

Local edge to interface stackIllustrative five-layer system. 2 edge nodes keep inspection and local decisions operating; the upstream link is unavailable for 4 hours each day.INTERFACEOperator decisions stay localANALYTICSDeferred central reportingCOREStore and forwardVISIONInspection stays on the LANEDGEProtected power and storagePROTECTED LANVARIABLE UPLINK
Stream / online linkExceeds link
Events / online link0.38%

Bars cap at link capacity. Text retains the true result. Packet motion is schematic, not a measured traffic rate.

Streaming demand38.40Mbps · derived / illustrative inputs
Edge event demand0.0384Mbps · derived / illustrative inputs
Event buffer, outage only0.069GB · derived / illustrative inputs
Video buffer, outage only69.12GB · derived / illustrative inputs
Nominal battery energy720Wh · derived / illustrative inputs
Protected local load108W · derived / illustrative inputs

The architecture decision

Keep the inspection loop inside the factory.

Keep acquisition, inference, local event storage and the operator interface on the protected local network. The event queue can recover in the connected window under these assumptions. Central analytics and model updates can wait.

Local buffer time
4.00 h
Backlog drain after reconnection
0.02 h
Single-node model transfer, exclusive link
2.67 min
Event demand versus streaming
99.9% lower
Daily event generation / link budget
0.415 / 90.00 GB
Factory to edge topologyCameras feed an edge node on the protected local network. The node holds events in a local buffer and forwards them over a variable uplink to central reporting. Illustrative; it follows the instrument's inputs.FACTORY LANEDGELOCAL BUFFERCENTRAL8 camerasNODE2 edge nodesdecides locally4.0 hheld on site0.069 GB queuedHQ10.00 Mbps uplinklink cut for the outageQueue clears
Illustrative topology. Camera slots dim as the camera count falls. The tank is the outage as a share of the day.

The model stays visible.

Encoded video: cameras × frames per second × kB per frame × 8 ÷ 1,000. Events: cameras × frames per second × event fraction × event kB × 8 ÷ 1,000. Storage and battery use the entered contiguous outage; the daily queue must clear in the remaining connected window.

Method source

GFT dimensional arithmetic; Microsoft Learn on offline message storage. All sizing defaults are illustrative. ↗

A planning aid. Not investment, legal or tax advice. Verify inputs and boundaries before an operational decision.

Compare / two scenarios, one verdict each

Change the failure boundary and the verdict flips.

Presets are illustrative test scenarios, not GFT programme measurements. Save the instrument's current inputs as A to compare them.

A Scenario A
B Scenario B
50% from A to B

Hover or focus a row to read A against B.

A: BaselineB: Weak link
38.40
Streaming demand Mbps+100%
76.80
0.0384
Event demand Mbps+100%
0.0768
0.069
Event buffer, outage only GB+300%
0.276
0.415
Events generated per day GB+100%
0.829
90.00
Link budget per day GB-100%
0.36
0.02
Backlog drain hn/a
No recovery
720
Nominal battery Wh+300%
2,880
Queue clears
Daily queue verdict2 to 4 edge nodes (+100%)
Queue does not clear
Outage
6.0 h
Event demand
0.0576 Mbps
Event buffer
0.156 GB
Battery
1,620 Wh
Uplink available
Queue clears

Blended inputs between A and B. Whole-number counts are rounded. Illustrative.

The calculation boundary

A daily balance decides whether the queue survives.

Five steps turn the inputs into a verdict. Scroll them and the picture changes with each. Every figure follows the instrument above and stays illustrative.

Daily balance of the edge queueFive states follow the five calculation steps: traffic, outage buffer, daily balance, backlog drain and battery energy. Values follow the scenario in the instrument and are illustrative.TRAFFIC / MbpsStreaming, V38.40 MbpsEvent metadata, E0.0384 MbpsUsable uplink, L10.00 MbpsEvents are 99.9% lower than streamingOUTAGE BUFFER / ONE DAY STARTS AT OUTAGEHour 00 / 0.000 GB0h6h12h18h24hEvent buffer at reconnection0.069 GBVideo, if kept locally69.12 GBDAILY BALANCE / GB PER DAY0.415 GB90.00 GBEvents arriveLink servesService exceeds arrivals: the queue clearsSpare link budget 89.59 GB per dayBACKLOG OVER ONE DAY / SHAPE NORMALISED TO PEAKHour 00 / 0.000 GB0h6h12h18h24hClears at hour 4.02, 0.02 h after reconnectionBATTERY / NOMINAL ENERGY108 W protected load60 W nodes48 W cameras432 Wh need86 Wh reserve202 Wh derating720 Wh nominal battery

  1. Step 1 of 5 / Traffic

    Send the verdict, not the picture.

    V = C × F × S × 8 / 1,000 [Mbps]E = C × F × (p / 100) × e × 8 / 1,000 [Mbps]

    Only the chosen share of processed frames produces an event, so the link carries far less than the cameras produce.

    Notes and limits

    The arithmetic uses decimal bytes. Frame payload means the average encoded bytes allocated to each frame, not an uncompressed image. Measure it over representative motion and lighting. A frame rate alone cannot determine camera bandwidth.

