Scaling a Water Meter Monitoring System for 50+ Units: NE101 + NG4500 Architecture

When Does LTE Stop Making Sense at Scale?

When you deploy 5–10 NE101 cameras on individual LTE Cat.1 SIM cards, the architecture is simple: each camera connects independently to cellular, uploads images to an MQTT broker, and your OCR pipeline processes reads on a server. For scattered remote sites, this remains the right approach.

At 50+ meters concentrated on one building, campus, or industrial facility, the economics and operational complexity change. With IoT SIM pricing at $1–3 per device per month, 100 units adds $1,200–$3,600/year in recurring connectivity costs. Managing 100 APN configurations, SIM renewals, and independent reconnect cycles becomes a maintenance burden. If images must stay entirely on-premises for compliance or privacy reasons, routing through cellular may also be architecturally undesirable.

At this point, a hub-and-spoke gateway architecture often becomes more efficient. NE101 cameras connect locally via Ethernet, WiFi, or WiFi HaLow to a single NG4500 edge server, which aggregates images, runs OCR locally, and delivers structured readings to your existing monitoring platform, such as SCADA or BMS. This reduces recurring SIM costs, keeps data on-premises, and centralizes inference and fleet management.

System Architecture

In large-scale deployments, image capture and OCR inference are handled separately. NE101 units capture and transmit images at each meter, while the NG4500 processes OCR centrally for all devices on the site. This reduces per-device cost, simplifies maintenance, and keeps AI workloads off battery-powered field devices.

The architecture consists of three layers:
1. Capture nodes: NE101 cameras installed at each water meter to capture images on schedule or by event.
2. Local connectivity: WiFi HaLow, WiFi, or Ethernet links cameras to the gateway, depending on site layout and existing infrastructure.
3. Centralized edge inference: NG4500 runs OCR and NeoMind locally, then sends structured readings to your monitoring platform.

Local Connectivity Options:
NE101 cameras connect to the NG4500 over a local network. WiFi HaLow is ideal for basement meters and long-range deployments where wall penetration matters. WiFi works best when existing coverage already reaches meter locations. Ethernet offers the highest reliability for fixed installations near network drops. For geographically dispersed meters without local infrastructure, an LTE-based architecture may be more practical.

Choosing the Right NG4500 Model:
For OCR workloads on meter images, even the entry-level NG4510 (Orin Nano 4GB, 20 TOPS) provides enough compute for dozens of cameras. The NG4511 (Orin Nano 8GB, 40 TOPS) offers the best balance of cost and performance for most deployments up to 100 cameras, while the NG4521 (Orin NX 16GB, 100 – 157 TOPS) is better suited for larger fleets or multi-model AI pipelines.

Recommended Hardware for a 50-Unit Pilot

A representative hardware stack for deploying 50 NE101 cameras with a centralized gateway architecture. Scale the NE101 count to your site requirements and choose connectivity based on existing infrastructure.

NE101 Variant Selection

NG4500 Sizing

Choosing the Right NG4500 for Your Deployment

*Actual capacity depends on image frequency, OCR model size, and NeoMind services.

Cost Comparison: Gateway vs Individual LTE

For dense deployments, a gateway architecture often reaches break-even within 12–24 months, depending on SIM pricing and existing infrastructure.

3-Year Cost Comparison

Operational Comparison

Gateway break-even depends on SIM pricing, deployment duration, and existing infrastructure. At $2/SIM/month, crossover is typically around Year 1. At lower SIM prices, it may extend to Year 2–3.
WiFi is most cost-effective when existing coverage exists. WiFi HaLow adds cost but is justified for long-range or wall-penetration scenarios.

Deployment Workflow

Deploying 50-300+ cameras requires planning but doesn’t have to be complex. The workflow below separates preparation, installation, gateway setup, and validation into distinct phases. Each phase is designed to scale efficiently: batch device preparation reduces on-site time, bulk configuration accelerates setup, and early validation prevents costly rework.

MQTT Payload and OCR Integration

Each NE101 publishes the same JSON payload regardless of connectivity type, including WiFi HaLow, WiFi, Ethernet, or 4G LTE. Images are delivered to the NG4500’s MQTT broker as Base64-encoded JPEGs inside the payload. The OCR pipeline subscribes to the relevant topics, decodes the image, runs digit recognition locally on the NG4500, and publishes the structured readings to an output topic for your existing monitoring platform, such as SCADA or BMS.

