Skip to content
ConsystenceConsystence Docs

Edge ML (Planned)

Consystence edge nodes today run a .NET 10 single-silo Orleans runtime focused on deterministic PLC communication, store-and-forward buffering, and the IT/OT firewall — they do not run machine-learning inference. The longer-term plan is to add a constrained, advisory-only inference capability at the edge so that anomaly and condition hints can be surfaced even when a site is disconnected from its cell.

Design intent

Section titled “Design intent”

The planned model is deliberately narrow:

  • Cloud-trained, edge-served. Models are trained centrally (in the cloud, where the data and compute live) and exported to ONNX. The edge never trains; it only runs inference against a frozen exported model.
  • CPU baseline, GPU where the workload needs it. Baseline inference runs on the edge CPU so it works on the standard PLCnext VL3 hardware — GPU is never assumed. The Nvidia Orin variant is the design’s home for GPU-compiled inference (TensorRT) where a site’s workload demands it, such as vision models or complex autoencoders. Both hardware variants run the same .NET edge runtime, and Orin is not a standalone ML appliance. See Orin setup and VL3 edge setup.
  • Strictly advisory. Inference output is limited to hints — anomaly scores, condition/health classifications, early-failure indicators — surfaced to operators for acknowledgement. It is one more advisory signal alongside the site LLM, never a control input.

Advisory-only is non-negotiable

Section titled “Advisory-only is non-negotiable”

This constraint exists because industrial safety standards and mining regulation require a deterministic, auditable control path, and ML models are probabilistic and opaque. The same boundary that applies to the site LLM applies, unchanged, to edge inference: edge ML produces hints, an operator decides, and the controller acts. There is no closed-loop, model-in-the-loop control path — planned or otherwise.

flowchart LR
    subgraph cloud[Cloud / cell]
        T[Train model<br/>export to ONNX]
    end
    subgraph edge[Edge runtime - planned]
        I[CPU inference baseline<br/>TensorRT GPU on Orin]
        H[Anomaly / condition<br/>hints]
    end
    O[Operator<br/>acknowledges]
    C[Controller<br/>deterministic path]
    T -->|signed model artefact| I
    I --> H
    H -->|advisory only| O
    O -->|executes| C

Explicitly out of scope

Section titled “Explicitly out of scope”

To avoid overclaiming, the following are not part of the platform today:

  • No TensorRT, .engine artefacts, or other GPU-compiled inference engines ship on edge nodes today — TensorRT on Orin is part of the documented design for GPU-dependent workloads, but nothing is deployable.
  • No GPU-required inference on the baseline path — CPU ONNX is the design target for the VL3; GPU-dependent workloads are what the Orin variant exists for.
  • No fleet-wide hot-loading or live model push to running edge nodes.
  • No on-device training, fine-tuning, or feedback learning at the edge.

When edge ML does arrive, model artefacts are expected to follow the same account-signed distribution discipline used elsewhere on the platform, and to degrade rather than block — if a model is absent, stale, or fails to load, the edge continues its deterministic PLC and store-and-forward work unaffected.

Section titled “Related reading”