    The event branch sends a metadata payload only for the chosen proportion of processed frames. It does not include thumbnails, clips or continuous remote video. Add those explicitly to the payload assumption if the design requires them.

    C is the active camera count, F the frame rate, S the encoded kB per frame, p the event percentage, e the event kB, H the outage duration and L usable upstream Mbps.

  2. Step 2 of 5 / Buffer

    The outage buffer is rate times hours.

    Buffer = rate × H × 3,600 / 8,000 [GB]

    Storage sizes to the outage you enter. It is not sized to the day, and it carries no disk reserve.

    Notes and limits

    The video buffer is a separate local recording option, not an additional event transmission requirement.

    Storage also needs filesystem overhead, retention policy, write endurance, corruption recovery and headroom. The displayed buffer has none of these.

  3. Step 3 of 5 / Daily balance

    Arrivals against service: who wins the day.

    Daily arrival = E × 24 × 0.45 [GB]Daily service = L × (24 − H) × 0.45 [GB]

    Service must strictly exceed arrivals. Otherwise extra storage only delays the overflow.

    Notes and limits

    The factor 0.45 converts a Mbps-hour into decimal GB. The queue passes only when daily service strictly exceeds arrivals, except for an empty workload.

    A failed daily balance has no sustainable repeating buffer size. The displayed event buffer then covers the first outage only.

  4. Step 4 of 5 / Drain

    A marginal link can carry live traffic and still never drain yesterday.

    Drain = E × H / (L − E) [h], when L > E

    Recovery deducts the events still arriving from the link that is trying to clear the backlog.

    Notes and limits

    The drain calculation deducts ongoing event traffic from the recovering link. A marginal link can therefore carry live traffic while failing to clear yesterday's backlog.

    Model transfer time is model MB multiplied by eight, divided by usable Mbps and then by seconds per minute. It assumes one node receives a full model over an otherwise idle link. Concurrent traffic, multiple node downloads, protocol overhead and failed retries lengthen deployment.

  5. Step 5 of 5 / Battery

    Battery energy follows load, hours and derating.

    N = ceil(C / cameras per node)P = N × node W + C × allocated camera WBattery Wh = P × H × (1 + reserve / 100) / (η × D)

    Reserve and derating are added on top of the energy the load needs during the outage.

    Notes and limits

    Battery efficiency η and allowed discharge D are fractions. The energy reconciliation is nominal battery energy multiplied by efficiency and discharge fraction equals load watts multiplied by outage hours and reserve factor. These equations derive capacity from the entered assumptions; they do not assert vendor performance.

    Model size does not establish inference throughput, memory residency or electrical load. Benchmark the complete camera allocation on the intended node, and measure its input power before accepting the battery result.

Interlude / one day on the clock

The link can fray. The decision path holds.

The stack on the clockFive-layer stack. While the uplink is down its threads fray and events queue on the local network; local layers keep running. Illustrative.INTERFACEOperator decisions stay localANALYTICSDeferred central reportingCOREStore and forwardVISIONInspection stays on the LANEDGEProtected power and storagePROTECTED LANVARIABLE UPLINK

Queue depth0.069 GB

Hour 04 of 24

Uplink Down

Link down. Events queue on the local network. Operator decisions continue.

Outage Replay Clear

Driven by the scenario in the instrument. Illustrative: a repeating day that starts at the outage. Not a measured traffic rate.

Assumptions / every default is replaceable

Every default is a replaceable assumption.

No sourced edge sizing values are supplied for GFT's Cameroon textile programme, so there is no programme-default preset. The baseline is an illustrative stress-test envelope. Limits bound this screening model; they are not equipment ratings.

Power6 inputs
Nominal battery energy as the outage lengthens0 h24 h
Nominal battery energy as the outage lengthens720 Wh at 4 h
  • Combined power and WAN outageIllustrative. One contiguous interruption in a repeating day.
    Default 4 h/day
  • Measured load per edge nodeIllustrative. Example protected load; replace with measured power.
    Default 30 W/node
  • Camera load, including allocated local networkIllustrative. Example protected load; replace with measured power.
    Default 6 W/camera
  • Battery-to-load efficiencyIllustrative. Example energy derating; replace with qualified battery data.
    Default 90 %
  • Allowed depth of dischargeIllustrative. Example energy derating; replace with qualified battery data.
    Default 80 %
  • Energy reserve above calculated needIllustrative. Example energy derating; replace with qualified battery data.
    Default 20 %
Network4 inputs
Spare link budget per day as the uplink grows0 Mbps40 Mbps
Spare link budget per day as the uplink grows89.6 GB per day spare at 10 Mbps
  • Usable upstream when connectedIllustrative. Available application throughput after other traffic.
    Default 10 Mbps
  • Model download per nodeIllustrative. Example full-model deployment artefact, not memory use.
    Default 200 MB
  • Frames producing an upstream eventIllustrative. Example payload budget; replace with captured traffic.
    Default 2 %
  • Payload per eventIllustrative. Example payload budget; replace with captured traffic.
    Default 2 kB/event
Capture and storage4 inputs
Video buffer, if kept locally, as cameras are added1 camera64 cameras
Video buffer, if kept locally, as cameras are added69.12 GB buffer at 8 cameras
  • Active camerasIllustrative. Example inspection allocation; benchmark before use.
    Default 8 cameras
  • Frames processed per cameraIllustrative. Example inspection allocation; benchmark before use.
    Default 15 frames/s
  • Camera allocation per nodeIllustrative. Example inspection allocation; benchmark before use.
    Default 4 cameras/node
  • Mean encoded frame payloadIllustrative. Example payload budget; replace with captured traffic.
    Default 40 kB/frame

Limits / what the number does not decide

Capacity is necessary. Recovery must be demonstrated.