{
  "ts": 1740640441620,
  "values": {
    "devName":     "NE101-WM-045",       // Configurable device name
    "devMac":      "D8:3B:DA:4D:10:2C",
    "battery":     84,                  // Battery % — monitor for replacement planning
    "snapType":    "Scheduled",          // Scheduled | Button | PIR | Alarm
    "localtime":   "2026-04-22 06:00:00",
    "imageSize":   74371,               // bytes
    "image":       "data:image/jpeg;base64,..."
  }
}

{
  "device_id":   "ne101-wm-045",
  "site":        "building-b-level2",
  "meter_id":    "WM-045",
  "timestamp":   "2026-04-22T06:00:05Z",
  "ocr_value":   "01027.8",
  "unit":        "m3",
  "battery_pct": 84,
  "source_topic": "meters/building-b/wm-045"
}

For lightweight OCR models running on the NG4510, inference is typically sub-second per image. End-to-end latency depends on model complexity, queue depth, and the selected Jetson Orin variant.

Fleet Operations at Scale

Managing 50-300+ cameras requires operational processes beyond simple installation. NeoMind provides the management layer, but long-term success depends on disciplined maintenance workflows. Key operational areas include:

Battery Replacement Planning

At 3+ years battery life (10 captures/day in WiFi mode), replacement is infrequent but predictable. For 100+ deployments, schedule replacements proactively rather than reactively. Use NeoMind alerts at 20% battery remaining, then plan bulk replacement events within a 3–6 month window to reduce site visits.

Automated Alerting Strategy

Use tiered alerts to avoid alarm fatigue while catching issues early:
Critical — immediate action required
device offline >24h, battery < 10%, OCR confidence below threshold
Warning — maintenance planning
Battery <20%, device offline >12h, image quality degradation
Informational — trend monitoring
Monthly battery reports, read success trends, anomaly detection

OTA Firmware Updates at Scale

NeoMind supports OTA updates across the NE101 fleet. For 50+ devices, use staged rollout:
1. update 5-10 devices first
2. monitor for issues for 24-48 hours
3. deploy fleet-wide after validation
For mission-critical deployments, maintain a fallback plan (e.g., retain ability to revert to previous firmware if issues arise).

Device Grouping and Access Control

For deployments across multiple buildings or sites, organize devices by building, floor, meter type, or deployment phase. This simplifies monitoring by making it easier to isolate issues and view relevant metrics without filtering through hundreds of devices.
Grouping also enables granular access control, so maintenance teams only see and manage the devices relevant to their assigned location or responsibility.

Storage and Retention Planning

At 4–6 reads/day with ~50–100KB images, a 100-device deployment generates roughly 20–30GB/year of image data. This is manageable for most NG4500 configurations, but longer retention periods increase storage requirements.
A common strategy is to retain raw images locally for 6–12 months, then archive or delete older data automatically based on compliance or audit requirements. NeoMind can automate retention and cleanup policies at scale.

Backup and Disaster Recovery

For production deployments, back up NeoMind configuration regularly, including device registration, automation rules, and dashboard layouts. Store backups externally or on a network location, and document the restore process so recovery is fast in the event of failure. For redundancy-critical sites, consider a warm-standby NG4500 with replicated configuration to minimize downtime.

When the Gateway Architecture Doesn’t Make Sense

The centralized gateway architecture is optimized for concentrated deployments on a single site with available power, network, and mounting infrastructure. In the following cases, a different architecture may be more practical:

Geographically Scattered Meters
If meters are spread across a wide area — such as rural water meters across 50 km² service area — there may be no practical location to host a gateway. In these cases, per-device LTE is usually the better option.

Small Deployments (<20 units)
For smaller deployments on one site, the upfront gateway investment may not justify the cost savings over individual LTE SIMs, especially for short-term projects. However, on-premises processing may still justify the gateway approach when data privacy is critical.

Existing WiFi Already Covers the Site
If your site already has reliable WiFi at all meter locations, a WiFi-based gateway architecture is often more cost-effective than deploying new WiFi HaLow infrastructure. Use HaLow only when range, walls, or interference prevent standard WiFi coverage.

Frequently Asked Questions