AI Models

AI Models (Sobek Neural Network Suite)

Overview

NexusEdge embeds 98 pre-trained neural models compiled for Hailo-8 and Hailo-10H silicon. All models run on-device with sub-millisecond inference latency.

Model Families

Nehebkau (LSTM)

• Purpose: Fault Detection and Diagnostics (FDD)
• Architecture: LSTM recurrent neural network
• Models: 24 equipment-specific variants (H8 + H10)
• Output: Health score (0-100%), 6 fault category scores

Medjed (GRU)

• Purpose: Anomaly detection and predictive maintenance
• Architecture: GRU recurrent neural network
• Models: 24 equipment-specific variants (H8 + H10)
• Output: Anomaly score (0.0-1.0), confidence, prediction horizon

Sage BOW

• Purpose: NexusOracle knowledge retrieval
• Architecture: Bag-of-Words retriever
• Corpus: 617 HVAC diagnostic scenarios
• Output: Top-5 ranked knowledge base entries per query

Equipment Coverage

Each equipment type has both a Nehebkau and Medjed model: AHU, RTU, Boiler, Cascade Boiler, Residential Boiler, Residential Furnace, Chiller, Water-Cooled Chiller, Cooling Tower, Pump, Booster Pump, VFD Pump Pack, Heat Pump, Residential Heat Pump, DOAS, Zone Reheat, Commercial Lighting, Residential Electrical, Garden, Smart Home, Pool Gas Heater, Pool Heat Pump, City Water, Well Water

Inference Pipeline

1. Sensor data collected every 5 seconds via I2C
2. Feature vector built from equipment inputs
3. Nehebkau + Medjed inference every 10 seconds
4. Results stored in local AegisDB
5. Results pushed to AN Console AegisDB every 60 seconds (Pro tier)

Hailo NPU

• Hailo-8: 26 TOPS, PCIe Gen3 x4
• Hailo-10H: 40 TOPS, PCIe Gen4 x4
• Fallback: CPU inference when no NPU detected
• Runtime: HailoRT 4.20.0

On-Controller Training

The AxonML training pipeline allows fine-tuning models with your own operational data. Training runs on-device and the updated weights stay on the controller.

Whitepapers

• Nehebkau FDD Whitepaper
• Medjed Anomaly Detection Whitepaper