TORUS AI/ML engine¶

Facility AI at every tier Doc TORUS-AI-001 Class Brochure Edge · gateway · facility operations

TORUS turns raw signal into a smaller, more certain facility event at each stage, so what travels the network and reaches the operator is a validated track, not a raw feed or a bare alarm.

Tier 1

Node

STM32U575 INT8 / CMSIS-NN

Tier 2

Gateway

Jetson Orin thermal / visual / fusion

Tier 3

CCISRT

Correlation and tracks

Privacy

Local raw data

Events cross the network

Capability brief TORUS-MEG gateway

01 · Why AI, not thresholds¶

A tripwire counts; an engine understands.

Capability	Detail
Classify by type, not just presence	Footfall, vehicle, or digging; voice, engine, or breaking glass. The system reports what it is, with a confidence, not merely that something happened.
Corroborate by fusing across domains	A seismic cue plus an acoustic cue from a neighbouring node is a stronger track than either alone. Fusion is where confidence is earned.
Suppress false alarms	Wind, rain, traffic, cooling equipment, and wildlife are the main sources of nuisance alarms for an unattended sensor. On-device models learn the background and suppress false alarms.

Diagram of TORUS detection classes: human, vehicle, drone, animal — Concept: representative planning classes and alert policies used for coverage design and simulation scenarios.

02 · Three tiers of facility AI¶

AI is distributed, not centralised. Every tier runs the inference its power and vantage allow, and passes a smaller, richer product to the next.

Tier	Runtime	Role	Status
Tier 1 · edge / node	STM32U575, generated INT8 / CMSIS-NN	Seismic and acoustic classification on the STM32U575 baseline, quantized for a battery budget. Hardware wake on seismic threshold or microphone Acoustic Activity Detect, then an ML pass confirms and classifies before any radio use. Audio is classified at the node; raw audio never leaves it.	transition
Tier 2 · gateway / MEG	NVIDIA Jetson Orin	Thermal and visual inference on cued imagery from the masts. Detect, track, and classify at the gateway, so backhaul carries decisions rather than video. Multi-node fusion across the local ring happens before anything leaves the site.	concept
Tier 3 · facility ops / CCISRT	CCISRT server	Clusters corroborating events across nodes into a single track with aggregate confidence. Explainable rule-based correlator today; a learned model drops in behind the same interface.	runnable

03 · Edge inference at the node¶

At the node, TinyML confirms and classifies before any radio use. Event versus background discrimination cuts the message rate and the power bill. The node path is a statically compiled INT8 inference pipeline, not a general-purpose tensor runtime: PyTorch training, fixed-shape ONNX interchange, INT8 quantization with representative site data, STM32Cube AI Studio generation, and CMSIS-NN kernels where supported.

Concept scene of edge detection across a perimeter field — Concept: seismic and acoustic detection field across an approach, with edge-tier event classification before uplink.

Vector math is a measured design gate

A Lattice iCE40 UP5K vector-math unit is carried as a measured design gate, not a production dependency. Retaining it requires a written non-power justification. Production node inference targets statically generated INT8 code with CMSIS-NN kernels where supported.

04 · Gateway-tier fusion and visual inference¶

The gateway receives the facility field, runs edge fusion, and where visual masts are attached, performs thermal and visual inference at the edge. This keeps raw imagery local while the backhaul carries compact decisions and fused events.

Procedural pole-mounted TORUS-MEG gateway with solar panel — Simulation asset: MEG gateway with pole mounting, solar power, local display and controls, dual radio antennas, and battery module.

Illustrative false-colour TORUS mast thermal camera feed — Simulation output: illustrative ironbow thermal display from MAST-01 with the selected actor in view. Not calibrated radiometric output.

05 · Facility operations correlation¶

CCISRT forms tracks across the facility and suppresses false alarms. The correlator ships as transparent rules so an operator can see why a track formed. Learned models are introduced behind that interface, with the rule-based path retained as a fallback and a sanity check.

Concept scene of a perimeter ring breach with multi-node detection — Concept: ring breach with multi-node detection, classification, confidence, and gateway correlation.

Concept render of the CCISRT operations room — Concept: operations room with common facility picture, event queue, thermal cues, and ring health.

06 · Facility-event pipeline¶

Raw signal in, decision out. Each stage reduces data and raises certainty.

Stage	Output
Sense	Raw signal from geophone, microphone, thermal, LiDAR, and RF at the node.
Classify	Edge TinyML labels the event and its confidence, and discards background.
Fuse	The gateway corroborates across the local ring and runs visual inference on cue.
Correlate	CCISRT forms tracks across the facility and suppresses false alarms.
Decide	The operator gets a validated track with location and confidence, ready to act on.

Data reduction

A raw waveform becomes a labelled event, a set of events becomes a track, and a track becomes a decision, so the network carries kilobytes of event data rather than megabytes of feed.

07 · Models and method¶

Tier	Runtime	Approach
Edge	STM32U575, generated INT8	PyTorch training, fixed-shape ONNX interchange, INT8 generation with STM32Cube AI Studio, CMSIS-NN kernels where supported, and wake-gated inference. The optional iCE40 UP5K remains a measured design gate, not a production dependency.
Gateway	NVIDIA Jetson Orin	Accelerated detection and tracking on thermal and visual streams, on-site so raw imagery stays local.
Facility ops	CCISRT server	Rule-based correlation today, explainable and auditable; learned multi-target tracking and classification fusion drop in behind the same track interface; northbound REST and webhooks publish events to facility systems.
All tiers	Isaac Sim	Scenarios generate labelled data and validate detection and false-alarm rates before facility trials.

08 · Data governance and roadmap¶

Classify at the source, keep raw data local.

Area	Principle	Detail
Privacy	Raw stays put	Audio and imagery are processed where captured; only compact labelled events cross the network.
Signature	Less on the wire	Sending decisions instead of feeds cuts bandwidth, power, and RF signature, which supports passive facility monitoring.
Assurance	Auditable inference	Track formation and command actions are logged, so an analyst can reconstruct why the system decided what it did.

Now	Next
On-device seismic and acoustic classification at the node	Larger vocabularies and per-site adaptation via OTA model updates
Rule-based correlation and track formation at CCISRT	Learned multi-target tracking and cross-domain classification fusion behind the same interface
Isaac Sim scenario data for validation	Active learning from operator confirmations to cut false alarms over time

Capability brief Deployment rehearsal