Layers under the selected limitFive-layer stack. It changes with the selected limit and with the simulation beside it. Illustrative design rule.INTERFACEOperator decisions stay localRUNNINGANALYTICSDeferred central reportingSTALECOREStore and forwardQUEUEINGVISIONInspection stays on the LANRUNNINGEDGEProtected power and storageRUNNINGPROTECTED LANVARIABLE UPLINK
Clean failure: the uplink drops, mains holds.

The model assumes a clean failure.

Continuous acquisition, constant traffic, constant power and a fully charged battery at the start. Real failures may hit only power, only the WAN, or different local circuits.

More on this limit

The combined outage envelope is deliberately explicit. Split it into separate measured cases before procurement. Cameras and allocated local network power are included; central servers, displays, cooling and site-wide loads are excluded unless you allocate them into the measured load inputs.

Simulation / what fails?

  • InterfaceRUNNING
  • AnalyticsSTALE
  • CoreQUEUEING
  • VisionRUNNING
  • EdgeRUNNING

4 of 5 layers keep running. Clean: the uplink drops, mains holds.

Dirty: mains and uplink both fail with no protected supply. Qualitative design rule, illustrative.

Acceptance test / recorded in this tab only

Tick what a witnessed outage test has shown.

0 of 8 shown

A tracker, not a certificate. It stores nothing and proves nothing; the test itself must be witnessed with the actual equipment. Each layer lights when every check mapped to it has been shown.

Layers shown by the outage testFive-layer stack. A layer lights when its checks have been shown in a witnessed outage test. Nothing is shown until you tick it.INTERFACEOperator decisions stay localNOT SHOWNANALYTICSDeferred central reportingNOT SHOWNCOREStore and forwardNOT SHOWNVISIONInspection stays on the LANNOT SHOWNEDGEProtected power and storageNOT SHOWNPROTECTED LANVARIABLE UPLINK

Sources / the receipt stays beside the design

Keep the provenance beside the design.

Microsoft Learn explains that offline message retention depends on configured retention time and available disk space. Source opened on 30 October 2026; the exact excerpt and source date are in the local receipt.

Microsoft Learn: extended offline capabilities ↗ Download the source receipt

Layers the source coversFive-layer stack. The layer a ledger row concerns is lit. Supported rows light green; rows the source does not support light amber. Illustrative mapping.INTERFACEOperator decisions stay localNOT CITEDANALYTICSDeferred central reportingNOT CITEDCOREStore and forwardSUPPORTEDVISIONInspection stays on the LANNOT CITEDEDGEProtected power and storageSUPPORTEDPROTECTED LANVARIABLE UPLINK
The lit layer is the one the row concerns. Illustrative mapping, not a vendor claim.
  • SupportsMessages wait locally
  • SupportsThe wait has a limit
Supports

Hover or focus a row to read what the source does and does not cover.

Continue the architecture review.

  1. Library subjectsOpen the library →
  2. Production timeOpen OEE line analysis →
  3. More instrumentsOpen the instrument directory →

Related library subjects: wp07, Factories on a Weak Grid; wp14, The Industrial Data Spine; and wp18, Edge to Dashboard. These titles come from the GFT paper register. Publication and review status must be checked in the library. A title is not evidence of a released paper.

Closing instrument / choose the failure boundary

Protect the whole local decision path.

An edge node alone cannot preserve inspection. Cameras, local networking, storage and the operator interface must remain available together. The calculated capacity is a starting point for a witnessed outage test.

What fails?

4 of 5 layers keep running

The decision stays local on battery for 4.0 h. Events queue; reporting goes stale.

Outage
4.0 h
Battery
720 Wh
Event queue
0.069 GB

A queue is a promise to deliver later.

A buffer is not resilience if it fills faster than it drains. Confirm replay, duplicate handling, clock recovery and safe restart with the actual equipment before accepting a design.

Read the measurement approachTest the production-time boundaryBack to the controls ↑
Failure boundary stackFive-layer stack. Each layer shows whether it keeps running under the chosen failure boundary. Illustrative design rule.INTERFACEOperator decisions stay localON BATTERYANALYTICSDeferred central reportingSTALECOREStore and forwardQUEUEINGVISIONInspection stays on the LANON BATTERYEDGEProtected power and storageON BATTERYPROTECTED LANVARIABLE UPLINK

Mains and uplink both down

Find a thread.

Type to search.